//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "phielim"
-#include "PHIElimination.h"
-#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/Passes.h"
+#include "PHIEliminationUtils.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/LiveVariables.h"
#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/Function.h"
-#include "llvm/Target/TargetMachine.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/Statistic.h"
+#include "llvm/IR/Function.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
#include <algorithm>
-#include <map>
using namespace llvm;
-STATISTIC(NumAtomic, "Number of atomic phis lowered");
-STATISTIC(NumSplits, "Number of critical edges split on demand");
+#define DEBUG_TYPE "phielim"
+
+static cl::opt<bool>
+DisableEdgeSplitting("disable-phi-elim-edge-splitting", cl::init(false),
+ cl::Hidden, cl::desc("Disable critical edge splitting "
+ "during PHI elimination"));
+
+static cl::opt<bool>
+SplitAllCriticalEdges("phi-elim-split-all-critical-edges", cl::init(false),
+ cl::Hidden, cl::desc("Split all critical edges during "
+ "PHI elimination"));
+
+static cl::opt<bool> NoPhiElimLiveOutEarlyExit(
+ "no-phi-elim-live-out-early-exit", cl::init(false), cl::Hidden,
+ cl::desc("Do not use an early exit if isLiveOutPastPHIs returns true."));
+
+namespace {
+ class PHIElimination : public MachineFunctionPass {
+ MachineRegisterInfo *MRI; // Machine register information
+ LiveVariables *LV;
+ LiveIntervals *LIS;
+
+ public:
+ static char ID; // Pass identification, replacement for typeid
+ PHIElimination() : MachineFunctionPass(ID) {
+ initializePHIEliminationPass(*PassRegistry::getPassRegistry());
+ }
+
+ bool runOnMachineFunction(MachineFunction &Fn) override;
+ void getAnalysisUsage(AnalysisUsage &AU) const override;
+
+ private:
+ /// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions
+ /// in predecessor basic blocks.
+ ///
+ bool EliminatePHINodes(MachineFunction &MF, MachineBasicBlock &MBB);
+ void LowerPHINode(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator LastPHIIt);
+
+ /// analyzePHINodes - Gather information about the PHI nodes in
+ /// here. In particular, we want to map the number of uses of a virtual
+ /// register which is 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 analyzePHINodes(const MachineFunction& Fn);
+
+ /// Split critical edges where necessary for good coalescer performance.
+ bool SplitPHIEdges(MachineFunction &MF, MachineBasicBlock &MBB,
+ MachineLoopInfo *MLI);
+
+ // These functions are temporary abstractions around LiveVariables and
+ // LiveIntervals, so they can go away when LiveVariables does.
+ bool isLiveIn(unsigned Reg, const MachineBasicBlock *MBB);
+ bool isLiveOutPastPHIs(unsigned Reg, const MachineBasicBlock *MBB);
+
+ typedef std::pair<unsigned, unsigned> BBVRegPair;
+ typedef DenseMap<BBVRegPair, unsigned> VRegPHIUse;
+
+ VRegPHIUse VRegPHIUseCount;
+
+ // Defs of PHI sources which are implicit_def.
+ SmallPtrSet<MachineInstr*, 4> ImpDefs;
+
+ // Map reusable lowered PHI node -> incoming join register.
+ typedef DenseMap<MachineInstr*, unsigned,
+ MachineInstrExpressionTrait> LoweredPHIMap;
+ LoweredPHIMap LoweredPHIs;
+ };
+}
+
+STATISTIC(NumLowered, "Number of phis lowered");
+STATISTIC(NumCriticalEdgesSplit, "Number of critical edges split");
+STATISTIC(NumReused, "Number of reused lowered phis");
char PHIElimination::ID = 0;
-static RegisterPass<PHIElimination>
-X("phi-node-elimination", "Eliminate PHI nodes for register allocation");
+char& llvm::PHIEliminationID = PHIElimination::ID;
-const PassInfo *const llvm::PHIEliminationID = &X;
+INITIALIZE_PASS_BEGIN(PHIElimination, "phi-node-elimination",
+ "Eliminate PHI nodes for register allocation",
+ false, false)
+INITIALIZE_PASS_DEPENDENCY(LiveVariables)
+INITIALIZE_PASS_END(PHIElimination, "phi-node-elimination",
+ "Eliminate PHI nodes for register allocation", false, false)
-void llvm::PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const {
+void PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addPreserved<LiveVariables>();
+ AU.addPreserved<SlotIndexes>();
+ AU.addPreserved<LiveIntervals>();
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 PHIElimination::runOnMachineFunction(MachineFunction &MF) {
+ MRI = &MF.getRegInfo();
+ LV = getAnalysisIfAvailable<LiveVariables>();
+ LIS = getAnalysisIfAvailable<LiveIntervals>();
- 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);
+ // This pass takes the function out of SSA form.
+ MRI->leaveSSA();
+
+ // Split critical edges to help the coalescer. This does not yet support
+ // updating LiveIntervals, so we disable it.
+ if (!DisableEdgeSplitting && (LV || LIS)) {
+ MachineLoopInfo *MLI = getAnalysisIfAvailable<MachineLoopInfo>();
+ for (auto &MBB : MF)
+ Changed |= SplitPHIEdges(MF, MBB, 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 (auto &MBB : MF)
+ Changed |= EliminatePHINodes(MF, MBB);
// Remove dead IMPLICIT_DEF instructions.
- for (SmallPtrSet<MachineInstr*,4>::iterator I = ImpDefs.begin(),
- E = ImpDefs.end(); I != E; ++I) {
- MachineInstr *DefMI = *I;
+ for (MachineInstr *DefMI : ImpDefs) {
unsigned DefReg = DefMI->getOperand(0).getReg();
- if (MRI->use_empty(DefReg))
+ if (MRI->use_nodbg_empty(DefReg)) {
+ if (LIS)
+ LIS->RemoveMachineInstrFromMaps(DefMI);
DefMI->eraseFromParent();
+ }
+ }
+
+ // Clean up the lowered PHI instructions.
+ for (LoweredPHIMap::iterator I = LoweredPHIs.begin(), E = LoweredPHIs.end();
+ I != E; ++I) {
+ if (LIS)
+ LIS->RemoveMachineInstrFromMaps(I->first);
+ MF.DeleteMachineInstr(I->first);
}
+ LoweredPHIs.clear();
ImpDefs.clear();
VRegPHIUseCount.clear();
+
return Changed;
}
/// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions in
/// predecessor basic blocks.
///
-bool llvm::PHIElimination::EliminatePHINodes(MachineFunction &MF,
+bool 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 LastPHIIt =
+ std::prev(MBB.SkipPHIsAndLabels(MBB.begin()));
+
+ while (MBB.front().isPHI())
+ LowerPHINode(MBB, LastPHIIt);
- while (MBB.front().getOpcode() == TargetInstrInfo::PHI)
- LowerAtomicPHINode(MBB, AfterPHIsIt);
+ return true;
+}
+/// isImplicitlyDefined - Return true if all defs of VirtReg are implicit-defs.
+/// This includes registers with no defs.
+static bool isImplicitlyDefined(unsigned VirtReg,
+ const MachineRegisterInfo *MRI) {
+ for (MachineInstr &DI : MRI->def_instructions(VirtReg))
+ if (!DI.isImplicitDef())
+ return false;
return true;
}
/// are implicit_def's.
static bool isSourceDefinedByImplicitDef(const MachineInstr *MPhi,
const MachineRegisterInfo *MRI) {
- 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)
+ for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2)
+ if (!isImplicitlyDefined(MPhi->getOperand(i).getReg(), MRI))
return false;
- }
return true;
}
-// FindCopyInsertPoint - Find a safe place in MBB to insert a copy from SrcReg
-// when following the CFG edge to SuccMBB. This needs to be after any def of
-// SrcReg, but before any subsequent point where control flow might jump out of
-// the basic block.
-MachineBasicBlock::iterator
-llvm::PHIElimination::FindCopyInsertPoint(MachineBasicBlock &MBB,
- MachineBasicBlock &SuccMBB,
- unsigned SrcReg) {
- // Handle the trivial case trivially.
- if (MBB.empty())
- return MBB.begin();
-
- // Usually, we just want to insert the copy before the first terminator
- // instruction. However, for the edge going to a landing pad, we must insert
- // the copy before the call/invoke instruction.
- if (!SuccMBB.isLandingPad())
- return MBB.getFirstTerminator();
-
- // Discover any defs/uses in this basic block.
- SmallPtrSet<MachineInstr*, 8> DefUsesInMBB;
- for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg),
- RE = MRI->reg_end(); RI != RE; ++RI) {
- MachineInstr *DefUseMI = &*RI;
- if (DefUseMI->getParent() == &MBB)
- DefUsesInMBB.insert(DefUseMI);
- }
- MachineBasicBlock::iterator InsertPoint;
- if (DefUsesInMBB.empty()) {
- // No defs. Insert the copy at the start of the basic block.
- InsertPoint = MBB.begin();
- } else if (DefUsesInMBB.size() == 1) {
- // Insert the copy immediately after the def/use.
- InsertPoint = *DefUsesInMBB.begin();
- ++InsertPoint;
- } else {
- // Insert the copy immediately after the last def/use.
- InsertPoint = MBB.end();
- while (!DefUsesInMBB.count(&*--InsertPoint)) {}
- ++InsertPoint;
- }
+/// LowerPHINode - Lower the PHI node at the top of the specified block,
+///
+void PHIElimination::LowerPHINode(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator LastPHIIt) {
+ ++NumLowered;
- // Make sure the copy goes after any phi nodes however.
- return SkipPHIsAndLabels(MBB, InsertPoint);
-}
+ MachineBasicBlock::iterator AfterPHIsIt = std::next(LastPHIIt);
-/// LowerAtomicPHINode - Lower the PHI node at the top of the specified block,
-/// under the assuption that it needs to be lowered in a way that supports
-/// atomic execution of PHIs. This lowering method is always correct all of the
-/// time.
-///
-void llvm::PHIElimination::LowerAtomicPHINode(
- MachineBasicBlock &MBB,
- MachineBasicBlock::iterator AfterPHIsIt) {
// 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
// into the phi node destination.
- const TargetInstrInfo *TII = MF.getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
if (isSourceDefinedByImplicitDef(MPhi, MRI))
// 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 " << PrintReg(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);
+ MachineInstr *PHICopy = std::prev(AfterPHIsIt);
if (IncomingReg) {
+ LiveVariables::VarInfo &VI = LV->getVarInfo(IncomingReg);
+
// Increment use count of the newly created virtual register.
- LV->getVarInfo(IncomingReg).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
}
}
+ // Update LiveIntervals for the new copy or implicit def.
+ if (LIS) {
+ MachineInstr *NewInstr = std::prev(AfterPHIsIt);
+ SlotIndex DestCopyIndex = LIS->InsertMachineInstrInMaps(NewInstr);
+
+ SlotIndex MBBStartIndex = LIS->getMBBStartIdx(&MBB);
+ if (IncomingReg) {
+ // Add the region from the beginning of MBB to the copy instruction to
+ // IncomingReg's live interval.
+ LiveInterval &IncomingLI = LIS->createEmptyInterval(IncomingReg);
+ VNInfo *IncomingVNI = IncomingLI.getVNInfoAt(MBBStartIndex);
+ if (!IncomingVNI)
+ IncomingVNI = IncomingLI.getNextValue(MBBStartIndex,
+ LIS->getVNInfoAllocator());
+ IncomingLI.addSegment(LiveInterval::Segment(MBBStartIndex,
+ DestCopyIndex.getRegSlot(),
+ IncomingVNI));
+ }
+
+ LiveInterval &DestLI = LIS->getInterval(DestReg);
+ assert(DestLI.begin() != DestLI.end() &&
+ "PHIs should have nonempty LiveIntervals.");
+ if (DestLI.endIndex().isDead()) {
+ // A dead PHI's live range begins and ends at the start of the MBB, but
+ // the lowered copy, which will still be dead, needs to begin and end at
+ // the copy instruction.
+ VNInfo *OrigDestVNI = DestLI.getVNInfoAt(MBBStartIndex);
+ assert(OrigDestVNI && "PHI destination should be live at block entry.");
+ DestLI.removeSegment(MBBStartIndex, MBBStartIndex.getDeadSlot());
+ DestLI.createDeadDef(DestCopyIndex.getRegSlot(),
+ LIS->getVNInfoAllocator());
+ DestLI.removeValNo(OrigDestVNI);
+ } else {
+ // Otherwise, remove the region from the beginning of MBB to the copy
+ // instruction from DestReg's live interval.
+ DestLI.removeSegment(MBBStartIndex, DestCopyIndex.getRegSlot());
+ VNInfo *DestVNI = DestLI.getVNInfoAt(DestCopyIndex.getRegSlot());
+ assert(DestVNI && "PHI destination should be live at its definition.");
+ DestVNI->def = DestCopyIndex.getRegSlot();
+ }
+ }
+
// 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();
+ bool SrcUndef = MPhi->getOperand(i*2+1).isUndef() ||
+ isImplicitlyDefined(SrcReg, MRI);
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) {
- ImpDefs.insert(DefMI);
- continue;
- }
-
// Check to make sure we haven't already emitted the copy for this block.
// This can happen because PHI nodes may have multiple entries for the same
// basic block.
- if (!MBBsInsertedInto.insert(&opBlock))
+ if (!MBBsInsertedInto.insert(&opBlock).second)
continue; // If the copy has already been emitted, we're done.
// Find a safe location to insert the copy, this may be the first terminator
// in the block (or end()).
MachineBasicBlock::iterator InsertPos =
- FindCopyInsertPoint(opBlock, MBB, SrcReg);
+ findPHICopyInsertPoint(&opBlock, &MBB, SrcReg);
// Insert the copy.
- TII->copyRegToReg(opBlock, InsertPos, IncomingReg, SrcReg, RC, RC);
-
- // Now update live variable information if we have it. Otherwise we're done
- if (!LV) continue;
-
- // We want to be able to insert a kill of the register if this PHI (aka, the
- // copy we just inserted) is the last use of the source value. Live
- // variable analysis conservatively handles this by saying that the value is
- // live until the end of the block the PHI entry lives in. If the value
- // really is dead at the PHI copy, there will be no successor blocks which
- // have the value live-in.
-
- // Also check to see if this register is in use by another PHI node which
- // has not yet been eliminated. If so, it will be killed at an appropriate
- // point later.
+ MachineInstr *NewSrcInstr = nullptr;
+ if (!reusedIncoming && IncomingReg) {
+ if (SrcUndef) {
+ // The source register is undefined, so there is no need for a real
+ // COPY, but we still need to ensure joint dominance by defs.
+ // Insert an IMPLICIT_DEF instruction.
+ NewSrcInstr = BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(),
+ TII->get(TargetOpcode::IMPLICIT_DEF),
+ IncomingReg);
+
+ // Clean up the old implicit-def, if there even was one.
+ if (MachineInstr *DefMI = MRI->getVRegDef(SrcReg))
+ if (DefMI->isImplicitDef())
+ ImpDefs.insert(DefMI);
+ } else {
+ NewSrcInstr = BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), IncomingReg)
+ .addReg(SrcReg, 0, SrcSubReg);
+ }
+ }
- // Is it used by any PHI instructions in this block?
- bool ValueIsUsed = VRegPHIUseCount[BBVRegPair(&opBlock, SrcReg)] != 0;
+ // We only need to update the LiveVariables kill of SrcReg if this was the
+ // last PHI use of SrcReg to be lowered on this CFG edge and it is not live
+ // out of the predecessor. We can also ignore undef sources.
+ if (LV && !SrcUndef &&
+ !VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)] &&
+ !LV->isLiveOut(SrcReg, opBlock)) {
+ // We want to be able to insert a kill of the register if this PHI (aka,
+ // the copy we just inserted) is the last use of the source value. Live
+ // variable analysis conservatively handles this by saying that the value
+ // is live until the end of the block the PHI entry lives in. If the value
+ // really is dead at the PHI copy, there will be no successor blocks which
+ // have the value live-in.
+
+ // 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!
- // 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!
- if (!ValueIsUsed && !isLiveOut(SrcReg, opBlock, *LV)) {
// In our final twist, we have to decide which instruction kills the
- // register. In most cases this is the copy, however, the first
- // 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 Term = opBlock.getFirstTerminator();
- if (Term != opBlock.end()) {
+ // register. In most cases this is the copy, however, terminator
+ // instructions at the end of the block may also use the value. In this
+ // case, we should mark the last such terminator as being the killing
+ // block, not the copy.
+ MachineBasicBlock::iterator KillInst = opBlock.end();
+ MachineBasicBlock::iterator FirstTerm = opBlock.getFirstTerminator();
+ for (MachineBasicBlock::iterator Term = FirstTerm;
+ Term != opBlock.end(); ++Term) {
if (Term->readsRegister(SrcReg))
KillInst = Term;
+ }
- // Check that no other terminators use values.
-#ifndef NDEBUG
- for (MachineBasicBlock::iterator TI = 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!");
+ if (KillInst == opBlock.end()) {
+ // No terminator uses the register.
+
+ if (reusedIncoming || !IncomingReg) {
+ // We may have to rewind a bit if we didn't insert a copy this time.
+ KillInst = FirstTerm;
+ while (KillInst != opBlock.begin()) {
+ --KillInst;
+ if (KillInst->isDebugValue())
+ continue;
+ if (KillInst->readsRegister(SrcReg))
+ break;
+ }
+ } else {
+ // We just inserted this copy.
+ KillInst = std::prev(InsertPos);
}
-#endif
}
+ assert(KillInst->readsRegister(SrcReg) && "Cannot find kill instruction");
// Finally, mark it killed.
LV->addVirtualRegisterKilled(SrcReg, KillInst);
unsigned opBlockNum = opBlock.getNumber();
LV->getVarInfo(SrcReg).AliveBlocks.reset(opBlockNum);
}
+
+ if (LIS) {
+ if (NewSrcInstr) {
+ LIS->InsertMachineInstrInMaps(NewSrcInstr);
+ LIS->addSegmentToEndOfBlock(IncomingReg, NewSrcInstr);
+ }
+
+ if (!SrcUndef &&
+ !VRegPHIUseCount[BBVRegPair(opBlock.getNumber(), SrcReg)]) {
+ LiveInterval &SrcLI = LIS->getInterval(SrcReg);
+
+ bool isLiveOut = false;
+ for (MachineBasicBlock::succ_iterator SI = opBlock.succ_begin(),
+ SE = opBlock.succ_end(); SI != SE; ++SI) {
+ SlotIndex startIdx = LIS->getMBBStartIdx(*SI);
+ VNInfo *VNI = SrcLI.getVNInfoAt(startIdx);
+
+ // Definitions by other PHIs are not truly live-in for our purposes.
+ if (VNI && VNI->def != startIdx) {
+ isLiveOut = true;
+ break;
+ }
+ }
+
+ if (!isLiveOut) {
+ MachineBasicBlock::iterator KillInst = opBlock.end();
+ MachineBasicBlock::iterator FirstTerm = opBlock.getFirstTerminator();
+ for (MachineBasicBlock::iterator Term = FirstTerm;
+ Term != opBlock.end(); ++Term) {
+ if (Term->readsRegister(SrcReg))
+ KillInst = Term;
+ }
+
+ if (KillInst == opBlock.end()) {
+ // No terminator uses the register.
+
+ if (reusedIncoming || !IncomingReg) {
+ // We may have to rewind a bit if we didn't just insert a copy.
+ KillInst = FirstTerm;
+ while (KillInst != opBlock.begin()) {
+ --KillInst;
+ if (KillInst->isDebugValue())
+ continue;
+ if (KillInst->readsRegister(SrcReg))
+ break;
+ }
+ } else {
+ // We just inserted this copy.
+ KillInst = std::prev(InsertPos);
+ }
+ }
+ assert(KillInst->readsRegister(SrcReg) &&
+ "Cannot find kill instruction");
+
+ SlotIndex LastUseIndex = LIS->getInstructionIndex(KillInst);
+ SrcLI.removeSegment(LastUseIndex.getRegSlot(),
+ LIS->getMBBEndIdx(&opBlock));
+ }
+ }
+ }
}
- // 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) {
+ if (LIS)
+ LIS->RemoveMachineInstrFromMaps(MPhi);
+ 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();
- I != E; ++I)
- for (MachineBasicBlock::const_iterator BBI = I->begin(), BBE = I->end();
- BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI)
- for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2)
- ++VRegPHIUseCount[BBVRegPair(BBI->getOperand(i + 1).getMBB(),
- BBI->getOperand(i).getReg())];
+void PHIElimination::analyzePHINodes(const MachineFunction& MF) {
+ for (const auto &MBB : MF)
+ for (const auto &BBI : MBB) {
+ if (!BBI.isPHI())
+ break;
+ for (unsigned i = 1, e = BBI.getNumOperands(); i != e; i += 2)
+ ++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)
+bool PHIElimination::SplitPHIEdges(MachineFunction &MF,
+ MachineBasicBlock &MBB,
+ MachineLoopInfo *MLI) {
+ if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad())
return false; // Quick exit for basic blocks without PHIs.
- for (MachineBasicBlock::const_iterator BBI = MBB.begin(), BBE = MBB.end();
- BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI) {
+ const MachineLoop *CurLoop = MLI ? MLI->getLoopFor(&MBB) : nullptr;
+ bool IsLoopHeader = CurLoop && &MBB == CurLoop->getHeader();
+
+ bool Changed = false;
+ for (MachineBasicBlock::iterator BBI = MBB.begin(), BBE = MBB.end();
+ 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 (isLiveOut(Reg, *PreMBB, LV) && !isLiveIn(Reg, MBB, LV))
- SplitCriticalEdge(PreMBB, &MBB);
+ // Is there a critical edge from PreMBB to MBB?
+ if (PreMBB->succ_size() == 1)
+ continue;
+
+ // 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 && !SplitAllCriticalEdges)
+ continue;
+ const MachineLoop *PreLoop = MLI ? MLI->getLoopFor(PreMBB) : nullptr;
+ if (IsLoopHeader && PreLoop == CurLoop && !SplitAllCriticalEdges)
+ continue;
+
+ // LV doesn't consider a phi use live-out, so isLiveOut only returns true
+ // when the source register is live-out for some other reason than a phi
+ // use. That means the copy we will insert in PreMBB won't be a kill, and
+ // there is a risk it may not be coalesced away.
+ //
+ // If the copy would be a kill, there is no need to split the edge.
+ bool ShouldSplit = isLiveOutPastPHIs(Reg, PreMBB);
+ if (!ShouldSplit && !NoPhiElimLiveOutEarlyExit)
+ continue;
+ if (ShouldSplit) {
+ DEBUG(dbgs() << PrintReg(Reg) << " live-out before critical edge BB#"
+ << PreMBB->getNumber() << " -> BB#" << MBB.getNumber()
+ << ": " << *BBI);
+ }
+
+ // If Reg is not live-in to MBB, it means it must be live-in to some
+ // other PreMBB successor, and we can avoid the interference by splitting
+ // the edge.
+ //
+ // If Reg *is* live-in to MBB, the interference is inevitable and a copy
+ // is likely to be left after coalescing. If we are looking at a loop
+ // exiting edge, split it so we won't insert code in the loop, otherwise
+ // don't bother.
+ ShouldSplit = ShouldSplit && !isLiveIn(Reg, &MBB);
+
+ // Check for a loop exiting edge.
+ if (!ShouldSplit && CurLoop != PreLoop) {
+ DEBUG({
+ dbgs() << "Split wouldn't help, maybe avoid loop copies?\n";
+ if (PreLoop) dbgs() << "PreLoop: " << *PreLoop;
+ if (CurLoop) dbgs() << "CurLoop: " << *CurLoop;
+ });
+ // This edge could be entering a loop, exiting a loop, or it could be
+ // both: Jumping directly form one loop to the header of a sibling
+ // loop.
+ // Split unless this edge is entering CurLoop from an outer loop.
+ ShouldSplit = PreLoop && !PreLoop->contains(CurLoop);
+ }
+ if (!ShouldSplit && !SplitAllCriticalEdges)
+ continue;
+ if (!PreMBB->SplitCriticalEdge(&MBB, this)) {
+ DEBUG(dbgs() << "Failed to split critical edge.\n");
+ continue;
+ }
+ Changed = true;
+ ++NumCriticalEdgesSplit;
}
}
- return true;
+ return Changed;
}
-bool llvm::PHIElimination::isLiveOut(unsigned Reg, const MachineBasicBlock &MBB,
- LiveVariables &LV) {
- LiveVariables::VarInfo &VI = LV.getVarInfo(Reg);
-
- // Loop over all of the successors of the basic block, checking to see if
- // the value is either live in the block, or if it is killed in the block.
- std::vector<MachineBasicBlock*> OpSuccBlocks;
- for (MachineBasicBlock::const_succ_iterator SI = MBB.succ_begin(),
- E = MBB.succ_end(); SI != E; ++SI) {
- MachineBasicBlock *SuccMBB = *SI;
-
- // Is it alive in this successor?
- unsigned SuccIdx = SuccMBB->getNumber();
- if (VI.AliveBlocks.test(SuccIdx))
- return true;
- OpSuccBlocks.push_back(SuccMBB);
- }
-
- // Check to see if this value is live because there is a use in a successor
- // that kills it.
- switch (OpSuccBlocks.size()) {
- case 1: {
- MachineBasicBlock *SuccMBB = OpSuccBlocks[0];
- for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
- if (VI.Kills[i]->getParent() == SuccMBB)
- return true;
- break;
- }
- case 2: {
- MachineBasicBlock *SuccMBB1 = OpSuccBlocks[0], *SuccMBB2 = OpSuccBlocks[1];
- for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
- if (VI.Kills[i]->getParent() == SuccMBB1 ||
- VI.Kills[i]->getParent() == SuccMBB2)
- return true;
- break;
- }
- default:
- std::sort(OpSuccBlocks.begin(), OpSuccBlocks.end());
- for (unsigned i = 0, e = VI.Kills.size(); i != e; ++i)
- if (std::binary_search(OpSuccBlocks.begin(), OpSuccBlocks.end(),
- VI.Kills[i]->getParent()))
- return true;
- }
- return false;
+bool PHIElimination::isLiveIn(unsigned Reg, const MachineBasicBlock *MBB) {
+ assert((LV || LIS) &&
+ "isLiveIn() requires either LiveVariables or LiveIntervals");
+ if (LIS)
+ return LIS->isLiveInToMBB(LIS->getInterval(Reg), MBB);
+ else
+ return LV->isLiveIn(Reg, *MBB);
}
-bool llvm::PHIElimination::isLiveIn(unsigned Reg, const MachineBasicBlock &MBB,
- LiveVariables &LV) {
- LiveVariables::VarInfo &VI = LV.getVarInfo(Reg);
-
- if (VI.AliveBlocks.test(MBB.getNumber()))
- return true;
-
- // defined in MBB?
- const MachineInstr *Def = MRI->getVRegDef(Reg);
- if (Def && Def->getParent() == &MBB)
+bool PHIElimination::isLiveOutPastPHIs(unsigned Reg,
+ const MachineBasicBlock *MBB) {
+ assert((LV || LIS) &&
+ "isLiveOutPastPHIs() requires either LiveVariables or LiveIntervals");
+ // LiveVariables considers uses in PHIs to be in the predecessor basic block,
+ // so that a register used only in a PHI is not live out of the block. In
+ // contrast, LiveIntervals considers uses in PHIs to be on the edge rather than
+ // in the predecessor basic block, so that a register used only in a PHI is live
+ // out of the block.
+ if (LIS) {
+ const LiveInterval &LI = LIS->getInterval(Reg);
+ for (const MachineBasicBlock *SI : MBB->successors())
+ if (LI.liveAt(LIS->getMBBStartIdx(SI)))
+ return true;
return false;
-
- // killed in MBB?
- return VI.findKill(&MBB);
-}
-
-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->push_back(NMBB);
- DEBUG(errs() << "PHIElimination splitting critical edge:"
- " BB#" << A->getNumber()
- << " -- BB#" << NMBB->getNumber()
- << " -- BB#" << B->getNumber() << '\n');
-
- A->ReplaceUsesOfBlockWith(B, NMBB);
- // If A may fall through to B, we may have to insert a branch.
- if (A->isLayoutSuccessor(B))
- A->updateTerminator();
-
- // Insert unconditional "jump B" instruction in NMBB.
- NMBB->addSuccessor(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);
-
- if (MachineDominatorTree *MDT=getAnalysisIfAvailable<MachineDominatorTree>())
- MDT->addNewBlock(NMBB, A);
-
- return NMBB;
+ } else {
+ return LV->isLiveOut(Reg, *MBB);
+ }
}