#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Function.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include <algorithm>
using namespace llvm;
STATISTIC(NumAtomic, "Number of atomic phis lowered");
-STATISTIC(NumSplits, "Number of critical edges split on demand");
+STATISTIC(NumReused, "Number of reused lowered phis");
char PHIElimination::ID = 0;
-static RegisterPass<PHIElimination>
-X("phi-node-elimination", "Eliminate PHI nodes for register allocation");
+INITIALIZE_PASS(PHIElimination, "phi-node-elimination",
+ "Eliminate PHI nodes for register allocation", false, false)
-const PassInfo *const llvm::PHIEliminationID = &X;
+char &llvm::PHIEliminationID = PHIElimination::ID;
void llvm::PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addPreserved<LiveVariables>();
AU.addPreserved<MachineDominatorTree>();
- // rdar://7401784 This would be nice:
- // AU.addPreservedID(MachineLoopInfoID);
+ AU.addPreserved<MachineLoopInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
-bool llvm::PHIElimination::runOnMachineFunction(MachineFunction &Fn) {
- MRI = &Fn.getRegInfo();
+bool llvm::PHIElimination::runOnMachineFunction(MachineFunction &MF) {
+ MRI = &MF.getRegInfo();
- PHIDefs.clear();
- PHIKills.clear();
bool Changed = false;
// Split critical edges to help the coalescer
- if (LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>())
- for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
- Changed |= SplitPHIEdges(Fn, *I, *LV);
+ if (LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>()) {
+ MachineLoopInfo *MLI = getAnalysisIfAvailable<MachineLoopInfo>();
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
+ Changed |= SplitPHIEdges(MF, *I, *LV, MLI);
+ }
// Populate VRegPHIUseCount
- analyzePHINodes(Fn);
+ analyzePHINodes(MF);
// Eliminate PHI instructions by inserting copies into predecessor blocks.
- for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
- Changed |= EliminatePHINodes(Fn, *I);
+ for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
+ Changed |= EliminatePHINodes(MF, *I);
// Remove dead IMPLICIT_DEF instructions.
- for (SmallPtrSet<MachineInstr*,4>::iterator I = ImpDefs.begin(),
+ for (SmallPtrSet<MachineInstr*, 4>::iterator I = ImpDefs.begin(),
E = ImpDefs.end(); I != E; ++I) {
MachineInstr *DefMI = *I;
unsigned DefReg = DefMI->getOperand(0).getReg();
- if (MRI->use_empty(DefReg))
+ if (MRI->use_nodbg_empty(DefReg))
DefMI->eraseFromParent();
}
+ // Clean up the lowered PHI instructions.
+ for (LoweredPHIMap::iterator I = LoweredPHIs.begin(), E = LoweredPHIs.end();
+ I != E; ++I)
+ MF.DeleteMachineInstr(I->first);
+
+ LoweredPHIs.clear();
ImpDefs.clear();
VRegPHIUseCount.clear();
+
return Changed;
}
///
bool llvm::PHIElimination::EliminatePHINodes(MachineFunction &MF,
MachineBasicBlock &MBB) {
- if (MBB.empty() || MBB.front().getOpcode() != TargetInstrInfo::PHI)
+ if (MBB.empty() || !MBB.front().isPHI())
return false; // Quick exit for basic blocks without PHIs.
// Get an iterator to the first instruction after the last PHI node (this may
// also be the end of the basic block).
- MachineBasicBlock::iterator AfterPHIsIt = SkipPHIsAndLabels(MBB, MBB.begin());
+ MachineBasicBlock::iterator AfterPHIsIt = MBB.SkipPHIsAndLabels(MBB.begin());
- while (MBB.front().getOpcode() == TargetInstrInfo::PHI)
+ while (MBB.front().isPHI())
LowerAtomicPHINode(MBB, AfterPHIsIt);
return true;
for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2) {
unsigned SrcReg = MPhi->getOperand(i).getReg();
const MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
- if (!DefMI || DefMI->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
+ if (!DefMI || !DefMI->isImplicitDef())
return false;
}
return true;
}
// Make sure the copy goes after any phi nodes however.
- return SkipPHIsAndLabels(MBB, InsertPoint);
+ return MBB.SkipPHIsAndLabels(InsertPoint);
}
/// LowerAtomicPHINode - Lower the PHI node at the top of the specified block,
void llvm::PHIElimination::LowerAtomicPHINode(
MachineBasicBlock &MBB,
MachineBasicBlock::iterator AfterPHIsIt) {
+ ++NumAtomic;
// Unlink the PHI node from the basic block, but don't delete the PHI yet.
MachineInstr *MPhi = MBB.remove(MBB.begin());
unsigned NumSrcs = (MPhi->getNumOperands() - 1) / 2;
unsigned DestReg = MPhi->getOperand(0).getReg();
+ assert(MPhi->getOperand(0).getSubReg() == 0 && "Can't handle sub-reg PHIs");
bool isDead = MPhi->getOperand(0).isDead();
// Create a new register for the incoming PHI arguments.
MachineFunction &MF = *MBB.getParent();
- const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg);
unsigned IncomingReg = 0;
+ bool reusedIncoming = false; // Is IncomingReg reused from an earlier PHI?
// Insert a register to register copy at the top of the current block (but
// after any remaining phi nodes) which copies the new incoming register
// If all sources of a PHI node are implicit_def, just emit an
// implicit_def instead of a copy.
BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
- TII->get(TargetInstrInfo::IMPLICIT_DEF), DestReg);
+ TII->get(TargetOpcode::IMPLICIT_DEF), DestReg);
else {
- IncomingReg = MF.getRegInfo().createVirtualRegister(RC);
- TII->copyRegToReg(MBB, AfterPHIsIt, DestReg, IncomingReg, RC, RC);
+ // Can we reuse an earlier PHI node? This only happens for critical edges,
+ // typically those created by tail duplication.
+ unsigned &entry = LoweredPHIs[MPhi];
+ if (entry) {
+ // An identical PHI node was already lowered. Reuse the incoming register.
+ IncomingReg = entry;
+ reusedIncoming = true;
+ ++NumReused;
+ DEBUG(dbgs() << "Reusing %reg" << IncomingReg << " for " << *MPhi);
+ } else {
+ const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg);
+ entry = IncomingReg = MF.getRegInfo().createVirtualRegister(RC);
+ }
+ BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), DestReg)
+ .addReg(IncomingReg);
}
- // Record PHI def.
- assert(!hasPHIDef(DestReg) && "Vreg has multiple phi-defs?");
- PHIDefs[DestReg] = &MBB;
-
// Update live variable information if there is any.
LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>();
if (LV) {
MachineInstr *PHICopy = prior(AfterPHIsIt);
if (IncomingReg) {
+ LiveVariables::VarInfo &VI = LV->getVarInfo(IncomingReg);
+
// Increment use count of the newly created virtual register.
- LV->getVarInfo(IncomingReg).NumUses++;
+ VI.NumUses++;
+ LV->setPHIJoin(IncomingReg);
+
+ // When we are reusing the incoming register, it may already have been
+ // killed in this block. The old kill will also have been inserted at
+ // AfterPHIsIt, so it appears before the current PHICopy.
+ if (reusedIncoming)
+ if (MachineInstr *OldKill = VI.findKill(&MBB)) {
+ DEBUG(dbgs() << "Remove old kill from " << *OldKill);
+ LV->removeVirtualRegisterKilled(IncomingReg, OldKill);
+ DEBUG(MBB.dump());
+ }
// Add information to LiveVariables to know that the incoming value is
// killed. Note that because the value is defined in several places (once
// Adjust the VRegPHIUseCount map to account for the removal of this PHI node.
for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2)
- --VRegPHIUseCount[BBVRegPair(MPhi->getOperand(i + 1).getMBB(),
+ --VRegPHIUseCount[BBVRegPair(MPhi->getOperand(i+1).getMBB()->getNumber(),
MPhi->getOperand(i).getReg())];
// Now loop over all of the incoming arguments, changing them to copy into the
SmallPtrSet<MachineBasicBlock*, 8> MBBsInsertedInto;
for (int i = NumSrcs - 1; i >= 0; --i) {
unsigned SrcReg = MPhi->getOperand(i*2+1).getReg();
+ unsigned SrcSubReg = MPhi->getOperand(i*2+1).getSubReg();
+
assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
"Machine PHI Operands must all be virtual registers!");
// path the PHI.
MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB();
- // Record the kill.
- PHIKills[SrcReg].insert(&opBlock);
-
// If source is defined by an implicit def, there is no need to insert a
// copy.
MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
- if (DefMI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF) {
+ if (DefMI->isImplicitDef()) {
ImpDefs.insert(DefMI);
continue;
}
FindCopyInsertPoint(opBlock, MBB, SrcReg);
// Insert the copy.
- TII->copyRegToReg(opBlock, InsertPos, IncomingReg, SrcReg, RC, RC);
+ if (!reusedIncoming && IncomingReg)
+ BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), IncomingReg).addReg(SrcReg, 0, SrcSubReg);
// Now update live variable information if we have it. Otherwise we're done
if (!LV) continue;
// point later.
// Is it used by any PHI instructions in this block?
- bool ValueIsUsed = VRegPHIUseCount[BBVRegPair(&opBlock, SrcReg)] != 0;
+ bool ValueIsUsed = VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)];
// Okay, if we now know that the value is not live out of the block, we can
// add a kill marker in this block saying that it kills the incoming value!
// terminator instruction at the end of the block may also use the value.
// In this case, we should mark *it* as being the killing block, not the
// copy.
- MachineBasicBlock::iterator KillInst = prior(InsertPos);
+ MachineBasicBlock::iterator KillInst;
MachineBasicBlock::iterator Term = opBlock.getFirstTerminator();
- if (Term != opBlock.end()) {
- if (Term->readsRegister(SrcReg))
- KillInst = Term;
+ if (Term != opBlock.end() && Term->readsRegister(SrcReg)) {
+ KillInst = Term;
// Check that no other terminators use values.
#ifndef NDEBUG
- for (MachineBasicBlock::iterator TI = next(Term); TI != opBlock.end();
- ++TI) {
+ for (MachineBasicBlock::iterator TI = llvm::next(Term);
+ TI != opBlock.end(); ++TI) {
assert(!TI->readsRegister(SrcReg) &&
"Terminator instructions cannot use virtual registers unless"
"they are the first terminator in a block!");
}
#endif
+ } else if (reusedIncoming || !IncomingReg) {
+ // We may have to rewind a bit if we didn't insert a copy this time.
+ KillInst = Term;
+ while (KillInst != opBlock.begin())
+ if ((--KillInst)->readsRegister(SrcReg))
+ break;
+ } else {
+ // We just inserted this copy.
+ KillInst = prior(InsertPos);
}
+ assert(KillInst->readsRegister(SrcReg) && "Cannot find kill instruction");
// Finally, mark it killed.
LV->addVirtualRegisterKilled(SrcReg, KillInst);
}
}
- // Really delete the PHI instruction now!
- MF.DeleteMachineInstr(MPhi);
- ++NumAtomic;
+ // Really delete the PHI instruction now, if it is not in the LoweredPHIs map.
+ if (reusedIncoming || !IncomingReg)
+ MF.DeleteMachineInstr(MPhi);
}
/// analyzePHINodes - Gather information about the PHI nodes in here. In
/// used in a PHI node. We map that to the BB the vreg is coming from. This is
/// used later to determine when the vreg is killed in the BB.
///
-void llvm::PHIElimination::analyzePHINodes(const MachineFunction& Fn) {
- for (MachineFunction::const_iterator I = Fn.begin(), E = Fn.end();
+void llvm::PHIElimination::analyzePHINodes(const MachineFunction& MF) {
+ for (MachineFunction::const_iterator I = MF.begin(), E = MF.end();
I != E; ++I)
for (MachineBasicBlock::const_iterator BBI = I->begin(), BBE = I->end();
- BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI)
+ BBI != BBE && BBI->isPHI(); ++BBI)
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2)
- ++VRegPHIUseCount[BBVRegPair(BBI->getOperand(i + 1).getMBB(),
+ ++VRegPHIUseCount[BBVRegPair(BBI->getOperand(i+1).getMBB()->getNumber(),
BBI->getOperand(i).getReg())];
}
bool llvm::PHIElimination::SplitPHIEdges(MachineFunction &MF,
MachineBasicBlock &MBB,
- LiveVariables &LV) {
- if (MBB.empty() || MBB.front().getOpcode() != TargetInstrInfo::PHI)
+ LiveVariables &LV,
+ MachineLoopInfo *MLI) {
+ if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad())
return false; // Quick exit for basic blocks without PHIs.
+ bool Changed = false;
for (MachineBasicBlock::const_iterator BBI = MBB.begin(), BBE = MBB.end();
- BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI) {
+ BBI != BBE && BBI->isPHI(); ++BBI) {
for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) {
unsigned Reg = BBI->getOperand(i).getReg();
MachineBasicBlock *PreMBB = BBI->getOperand(i+1).getMBB();
// We break edges when registers are live out from the predecessor block
// (not considering PHI nodes). If the register is live in to this block
// anyway, we would gain nothing from splitting.
- if (!LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB))
- SplitCriticalEdge(PreMBB, &MBB);
+ // Avoid splitting backedges of loops. It would introduce small
+ // out-of-line blocks into the loop which is very bad for code placement.
+ if (PreMBB != &MBB &&
+ !LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB)) {
+ if (!MLI ||
+ !(MLI->getLoopFor(PreMBB) == MLI->getLoopFor(&MBB) &&
+ MLI->isLoopHeader(&MBB)))
+ Changed |= PreMBB->SplitCriticalEdge(&MBB, this) != 0;
+ }
}
}
return true;
}
-
-MachineBasicBlock *PHIElimination::SplitCriticalEdge(MachineBasicBlock *A,
- MachineBasicBlock *B) {
- assert(A && B && "Missing MBB end point");
-
- MachineFunction *MF = A->getParent();
-
- // We may need to update A's terminator, but we can't do that if AnalyzeBranch
- // fails. If A uses a jump table, we won't touch it.
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
- MachineBasicBlock *TBB = 0, *FBB = 0;
- SmallVector<MachineOperand, 4> Cond;
- if (TII->AnalyzeBranch(*A, TBB, FBB, Cond))
- return NULL;
-
- ++NumSplits;
-
- MachineBasicBlock *NMBB = MF->CreateMachineBasicBlock();
- MF->insert(next(MachineFunction::iterator(A)), NMBB);
- DEBUG(errs() << "PHIElimination splitting critical edge:"
- " BB#" << A->getNumber()
- << " -- BB#" << NMBB->getNumber()
- << " -- BB#" << B->getNumber() << '\n');
-
- A->ReplaceUsesOfBlockWith(B, NMBB);
- A->updateTerminator();
-
- // Insert unconditional "jump B" instruction in NMBB if necessary.
- NMBB->addSuccessor(B);
- if (!NMBB->isLayoutSuccessor(B)) {
- Cond.clear();
- MF->getTarget().getInstrInfo()->InsertBranch(*NMBB, B, NULL, Cond);
- }
-
- // Fix PHI nodes in B so they refer to NMBB instead of A
- for (MachineBasicBlock::iterator i = B->begin(), e = B->end();
- i != e && i->getOpcode() == TargetInstrInfo::PHI; ++i)
- for (unsigned ni = 1, ne = i->getNumOperands(); ni != ne; ni += 2)
- if (i->getOperand(ni+1).getMBB() == A)
- i->getOperand(ni+1).setMBB(NMBB);
-
- if (LiveVariables *LV=getAnalysisIfAvailable<LiveVariables>())
- LV->addNewBlock(NMBB, A, B);
-
- if (MachineDominatorTree *MDT=getAnalysisIfAvailable<MachineDominatorTree>())
- MDT->addNewBlock(NMBB, A);
-
- return NMBB;
-}