STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
STATISTIC(NumBranchOpts, "Number of branches optimized");
STATISTIC(NumTailMerge , "Number of block tails merged");
-STATISTIC(NumTailDups , "Number of tail duplicated blocks");
-STATISTIC(NumInstrDups , "Additional instructions due to tail duplication");
static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
cl::init(cl::BOU_UNSET), cl::Hidden);
public:
static char ID;
explicit BranchFolderPass(bool defaultEnableTailMerge)
- : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {}
+ : MachineFunctionPass(ID), BranchFolder(defaultEnableTailMerge) {}
virtual bool runOnMachineFunction(MachineFunction &MF);
virtual const char *getPassName() const { return "Control Flow Optimizer"; }
/// function, updating the CFG.
void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
assert(MBB->pred_empty() && "MBB must be dead!");
- DEBUG(errs() << "\nRemoving MBB: " << *MBB);
+ DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
MachineFunction *MF = MBB->getParent();
// drop all successors.
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());
- }
- }
-
// Remove the block.
MF->erase(MBB);
}
SmallSet<unsigned, 4> ImpDefRegs;
MachineBasicBlock::iterator I = MBB->begin();
while (I != MBB->end()) {
- if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
+ if (!I->isImplicitDef())
break;
unsigned Reg = I->getOperand(0).getReg();
ImpDefRegs.insert(Reg);
MadeChange |= MadeChangeThisIteration;
}
- // Do tail duplication after tail merging is done. Otherwise it is
- // tough to avoid situations where tail duplication and tail merging undo
- // each other's transformations ad infinitum.
- MadeChangeThisIteration = true;
- while (MadeChangeThisIteration) {
- MadeChangeThisIteration = false;
- MadeChangeThisIteration |= TailDuplicateBlocks(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();
- const std::vector<MachineJumpTableEntry> &JTs = JTI->getJumpTables();
- if (!JTs.empty()) {
- // 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());
- for (MachineFunction::iterator BB = MF.begin(), E = MF.end();
- BB != E; ++BB) {
- for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end();
- I != E; ++I)
- 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);
- }
- }
-
- // 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.
- for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i)
- if (!JTIsLive.test(i)) {
- JTI->RemoveJumpTable(i);
- MadeChange = true;
+ if (JTI == 0) {
+ delete RS;
+ return MadeChange;
+ }
+
+ // 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();
+ I != E; ++I)
+ for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
+ MachineOperand &Op = I->getOperand(op);
+ if (!Op.isJTI()) continue;
+
+ // Remember that this JT is live.
+ JTIsLive.set(Op.getIndex());
}
}
+ // 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);
+ MadeChange = true;
+ }
+
delete RS;
return MadeChange;
}
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
// relative order. This is untenable because normal compiler
// optimizations (like this one) may reorder and/or merge these
// directives.
- I1->getOpcode() == TargetInstrInfo::INLINEASM) {
+ I1->isInlineAsm()) {
++I1; ++I2;
break;
}
++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;
}
/// 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();
-
- // Remove all the old successors of OldBB from the CFG.
- while (!OldBB->succ_empty())
- OldBB->removeSuccessor(OldBB->succ_begin());
-
- // Remove all the dead instructions from the end of OldBB.
- OldBB->erase(OldInst, OldBB->end());
-
- // 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);
+ TII->ReplaceTailWithBranchTo(OldInst, 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.
MachineBasicBlock::iterator E) {
unsigned Time = 0;
for (; I != E; ++I) {
+ if (I->isDebugValue())
+ continue;
const TargetInstrDesc &TID = I->getDesc();
if (TID.isCall())
Time += 10;
static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
const TargetInstrInfo *TII) {
MachineFunction *MF = CurMBB->getParent();
- MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB));
+ 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
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.
// 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();
- DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size "
+ // 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
// this many instructions in common.
unsigned minCommonTailLength = TailMergeSize;
- DEBUG(errs() << "\nTryTailMergeBlocks: ";
+ DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
- errs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
+ dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
<< (i == e-1 ? "" : ", ");
- errs() << "\n";
+ dbgs() << "\n";
if (SuccBB) {
- errs() << " with successor BB#" << SuccBB->getNumber() << '\n';
+ dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
if (PredBB)
- errs() << " which has fall-through from BB#"
+ dbgs() << " which has fall-through from BB#"
<< PredBB->getNumber() << "\n";
}
- errs() << "Looking for common tails of at least "
+ dbgs() << "Looking for common tails of at least "
<< minCommonTailLength << " instruction"
<< (minCommonTailLength == 1 ? "" : "s") << '\n';
);
!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();
// MBB is common tail. Adjust all other BB's to jump to this one.
// Traversal must be forwards so erases work.
- DEBUG(errs() << "\nUsing common tail in BB#" << MBB->getNumber()
+ DEBUG(dbgs() << "\nUsing common tail in BB#" << MBB->getNumber()
<< " for ");
for (unsigned int i=0, e = SameTails.size(); i != e; ++i) {
if (commonTailIndex == i)
continue;
- DEBUG(errs() << "BB#" << SameTails[i].getBlock()->getNumber()
+ DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber()
<< (i == e-1 ? "" : ", "));
// Hack the end off BB i, making it jump to BB commonTailIndex instead.
ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB);
// BB i is no longer a predecessor of SuccBB; remove it from the worklist.
MergePotentials.erase(SameTails[i].getMPIter());
}
- DEBUG(errs() << "\n");
+ DEBUG(dbgs() << "\n");
// We leave commonTailIndex in the worklist in case there are other blocks
// that match it with a smaller number of instructions.
MadeChange = true;
MergePotentials.clear();
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
if (I->succ_empty())
- MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I, 2U), I));
+ MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(I), I));
}
// See if we can do any tail merging on those.
// a compile-time infinite loop repeatedly doing and undoing the same
// transformations.)
- for (MachineFunction::iterator I = next(MF.begin()), E = MF.end();
+ for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
I != E; ++I) {
if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) {
SmallPtrSet<MachineBasicBlock *, 8> UniquePreds;
continue;
// This is the QBB case described above
if (!FBB)
- FBB = next(MachineFunction::iterator(PBB));
+ FBB = llvm::next(MachineFunction::iterator(PBB));
}
// Failing case: the only way IBB can be reached from PBB is via
// exception handling. Happens for landing pads. Would be nice
}
// 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 (MergePotentials.size() >= 2)
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
-/// block after it, but it uses an explicit branch to do so (e.g. a table jump).
-///
-/// True is a conservative answer.
-///
-bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB,
- bool BranchUnAnalyzable,
- MachineBasicBlock *TBB,
- MachineBasicBlock *FBB,
- const SmallVectorImpl<MachineOperand> &Cond) {
- MachineFunction::iterator Fallthrough = CurBB;
- ++Fallthrough;
- // If FallthroughBlock is off the end of the function, it can't fall through.
- if (Fallthrough == CurBB->getParent()->end())
- return false;
-
- // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible.
- if (!CurBB->isSuccessor(Fallthrough))
- return false;
-
- // If we couldn't analyze the branch, examine the last instruction.
- // If the block doesn't end in a known control barrier, assume fallthrough
- // is possible. The isPredicable check is needed because this code can be
- // called during IfConversion, where an instruction which is normally a
- // Barrier is predicated and thus no longer an actual control barrier. This
- // is over-conservative though, because if an instruction isn't actually
- // predicated we could still treat it like a barrier.
- if (BranchUnAnalyzable)
- return CurBB->empty() || !CurBB->back().getDesc().isBarrier() ||
- CurBB->back().getDesc().isPredicable();
-
- // If there is no branch, control always falls through.
- if (TBB == 0) return true;
-
- // If there is some explicit branch to the fallthrough block, it can obviously
- // reach, even though the branch should get folded to fall through implicitly.
- if (MachineFunction::iterator(TBB) == Fallthrough ||
- MachineFunction::iterator(FBB) == Fallthrough)
+// 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;
-
- // If it's an unconditional branch to some block not the fall through, it
- // doesn't fall through.
- if (Cond.empty()) return false;
-
- // Otherwise, if it is conditional and has no explicit false block, it falls
- // through.
- return FBB == 0;
+ for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end();
+ MBBI!=MBBE; ++MBBI) {
+ if (!MBBI->isDebugValue())
+ return false;
+ }
+ return true;
}
-/// CanFallThrough - Return true if the specified 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 block after it, but it uses an explicit
-/// branch to do so (e.g. a table jump).
-///
-/// True is a conservative answer.
-///
-bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) {
- MachineBasicBlock *TBB = 0, *FBB = 0;
- SmallVector<MachineOperand, 4> Cond;
- bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true);
- return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
+// 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->getDesc().isBranch());
}
/// IsBetterFallthrough - Return true if it would be clearly better to
// 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();
+ // 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->getDesc().isCall() && !MBB1I->getDesc().isCall();
}
-/// TailDuplicateBlocks - Look for small blocks that are unconditionally
-/// branched to and do not fall through. Tail-duplicate their instructions
-/// into their predecessors to eliminate (dynamic) branches.
-bool BranchFolder::TailDuplicateBlocks(MachineFunction &MF) {
- bool MadeChange = false;
-
- for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
- MachineBasicBlock *MBB = I++;
-
- // Only duplicate blocks that end with unconditional branches.
- if (CanFallThrough(MBB))
- continue;
-
- MadeChange |= TailDuplicate(MBB, MF);
-
- // If it is dead, remove it.
- if (MBB->pred_empty()) {
- NumInstrDups -= MBB->size();
- RemoveDeadBlock(MBB);
- MadeChange = true;
- ++NumDeadBlocks;
- }
- }
- return MadeChange;
-}
-
-/// TailDuplicate - If it is profitable, duplicate TailBB's contents in each
-/// of its predecessors.
-bool BranchFolder::TailDuplicate(MachineBasicBlock *TailBB,
- MachineFunction &MF) {
- // Don't try to tail-duplicate single-block loops.
- if (TailBB->isSuccessor(TailBB))
- return false;
-
- // Set the limit on the number of instructions to duplicate, with a default
- // of one less than the tail-merge threshold. When optimizing for size,
- // duplicate only one, because one branch instruction can be eliminated to
- // compensate for the duplication.
- unsigned MaxDuplicateCount =
- MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize) ?
- 1 : TII->TailDuplicationLimit(*TailBB, TailMergeSize - 1);
-
- // Check the instructions in the block to determine whether tail-duplication
- // is invalid or unlikely to be profitable.
- unsigned i = 0;
- bool HasCall = false;
- for (MachineBasicBlock::iterator I = TailBB->begin();
- I != TailBB->end(); ++I, ++i) {
- // Non-duplicable things shouldn't be tail-duplicated.
- if (I->getDesc().isNotDuplicable()) return false;
- // Don't duplicate more than the threshold.
- if (i == MaxDuplicateCount) return false;
- // Remember if we saw a call.
- if (I->getDesc().isCall()) HasCall = true;
- }
- // Heuristically, don't tail-duplicate calls if it would expand code size,
- // as it's less likely to be worth the extra cost.
- if (i > 1 && HasCall)
- return false;
-
- // Iterate through all the unique predecessors and tail-duplicate this
- // block into them, if possible. Copying the list ahead of time also
- // avoids trouble with the predecessor list reallocating.
- bool Changed = false;
- SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
- TailBB->pred_end());
- for (SmallSetVector<MachineBasicBlock *, 8>::iterator PI = Preds.begin(),
- PE = Preds.end(); PI != PE; ++PI) {
- MachineBasicBlock *PredBB = *PI;
-
- assert(TailBB != PredBB &&
- "Single-block loop should have been rejected earlier!");
- if (PredBB->succ_size() > 1) continue;
-
- MachineBasicBlock *PredTBB, *PredFBB;
- SmallVector<MachineOperand, 4> PredCond;
- if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true))
- continue;
- if (!PredCond.empty())
- continue;
- // EH edges are ignored by AnalyzeBranch.
- if (PredBB->succ_size() != 1)
- continue;
- // Don't duplicate into a fall-through predecessor (at least for now).
- if (PredBB->isLayoutSuccessor(TailBB) && CanFallThrough(PredBB))
- continue;
-
- DEBUG(errs() << "\nTail-duplicating into PredBB: " << *PredBB
- << "From Succ: " << *TailBB);
-
- // Remove PredBB's unconditional branch.
- TII->RemoveBranch(*PredBB);
- // Clone the contents of TailBB into PredBB.
- for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
- I != E; ++I) {
- MachineInstr *NewMI = MF.CloneMachineInstr(I);
- PredBB->insert(PredBB->end(), NewMI);
- }
- NumInstrDups += TailBB->size() - 1; // subtract one for removed branch
-
- // Update the CFG.
- PredBB->removeSuccessor(PredBB->succ_begin());
- assert(PredBB->succ_empty() &&
- "TailDuplicate called on block with multiple successors!");
- for (MachineBasicBlock::succ_iterator I = TailBB->succ_begin(),
- E = TailBB->succ_end(); I != E; ++I)
- PredBB->addSuccessor(*I);
-
- Changed = true;
- ++NumTailDups;
- }
-
- // If TailBB was duplicated into all its predecessors except for the prior
- // block, which falls through unconditionally, move the contents of this
- // block into the prior block.
- MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(TailBB));
- MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
- SmallVector<MachineOperand, 4> PriorCond;
- bool PriorUnAnalyzable =
- TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
- // This has to check PrevBB->succ_size() because EH edges are ignored by
- // AnalyzeBranch.
- if (!PriorUnAnalyzable && PriorCond.empty() && !PriorTBB &&
- TailBB->pred_size() == 1 && PrevBB.succ_size() == 1 &&
- !TailBB->hasAddressTaken()) {
- DEBUG(errs() << "\nMerging into block: " << PrevBB
- << "From MBB: " << *TailBB);
- PrevBB.splice(PrevBB.end(), TailBB, TailBB->begin(), TailBB->end());
- PrevBB.removeSuccessor(PrevBB.succ_begin());;
- assert(PrevBB.succ_empty());
- PrevBB.transferSuccessors(TailBB);
- Changed = true;
- }
-
- return Changed;
-}
-
/// OptimizeBlock - Analyze and optimize control flow related to the specified
/// block. This is never called on the entry block.
bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
bool MadeChange = false;
MachineFunction &MF = *MBB->getParent();
+ DebugLoc dl; // FIXME: this is nowhere
ReoptimizeBlock:
MachineFunction::iterator FallThrough = MBB;
// 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;
}
// If MBB was the target of a jump table, update jump tables to go to the
// fallthrough instead.
- MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough);
+ if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
+ MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
MadeChange = true;
}
return MadeChange;
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;
if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
PrevBB.succ_size() == 1 &&
!MBB->hasAddressTaken()) {
- DEBUG(errs() << "\nMerging into block: " << PrevBB
+ DEBUG(dbgs() << "\nMerging into block: " << PrevBB
<< "From MBB: " << *MBB);
PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end());
PrevBB.removeSuccessor(PrevBB.succ_begin());;
// the condition is false, remove the uncond second branch.
if (PriorFBB == MBB) {
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
if (!TII->ReverseBranchCondition(NewPriorCond)) {
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond);
+ TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond, dl);
MadeChange = true;
++NumBranchOpts;
goto ReoptimizeBlock;
// the assert condition out of the loop body.
if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
MachineFunction::iterator(PriorTBB) == FallThrough &&
- !CanFallThrough(MBB)) {
+ !MBB->canFallThrough()) {
bool DoTransform = true;
// We have to be careful that the succs of PredBB aren't both no-successor
!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() &&
- (!CanFallThrough(PriorTBB) || PriorTBB->empty()))
- DoTransform = false;
-
-
if (DoTransform) {
// Reverse the branch so we will fall through on the previous true cond.
SmallVector<MachineOperand, 4> NewPriorCond(PriorCond);
if (!TII->ReverseBranchCondition(NewPriorCond)) {
- DEBUG(errs() << "\nMoving MBB: " << *MBB
+ DEBUG(dbgs() << "\nMoving MBB: " << *MBB
<< "To make fallthrough to: " << *PriorTBB << "\n");
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());
SmallVector<MachineOperand, 4> NewCond(CurCond);
if (!TII->ReverseBranchCondition(NewCond)) {
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()) {
// 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
// falls through into MBB and we can't understand the prior block's branch
// condition.
if (MBB->empty()) {
- bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB);
+ bool PredHasNoFallThrough = !PrevBB.canFallThrough();
if (PredHasNoFallThrough || !PriorUnAnalyzable ||
!PrevBB.isSuccessor(MBB)) {
// If the prior block falls through into us, turn it into an
PriorFBB = MBB;
}
TII->RemoveBranch(PrevBB);
- TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, dl);
}
// Iterate through all the predecessors, revectoring each in-turn.
if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) {
TII->RemoveBranch(*PMBB);
NewCurCond.clear();
- TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond);
+ TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond, dl);
MadeChange = true;
++NumBranchOpts;
PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
}
// Change any jumptables to go to the new MBB.
- MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB);
+ if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
+ MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
if (DidChange) {
++NumBranchOpts;
MadeChange = true;
}
// 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 the prior block doesn't fall through into this block, and if this
// block doesn't fall through into some other block, see if we can find a
// place to move this block where a fall-through will happen.
- if (!CanFallThrough(&PrevBB, PriorUnAnalyzable,
- PriorTBB, PriorFBB, PriorCond)) {
+ if (!PrevBB.canFallThrough()) {
// Now we know that there was no fall-through into this block, check to
// see if it has a fall-through into its successor.
- bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB,
- CurCond);
+ bool CurFallsThru = MBB->canFallThrough();
if (!MBB->isLandingPad()) {
// Check all the predecessors of this block. If one of them has no fall
// Analyze the branch at the end of the pred.
MachineBasicBlock *PredBB = *PI;
MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
- MachineBasicBlock *PredTBB, *PredFBB;
+ MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
SmallVector<MachineOperand, 4> PredCond;
- if (PredBB != MBB && !CanFallThrough(PredBB) &&
+ if (PredBB != MBB && !PredBB->canFallThrough() &&
!TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
&& (!CurFallsThru || !CurTBB || !CurFBB)
&& (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
// B elsewhere
// next:
if (CurFallsThru) {
- MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB));
+ MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
CurCond.clear();
- TII->InsertBranch(*MBB, NextBB, 0, CurCond);
+ TII->InsertBranch(*MBB, NextBB, 0, CurCond, dl);
}
MBB->moveAfter(PredBB);
MadeChange = true;
// and if the successor isn't an EH destination, we can arrange for the
// fallthrough to happen.
if (SuccBB != MBB && &*SuccPrev != MBB &&
- !CanFallThrough(SuccPrev) && !CurUnAnalyzable &&
+ !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
!SuccBB->isLandingPad()) {
MBB->moveBefore(SuccBB);
MadeChange = true;
// Okay, there is no really great place to put this block. If, however,
// the block before this one would be a fall-through if this block were
// removed, move this block to the end of the function.
- MachineBasicBlock *PrevTBB, *PrevFBB;
+ MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
SmallVector<MachineOperand, 4> PrevCond;
if (FallThrough != MF.end() &&
!TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
projects / oota-llvm.git / blobdiff