//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/MRegisterInfo.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
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) :
+ explicit BranchFolder(bool defaultEnableTailMerge) :
MachineFunctionPass((intptr_t)&ID) {
switch (FlagEnableTailMerge) {
case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break;
MachineBasicBlock::iterator BBI1);
std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
- const MRegisterInfo *RegInfo;
+ const TargetRegisterInfo *RegInfo;
RegScavenger *RS;
// Branch optzn.
bool OptimizeBranches(MachineFunction &MF);
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();
for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
MachineOperand &Op = I->getOperand(op);
if (!Op.isJumpTableIndex()) continue;
- unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
- Op.setJumpTableIndex(NewIdx);
+ unsigned NewIdx = JTMapping[Op.getIndex()];
+ Op.setIndex(NewIdx);
// Remember that this JT is live.
JTIsLive[NewIdx] = true;
case MachineOperand::MO_Register: OperandHash = Op.getReg(); break;
case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break;
case MachineOperand::MO_MachineBasicBlock:
- OperandHash = Op.getMachineBasicBlock()->getNumber();
+ OperandHash = Op.getMBB()->getNumber();
break;
- case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break;
+ case MachineOperand::MO_FrameIndex:
case MachineOperand::MO_ConstantPoolIndex:
- OperandHash = Op.getConstantPoolIndex();
- break;
case MachineOperand::MO_JumpTableIndex:
- OperandHash = Op.getJumpTableIndex();
+ OperandHash = Op.getIndex();
break;
case MachineOperand::MO_GlobalAddress:
case MachineOperand::MO_ExternalSymbol:
unsigned TailLen = 0;
while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
--I1; --I2;
- if (!I1->isIdenticalTo(I2)) {
+ 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; ++I2;
break;
}
/// EstimateRuntime - Make a rough estimate for how long it will take to run
/// the specified code.
static unsigned EstimateRuntime(MachineBasicBlock::iterator I,
- MachineBasicBlock::iterator E,
- const TargetInstrInfo *TII) {
+ MachineBasicBlock::iterator E) {
unsigned Time = 0;
for (; I != E; ++I) {
- const TargetInstrDescriptor &TID = TII->get(I->getOpcode());
- if (TID.Flags & M_CALL_FLAG)
+ const TargetInstrDesc &TID = I->getDesc();
+ if (TID.isCall())
Time += 10;
- else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
+ else if (TID.isSimpleLoad() || TID.mayStore())
Time += 2;
else
++Time;
MachineBasicBlock::iterator MBB1I,
MachineBasicBlock *MBB2,
MachineBasicBlock::iterator MBB2I,
- const TargetInstrInfo *TII,
MachineBasicBlock *PredBB) {
// If one block is the entry block, split the other one; we can't generate
// a branch to the entry block, as its label is not emitted.
// TODO: if we had some notion of which block was hotter, we could split
// the hot block, so it is the fall-through. Since we don't have profile info
// make a decision based on which will hurt most to split.
- unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII);
- unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII);
+ unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I);
+ unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I);
// If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the
// MBB1 block so it falls through. This will penalize the MBB2 path, but will
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
bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB,
MachineBasicBlock* PredBB) {
- unsigned minCommonTailLength = (SuccBB ? 1 : 2);
+ // It doesn't make sense to save a single instruction since tail merging
+ // will add a jump.
+ // FIXME: Ask the target to provide the threshold?
+ unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1;
MadeChange = false;
// Sort by hash value so that blocks with identical end sequences sort
continue;
}
- // Look through all the blocks that have the same hash as this one, and
- // find the one that has the largest number of instructions in common.
+ // Look through all the pairs of blocks that have the same hash as this
+ // one, and find the pair that has the largest number of instructions in
+ // common.
// Since instructions may get combined later (e.g. single stores into
// store multiple) this measure is not particularly accurate.
MachineBasicBlock::iterator BBI1, BBI2;
- unsigned FoundMatch = ~0U;
+ unsigned FoundI = ~0U, FoundJ = ~0U;
unsigned maxCommonTailLength = 0U;
- for (int i = MergePotentials.size()-2;
+ for (int i = MergePotentials.size()-1;
i != -1 && MergePotentials[i].first == CurHash; --i) {
- MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
- unsigned CommonTailLen = ComputeCommonTailLength(CurMBB,
- MergePotentials[i].second,
- TrialBBI1, TrialBBI2);
- if (CommonTailLen >= minCommonTailLength &&
- CommonTailLen >= maxCommonTailLength) {
- FoundMatch = i;
- maxCommonTailLength = CommonTailLen;
- BBI1 = TrialBBI1;
- BBI2 = TrialBBI2;
+ for (int j = i-1;
+ j != -1 && MergePotentials[j].first == CurHash; --j) {
+ MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
+ unsigned CommonTailLen = ComputeCommonTailLength(
+ MergePotentials[i].second,
+ MergePotentials[j].second,
+ TrialBBI1, TrialBBI2);
+ if (CommonTailLen >= minCommonTailLength &&
+ CommonTailLen > maxCommonTailLength) {
+ FoundI = i;
+ FoundJ = j;
+ maxCommonTailLength = CommonTailLen;
+ BBI1 = TrialBBI1;
+ BBI2 = TrialBBI2;
+ }
}
- }
+ }
- // If we didn't find anything that has at least minCommonTailLength
- // instructions matching this one, bail out.
- if (FoundMatch == ~0U) {
- // Put the unconditional branch back, if we need one.
- if (SuccBB && CurMBB != PredBB)
- FixTail(CurMBB, SuccBB, TII);
- MergePotentials.pop_back();
+ // If we didn't find any pair that has at least minCommonTailLength
+ // instructions in common, bail out. All entries with this
+ // hash code can go away now.
+ if (FoundI == ~0U) {
+ for (int i = MergePotentials.size()-1;
+ i != -1 && MergePotentials[i].first == CurHash; --i) {
+ // Put the unconditional branch back, if we need one.
+ CurMBB = MergePotentials[i].second;
+ if (SuccBB && CurMBB != PredBB)
+ FixTail(CurMBB, SuccBB, TII);
+ MergePotentials.pop_back();
+ }
continue;
}
-
- // Otherwise, move the matching block to the right position.
- if (FoundMatch != MergePotentials.size()-2)
- std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
+ // Otherwise, move the block(s) to the right position(s). So that
+ // BBI1/2 will be valid, the last must be I and the next-to-last J.
+ if (FoundI != MergePotentials.size()-1)
+ std::swap(MergePotentials[FoundI], *(MergePotentials.end()-1));
+ if (FoundJ != MergePotentials.size()-2)
+ std::swap(MergePotentials[FoundJ], *(MergePotentials.end()-2));
+
+ CurMBB = (MergePotentials.end()-1)->second;
MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
// If neither block is the entire common tail, split the tail of one block
continue;
}
+ MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
+ unsigned CommonTailLen = ComputeCommonTailLength(CurMBB, MBB2,
+ TrialBBI1, TrialBBI2);
+ if (CommonTailLen < minCommonTailLength)
+ continue;
+
// Decide whether we want to split CurMBB or MBB2.
- if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) {
+ if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, PredBB)) {
CurMBB = SplitMBBAt(*CurMBB, BBI1);
BBI1 = CurMBB->begin();
MergePotentials.back().second = CurMBB;
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.
if (!FBB)
FBB = 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
+ // to have a bit in the edge so we didn't have to do all this.
+ if (IBB->isLandingPad()) {
+ MachineFunction::iterator IP = PBB; IP++;
+ MachineBasicBlock* PredNextBB = NULL;
+ if (IP!=MF.end())
+ PredNextBB = IP;
+ if (TBB==NULL) {
+ if (IBB!=PredNextBB) // fallthrough
+ continue;
+ } else if (FBB) {
+ if (TBB!=IBB && FBB!=IBB) // cbr then ubr
+ continue;
+ } else if (Cond.empty()) {
+ if (TBB!=IBB) // ubr
+ continue;
+ } else {
+ if (TBB!=IBB && IBB!=PredNextBB) // cbr
+ continue;
+ }
+ }
// Remove the unconditional branch at the end, if any.
if (TBB && (Cond.size()==0 || FBB)) {
TII->RemoveBranch(*PBB);
}
-/// 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.
-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::pred_iterator SI = MBB.succ_begin();
- 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;
- } 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;
-}
-
-
-/// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to
-/// 'Old', change the code and CFG so that it branches to 'New' instead.
-static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB,
- MachineBasicBlock *Old,
- MachineBasicBlock *New,
- const TargetInstrInfo *TII) {
- assert(Old != New && "Cannot replace self with self!");
-
- MachineBasicBlock::iterator I = BB->end();
- while (I != BB->begin()) {
- --I;
- if (!TII->isTerminatorInstr(I->getOpcode())) break;
-
- // Scan the operands of this machine instruction, replacing any uses of Old
- // with New.
- for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
- if (I->getOperand(i).isMachineBasicBlock() &&
- I->getOperand(i).getMachineBasicBlock() == Old)
- I->getOperand(i).setMachineBasicBlock(New);
- }
-
- // Update the successor information. If New was already a successor, just
- // remove the link to Old instead of creating another one. PR 1444.
- bool HadSuccessorNew = false;
- std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end());
- for (int i = Succs.size()-1; i >= 0; --i)
- if (Succs[i] == New) {
- HadSuccessorNew = true;
- break;
- }
- for (int i = Succs.size()-1; i >= 0; --i)
- if (Succs[i] == Old) {
- BB->removeSuccessor(Old);
- if (!HadSuccessorNew)
- BB->addSuccessor(New);
- }
-}
-
/// 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
/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
/// result in infinite loops.
static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
- MachineBasicBlock *MBB2,
- const TargetInstrInfo &TII) {
+ MachineBasicBlock *MBB2) {
// Right now, we use a simple heuristic. If MBB2 ends with a call, and
// 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 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 TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode());
+ return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall();
}
/// OptimizeBlock - Analyze and optimize control flow related to the specified
++FallThrough;
// If this block is empty, make everyone use its fall-through, not the block
- // explicitly.
- if (MBB->empty()) {
+ // explicitly. Landing pads should not do this since the landing-pad table
+ // points to this block.
+ if (MBB->empty() && !MBB->isLandingPad()) {
// Dead block? Leave for cleanup later.
if (MBB->pred_empty()) return;
// instead.
while (!MBB->pred_empty()) {
MachineBasicBlock *Pred = *(MBB->pred_end()-1);
- ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII);
+ Pred->ReplaceUsesOfBlockWith(MBB, FallThrough);
}
// If MBB was the target of a jump table, update jump tables to go to 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
// last. Only do the swap if one is clearly better to fall through than
// the other.
if (FallThrough == --MBB->getParent()->end() &&
- !IsBetterFallthrough(PriorTBB, MBB, *TII))
+ !IsBetterFallthrough(PriorTBB, MBB))
DoTransform = false;
// We don't want to do this transformation if we have control flow like:
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:
// 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 &&
- TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
+ MBB->begin()->getDesc().isBranch() && CurTBB != 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.
}
// 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;
- ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII);
+ PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB);
}
}