//
// 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.
//
//===----------------------------------------------------------------------===//
//
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "branchfolding"
+#include "BranchFolding.h"
+#include "llvm/Function.h"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/MachineDebugInfo.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include <algorithm>
using namespace llvm;
-static Statistic<> NumDeadBlocks("branchfold", "Number of dead blocks removed");
-static Statistic<> NumBranchOpts("branchfold", "Number of branches optimized");
-static Statistic<> NumTailMerge ("branchfold", "Number of block tails merged");
-static cl::opt<bool> EnableTailMerge("enable-tail-merge", cl::Hidden);
+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);
+
+// Throttle for huge numbers of predecessors (compile speed problems)
+static cl::opt<unsigned>
+TailMergeThreshold("tail-merge-threshold",
+ cl::desc("Max number of predecessors to consider tail merging"),
+ cl::init(150), cl::Hidden);
+
+// Heuristic for tail merging (and, inversely, tail duplication).
+// TODO: This should be replaced with a target query.
+static cl::opt<unsigned>
+TailMergeSize("tail-merge-size",
+ cl::desc("Min number of instructions to consider tail merging"),
+ cl::init(3), cl::Hidden);
namespace {
- struct BranchFolder : public MachineFunctionPass {
+ /// BranchFolderPass - Wrap branch folder in a machine function pass.
+ class BranchFolderPass : public MachineFunctionPass {
+ public:
+ static char ID;
+ explicit BranchFolderPass(): MachineFunctionPass(ID) {}
+
virtual bool runOnMachineFunction(MachineFunction &MF);
- virtual const char *getPassName() const { return "Control Flow Optimizer"; }
- const TargetInstrInfo *TII;
- MachineDebugInfo *MDI;
- bool MadeChange;
- private:
- // Tail Merging.
- bool TailMergeBlocks(MachineFunction &MF);
- void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
- MachineBasicBlock *NewDest);
- MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB,
- MachineBasicBlock::iterator BBI1);
-
- // Branch optzn.
- bool OptimizeBranches(MachineFunction &MF);
- void OptimizeBlock(MachineBasicBlock *MBB);
- void RemoveDeadBlock(MachineBasicBlock *MBB);
-
- bool CanFallThrough(MachineBasicBlock *CurBB);
- bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable,
- MachineBasicBlock *TBB, MachineBasicBlock *FBB,
- const std::vector<MachineOperand> &Cond);
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<TargetPassConfig>();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
};
}
-FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); }
+char BranchFolderPass::ID = 0;
+char &llvm::BranchFolderPassID = BranchFolderPass::ID;
+
+INITIALIZE_PASS(BranchFolderPass, "branch-folder",
+ "Control Flow Optimizer", false, false)
+
+bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) {
+ 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, 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
/// function, updating the CFG.
void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) {
assert(MBB->pred_empty() && "MBB must be dead!");
-
+ DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
+
MachineFunction *MF = MBB->getParent();
// drop all successors.
while (!MBB->succ_empty())
MBB->removeSuccessor(MBB->succ_end()-1);
-
- // If there is DWARF info to active, check to see if there are any DWARF_LABEL
- // records in the basic block. If so, unregister them from MachineDebugInfo.
- if (MDI && !MBB->empty()) {
- unsigned DWARF_LABELOpc = TII->getDWARF_LABELOpcode();
- assert(DWARF_LABELOpc &&
- "Target supports dwarf but didn't implement getDWARF_LABELOpcode!");
-
- for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
- I != E; ++I) {
- if ((unsigned)I->getOpcode() == DWARF_LABELOpc) {
- // The label ID # is always operand #0, an immediate.
- MDI->InvalidateLabel(I->getOperand(0).getImm());
- }
+
+ // Avoid matching if this pointer gets reused.
+ TriedMerging.erase(MBB);
+
+ // Remove the block.
+ MF->erase(MBB);
+}
+
+/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def
+/// followed by terminators, and if the implicitly defined registers are not
+/// used by the terminators, remove those implicit_def's. e.g.
+/// BB1:
+/// r0 = implicit_def
+/// r1 = implicit_def
+/// br
+/// This block can be optimized away later if the implicit instructions are
+/// removed.
+bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) {
+ SmallSet<unsigned, 4> ImpDefRegs;
+ MachineBasicBlock::iterator I = MBB->begin();
+ while (I != MBB->end()) {
+ if (!I->isImplicitDef())
+ break;
+ unsigned Reg = I->getOperand(0).getReg();
+ ImpDefRegs.insert(Reg);
+ for (const uint16_t *SubRegs = TRI->getSubRegisters(Reg);
+ unsigned SubReg = *SubRegs; ++SubRegs)
+ ImpDefRegs.insert(SubReg);
+ ++I;
+ }
+ if (ImpDefRegs.empty())
+ return false;
+
+ MachineBasicBlock::iterator FirstTerm = I;
+ while (I != MBB->end()) {
+ if (!TII->isUnpredicatedTerminator(I))
+ return false;
+ // See if it uses any of the implicitly defined registers.
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = I->getOperand(i);
+ if (!MO.isReg() || !MO.isUse())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (ImpDefRegs.count(Reg))
+ return false;
}
+ ++I;
}
-
- // Remove the block.
- MF->getBasicBlockList().erase(MBB);
+
+ I = MBB->begin();
+ while (I != FirstTerm) {
+ MachineInstr *ImpDefMI = &*I;
+ ++I;
+ MBB->erase(ImpDefMI);
+ }
+
+ return true;
}
-bool BranchFolder::runOnMachineFunction(MachineFunction &MF) {
- TII = MF.getTarget().getInstrInfo();
- if (!TII) return false;
+/// OptimizeFunction - Perhaps branch folding, tail merging and other
+/// CFG optimizations on the given function.
+bool BranchFolder::OptimizeFunction(MachineFunction &MF,
+ const TargetInstrInfo *tii,
+ const TargetRegisterInfo *tri,
+ MachineModuleInfo *mmi) {
+ if (!tii) return false;
+
+ TriedMerging.clear();
+
+ TII = tii;
+ TRI = tri;
+ MMI = mmi;
+ RS = NULL;
+
+ // Use a RegScavenger to help update liveness when required.
+ MachineRegisterInfo &MRI = MF.getRegInfo();
+ if (MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF))
+ RS = new RegScavenger();
+ else
+ MRI.invalidateLiveness();
+
+ // Fix CFG. The later algorithms expect it to be right.
+ bool MadeChange = false;
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) {
+ MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true))
+ MadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty());
+ MadeChange |= OptimizeImpDefsBlock(MBB);
+ }
- MDI = getAnalysisToUpdate<MachineDebugInfo>();
-
- bool EverMadeChange = false;
bool MadeChangeThisIteration = true;
while (MadeChangeThisIteration) {
- MadeChangeThisIteration = false;
- MadeChangeThisIteration |= TailMergeBlocks(MF);
- MadeChangeThisIteration |= OptimizeBranches(MF);
- EverMadeChange |= MadeChangeThisIteration;
+ 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();
- 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)
- 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.
- std::vector<bool> JTIsLive;
- JTIsLive.resize(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.isJumpTableIndex()) continue;
- unsigned NewIdx = JTMapping[Op.getJumpTableIndex()];
- Op.setJumpTableIndex(NewIdx);
-
- // Remember that this JT is live.
- JTIsLive[NewIdx] = true;
- }
- }
-
- // 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[i]) {
- JTI->RemoveJumpTable(i);
- EverMadeChange = 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());
}
}
-
- return EverMadeChange;
+
+ // 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;
}
//===----------------------------------------------------------------------===//
unsigned Hash = MI->getOpcode();
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
const MachineOperand &Op = MI->getOperand(i);
-
+
// Merge in bits from the operand if easy.
unsigned OperandHash = 0;
switch (Op.getType()) {
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:
break;
default: break;
}
-
+
Hash += ((OperandHash << 3) | Op.getType()) << (i&31);
}
return Hash;
}
-/// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two
-/// instructions, because cross-jumping only saves code when at least two
-/// instructions are removed (since a branch must be inserted).
+/// 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())
- return Hash; // Single instr MBB.
-
+
--I;
- // Hash in the second-to-last instruction.
- Hash ^= HashMachineInstr(I) << 2;
- return Hash;
+ // 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;
+ }
+
+ return HashMachineInstr(I);
}
/// ComputeCommonTailLength - Given two machine basic blocks, compute the number
MachineBasicBlock::iterator &I2) {
I1 = MBB1->end();
I2 = MBB2->end();
-
+
unsigned TailLen = 0;
while (I1 != MBB1->begin() && I2 != MBB2->begin()) {
--I1; --I2;
- if (!I1->isIdenticalTo(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->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;
}
+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();
-
- // 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, std::vector<MachineOperand>());
- OldBB->addSuccessor(NewDest);
+ MachineBasicBlock *CurMBB = OldInst->getParent();
+
+ TII->ReplaceTailWithBranchTo(OldInst, NewDest);
+
+ // For targets that use the register scavenger, we must maintain LiveIns.
+ MaintainLiveIns(CurMBB, 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.
MachineFunction::iterator MBBI = &CurMBB;
- MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock());
- CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB);
+ MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock());
+ CurMBB.getParent()->insert(++MBBI, NewMBB);
// Move all the successors of this block to the specified block.
- while (!CurMBB.succ_empty()) {
- MachineBasicBlock *S = *(CurMBB.succ_end()-1);
- NewMBB->addSuccessor(S);
- CurMBB.removeSuccessor(S);
- }
-
+ NewMBB->transferSuccessors(&CurMBB);
+
// Add an edge from CurMBB to NewMBB for the fall-through.
CurMBB.addSuccessor(NewMBB);
-
+
// Splice the code over.
NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end());
+
+ // For targets that use the register scavenger, we must maintain LiveIns.
+ MaintainLiveIns(&CurMBB, NewMBB);
+
return NewMBB;
}
/// 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)
+ if (I->isDebugValue())
+ continue;
+ if (I->isCall())
Time += 10;
- else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG))
+ else if (I->mayLoad() || I->mayStore())
Time += 2;
else
++Time;
return Time;
}
-/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at
-/// MBB2I and then insert an unconditional branch in the other block. Determine
-/// which is the best to split
-static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1,
- MachineBasicBlock::iterator MBB1I,
- MachineBasicBlock *MBB2,
- MachineBasicBlock::iterator MBB2I,
- const TargetInstrInfo *TII) {
- // 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);
-
- // 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
- // have a lower overall impact on the program execution.
- return MBB1Time < MBB2Time;
+// CurMBB needs to add an unconditional branch to SuccMBB (we removed these
+// branches temporarily for tail merging). In the case where CurMBB ends
+// with a conditional branch to the next block, optimize by reversing the
+// test and conditionally branching to SuccMBB instead.
+static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB,
+ const TargetInstrInfo *TII) {
+ MachineFunction *MF = CurMBB->getParent();
+ 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, dl);
+ return;
+ }
+ }
+ }
+ TII->InsertBranch(*CurMBB, SuccBB, NULL,
+ SmallVector<MachineOperand, 0>(), dl);
}
-bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
- MadeChange = false;
-
- if (!EnableTailMerge) return false;
-
- // Find blocks with no successors.
- std::vector<std::pair<unsigned,MachineBasicBlock*> > MergePotentials;
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
- if (I->succ_empty())
- MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I));
+bool
+BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const {
+ if (getHash() < o.getHash())
+ return true;
+ else if (getHash() > o.getHash())
+ return false;
+ else if (getBlock()->getNumber() < o.getBlock()->getNumber())
+ return true;
+ else if (getBlock()->getNumber() > o.getBlock()->getNumber())
+ return false;
+ else {
+ // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing
+ // an object with itself.
+#ifndef _GLIBCXX_DEBUG
+ llvm_unreachable("Predecessor appears twice");
+#else
+ return false;
+#endif
+ }
+}
+
+/// CountTerminators - Count the number of terminators in the given
+/// block and set I to the position of the first non-terminator, if there
+/// is one, or MBB->end() otherwise.
+static unsigned CountTerminators(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &I) {
+ I = MBB->end();
+ unsigned NumTerms = 0;
+ for (;;) {
+ if (I == MBB->begin()) {
+ I = MBB->end();
+ break;
+ }
+ --I;
+ if (!I->isTerminator()) break;
+ ++NumTerms;
+ }
+ return NumTerms;
+}
+
+/// ProfitableToMerge - Check if two machine basic blocks have a common tail
+/// and decide if it would be profitable to merge those tails. Return the
+/// length of the common tail and iterators to the first common instruction
+/// in each block.
+static bool ProfitableToMerge(MachineBasicBlock *MBB1,
+ MachineBasicBlock *MBB2,
+ unsigned minCommonTailLength,
+ unsigned &CommonTailLen,
+ MachineBasicBlock::iterator &I1,
+ MachineBasicBlock::iterator &I2,
+ MachineBasicBlock *SuccBB,
+ MachineBasicBlock *PredBB) {
+ CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2);
+ 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.
+ if (MBB1 == PredBB || MBB2 == PredBB) {
+ MachineBasicBlock::iterator I;
+ unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I);
+ if (CommonTailLen > NumTerms)
+ return true;
+ }
+
+ // If one of the blocks can be completely merged and happens to be in
+ // a position where the other could fall through into it, merge any number
+ // of instructions, because it can be done without a branch.
+ // TODO: If the blocks are not adjacent, move one of them so that they are?
+ if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin())
+ return true;
+ if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
+ return true;
+
+ // If both blocks have an unconditional branch temporarily stripped out,
+ // count that as an additional common instruction for the following
+ // heuristics.
+ unsigned EffectiveTailLen = CommonTailLen;
+ if (SuccBB && MBB1 != PredBB && MBB2 != PredBB &&
+ !MBB1->back().isBarrier() &&
+ !MBB2->back().isBarrier())
+ ++EffectiveTailLen;
+
+ // Check if the common tail is long enough to be worthwhile.
+ if (EffectiveTailLen >= minCommonTailLength)
+ return true;
+
+ // If we are optimizing for code size, 2 instructions in common is enough if
+ // 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()))
+ return true;
+
+ return false;
+}
+
+/// ComputeSameTails - Look through all the blocks in MergePotentials that have
+/// hash CurHash (guaranteed to match the last element). Build the vector
+/// SameTails of all those that have the (same) largest number of instructions
+/// in common of any pair of these blocks. SameTails entries contain an
+/// iterator into MergePotentials (from which the MachineBasicBlock can be
+/// found) and a MachineBasicBlock::iterator into that MBB indicating the
+/// instruction where the matching code sequence begins.
+/// Order of elements in SameTails is the reverse of the order in which
+/// those blocks appear in MergePotentials (where they are not necessarily
+/// consecutive).
+unsigned BranchFolder::ComputeSameTails(unsigned CurHash,
+ unsigned minCommonTailLength,
+ MachineBasicBlock *SuccBB,
+ MachineBasicBlock *PredBB) {
+ unsigned maxCommonTailLength = 0U;
+ SameTails.clear();
+ MachineBasicBlock::iterator TrialBBI1, TrialBBI2;
+ MPIterator HighestMPIter = prior(MergePotentials.end());
+ for (MPIterator CurMPIter = prior(MergePotentials.end()),
+ B = MergePotentials.begin();
+ CurMPIter != B && CurMPIter->getHash() == CurHash;
+ --CurMPIter) {
+ for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) {
+ unsigned CommonTailLen;
+ if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(),
+ minCommonTailLength,
+ CommonTailLen, TrialBBI1, TrialBBI2,
+ SuccBB, PredBB)) {
+ if (CommonTailLen > maxCommonTailLength) {
+ SameTails.clear();
+ maxCommonTailLength = CommonTailLen;
+ HighestMPIter = CurMPIter;
+ SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1));
+ }
+ if (HighestMPIter == CurMPIter &&
+ CommonTailLen == maxCommonTailLength)
+ SameTails.push_back(SameTailElt(I, TrialBBI2));
+ }
+ if (I == B)
+ break;
+ }
+ }
+ return maxCommonTailLength;
+}
+
+/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from
+/// MergePotentials, restoring branches at ends of blocks as appropriate.
+void BranchFolder::RemoveBlocksWithHash(unsigned CurHash,
+ MachineBasicBlock *SuccBB,
+ MachineBasicBlock *PredBB) {
+ MPIterator CurMPIter, B;
+ for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin();
+ CurMPIter->getHash() == CurHash;
+ --CurMPIter) {
+ // Put the unconditional branch back, if we need one.
+ MachineBasicBlock *CurMBB = CurMPIter->getBlock();
+ if (SuccBB && CurMBB != PredBB)
+ FixTail(CurMBB, SuccBB, TII);
+ if (CurMPIter == B)
+ break;
}
-
+ if (CurMPIter->getHash() != CurHash)
+ CurMPIter++;
+ MergePotentials.erase(CurMPIter, MergePotentials.end());
+}
+
+/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist
+/// only of the common tail. Create a block that does by splitting one.
+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.
+ if (SameTails[i].getBlock() == PredBB) {
+ commonTailIndex = i;
+ break;
+ }
+ // Otherwise, make a (fairly bogus) choice based on estimate of
+ // how long it will take the various blocks to execute.
+ unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(),
+ SameTails[i].getTailStartPos());
+ if (t <= TimeEstimate) {
+ TimeEstimate = t;
+ commonTailIndex = i;
+ }
+ }
+
+ MachineBasicBlock::iterator BBI =
+ 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 we split PredBB, newMBB is the new predecessor.
+ if (PredBB == MBB)
+ PredBB = newMBB;
+
+ return true;
+}
+
+// See if any of the blocks in MergePotentials (which all have a common single
+// successor, or all have no successor) can be tail-merged. If there is a
+// successor, any blocks in MergePotentials that are not tail-merged and
+// are not immediately before Succ must have an unconditional branch to
+// Succ added (but the predecessor/successor lists need no adjustment).
+// The lone predecessor of Succ that falls through into Succ,
+// if any, is given in PredBB.
+
+bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB,
+ MachineBasicBlock *PredBB) {
+ bool MadeChange = false;
+
+ // Except for the special cases below, tail-merge if there are at least
+ // this many instructions in common.
+ unsigned minCommonTailLength = TailMergeSize;
+
+ DEBUG(dbgs() << "\nTryTailMergeBlocks: ";
+ for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i)
+ dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber()
+ << (i == e-1 ? "" : ", ");
+ dbgs() << "\n";
+ if (SuccBB) {
+ dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n';
+ if (PredBB)
+ dbgs() << " which has fall-through from BB#"
+ << PredBB->getNumber() << "\n";
+ }
+ dbgs() << "Looking for common tails of at least "
+ << minCommonTailLength << " instruction"
+ << (minCommonTailLength == 1 ? "" : "s") << '\n';
+ );
+
// Sort by hash value so that blocks with identical end sequences sort
// together.
std::stable_sort(MergePotentials.begin(), MergePotentials.end());
// Walk through equivalence sets looking for actual exact matches.
while (MergePotentials.size() > 1) {
- unsigned CurHash = (MergePotentials.end()-1)->first;
- unsigned PrevHash = (MergePotentials.end()-2)->first;
- MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second;
-
- // If there is nothing that matches the hash of the current basic block,
- // give up.
- if (CurHash != PrevHash) {
- MergePotentials.pop_back();
+ unsigned CurHash = MergePotentials.back().getHash();
+
+ // Build SameTails, identifying the set of blocks with this hash code
+ // and with the maximum number of instructions in common.
+ unsigned maxCommonTailLength = ComputeSameTails(CurHash,
+ minCommonTailLength,
+ SuccBB, PredBB);
+
+ // If we didn't find any pair that has at least minCommonTailLength
+ // instructions in common, remove all blocks with this hash code and retry.
+ if (SameTails.empty()) {
+ RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
continue;
}
-
- // Determine the actual length of the shared tail between these two basic
- // blocks. Because the hash can have collisions, it's possible that this is
- // less than 2.
- MachineBasicBlock::iterator BBI1, BBI2;
- unsigned CommonTailLen =
- ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second,
- BBI1, BBI2);
-
- // If the tails don't have at least two instructions in common, see if there
- // is anything else in the equivalence class that does match.
- if (CommonTailLen < 2) {
- unsigned FoundMatch = ~0U;
- for (int i = MergePotentials.size()-2;
- i != -1 && MergePotentials[i].first == CurHash; --i) {
- CommonTailLen = ComputeCommonTailLength(CurMBB,
- MergePotentials[i].second,
- BBI1, BBI2);
- if (CommonTailLen >= 2) {
- FoundMatch = i;
+
+ // If one of the blocks is the entire common tail (and not the entry
+ // block, which we can't jump to), we can treat all blocks with this same
+ // tail at once. Use PredBB if that is one of the possibilities, as that
+ // will not introduce any extra branches.
+ MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()->
+ getParent()->begin();
+ unsigned commonTailIndex = SameTails.size();
+ // If there are two blocks, check to see if one can be made to fall through
+ // into the other.
+ if (SameTails.size() == 2 &&
+ SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) &&
+ SameTails[1].tailIsWholeBlock())
+ commonTailIndex = 1;
+ else if (SameTails.size() == 2 &&
+ SameTails[1].getBlock()->isLayoutSuccessor(
+ SameTails[0].getBlock()) &&
+ SameTails[0].tailIsWholeBlock())
+ commonTailIndex = 0;
+ else {
+ // Otherwise just pick one, favoring the fall-through predecessor if
+ // there is one.
+ for (unsigned i = 0, e = SameTails.size(); i != e; ++i) {
+ MachineBasicBlock *MBB = SameTails[i].getBlock();
+ if (MBB == EntryBB && SameTails[i].tailIsWholeBlock())
+ continue;
+ if (MBB == PredBB) {
+ commonTailIndex = i;
break;
}
+ if (SameTails[i].tailIsWholeBlock())
+ commonTailIndex = i;
}
-
- // If we didn't find anything that has at least two instructions matching
- // this one, bail out.
- if (FoundMatch == ~0U) {
- MergePotentials.pop_back();
- continue;
- }
-
- // Otherwise, move the matching block to the right position.
- std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2));
}
- MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second;
-
- // If neither block is the entire common tail, split the tail of one block
- // to make it redundant with the other tail.
- if (CurMBB->begin() != BBI1 && MBB2->begin() != BBI2) {
- if (0) { // Enable this to disable partial tail merges.
- MergePotentials.pop_back();
+ if (commonTailIndex == SameTails.size() ||
+ (SameTails[commonTailIndex].getBlock() == PredBB &&
+ !SameTails[commonTailIndex].tailIsWholeBlock())) {
+ // None of the blocks consist entirely of the common tail.
+ // Split a block so that one does.
+ if (!CreateCommonTailOnlyBlock(PredBB,
+ maxCommonTailLength, commonTailIndex)) {
+ RemoveBlocksWithHash(CurHash, SuccBB, PredBB);
continue;
}
-
- // Decide whether we want to split CurMBB or MBB2.
- if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII)) {
- CurMBB = SplitMBBAt(*CurMBB, BBI1);
- BBI1 = CurMBB->begin();
- MergePotentials.back().second = CurMBB;
- } else {
- MBB2 = SplitMBBAt(*MBB2, BBI2);
- BBI2 = MBB2->begin();
- (MergePotentials.end()-2)->second = MBB2;
- }
}
-
- if (MBB2->begin() == BBI2) {
- // Hack the end off CurMBB, making it jump to MBBI@ instead.
- ReplaceTailWithBranchTo(BBI1, MBB2);
- // This modifies CurMBB, so remove it from the worklist.
- MergePotentials.pop_back();
- } else {
- assert(CurMBB->begin() == BBI1 && "Didn't split block correctly?");
- // Hack the end off MBB2, making it jump to CurMBB instead.
- ReplaceTailWithBranchTo(BBI2, CurMBB);
- // This modifies MBB2, so remove it from the worklist.
- MergePotentials.erase(MergePotentials.end()-2);
+
+ 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(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(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(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;
}
-
return MadeChange;
}
+bool BranchFolder::TailMergeBlocks(MachineFunction &MF) {
+ bool MadeChange = false;
+ if (!EnableTailMerge) return MadeChange;
+
+ // First find blocks with no successors.
+ MergePotentials.clear();
+ 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), 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() >= 2)
+ MadeChange |= TryTailMergeBlocks(NULL, NULL);
+
+ // Look at blocks (IBB) with multiple predecessors (PBB).
+ // We change each predecessor to a canonical form, by
+ // (1) temporarily removing any unconditional branch from the predecessor
+ // to IBB, and
+ // (2) alter conditional branches so they branch to the other block
+ // not IBB; this may require adding back an unconditional branch to IBB
+ // later, where there wasn't one coming in. E.g.
+ // Bcc IBB
+ // fallthrough to QBB
+ // here becomes
+ // Bncc QBB
+ // with a conceptual B to IBB after that, which never actually exists.
+ // With those changes, we see whether the predecessors' tails match,
+ // and merge them if so. We change things out of canonical form and
+ // back to the way they were later in the process. (OptimizeBranches
+ // would undo some of this, but we can't use it, because we'd get into
+ // a compile-time infinite loop repeatedly doing and undoing the same
+ // transformations.)
+
+ for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
+ I != E; ++I) {
+ if (I->pred_size() < 2) continue;
+ 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 && 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)) {
+ // Failing case: IBB is the target of a cbr, and we cannot reverse the
+ // branch.
+ SmallVector<MachineOperand, 4> NewCond(Cond);
+ if (!Cond.empty() && TBB == IBB) {
+ if (TII->ReverseBranchCondition(NewCond))
+ continue;
+ // This is the QBB case described above
+ if (!FBB)
+ 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 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.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, dl);
+ }
+
+ 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
+ if (MergePotentials.size() == 1 &&
+ MergePotentials.begin()->getBlock() != PredBB)
+ FixTail(MergePotentials.begin()->getBlock(), IBB, TII);
+ }
+
+ return MadeChange;
+}
//===----------------------------------------------------------------------===//
// Branch Optimization
//===----------------------------------------------------------------------===//
bool BranchFolder::OptimizeBranches(MachineFunction &MF) {
- MadeChange = false;
-
- for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) {
+ bool MadeChange = false;
+
+ // Make sure blocks are numbered in order
+ MF.RenumberBlocks();
+
+ for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end();
+ I != E; ) {
MachineBasicBlock *MBB = I++;
- OptimizeBlock(MBB);
-
+ MadeChange |= OptimizeBlock(MBB);
+
// If it is dead, remove it.
if (MBB->pred_empty()) {
RemoveDeadBlock(MBB);
return MadeChange;
}
-
-/// 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 = 0;
- ++SI;
- } else if (*SI == DestB) {
- DestB = 0;
- ++SI;
- } else {
- // Otherwise, this is a superfluous edge, remove it.
- MBB.removeSuccessor(SI);
- MadeChange = true;
- }
+// 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;
}
- if (!AddedFallThrough) {
- assert(DestA == 0 && DestB == 0 &&
- "MachineCFG is missing edges!");
- } else if (isCond) {
- assert(DestA == 0 && "MachineCFG is missing edges!");
+ 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 MadeChange;
+ 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
+/// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will
+/// result in infinite loops.
+static bool IsBetterFallthrough(MachineBasicBlock *MBB1,
+ 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 (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;
-/// 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!");
+ // 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();
+}
- MachineBasicBlock::iterator I = BB->end();
- while (I != BB->begin()) {
+/// 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 (!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.
- std::vector<MachineBasicBlock*> Succs(BB->succ_begin(), BB->succ_end());
- for (int i = Succs.size()-1; i >= 0; --i)
- if (Succs[i] == Old) {
- BB->removeSuccessor(Old);
- 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
-/// 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 std::vector<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, assume it could fall through.
- if (BranchUnAnalyzable) return true;
-
- // 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)
- 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;
-}
-
-/// 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;
- std::vector<MachineOperand> Cond;
- bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond);
- return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond);
+ if (I->isBranch())
+ return I->getDebugLoc();
+ return DebugLoc();
}
/// OptimizeBlock - Analyze and optimize control flow related to the specified
/// block. This is never called on the entry block.
-void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
+bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
+ bool MadeChange = false;
+ MachineFunction &MF = *MBB->getParent();
+ReoptimizeBlock:
+
MachineFunction::iterator FallThrough = MBB;
++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. Blocks with their addresses taken shouldn't be
+ // optimized away.
+ if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) {
// Dead block? Leave for cleanup later.
- if (MBB->pred_empty()) return;
-
- if (FallThrough == MBB->getParent()->end()) {
+ if (MBB->pred_empty()) return MadeChange;
+
+ if (FallThrough == MF.end()) {
// TODO: Simplify preds to not branch here if possible!
} else {
// Rewrite all predecessors of the old block to go to the fallthrough
// 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
// fallthrough instead.
- MBB->getParent()->getJumpTableInfo()->
- ReplaceMBBInJumpTables(MBB, FallThrough);
+ if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
+ MJTI->ReplaceMBBInJumpTables(MBB, FallThrough);
MadeChange = true;
}
- return;
+ return MadeChange;
}
// Check to see if we can simplify the terminator of the block before this
MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB));
MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0;
- std::vector<MachineOperand> PriorCond;
+ SmallVector<MachineOperand, 4> PriorCond;
bool PriorUnAnalyzable =
- TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond);
+ TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true);
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
// a fall-through.
if (PriorTBB && PriorTBB == PriorFBB) {
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
TII->RemoveBranch(PrevBB);
- PriorCond.clear();
+ PriorCond.clear();
if (PriorTBB != MBB)
- TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond);
+ TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond, dl);
MadeChange = true;
++NumBranchOpts;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
+ }
+
+ // If the previous block unconditionally falls through to this block and
+ // this block has no other predecessors, move the contents of this block
+ // into the prior block. This doesn't usually happen when SimplifyCFG
+ // has been used, but it can happen if tail merging splits a fall-through
+ // predecessor of a block.
+ // This has to check PrevBB->succ_size() because EH edges are ignored by
+ // AnalyzeBranch.
+ if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 &&
+ PrevBB.succ_size() == 1 &&
+ !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());
+ assert(PrevBB.succ_empty());
+ PrevBB.transferSuccessors(MBB);
+ MadeChange = true;
+ return MadeChange;
}
-
+
// If the previous branch *only* branches to *this* block (conditional or
// not) remove the branch.
if (PriorTBB == MBB && PriorFBB == 0) {
TII->RemoveBranch(PrevBB);
MadeChange = true;
++NumBranchOpts;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
}
-
+
// 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;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
}
-
+
// If the prior block branches here on true and somewhere else on false, and
// if the branch condition is reversible, reverse the branch to create a
// fall-through.
if (PriorTBB == MBB) {
- std::vector<MachineOperand> NewPriorCond(PriorCond);
+ 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;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
+ }
+ }
+
+ // If this block has no successors (e.g. it is a return block or ends with
+ // a call to a no-return function like abort or __cxa_throw) and if the pred
+ // falls through into this block, and if it would otherwise fall through
+ // into the block after this, move this block to the end of the function.
+ //
+ // We consider it more likely that execution will stay in the function (e.g.
+ // due to loops) than it is to exit it. This asserts in loops etc, moving
+ // the assert condition out of the loop body.
+ if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
+ MachineFunction::iterator(PriorTBB) == FallThrough &&
+ !MBB->canFallThrough()) {
+ bool DoTransform = true;
+
+ // We have to be careful that the succs of PredBB aren't both no-successor
+ // blocks. If neither have successors and if PredBB is the second from
+ // last block in the function, we'd just keep swapping the two blocks for
+ // last. Only do the swap if one is clearly better to fall through than
+ // the other.
+ if (FallThrough == --MF.end() &&
+ !IsBetterFallthrough(PriorTBB, MBB))
+ 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(dbgs() << "\nMoving MBB: " << *MBB
+ << "To make fallthrough to: " << *PriorTBB << "\n");
+
+ DebugLoc dl = getBranchDebugLoc(PrevBB);
+ TII->RemoveBranch(PrevBB);
+ TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond, dl);
+
+ // Move this block to the end of the function.
+ MBB->moveAfter(--MF.end());
+ MadeChange = true;
+ ++NumBranchOpts;
+ return MadeChange;
+ }
}
}
}
-
+
// Analyze the branch in the current block.
MachineBasicBlock *CurTBB = 0, *CurFBB = 0;
- std::vector<MachineOperand> CurCond;
- bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond);
+ SmallVector<MachineOperand, 4> CurCond;
+ bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true);
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
+ // 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:
// Loop: xxx; jcc Out; jmp Loop
// we want:
// Loop: xxx; jncc Loop; jmp Out
if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) {
- std::vector<MachineOperand> NewCond(CurCond);
+ 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;
- return OptimizeBlock(MBB);
+ 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 &&
- TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) {
+ if (CurTBB && CurCond.empty() && CurFBB == 0 &&
+ 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
// 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
// explicit branch to us to make updates simpler.
- if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
+ if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) &&
PriorTBB != MBB && PriorFBB != MBB) {
if (PriorTBB == 0) {
assert(PriorCond.empty() && PriorFBB == 0 &&
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.
- 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);
+ // If this change resulted in PMBB ending in a conditional
+ // branch where both conditions go to the same destination,
+ // change this to an unconditional branch (and fix the CFG).
+ MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0;
+ SmallVector<MachineOperand, 4> NewCurCond;
+ 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, pdl);
+ MadeChange = true;
+ ++NumBranchOpts;
+ PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false);
+ }
}
}
// Change any jumptables to go to the new MBB.
- MBB->getParent()->getJumpTableInfo()->
- ReplaceMBBInJumpTables(MBB, CurTBB);
+ if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo())
+ MJTI->ReplaceMBBInJumpTables(MBB, CurTBB);
if (DidChange) {
++NumBranchOpts;
MadeChange = true;
- if (!HasBranchToSelf) return;
+ if (!HasBranchToSelf) return MadeChange;
}
}
}
-
+
// 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.
- if (!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
// throughs, move this block right after it.
for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
// Analyze the branch at the end of the pred.
MachineBasicBlock *PredBB = *PI;
MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough;
- std::vector<MachineOperand> PredCond;
- if (PredBB != MBB && !CanFallThrough(PredBB)) {
+ MachineBasicBlock *PredTBB = 0, *PredFBB = 0;
+ SmallVector<MachineOperand, 4> PredCond;
+ if (PredBB != MBB && !PredBB->canFallThrough() &&
+ !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)
+ && (!CurFallsThru || !CurTBB || !CurFBB)
+ && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) {
+ // If the current block doesn't fall through, just move it.
+ // If the current block can fall through and does not end with a
+ // conditional branch, we need to append an unconditional jump to
+ // the (current) next block. To avoid a possible compile-time
+ // infinite loop, move blocks only backward in this case.
+ // Also, if there are already 2 branches here, we cannot add a third;
+ // this means we have the case
+ // Bcc next
+ // B elsewhere
+ // next:
+ if (CurFallsThru) {
+ MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB));
+ CurCond.clear();
+ TII->InsertBranch(*MBB, NextBB, 0, CurCond, DebugLoc());
+ }
MBB->moveAfter(PredBB);
MadeChange = true;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
}
}
-
+ }
+
+ if (!CurFallsThru) {
// Check all successors to see if we can move this block before it.
for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
E = MBB->succ_end(); SI != E; ++SI) {
// Analyze the branch at the end of the block before the succ.
MachineBasicBlock *SuccBB = *SI;
MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev;
- std::vector<MachineOperand> SuccPrevCond;
- if (SuccBB != MBB && !CanFallThrough(SuccPrev)) {
+
+ // If this block doesn't already fall-through to that successor, and if
+ // the succ doesn't already have a block that can fall through into it,
+ // and if the successor isn't an EH destination, we can arrange for the
+ // fallthrough to happen.
+ if (SuccBB != MBB && &*SuccPrev != MBB &&
+ !SuccPrev->canFallThrough() && !CurUnAnalyzable &&
+ !SuccBB->isLandingPad()) {
MBB->moveBefore(SuccBB);
MadeChange = true;
- return OptimizeBlock(MBB);
+ goto ReoptimizeBlock;
}
}
-
+
// 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.
- if (FallThrough != MBB->getParent()->end() &&
+ MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0;
+ SmallVector<MachineOperand, 4> PrevCond;
+ if (FallThrough != MF.end() &&
+ !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) &&
PrevBB.isSuccessor(FallThrough)) {
- MBB->moveAfter(--MBB->getParent()->end());
+ MBB->moveAfter(--MF.end());
MadeChange = true;
- return;
+ return MadeChange;
}
}
}
+
+ 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;
}
projects / oota-llvm.git / blobdiff