STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
STATISTIC(NumBranchOpts, "Number of branches optimized");
STATISTIC(NumTailMerge , "Number of block tails merged");
+STATISTIC(NumHoist , "Number of times common instructions are hoisted");
static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
cl::init(cl::BOU_UNSET), cl::Hidden);
namespace {
/// BranchFolderPass - Wrap branch folder in a machine function pass.
- class BranchFolderPass : public MachineFunctionPass,
- public BranchFolder {
+ class BranchFolderPass : public MachineFunctionPass {
public:
static char ID;
- explicit BranchFolderPass(bool defaultEnableTailMerge)
- : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
+ explicit BranchFolderPass(): MachineFunctionPass(ID) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
- virtual const char *getPassName() const { return "Control Flow Optimizer"; }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<TargetPassConfig>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
};
}
char BranchFolderPass::ID = 0;
+char &llvm::BranchFolderPassID = BranchFolderPass::ID;
-FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) {
- return new BranchFolderPass(DefaultEnableTailMerge);
-}
+INITIALIZE_PASS(BranchFolderPass, "branch-folder",
+ "Control Flow Optimizer", false, false)
bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
- return OptimizeFunction(MF,
- MF.getTarget().getInstrInfo(),
- MF.getTarget().getRegisterInfo(),
- getAnalysisIfAvailable<MachineModuleInfo>());
+ TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
+ BranchFolder Folder(PassConfig->getEnableTailMerge(), /*CommonHoist=*/true);
+ return Folder.OptimizeFunction(MF,
+ MF.getTarget().getInstrInfo(),
+ MF.getTarget().getRegisterInfo(),
+ getAnalysisIfAvailable<MachineModuleInfo>());
}
-BranchFolder::BranchFolder(bool defaultEnableTailMerge) {
+BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist) {
switch (FlagEnableTailMerge) {
case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
case cl::BOU_TRUE: EnableTailMerge = true; break;
case cl::BOU_FALSE: EnableTailMerge = false; break;
}
+
+ EnableHoistCommonCode = CommonHoist;
}
/// RemoveDeadBlock - Remove the specified dead machine basic block from the
while (!MBB->succ_empty())
MBB->removeSuccessor(MBB->succ_end()-1);
- // If there are any labels in the basic block, unregister them from
- // MachineModuleInfo.
- if (MMI && !MBB->empty()) {
- for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
- I != E; ++I) {
- if (I->isLabel())
- // The label ID # is always operand #0, an immediate.
- MMI->InvalidateLabel(I->getOperand(0).getImm());
- }
- }
+ // Avoid matching if this pointer gets reused.
+ TriedMerging.erase(MBB);
// Remove the block.
MF->erase(MBB);
break;
unsigned Reg = I->getOperand(0).getReg();
ImpDefRegs.insert(Reg);
- for (const unsigned *SubRegs = TRI->getSubRegisters(Reg);
+ for (const uint16_t *SubRegs = TRI->getSubRegisters(Reg);
unsigned SubReg = *SubRegs; ++SubRegs)
ImpDefRegs.insert(SubReg);
++I;
MachineModuleInfo *mmi) {
if (!tii) return false;
+ TriedMerging.clear();
+
TII = tii;
TRI = tri;
MMI = mmi;
bool MadeChangeThisIteration = true;
while (MadeChangeThisIteration) {
- MadeChangeThisIteration = false;
- MadeChangeThisIteration |= TailMergeBlocks(MF);
- MadeChangeThisIteration |= OptimizeBranches(MF);
+ MadeChangeThisIteration = TailMergeBlocks(MF);
+ MadeChangeThisIteration |= OptimizeBranches(MF);
+ if (EnableHoistCommonCode)
+ MadeChangeThisIteration |= HoistCommonCode(MF);
MadeChange |= MadeChangeThisIteration;
}
- // See if any jump tables have become mergable or dead as the code generator
+ // See if any jump tables have become dead as the code generator
// did its thing.
MachineJumpTableInfo *JTI = MF.getJumpTableInfo();
if (JTI == 0) {
delete RS;
return MadeChange;
}
-
- const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
- // Figure out how these jump tables should be merged.
- std::vector<unsigned> JTMapping;
- JTMapping.reserve(JTs.size());
-
- // We always keep the 0th jump table.
- JTMapping.push_back(0);
-
- // Scan the jump tables, seeing if there are any duplicates. Note that this
- // is N^2, which should be fixed someday.
- for (unsigned i = 1, e = JTs.size(); i != e; ++i) {
- if (JTs[i].MBBs.empty())
- JTMapping.push_back(i);
- else
- JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs));
- }
- // If a jump table was merge with another one, walk the function rewriting
- // references to jump tables to reference the new JT ID's. Keep track of
- // whether we see a jump table idx, if not, we can delete the JT.
- BitVector JTIsLive(JTs.size());
+ // Walk the function to find jump tables that are live.
+ BitVector JTIsLive(JTI->getJumpTables().size());
for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
BB != E; ++BB) {
for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
MachineOperand &Op = I->getOperand(op);
if (!Op.isJTI()) continue;
- unsigned NewIdx = JTMapping[Op.getIndex()];
- Op.setIndex(NewIdx);
// Remember that this JT is live.
- JTIsLive.set(NewIdx);
+ JTIsLive.set(Op.getIndex());
}
}
- // Finally, remove dead jump tables. This happens either because the
- // indirect jump was unreachable (and thus deleted) or because the jump
- // table was merged with some other one.
+ // Finally, remove dead jump tables. This happens when the
+ // indirect jump was unreachable (and thus deleted).
for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
if (!JTIsLive.test(i)) {
JTI->RemoveJumpTable(i);
return Hash;
}
-/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks
-/// with no successors, we hash two instructions, because cross-jumping
-/// only saves code when at least two instructions are removed (since a
-/// branch must be inserted). For blocks with a successor, one of the
-/// two blocks to be tail-merged will end with a branch already, so
-/// it gains to cross-jump even for one instruction.
-static unsigned HashEndOfMBB(const MachineBasicBlock *MBB,
- unsigned minCommonTailLength) {
+/// HashEndOfMBB - Hash the last instruction in the MBB.
+static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) {
MachineBasicBlock::const_iterator I = MBB->end();
if (I == MBB->begin())
return 0; // Empty MBB.
--I;
- unsigned Hash = HashMachineInstr(I);
-
- if (I == MBB->begin() || minCommonTailLength == 1)
- return Hash; // Single instr MBB.
+ // Skip debug info so it will not affect codegen.
+ while (I->isDebugValue()) {
+ if (I==MBB->begin())
+ return 0; // MBB empty except for debug info.
+ --I;
+ }
- --I;
- // Hash in the second-to-last instruction.
- Hash ^= HashMachineInstr(I) << 2;
- return Hash;
+ return HashMachineInstr(I);
}
/// ComputeCommonTailLength - Given two machine basic blocks, compute the number
unsigned TailLen = 0;
while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
--I1; --I2;
+ // Skip debugging pseudos; necessary to avoid changing the code.
+ while (I1->isDebugValue()) {
+ if (I1==MBB1->begin()) {
+ while (I2->isDebugValue()) {
+ if (I2==MBB2->begin())
+ // I1==DBG at begin; I2==DBG at begin
+ return TailLen;
+ --I2;
+ }
+ ++I2;
+ // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin
+ return TailLen;
+ }
+ --I1;
+ }
+ // I1==first (untested) non-DBG preceding known match
+ while (I2->isDebugValue()) {
+ if (I2==MBB2->begin()) {
+ ++I1;
+ // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin
+ return TailLen;
+ }
+ --I2;
+ }
+ // I1, I2==first (untested) non-DBGs preceding known match
if (!I1->isIdenticalTo(I2) ||
// FIXME: This check is dubious. It's used to get around a problem where
// people incorrectly expect inline asm directives to remain in the same
}
++TailLen;
}
+ // Back past possible debugging pseudos at beginning of block. This matters
+ // when one block differs from the other only by whether debugging pseudos
+ // are present at the beginning. (This way, the various checks later for
+ // I1==MBB1->begin() work as expected.)
+ if (I1 == MBB1->begin() && I2 != MBB2->begin()) {
+ --I2;
+ while (I2->isDebugValue()) {
+ if (I2 == MBB2->begin()) {
+ return TailLen;
+ }
+ --I2;
+ }
+ ++I2;
+ }
+ if (I2 == MBB2->begin() && I1 != MBB1->begin()) {
+ --I1;
+ while (I1->isDebugValue()) {
+ if (I1 == MBB1->begin())
+ return TailLen;
+ --I1;
+ }
+ ++I1;
+ }
return TailLen;
}
+void BranchFolder::MaintainLiveIns(MachineBasicBlock *CurMBB,
+ MachineBasicBlock *NewMBB) {
+ if (RS) {
+ RS->enterBasicBlock(CurMBB);
+ if (!CurMBB->empty())
+ RS->forward(prior(CurMBB->end()));
+ BitVector RegsLiveAtExit(TRI->getNumRegs());
+ RS->getRegsUsed(RegsLiveAtExit, false);
+ for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
+ if (RegsLiveAtExit[i])
+ NewMBB->addLiveIn(i);
+ }
+}
+
/// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
-/// after it, replacing it with an unconditional branch to NewDest. This
-/// returns true if OldInst's block is modified, false if NewDest is modified.
+/// after it, replacing it with an unconditional branch to NewDest.
void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
MachineBasicBlock *NewDest) {
- MachineBasicBlock *OldBB = OldInst->getParent();
+ MachineBasicBlock *CurMBB = OldInst->getParent();
- // Remove all the old successors of OldBB from the CFG.
- while (!OldBB->succ_empty())
- OldBB->removeSuccessor(OldBB->succ_begin());
+ TII->ReplaceTailWithBranchTo(OldInst, NewDest);
- // Remove all the dead instructions from the end of OldBB.
- OldBB->erase(OldInst, OldBB->end());
+ // For targets that use the register scavenger, we must maintain LiveIns.
+ MaintainLiveIns(CurMBB, NewDest);
- // If OldBB isn't immediately before OldBB, insert a branch to it.
- if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
- TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>());
- OldBB->addSuccessor(NewDest);
++NumTailMerge;
}
/// iterator. This returns the new MBB.
MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB,
MachineBasicBlock::iterator BBI1) {
+ if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1))
+ return 0;
+
MachineFunction &MF = *CurMBB.getParent();
// Create the fall-through block.
NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
// For targets that use the register scavenger, we must maintain LiveIns.
- if (RS) {
- RS->enterBasicBlock(&CurMBB);
- if (!CurMBB.empty())
- RS->forward(prior(CurMBB.end()));
- BitVector RegsLiveAtExit(TRI->getNumRegs());
- RS->getRegsUsed(RegsLiveAtExit, false);
- for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++)
- if (RegsLiveAtExit[i])
- NewMBB->addLiveIn(i);
- }
+ MaintainLiveIns(&CurMBB, NewMBB);
return NewMBB;
}
MachineBasicBlock::iterator E) {
unsigned Time = 0;
for (; I != E; ++I) {
- const TargetInstrDesc &TID = I->getDesc();
- if (TID.isCall())
+ if (I->isDebugValue())
+ continue;
+ if (I->isCall())
Time += 10;
- else if (TID.mayLoad() || TID.mayStore())
+ else if (I->mayLoad() || I->mayStore())
Time += 2;
else
++Time;
MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB));
MachineBasicBlock *TBB = 0, *FBB = 0;
SmallVector<MachineOperand, 4> Cond;
+ DebugLoc dl; // FIXME: this is nowhere
if (I != MF->end() &&
!TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) {
MachineBasicBlock *NextBB = I;
if (TBB == NextBB && !Cond.empty() && !FBB) {
if (!TII->ReverseBranchCondition(Cond)) {
TII->RemoveBranch(*CurMBB);
- TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond);
+ TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond, dl);
return;
}
}
}
- TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector<MachineOperand, 0>());
+ TII->InsertBranch(*CurMBB, SuccBB, NULL,
+ SmallVector<MachineOperand, 0>(), dl);
}
bool
// an object with itself.
#ifndef _GLIBCXX_DEBUG
llvm_unreachable("Predecessor appears twice");
-#endif
+#else
return false;
+#endif
}
}
break;
}
--I;
- if (!I->getDesc().isTerminator()) break;
+ if (!I->isTerminator()) break;
++NumTerms;
}
return NumTerms;
MachineBasicBlock *SuccBB,
MachineBasicBlock *PredBB) {
CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
- MachineFunction *MF = MBB1->getParent();
-
if (CommonTailLen == 0)
return false;
+ DEBUG(dbgs() << "Common tail length of BB#" << MBB1->getNumber()
+ << " and BB#" << MBB2->getNumber() << " is " << CommonTailLen
+ << '\n');
// It's almost always profitable to merge any number of non-terminator
// instructions with the block that falls through into the common successor.
// heuristics.
unsigned EffectiveTailLen = CommonTailLen;
if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
- !MBB1->back().getDesc().isBarrier() &&
- !MBB2->back().getDesc().isBarrier())
+ !MBB1->back().isBarrier() &&
+ !MBB2->back().isBarrier())
++EffectiveTailLen;
// Check if the common tail is long enough to be worthwhile.
// we don't have to split a block. At worst we will be introducing 1 new
// branch instruction, which is likely to be smaller than the 2
// instructions that would be deleted in the merge.
+ MachineFunction *MF = MBB1->getParent();
if (EffectiveTailLen >= 2 &&
MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) &&
(I1 == MBB1->begin() || I2 == MBB2->begin()))
/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
/// only of the common tail. Create a block that does by splitting one.
-unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
- unsigned maxCommonTailLength) {
- unsigned commonTailIndex = 0;
+bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB,
+ unsigned maxCommonTailLength,
+ unsigned &commonTailIndex) {
+ commonTailIndex = 0;
unsigned TimeEstimate = ~0U;
for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
// Use PredBB if possible; that doesn't require a new branch.
SameTails[commonTailIndex].getTailStartPos();
MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
+ // If the common tail includes any debug info we will take it pretty
+ // randomly from one of the inputs. Might be better to remove it?
DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
<< maxCommonTailLength);
MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI);
+ if (!newMBB) {
+ DEBUG(dbgs() << "... failed!");
+ return false;
+ }
+
SameTails[commonTailIndex].setBlock(newMBB);
SameTails[commonTailIndex].setTailStartPos(newMBB->begin());
if (PredBB == MBB)
PredBB = newMBB;
- return commonTailIndex;
+ return true;
}
// See if any of the blocks in MergePotentials (which all have a common single
!SameTails[commonTailIndex].tailIsWholeBlock())) {
// None of the blocks consist entirely of the common tail.
// Split a block so that one does.
- commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength);
+ if (!CreateCommonTailOnlyBlock(PredBB,
+ maxCommonTailLength, commonTailIndex)) {
+ RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
+ continue;
+ }
}
MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock();
// First find blocks with no successors.
MergePotentials.clear();
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end();
+ I != E && MergePotentials.size() < TailMergeThreshold; ++I) {
+ if (TriedMerging.count(I))
+ continue;
if (I->succ_empty())
- MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
+ MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
}
+ // If this is a large problem, avoid visiting the same basic blocks
+ // multiple times.
+ if (MergePotentials.size() == TailMergeThreshold)
+ for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
+ TriedMerging.insert(MergePotentials[i].getBlock());
// See if we can do any tail merging on those.
- if (MergePotentials.size() < TailMergeThreshold &&
- MergePotentials.size() >= 2)
+ if (MergePotentials.size() >= 2)
MadeChange |= TryTailMergeBlocks(NULL, NULL);
// Look at blocks (IBB) with multiple predecessors (PBB).
for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
I != E; ++I) {
- if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
+ if (I->pred_size() >= 2) {
SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
MachineBasicBlock *IBB = I;
MachineBasicBlock *PredBB = prior(I);
MergePotentials.clear();
for (MachineBasicBlock::pred_iterator P = I->pred_begin(),
E2 = I->pred_end();
- P != E2; ++P) {
+ P != E2 && MergePotentials.size() < TailMergeThreshold; ++P) {
MachineBasicBlock *PBB = *P;
+ if (TriedMerging.count(PBB))
+ continue;
// Skip blocks that loop to themselves, can't tail merge these.
if (PBB == IBB)
continue;
// Visit each predecessor only once.
if (!UniquePreds.insert(PBB))
continue;
+ // Skip blocks which may jump to a landing pad. Can't tail merge these.
+ if (PBB->getLandingPadSuccessor())
+ continue;
MachineBasicBlock *TBB = 0, *FBB = 0;
SmallVector<MachineOperand, 4> Cond;
if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) {
}
// Remove the unconditional branch at the end, if any.
if (TBB && (Cond.empty() || FBB)) {
+ DebugLoc dl; // FIXME: this is nowhere
TII->RemoveBranch(*PBB);
if (!Cond.empty())
// reinsert conditional branch only, for now
- TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond);
+ TII->InsertBranch(*PBB, (TBB == IBB) ? FBB : TBB, 0, NewCond, dl);
}
- MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB, 1U),
- *P));
+ MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(PBB), *P));
}
}
+ // If this is a large problem, avoid visiting the same basic blocks
+ // multiple times.
+ if (MergePotentials.size() == TailMergeThreshold)
+ for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
+ TriedMerging.insert(MergePotentials[i].getBlock());
if (MergePotentials.size() >= 2)
MadeChange |= TryTailMergeBlocks(IBB, PredBB);
// Reinsert an unconditional branch if needed.
- // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks.
- PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
+ // The 1 below can occur as a result of removing blocks in
+ // TryTailMergeBlocks.
+ PredBB = prior(I); // this may have been changed in TryTailMergeBlocks
if (MergePotentials.size() == 1 &&
MergePotentials.begin()->getBlock() != PredBB)
FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
// Make sure blocks are numbered in order
MF.RenumberBlocks();
- for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
+ for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
+ I != E; ) {
MachineBasicBlock *MBB = I++;
MadeChange |= OptimizeBlock(MBB);
return MadeChange;
}
+// Blocks should be considered empty if they contain only debug info;
+// else the debug info would affect codegen.
+static bool IsEmptyBlock(MachineBasicBlock *MBB) {
+ if (MBB->empty())
+ return true;
+ for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
+ MBBI!=MBBE; ++MBBI) {
+ if (!MBBI->isDebugValue())
+ return false;
+ }
+ return true;
+}
+
+// Blocks with only debug info and branches should be considered the same
+// as blocks with only branches.
+static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) {
+ MachineBasicBlock::iterator MBBI, MBBE;
+ for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) {
+ if (!MBBI->isDebugValue())
+ break;
+ }
+ return (MBBI->isBranch());
+}
/// IsBetterFallthrough - Return true if it would be clearly better to
/// fall-through to MBB1 than to fall through into MBB2. This has to return
// MBB1 doesn't, we prefer to fall through into MBB1. This allows us to
// optimize branches that branch to either a return block or an assert block
// into a fallthrough to the return.
- if (MBB1->empty() || MBB2->empty()) return false;
+ if (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) return false;
// If there is a clear successor ordering we make sure that one block
// will fall through to the next
if (MBB1->isSuccessor(MBB2)) return true;
if (MBB2->isSuccessor(MBB1)) return false;
- MachineInstr *MBB1I = --MBB1->end();
- MachineInstr *MBB2I = --MBB2->end();
- return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
+ // Neither block consists entirely of debug info (per IsEmptyBlock check),
+ // so we needn't test for falling off the beginning here.
+ MachineBasicBlock::iterator MBB1I = --MBB1->end();
+ while (MBB1I->isDebugValue())
+ --MBB1I;
+ MachineBasicBlock::iterator MBB2I = --MBB2->end();
+ while (MBB2I->isDebugValue())
+ --MBB2I;
+ return MBB2I->isCall() && !MBB1I->isCall();
+}
+
+/// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch
+/// instructions on the block. Always use the DebugLoc of the first
+/// branching instruction found unless its absent, in which case use the
+/// DebugLoc of the second if present.
+static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) {
+ MachineBasicBlock::iterator I = MBB.end();
+ if (I == MBB.begin())
+ return DebugLoc();
+ --I;
+ while (I->isDebugValue() && I != MBB.begin())
+ --I;
+ if (I->isBranch())
+ return I->getDebugLoc();
+ return DebugLoc();
}
/// OptimizeBlock - Analyze and optimize control flow related to the specified
// explicitly. Landing pads should not do this since the landing-pad table
// points to this block. Blocks with their addresses taken shouldn't be
// optimized away.
- if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
+ if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
// Dead block? Leave for cleanup later.
if (MBB->pred_empty()) return MadeChange;
// destination, remove the branch, replacing it with an unconditional one or
// a fall-through.
if (PriorTBB && PriorTBB == PriorFBB) {
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
PriorCond.clear();
if (PriorTBB != MBB)
- TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
// AnalyzeBranch.
if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
PrevBB.succ_size() == 1 &&
- !MBB->hasAddressTaken()) {
+ !MBB->hasAddressTaken() && !MBB->isLandingPad()) {
DEBUG(dbgs() << "\nMerging into block: " << PrevBB
<< "From MBB: " << *MBB);
+ // Remove redundant DBG_VALUEs first.
+ if (PrevBB.begin() != PrevBB.end()) {
+ MachineBasicBlock::iterator PrevBBIter = PrevBB.end();
+ --PrevBBIter;
+ MachineBasicBlock::iterator MBBIter = MBB->begin();
+ // Check if DBG_VALUE at the end of PrevBB is identical to the
+ // DBG_VALUE at the beginning of MBB.
+ while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end()
+ && PrevBBIter->isDebugValue() && MBBIter->isDebugValue()) {
+ if (!MBBIter->isIdenticalTo(PrevBBIter))
+ break;
+ MachineInstr *DuplicateDbg = MBBIter;
+ ++MBBIter; -- PrevBBIter;
+ DuplicateDbg->eraseFromParent();
+ }
+ }
PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
- PrevBB.removeSuccessor(PrevBB.succ_begin());;
+ PrevBB.removeSuccessor(PrevBB.succ_begin());
assert(PrevBB.succ_empty());
PrevBB.transferSuccessors(MBB);
MadeChange = true;
// If the prior block branches somewhere else on the condition and here if
// the condition is false, remove the uncond second branch.
if (PriorFBB == MBB) {
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
if (PriorTBB == MBB) {
SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
if (!TII->ReverseBranchCondition(NewPriorCond)) {
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
+ TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
!IsBetterFallthrough(PriorTBB, MBB))
DoTransform = false;
- // We don't want to do this transformation if we have control flow like:
- // br cond BB2
- // BB1:
- // ..
- // jmp BBX
- // BB2:
- // ..
- // ret
- //
- // In this case, we could actually be moving the return block *into* a
- // loop!
- if (DoTransform && !MBB->succ_empty() &&
- (!PriorTBB->canFallThrough() || PriorTBB->empty()))
- DoTransform = false;
-
-
if (DoTransform) {
// Reverse the branch so we will fall through on the previous true cond.
SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
DEBUG(dbgs() << "\nMoving MBB: " << *MBB
<< "To make fallthrough to: " << *PriorTBB << "\n");
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond);
+ TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
// Move this block to the end of the function.
MBB->moveAfter(--MF.end());
if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
SmallVector<MachineOperand, 4> NewCond(CurCond);
if (!TII->ReverseBranchCondition(NewCond)) {
+ DebugLoc dl = getBranchDebugLoc(*MBB);
TII->RemoveBranch(*MBB);
- TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond);
+ TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
// If this branch is the only thing in its block, see if we can forward
// other blocks across it.
if (CurTBB && CurCond.empty() && CurFBB == 0 &&
- MBB->begin()->getDesc().isBranch() && CurTBB != MBB &&
+ IsBranchOnlyBlock(MBB) && CurTBB != MBB &&
!MBB->hasAddressTaken()) {
+ DebugLoc dl = getBranchDebugLoc(*MBB);
// This block may contain just an unconditional branch. Because there can
// be 'non-branch terminators' in the block, try removing the branch and
// then seeing if the block is empty.
TII->RemoveBranch(*MBB);
-
+ // If the only things remaining in the block are debug info, remove these
+ // as well, so this will behave the same as an empty block in non-debug
+ // mode.
+ if (!MBB->empty()) {
+ bool NonDebugInfoFound = false;
+ for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
+ I != E; ++I) {
+ if (!I->isDebugValue()) {
+ NonDebugInfoFound = true;
+ break;
+ }
+ }
+ if (!NonDebugInfoFound)
+ // Make the block empty, losing the debug info (we could probably
+ // improve this in some cases.)
+ MBB->erase(MBB->begin(), MBB->end());
+ }
// If this block is just an unconditional branch to CurTBB, we can
// usually completely eliminate the block. The only case we cannot
// completely eliminate the block is when the block before this one
assert(PriorFBB == 0 && "Machine CFG out of date!");
PriorFBB = MBB;
}
+ DebugLoc pdl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl);
}
// Iterate through all the predecessors, revectoring each in-turn.
bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB,
NewCurFBB, NewCurCond, true);
if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
+ DebugLoc pdl = getBranchDebugLoc(*PMBB);
TII->RemoveBranch(*PMBB);
NewCurCond.clear();
- TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
+ TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, pdl);
MadeChange = true;
++NumBranchOpts;
PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
}
// Add the branch back if the block is more than just an uncond branch.
- TII->InsertBranch(*MBB, CurTBB, 0, CurCond);
+ TII->InsertBranch(*MBB, CurTBB, 0, CurCond, dl);
}
}
if (CurFallsThru) {
MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
CurCond.clear();
- TII->InsertBranch(*MBB, NextBB, 0, CurCond);
+ TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc());
}
MBB->moveAfter(PredBB);
MadeChange = true;
return MadeChange;
}
+
+//===----------------------------------------------------------------------===//
+// Hoist Common Code
+//===----------------------------------------------------------------------===//
+
+/// HoistCommonCode - Hoist common instruction sequences at the start of basic
+/// blocks to their common predecessor.
+bool BranchFolder::HoistCommonCode(MachineFunction &MF) {
+ bool MadeChange = false;
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) {
+ MachineBasicBlock *MBB = I++;
+ MadeChange |= HoistCommonCodeInSuccs(MBB);
+ }
+
+ return MadeChange;
+}
+
+/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given
+/// its 'true' successor.
+static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB,
+ MachineBasicBlock *TrueBB) {
+ for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
+ E = BB->succ_end(); SI != E; ++SI) {
+ MachineBasicBlock *SuccBB = *SI;
+ if (SuccBB != TrueBB)
+ return SuccBB;
+ }
+ return NULL;
+}
+
+/// findHoistingInsertPosAndDeps - Find the location to move common instructions
+/// in successors to. The location is ususally just before the terminator,
+/// however if the terminator is a conditional branch and its previous
+/// instruction is the flag setting instruction, the previous instruction is
+/// the preferred location. This function also gathers uses and defs of the
+/// instructions from the insertion point to the end of the block. The data is
+/// used by HoistCommonCodeInSuccs to ensure safety.
+static
+MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB,
+ const TargetInstrInfo *TII,
+ const TargetRegisterInfo *TRI,
+ SmallSet<unsigned,4> &Uses,
+ SmallSet<unsigned,4> &Defs) {
+ MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
+ if (!TII->isUnpredicatedTerminator(Loc))
+ return MBB->end();
+
+ for (unsigned i = 0, e = Loc->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = Loc->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isUse()) {
+ Uses.insert(Reg);
+ for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+ Uses.insert(*AS);
+ } else if (!MO.isDead())
+ // Don't try to hoist code in the rare case the terminator defines a
+ // register that is later used.
+ return MBB->end();
+ }
+
+ if (Uses.empty())
+ return Loc;
+ if (Loc == MBB->begin())
+ return MBB->end();
+
+ // The terminator is probably a conditional branch, try not to separate the
+ // branch from condition setting instruction.
+ MachineBasicBlock::iterator PI = Loc;
+ --PI;
+ while (PI != MBB->begin() && Loc->isDebugValue())
+ --PI;
+
+ bool IsDef = false;
+ for (unsigned i = 0, e = PI->getNumOperands(); !IsDef && i != e; ++i) {
+ const MachineOperand &MO = PI->getOperand(i);
+ // If PI has a regmask operand, it is probably a call. Separate away.
+ if (MO.isRegMask())
+ return Loc;
+ if (!MO.isReg() || MO.isUse())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (Uses.count(Reg))
+ IsDef = true;
+ }
+ if (!IsDef)
+ // The condition setting instruction is not just before the conditional
+ // branch.
+ return Loc;
+
+ // Be conservative, don't insert instruction above something that may have
+ // side-effects. And since it's potentially bad to separate flag setting
+ // instruction from the conditional branch, just abort the optimization
+ // completely.
+ // Also avoid moving code above predicated instruction since it's hard to
+ // reason about register liveness with predicated instruction.
+ bool DontMoveAcrossStore = true;
+ if (!PI->isSafeToMove(TII, 0, DontMoveAcrossStore) ||
+ TII->isPredicated(PI))
+ return MBB->end();
+
+
+ // Find out what registers are live. Note this routine is ignoring other live
+ // registers which are only used by instructions in successor blocks.
+ for (unsigned i = 0, e = PI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = PI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isUse()) {
+ Uses.insert(Reg);
+ for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+ Uses.insert(*AS);
+ } else {
+ if (Uses.count(Reg)) {
+ Uses.erase(Reg);
+ for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR)
+ Uses.erase(*SR); // Use getSubRegisters to be conservative
+ }
+ Defs.insert(Reg);
+ for (const uint16_t *AS = TRI->getAliasSet(Reg); *AS; ++AS)
+ Defs.insert(*AS);
+ }
+ }
+
+ return PI;
+}
+
+/// HoistCommonCodeInSuccs - If the successors of MBB has common instruction
+/// sequence at the start of the function, move the instructions before MBB
+/// terminator if it's legal.
+bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) {
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty())
+ return false;
+
+ if (!FBB) FBB = findFalseBlock(MBB, TBB);
+ if (!FBB)
+ // Malformed bcc? True and false blocks are the same?
+ return false;
+
+ // Restrict the optimization to cases where MBB is the only predecessor,
+ // it is an obvious win.
+ if (TBB->pred_size() > 1 || FBB->pred_size() > 1)
+ return false;
+
+ // Find a suitable position to hoist the common instructions to. Also figure
+ // out which registers are used or defined by instructions from the insertion
+ // point to the end of the block.
+ SmallSet<unsigned, 4> Uses, Defs;
+ MachineBasicBlock::iterator Loc =
+ findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs);
+ if (Loc == MBB->end())
+ return false;
+
+ bool HasDups = false;
+ SmallVector<unsigned, 4> LocalDefs;
+ SmallSet<unsigned, 4> LocalDefsSet;
+ MachineBasicBlock::iterator TIB = TBB->begin();
+ MachineBasicBlock::iterator FIB = FBB->begin();
+ MachineBasicBlock::iterator TIE = TBB->end();
+ MachineBasicBlock::iterator FIE = FBB->end();
+ while (TIB != TIE && FIB != FIE) {
+ // Skip dbg_value instructions. These do not count.
+ if (TIB->isDebugValue()) {
+ while (TIB != TIE && TIB->isDebugValue())
+ ++TIB;
+ if (TIB == TIE)
+ break;
+ }
+ if (FIB->isDebugValue()) {
+ while (FIB != FIE && FIB->isDebugValue())
+ ++FIB;
+ if (FIB == FIE)
+ break;
+ }
+ if (!TIB->isIdenticalTo(FIB, MachineInstr::CheckKillDead))
+ break;
+
+ if (TII->isPredicated(TIB))
+ // Hard to reason about register liveness with predicated instruction.
+ break;
+
+ bool IsSafe = true;
+ for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = TIB->getOperand(i);
+ // Don't attempt to hoist instructions with register masks.
+ if (MO.isRegMask()) {
+ IsSafe = false;
+ break;
+ }
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isDef()) {
+ if (Uses.count(Reg)) {
+ // Avoid clobbering a register that's used by the instruction at
+ // the point of insertion.
+ IsSafe = false;
+ break;
+ }
+
+ if (Defs.count(Reg) && !MO.isDead()) {
+ // Don't hoist the instruction if the def would be clobber by the
+ // instruction at the point insertion. FIXME: This is overly
+ // conservative. It should be possible to hoist the instructions
+ // in BB2 in the following example:
+ // BB1:
+ // r1, eflag = op1 r2, r3
+ // brcc eflag
+ //
+ // BB2:
+ // r1 = op2, ...
+ // = op3, r1<kill>
+ IsSafe = false;
+ break;
+ }
+ } else if (!LocalDefsSet.count(Reg)) {
+ if (Defs.count(Reg)) {
+ // Use is defined by the instruction at the point of insertion.
+ IsSafe = false;
+ break;
+ }
+
+ if (MO.isKill() && Uses.count(Reg))
+ // Kills a register that's read by the instruction at the point of
+ // insertion. Remove the kill marker.
+ MO.setIsKill(false);
+ }
+ }
+ if (!IsSafe)
+ break;
+
+ bool DontMoveAcrossStore = true;
+ if (!TIB->isSafeToMove(TII, 0, DontMoveAcrossStore))
+ break;
+
+ // Remove kills from LocalDefsSet, these registers had short live ranges.
+ for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = TIB->getOperand(i);
+ if (!MO.isReg() || !MO.isUse() || !MO.isKill())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg || !LocalDefsSet.count(Reg))
+ continue;
+ for (const uint16_t *OR = TRI->getOverlaps(Reg); *OR; ++OR)
+ LocalDefsSet.erase(*OR);
+ }
+
+ // Track local defs so we can update liveins.
+ for (unsigned i = 0, e = TIB->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = TIB->getOperand(i);
+ if (!MO.isReg() || !MO.isDef() || MO.isDead())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ LocalDefs.push_back(Reg);
+ for (const uint16_t *OR = TRI->getOverlaps(Reg); *OR; ++OR)
+ LocalDefsSet.insert(*OR);
+ }
+
+ HasDups = true;
+ ++TIB;
+ ++FIB;
+ }
+
+ if (!HasDups)
+ return false;
+
+ MBB->splice(Loc, TBB, TBB->begin(), TIB);
+ FBB->erase(FBB->begin(), FIB);
+
+ // Update livein's.
+ for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
+ unsigned Def = LocalDefs[i];
+ if (LocalDefsSet.count(Def)) {
+ TBB->addLiveIn(Def);
+ FBB->addLiveIn(Def);
+ }
+ }
+
+ ++NumHoist;
+ return true;
+}