//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "splitter"
+#define DEBUG_TYPE "regalloc"
#include "SplitKit.h"
+#include "LiveRangeEdit.h"
#include "VirtRegMap.h"
#include "llvm/CodeGen/CalcSpillWeights.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
MachineBasicBlock *MBB = MI->getParent();
if (usingBlocks_[MBB]++)
continue;
- if (MachineLoop *Loop = loops_.getLoopFor(MBB))
+ for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop;
+ Loop = Loop->getParentLoop())
usingLoops_[Loop]++;
}
DEBUG(dbgs() << " counted "
<< usingLoops_.size() << " loops.\n");
}
-/// removeUse - Update statistics by noting that MI no longer uses curli.
-void SplitAnalysis::removeUse(const MachineInstr *MI) {
- if (!usingInstrs_.erase(MI))
- return;
-
- // Decrement MBB count.
- const MachineBasicBlock *MBB = MI->getParent();
- BlockCountMap::iterator bi = usingBlocks_.find(MBB);
- assert(bi != usingBlocks_.end() && "MBB missing");
- assert(bi->second && "0 count in map");
- if (--bi->second)
- return;
- // No more uses in MBB.
- usingBlocks_.erase(bi);
-
- // Decrement loop count.
- MachineLoop *Loop = loops_.getLoopFor(MBB);
- if (!Loop)
- return;
- LoopCountMap::iterator li = usingLoops_.find(Loop);
- assert(li != usingLoops_.end() && "Loop missing");
- assert(li->second && "0 count in map");
- if (--li->second)
- return;
- // No more blocks in Loop.
- usingLoops_.erase(li);
+void SplitAnalysis::print(const BlockPtrSet &B, raw_ostream &OS) const {
+ for (BlockPtrSet::const_iterator I = B.begin(), E = B.end(); I != E; ++I) {
+ unsigned count = usingBlocks_.lookup(*I);
+ OS << " BB#" << (*I)->getNumber();
+ if (count)
+ OS << '(' << count << ')';
+ }
}
// Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
}
}
+void SplitAnalysis::print(const LoopBlocks &B, raw_ostream &OS) const {
+ OS << "Loop:";
+ print(B.Loop, OS);
+ OS << ", preds:";
+ print(B.Preds, OS);
+ OS << ", exits:";
+ print(B.Exits, OS);
+}
+
/// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
/// and around the Loop.
SplitAnalysis::LoopPeripheralUse SplitAnalysis::
if (Blocks.Loop.count(MBB))
continue;
// It must be an unrelated block.
+ DEBUG(dbgs() << ", outside: BB#" << MBB->getNumber());
return OutsideLoop;
}
return use;
}
/// getCriticalExits - It may be necessary to partially break critical edges
-/// leaving the loop if an exit block has phi uses of curli. Collect the exit
-/// blocks that need special treatment into CriticalExits.
+/// leaving the loop if an exit block has predecessors from outside the loop
+/// periphery.
void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
BlockPtrSet &CriticalExits) {
CriticalExits.clear();
- // A critical exit block contains a phi def of curli, and has a predecessor
- // that is not in the loop nor a loop predecessor.
- // For such an exit block, the edges carrying the new variable must be moved
- // to a new pre-exit block.
+ // A critical exit block has curli live-in, and has a predecessor that is not
+ // in the loop nor a loop predecessor. For such an exit block, the edges
+ // carrying the new variable must be moved to a new pre-exit block.
for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
I != E; ++I) {
- const MachineBasicBlock *Succ = *I;
- SlotIndex SuccIdx = lis_.getMBBStartIdx(Succ);
- VNInfo *SuccVNI = curli_->getVNInfoAt(SuccIdx);
- // This exit may not have curli live in at all. No need to split.
- if (!SuccVNI)
+ const MachineBasicBlock *Exit = *I;
+ // A single-predecessor exit block is definitely not a critical edge.
+ if (Exit->pred_size() == 1)
continue;
- // If this is not a PHI def, it is either using a value from before the
- // loop, or a value defined inside the loop. Both are safe.
- if (!SuccVNI->isPHIDef() || SuccVNI->def.getBaseIndex() != SuccIdx)
+ // This exit may not have curli live in at all. No need to split.
+ if (!lis_.isLiveInToMBB(*curli_, Exit))
continue;
- // This exit block does have a PHI. Does it also have a predecessor that is
- // not a loop block or loop predecessor?
- for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
- PE = Succ->pred_end(); PI != PE; ++PI) {
+ // Does this exit block have a predecessor that is not a loop block or loop
+ // predecessor?
+ for (MachineBasicBlock::const_pred_iterator PI = Exit->pred_begin(),
+ PE = Exit->pred_end(); PI != PE; ++PI) {
const MachineBasicBlock *Pred = *PI;
if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
continue;
// This is a critical exit block, and we need to split the exit edge.
- CriticalExits.insert(Succ);
+ CriticalExits.insert(Exit);
+ break;
+ }
+ }
+}
+
+void SplitAnalysis::getCriticalPreds(const SplitAnalysis::LoopBlocks &Blocks,
+ BlockPtrSet &CriticalPreds) {
+ CriticalPreds.clear();
+
+ // A critical predecessor block has curli live-out, and has a successor that
+ // has curli live-in and is not in the loop nor a loop exit block. For such a
+ // predecessor block, we must carry the value in both the 'inside' and
+ // 'outside' registers.
+ for (BlockPtrSet::iterator I = Blocks.Preds.begin(), E = Blocks.Preds.end();
+ I != E; ++I) {
+ const MachineBasicBlock *Pred = *I;
+ // Definitely not a critical edge.
+ if (Pred->succ_size() == 1)
+ continue;
+ // This block may not have curli live out at all if there is a PHI.
+ if (!lis_.isLiveOutOfMBB(*curli_, Pred))
+ continue;
+ // Does this block have a successor outside the loop?
+ for (MachineBasicBlock::const_pred_iterator SI = Pred->succ_begin(),
+ SE = Pred->succ_end(); SI != SE; ++SI) {
+ const MachineBasicBlock *Succ = *SI;
+ if (Blocks.Loop.count(Succ) || Blocks.Exits.count(Succ))
+ continue;
+ if (!lis_.isLiveInToMBB(*curli_, Succ))
+ continue;
+ // This is a critical predecessor block.
+ CriticalPreds.insert(Pred);
break;
}
}
if (usingLoops_.empty())
return 0;
- LoopPtrSet Loops, SecondLoops;
+ LoopPtrSet Loops;
LoopBlocks Blocks;
BlockPtrSet CriticalExits;
- // Find first-class and second class candidate loops.
- // We prefer to split around loops where curli is used outside the periphery.
+ // We split around loops where curli is used outside the periphery.
for (LoopCountMap::const_iterator I = usingLoops_.begin(),
E = usingLoops_.end(); I != E; ++I) {
const MachineLoop *Loop = I->first;
getLoopBlocks(Loop, Blocks);
+ DEBUG({ dbgs() << " "; print(Blocks, dbgs()); });
- // FIXME: We need an SSA updater to properly handle multiple exit blocks.
- if (Blocks.Exits.size() > 1) {
- DEBUG(dbgs() << " multiple exits from " << *Loop);
- continue;
- }
-
- LoopPtrSet *LPS = 0;
switch(analyzeLoopPeripheralUse(Blocks)) {
case OutsideLoop:
- LPS = &Loops;
break;
case MultiPeripheral:
- LPS = &SecondLoops;
+ // FIXME: We could split a live range with multiple uses in a peripheral
+ // block and still make progress. However, it is possible that splitting
+ // another live range will insert copies into a peripheral block, and
+ // there is a small chance we can enter an infinity loop, inserting copies
+ // forever.
+ // For safety, stick to splitting live ranges with uses outside the
+ // periphery.
+ DEBUG(dbgs() << ": multiple peripheral uses\n");
break;
case ContainedInLoop:
- DEBUG(dbgs() << " contained in " << *Loop);
+ DEBUG(dbgs() << ": fully contained\n");
continue;
case SinglePeripheral:
- DEBUG(dbgs() << " single peripheral use in " << *Loop);
+ DEBUG(dbgs() << ": single peripheral use\n");
continue;
}
// Will it be possible to split around this loop?
getCriticalExits(Blocks, CriticalExits);
- DEBUG(dbgs() << " " << CriticalExits.size() << " critical exits from "
- << *Loop);
+ DEBUG(dbgs() << ": " << CriticalExits.size() << " critical exits\n");
if (!canSplitCriticalExits(Blocks, CriticalExits))
continue;
// This is a possible split.
- assert(LPS);
- LPS->insert(Loop);
+ Loops.insert(Loop);
}
- DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size() << " + "
- << SecondLoops.size() << " candidate loops.\n");
-
- // If there are no first class loops available, look at second class loops.
- if (Loops.empty())
- Loops = SecondLoops;
+ DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size()
+ << " candidate loops.\n");
if (Loops.empty())
return 0;
return Best;
}
-/// getMultiUseBlocks - if curli has more than one use in a basic block, it
-/// may be an advantage to split curli for the duration of the block.
-bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
- // If curli is local to one block, there is no point to splitting it.
- if (usingBlocks_.size() <= 1)
- return false;
- // Add blocks with multiple uses.
- for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
- I != E; ++I)
- switch (I->second) {
- case 0:
- case 1:
- continue;
- case 2: {
- // It doesn't pay to split a 2-instr block if it redefines curli.
- VNInfo *VN1 = curli_->getVNInfoAt(lis_.getMBBStartIdx(I->first));
- VNInfo *VN2 =
- curli_->getVNInfoAt(lis_.getMBBEndIdx(I->first).getPrevIndex());
- // live-in and live-out with a different value.
- if (VN1 && VN2 && VN1 != VN2)
- continue;
- } // Fall through.
- default:
- Blocks.insert(I->first);
- }
- return !Blocks.empty();
-}
-
//===----------------------------------------------------------------------===//
// LiveIntervalMap
//===----------------------------------------------------------------------===//
void LiveIntervalMap::reset(LiveInterval *li) {
li_ = li;
valueMap_.clear();
+ liveOutCache_.clear();
}
bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
// Create a new value.
VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
+ // Preserve the PHIDef bit.
+ if (ParentVNI->isPHIDef() && Idx == ParentVNI->def)
+ VNI->setIsPHIDef(true);
+
// Use insert for lookup, so we can add missing values with a second lookup.
std::pair<ValueMap::iterator,bool> InsP =
valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
// Now for the fun part. We know that ParentVNI potentially has multiple defs,
// and we may need to create even more phi-defs to preserve VNInfo SSA form.
- // Perform a depth-first search for predecessor blocks where we know the
- // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
-
- // Track MBBs where we have created or learned the dominating value.
- // This may change during the DFS as we create new phi-defs.
- typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
- MBBValueMap DomValue;
-
- for (idf_iterator<MachineBasicBlock*>
- IDFI = idf_begin(IdxMBB),
- IDFE = idf_end(IdxMBB); IDFI != IDFE;) {
- MachineBasicBlock *MBB = *IDFI;
- SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
-
- // We are operating on the restricted CFG where ParentVNI is live.
- if (parentli_.getVNInfoAt(End) != ParentVNI) {
- IDFI.skipChildren();
- continue;
- }
-
- // Do we have a dominating value in this block?
- VNInfo *VNI = extendTo(MBB, End);
- if (!VNI) {
- ++IDFI;
- continue;
+ // Perform a search for all predecessor blocks where we know the dominating
+ // VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
+ DEBUG(dbgs() << "\n Reaching defs for BB#" << IdxMBB->getNumber()
+ << " at " << Idx << " in " << *li_ << '\n');
+
+ // Blocks where li_ should be live-in.
+ SmallVector<MachineDomTreeNode*, 16> LiveIn;
+ LiveIn.push_back(mdt_[IdxMBB]);
+
+ // Using liveOutCache_ as a visited set, perform a BFS for all reaching defs.
+ for (unsigned i = 0; i != LiveIn.size(); ++i) {
+ MachineBasicBlock *MBB = LiveIn[i]->getBlock();
+ for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
+ PE = MBB->pred_end(); PI != PE; ++PI) {
+ MachineBasicBlock *Pred = *PI;
+ // Is this a known live-out block?
+ std::pair<LiveOutMap::iterator,bool> LOIP =
+ liveOutCache_.insert(std::make_pair(Pred, LiveOutPair()));
+ // Yes, we have been here before.
+ if (!LOIP.second) {
+ DEBUG(if (VNInfo *VNI = LOIP.first->second.first)
+ dbgs() << " known valno #" << VNI->id
+ << " at BB#" << Pred->getNumber() << '\n');
+ continue;
+ }
+
+ // Does Pred provide a live-out value?
+ SlotIndex Last = lis_.getMBBEndIdx(Pred).getPrevSlot();
+ if (VNInfo *VNI = extendTo(Pred, Last)) {
+ MachineBasicBlock *DefMBB = lis_.getMBBFromIndex(VNI->def);
+ DEBUG(dbgs() << " found valno #" << VNI->id
+ << " from BB#" << DefMBB->getNumber()
+ << " at BB#" << Pred->getNumber() << '\n');
+ LiveOutPair &LOP = LOIP.first->second;
+ LOP.first = VNI;
+ LOP.second = mdt_[DefMBB];
+ continue;
+ }
+ // No, we need a live-in value for Pred as well
+ if (Pred != IdxMBB)
+ LiveIn.push_back(mdt_[Pred]);
}
+ }
- // Yes, VNI dominates MBB. Track the path back to IdxMBB, creating phi-defs
- // as needed along the way.
- for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
- // Start from MBB's immediate successor. End at IdxMBB.
- MachineBasicBlock *Succ = IDFI.getPath(PI-1);
- std::pair<MBBValueMap::iterator, bool> InsP =
- DomValue.insert(MBBValueMap::value_type(Succ, VNI));
-
- // This is the first time we backtrack to Succ.
- if (InsP.second)
- continue;
-
- // We reached Succ again with the same VNI. Nothing is going to change.
- VNInfo *OVNI = InsP.first->second;
- if (OVNI == VNI)
- break;
+ // We may need to add phi-def values to preserve the SSA form.
+ // This is essentially the same iterative algorithm that SSAUpdater uses,
+ // except we already have a dominator tree, so we don't have to recompute it.
+ VNInfo *IdxVNI = 0;
+ unsigned Changes;
+ do {
+ Changes = 0;
+ DEBUG(dbgs() << " Iterating over " << LiveIn.size() << " blocks.\n");
+ // Propagate live-out values down the dominator tree, inserting phi-defs when
+ // necessary. Since LiveIn was created by a BFS, going backwards makes it more
+ // likely for us to visit immediate dominators before their children.
+ for (unsigned i = LiveIn.size(); i; --i) {
+ MachineDomTreeNode *Node = LiveIn[i-1];
+ MachineBasicBlock *MBB = Node->getBlock();
+ MachineDomTreeNode *IDom = Node->getIDom();
+ LiveOutPair IDomValue;
+ // We need a live-in value to a block with no immediate dominator?
+ // This is probably an unreachable block that has survived somehow.
+ bool needPHI = !IDom;
+
+ // Get the IDom live-out value.
+ if (!needPHI) {
+ LiveOutMap::iterator I = liveOutCache_.find(IDom->getBlock());
+ if (I != liveOutCache_.end())
+ IDomValue = I->second;
+ else
+ // If IDom is outside our set of live-out blocks, there must be new
+ // defs, and we need a phi-def here.
+ needPHI = true;
+ }
- // Succ already has a phi-def. No need to continue.
- SlotIndex Start = lis_.getMBBStartIdx(Succ);
- if (OVNI->def == Start)
- break;
+ // IDom dominates all of our predecessors, but it may not be the immediate
+ // dominator. Check if any of them have live-out values that are properly
+ // dominated by IDom. If so, we need a phi-def here.
+ if (!needPHI) {
+ for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
+ PE = MBB->pred_end(); PI != PE; ++PI) {
+ LiveOutPair Value = liveOutCache_[*PI];
+ if (!Value.first || Value.first == IDomValue.first)
+ continue;
+ // This predecessor is carrying something other than IDomValue.
+ // It could be because IDomValue hasn't propagated yet, or it could be
+ // because MBB is in the dominance frontier of that value.
+ if (mdt_.dominates(IDom, Value.second)) {
+ needPHI = true;
+ break;
+ }
+ }
+ }
- // We have a collision between the old and new VNI at Succ. That means
- // neither dominates and we need a new phi-def.
- VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
- VNI->setIsPHIDef(true);
- InsP.first->second = VNI;
-
- // Replace OVNI with VNI in the remaining path.
- for (; PI > 1 ; --PI) {
- MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
- if (I == DomValue.end() || I->second != OVNI)
- break;
- I->second = VNI;
+ // Create a phi-def if required.
+ if (needPHI) {
+ ++Changes;
+ SlotIndex Start = lis_.getMBBStartIdx(MBB);
+ VNInfo *VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
+ VNI->setIsPHIDef(true);
+ DEBUG(dbgs() << " - BB#" << MBB->getNumber()
+ << " phi-def #" << VNI->id << " at " << Start << '\n');
+ // We no longer need li_ to be live-in.
+ LiveIn.erase(LiveIn.begin()+(i-1));
+ // Blocks in LiveIn are either IdxMBB, or have a value live-through.
+ if (MBB == IdxMBB)
+ IdxVNI = VNI;
+ // Check if we need to update live-out info.
+ LiveOutMap::iterator I = liveOutCache_.find(MBB);
+ if (I == liveOutCache_.end() || I->second.second == Node) {
+ // We already have a live-out defined in MBB, so this must be IdxMBB.
+ assert(MBB == IdxMBB && "Adding phi-def to known live-out");
+ li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
+ } else {
+ // This phi-def is also live-out, so color the whole block.
+ li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
+ I->second = LiveOutPair(VNI, Node);
+ }
+ } else if (IDomValue.first) {
+ // No phi-def here. Remember incoming value for IdxMBB.
+ if (MBB == IdxMBB)
+ IdxVNI = IDomValue.first;
+ // Propagate IDomValue if needed:
+ // MBB is live-out and doesn't define its own value.
+ LiveOutMap::iterator I = liveOutCache_.find(MBB);
+ if (I != liveOutCache_.end() && I->second.second != Node &&
+ I->second.first != IDomValue.first) {
+ ++Changes;
+ I->second = IDomValue;
+ DEBUG(dbgs() << " - BB#" << MBB->getNumber()
+ << " idom valno #" << IDomValue.first->id
+ << " from BB#" << IDom->getBlock()->getNumber() << '\n');
+ }
}
}
+ DEBUG(dbgs() << " - made " << Changes << " changes.\n");
+ } while (Changes);
- // No need to search the children, we found a dominating value.
- IDFI.skipChildren();
- }
+ assert(IdxVNI && "Didn't find value for Idx");
- // The search should at least find a dominating value for IdxMBB.
- assert(!DomValue.empty() && "Couldn't find a reaching definition");
+#ifndef NDEBUG
+ // Check the liveOutCache_ invariants.
+ for (LiveOutMap::iterator I = liveOutCache_.begin(), E = liveOutCache_.end();
+ I != E; ++I) {
+ assert(I->first && "Null MBB entry in cache");
+ assert(I->second.first && "Null VNInfo in cache");
+ assert(I->second.second && "Null DomTreeNode in cache");
+ if (I->second.second->getBlock() == I->first)
+ continue;
+ for (MachineBasicBlock::pred_iterator PI = I->first->pred_begin(),
+ PE = I->first->pred_end(); PI != PE; ++PI)
+ assert(liveOutCache_.lookup(*PI) == I->second && "Bad invariant");
+ }
+#endif
- // Since we went through the trouble of a full DFS visiting all reaching defs,
- // the values in DomValue are now accurate. No more phi-defs are needed for
- // these blocks, so we can color the live ranges.
+ // Since we went through the trouble of a full BFS visiting all reaching defs,
+ // the values in LiveIn are now accurate. No more phi-defs are needed
+ // for these blocks, so we can color the live ranges.
// This makes the next mapValue call much faster.
- VNInfo *IdxVNI = 0;
- for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
- ++I) {
- MachineBasicBlock *MBB = I->first;
- VNInfo *VNI = I->second;
- SlotIndex Start = lis_.getMBBStartIdx(MBB);
- if (MBB == IdxMBB) {
- // Don't add full liveness to IdxMBB, stop at Idx.
- if (Start != Idx)
- li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
- // The caller had better add some liveness to IdxVNI, or it leaks.
- IdxVNI = VNI;
- } else
- li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
+ for (unsigned i = 0, e = LiveIn.size(); i != e; ++i) {
+ MachineBasicBlock *MBB = LiveIn[i]->getBlock();
+ SlotIndex Start = lis_.getMBBStartIdx(MBB);
+ if (MBB == IdxMBB) {
+ li_->addRange(LiveRange(Start, Idx.getNextSlot(), IdxVNI));
+ continue;
+ }
+ // Anything in LiveIn other than IdxMBB is live-through.
+ VNInfo *VNI = liveOutCache_.lookup(MBB).first;
+ assert(VNI && "Missing block value");
+ li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
}
- assert(IdxVNI && "Didn't find value for Idx");
return IdxVNI;
}
// extendTo - Find the last li_ value defined in MBB at or before Idx. The
// parentli_ is assumed to be live at Idx. Extend the live range to Idx.
// Return the found VNInfo, or NULL.
-VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
+VNInfo *LiveIntervalMap::extendTo(const MachineBasicBlock *MBB, SlotIndex Idx) {
assert(li_ && "call reset first");
LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
if (I == li_->begin())
// ParentVNI is a complex value. We must map per MBB.
MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
- MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End);
+ MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
if (MBB == MBBE) {
li_->addRange(LiveRange(Start, End, VNI));
addSimpleRange(I->start, std::min(End, I->end), I->valno);
}
-VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
- const VNInfo *ParentVNI,
- MachineBasicBlock &MBB,
- MachineBasicBlock::iterator I) {
+VNInfo *LiveIntervalMap::defByCopy(const VNInfo *ParentVNI,
+ MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I) {
const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
get(TargetOpcode::COPY);
- MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
+ MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg)
+ .addReg(parentli_.reg);
SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
VNInfo *VNI = defValue(ParentVNI, DefIdx);
VNI->setCopy(MI);
//===----------------------------------------------------------------------===//
/// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
-SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
- SmallVectorImpl<LiveInterval*> &intervals)
+SplitEditor::SplitEditor(SplitAnalysis &sa,
+ LiveIntervals &lis,
+ VirtRegMap &vrm,
+ MachineDominatorTree &mdt,
+ LiveRangeEdit &edit)
: sa_(sa), lis_(lis), vrm_(vrm),
mri_(vrm.getMachineFunction().getRegInfo()),
tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
- curli_(sa_.getCurLI()),
- dupli_(lis_, *curli_),
- openli_(lis_, *curli_),
- intervals_(intervals),
- firstInterval(intervals_.size())
+ edit_(edit),
+ dupli_(lis_, mdt, edit.getParent()),
+ openli_(lis_, mdt, edit.getParent())
{
- assert(curli_ && "SplitEditor created from empty SplitAnalysis");
-
- // Make sure curli_ is assigned a stack slot, so all our intervals get the
- // same slot as curli_.
- if (vrm_.getStackSlot(curli_->reg) == VirtRegMap::NO_STACK_SLOT)
- vrm_.assignVirt2StackSlot(curli_->reg);
-
-}
-
-LiveInterval *SplitEditor::createInterval() {
- unsigned Reg = mri_.createVirtualRegister(mri_.getRegClass(curli_->reg));
- LiveInterval &Intv = lis_.getOrCreateInterval(Reg);
- vrm_.grow();
- vrm_.assignVirt2StackSlot(Reg, vrm_.getStackSlot(curli_->reg));
- return &Intv;
}
bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
- for (int i = firstInterval, e = intervals_.size(); i != e; ++i)
- if (intervals_[i]->liveAt(Idx))
+ for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
+ if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
return true;
return false;
}
assert(!openli_.getLI() && "Previous LI not closed before openIntv");
if (!dupli_.getLI())
- dupli_.reset(createInterval());
+ dupli_.reset(&edit_.create(mri_, lis_, vrm_));
- openli_.reset(createInterval());
- intervals_.push_back(openli_.getLI());
+ openli_.reset(&edit_.create(mri_, lis_, vrm_));
}
/// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
/// not live before Idx, a COPY is not inserted.
void SplitEditor::enterIntvBefore(SlotIndex Idx) {
assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
- VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getUseIndex());
+ DEBUG(dbgs() << " enterIntvBefore " << Idx);
+ VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getUseIndex());
if (!ParentVNI) {
- DEBUG(dbgs() << " enterIntvBefore " << Idx << ": not live\n");
+ DEBUG(dbgs() << ": not live\n");
return;
}
+ DEBUG(dbgs() << ": valno " << ParentVNI->id);
truncatedValues.insert(ParentVNI);
MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
assert(MI && "enterIntvBefore called with invalid index");
- openli_.defByCopyFrom(curli_->reg, ParentVNI, *MI->getParent(), MI);
- DEBUG(dbgs() << " enterIntvBefore " << Idx << ": " << *openli_.getLI()
- << '\n');
+ VNInfo *VNI = openli_.defByCopy(ParentVNI, *MI->getParent(), MI);
+ openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
+ DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
}
/// enterIntvAtEnd - Enter openli at the end of MBB.
void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
SlotIndex End = lis_.getMBBEndIdx(&MBB);
- VNInfo *ParentVNI = curli_->getVNInfoAt(End.getPrevSlot());
+ DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
+ VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End.getPrevSlot());
if (!ParentVNI) {
- DEBUG(dbgs() << " enterIntvAtEnd " << End << ": not live\n");
+ DEBUG(dbgs() << ": not live\n");
return;
}
+ DEBUG(dbgs() << ": valno " << ParentVNI->id);
truncatedValues.insert(ParentVNI);
- VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
- MBB, MBB.getFirstTerminator());
+ VNInfo *VNI = openli_.defByCopy(ParentVNI, MBB, MBB.getFirstTerminator());
// Make sure openli is live out of MBB.
openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
- DEBUG(dbgs() << " enterIntvAtEnd: " << *openli_.getLI() << '\n');
+ DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
}
/// useIntv - indicate that all instructions in MBB should use openli.
/// leaveIntvAfter - Leave openli after the instruction at Idx.
void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
+ DEBUG(dbgs() << " leaveIntvAfter " << Idx);
// The interval must be live beyond the instruction at Idx.
- VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getBoundaryIndex());
+ VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getBoundaryIndex());
if (!ParentVNI) {
- DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": not live\n");
+ DEBUG(dbgs() << ": not live\n");
return;
}
+ DEBUG(dbgs() << ": valno " << ParentVNI->id);
MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
MachineBasicBlock *MBB = MII->getParent();
- VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
- llvm::next(MII));
+ VNInfo *VNI = dupli_.defByCopy(ParentVNI, *MBB, llvm::next(MII));
// Finally we must make sure that openli is properly extended from Idx to the
// new copy.
openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
- DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": " << *openli_.getLI()
- << '\n');
+ DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
}
/// leaveIntvAtTop - Leave the interval at the top of MBB.
/// Currently, only one value can leave the interval.
void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
-
SlotIndex Start = lis_.getMBBStartIdx(&MBB);
- VNInfo *ParentVNI = curli_->getVNInfoAt(Start);
+ DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
- // Is curli even live-in to MBB?
+ VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start);
if (!ParentVNI) {
- DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": not live\n");
+ DEBUG(dbgs() << ": not live\n");
return;
}
// We are going to insert a back copy, so we must have a dupli_.
- VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
- MBB, MBB.begin());
+ VNInfo *VNI = dupli_.defByCopy(ParentVNI, MBB,
+ MBB.SkipPHIsAndLabels(MBB.begin()));
// Finally we must make sure that openli is properly extended from Start to
// the new copy.
openli_.addSimpleRange(Start, VNI->def, ParentVNI);
- DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": " << *openli_.getLI()
- << '\n');
+ DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
}
/// closeIntv - Indicate that we are done editing the currently open
openli_.reset(0);
}
+/// rewrite - Rewrite all uses of reg to use the new registers.
+void SplitEditor::rewrite(unsigned reg) {
+ for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
+ RE = mri_.reg_end(); RI != RE;) {
+ MachineOperand &MO = RI.getOperand();
+ unsigned OpNum = RI.getOperandNo();
+ MachineInstr *MI = MO.getParent();
+ ++RI;
+ if (MI->isDebugValue()) {
+ DEBUG(dbgs() << "Zapping " << *MI);
+ // FIXME: We can do much better with debug values.
+ MO.setReg(0);
+ continue;
+ }
+ SlotIndex Idx = lis_.getInstructionIndex(MI);
+ Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
+ LiveInterval *LI = 0;
+ for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
+ ++I) {
+ LiveInterval *testli = *I;
+ if (testli->liveAt(Idx)) {
+ LI = testli;
+ break;
+ }
+ }
+ DEBUG(dbgs() << " rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx);
+ assert(LI && "No register was live at use");
+ MO.setReg(LI->reg);
+ if (MO.isUse() && !MI->isRegTiedToDefOperand(OpNum))
+ MO.setIsKill(LI->killedAt(Idx.getDefIndex()));
+ DEBUG(dbgs() << '\t' << *MI);
+ }
+}
+
void
SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
- SlotIndex sidx = Start;
+ // Build vector of iterator pairs from the intervals.
+ typedef std::pair<LiveInterval::const_iterator,
+ LiveInterval::const_iterator> IIPair;
+ SmallVector<IIPair, 8> Iters;
+ for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
+ ++LI) {
+ if (*LI == dupli_.getLI())
+ continue;
+ LiveInterval::const_iterator I = (*LI)->find(Start);
+ LiveInterval::const_iterator E = (*LI)->end();
+ if (I != E)
+ Iters.push_back(std::make_pair(I, E));
+ }
+ SlotIndex sidx = Start;
// Break [Start;End) into segments that don't overlap any intervals.
for (;;) {
SlotIndex next = sidx, eidx = End;
// Find overlapping intervals.
- for (int i = firstInterval, e = intervals_.size(); i != e && sidx < eidx;
- ++i) {
- LiveInterval::const_iterator I = intervals_[i]->find(sidx);
- LiveInterval::const_iterator E = intervals_[i]->end();
- if (I == E)
- continue;
+ for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
+ LiveInterval::const_iterator I = Iters[i].first;
// Interval I is overlapping [sidx;eidx). Trim sidx.
if (I->start <= sidx) {
sidx = I->end;
- if (++I == E)
+ // Move to the next run, remove iters when all are consumed.
+ I = ++Iters[i].first;
+ if (I == Iters[i].second) {
+ Iters.erase(Iters.begin() + i);
+ --i;
continue;
+ }
}
// Trim eidx too if needed.
if (I->start >= eidx)
continue;
eidx = I->start;
- if (I->end > next)
- next = I->end;
+ next = I->end;
}
// Now, [sidx;eidx) doesn't overlap anything in intervals_.
if (sidx < eidx)
}
}
-/// rewrite - after all the new live ranges have been created, rewrite
-/// instructions using curli to use the new intervals.
-bool SplitEditor::rewrite() {
- assert(!openli_.getLI() && "Previous LI not closed before rewrite");
-
+void SplitEditor::computeRemainder() {
// First we need to fill in the live ranges in dupli.
// If values were redefined, we need a full recoloring with SSA update.
// If values were truncated, we only need to truncate the ranges.
// If values were partially rematted, we should shrink to uses.
// If values were fully rematted, they should be omitted.
// FIXME: If a single value is redefined, just move the def and truncate.
+ LiveInterval &parent = edit_.getParent();
// Values that are fully contained in the split intervals.
SmallPtrSet<const VNInfo*, 8> deadValues;
-
// Map all curli values that should have live defs in dupli.
- for (LiveInterval::const_vni_iterator I = curli_->vni_begin(),
- E = curli_->vni_end(); I != E; ++I) {
+ for (LiveInterval::const_vni_iterator I = parent.vni_begin(),
+ E = parent.vni_end(); I != E; ++I) {
const VNInfo *VNI = *I;
+ // Don't transfer unused values to the new intervals.
+ if (VNI->isUnused())
+ continue;
// Original def is contained in the split intervals.
if (intervalsLiveAt(VNI->def)) {
// Did this value escape?
}
// Add all ranges to dupli.
- for (LiveInterval::const_iterator I = curli_->begin(), E = curli_->end();
+ for (LiveInterval::const_iterator I = parent.begin(), E = parent.end();
I != E; ++I) {
const LiveRange &LR = *I;
if (truncatedValues.count(LR.valno)) {
}
}
+ // Extend dupli_ to be live out of any critical loop predecessors.
+ // This means we have multiple registers live out of those blocks.
+ // The alternative would be to split the critical edges.
+ if (criticalPreds_.empty())
+ return;
+ for (SplitAnalysis::BlockPtrSet::iterator I = criticalPreds_.begin(),
+ E = criticalPreds_.end(); I != E; ++I)
+ dupli_.extendTo(*I, lis_.getMBBEndIdx(*I).getPrevSlot());
+ criticalPreds_.clear();
+}
- const LiveInterval *curli = sa_.getCurLI();
- for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(curli->reg),
- RE = mri_.reg_end(); RI != RE;) {
- MachineOperand &MO = RI.getOperand();
- MachineInstr *MI = MO.getParent();
- ++RI;
- if (MI->isDebugValue()) {
- DEBUG(dbgs() << "Zapping " << *MI);
- // FIXME: We can do much better with debug values.
- MO.setReg(0);
- continue;
- }
- SlotIndex Idx = lis_.getInstructionIndex(MI);
- Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
- LiveInterval *LI = dupli_.getLI();
- for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
- LiveInterval *testli = intervals_[i];
- if (testli->liveAt(Idx)) {
- LI = testli;
- break;
- }
- }
- if (LI) {
- MO.setReg(LI->reg);
- sa_.removeUse(MI);
- DEBUG(dbgs() << " rewrite " << Idx << '\t' << *MI);
- }
- }
+void SplitEditor::finish() {
+ assert(!openli_.getLI() && "Previous LI not closed before rewrite");
+ assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
- // dupli_ goes in last, after rewriting.
- if (dupli_.getLI()) {
- if (dupli_.getLI()->empty()) {
- DEBUG(dbgs() << " dupli became empty?\n");
- lis_.removeInterval(dupli_.getLI()->reg);
- dupli_.reset(0);
- } else {
- dupli_.getLI()->RenumberValues(lis_);
- intervals_.push_back(dupli_.getLI());
- }
+ // Complete dupli liveness.
+ computeRemainder();
+
+ // Get rid of unused values and set phi-kill flags.
+ for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
+ (*I)->RenumberValues(lis_);
+
+ // Rewrite instructions.
+ rewrite(edit_.getReg());
+
+ // Now check if any registers were separated into multiple components.
+ ConnectedVNInfoEqClasses ConEQ(lis_);
+ for (unsigned i = 0, e = edit_.size(); i != e; ++i) {
+ // Don't use iterators, they are invalidated by create() below.
+ LiveInterval *li = edit_.get(i);
+ unsigned NumComp = ConEQ.Classify(li);
+ if (NumComp <= 1)
+ continue;
+ DEBUG(dbgs() << " " << NumComp << " components: " << *li << '\n');
+ SmallVector<LiveInterval*, 8> dups;
+ dups.push_back(li);
+ for (unsigned i = 1; i != NumComp; ++i)
+ dups.push_back(&edit_.create(mri_, lis_, vrm_));
+ ConEQ.Distribute(&dups[0]);
+ // Rewrite uses to the new regs.
+ rewrite(li->reg);
}
// Calculate spill weight and allocation hints for new intervals.
VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
- for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
- LiveInterval &li = *intervals_[i];
+ for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
+ LiveInterval &li = **I;
vrai.CalculateRegClass(li.reg);
vrai.CalculateWeightAndHint(li);
DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
<< ":" << li << '\n');
}
- return dupli_.getLI();
}
// Loop Splitting
//===----------------------------------------------------------------------===//
-bool SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
+void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
SplitAnalysis::LoopBlocks Blocks;
sa_.getLoopBlocks(Loop, Blocks);
+ DEBUG({
+ dbgs() << " splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
+ });
+
// Break critical edges as needed.
SplitAnalysis::BlockPtrSet CriticalExits;
sa_.getCriticalExits(Blocks, CriticalExits);
assert(CriticalExits.empty() && "Cannot break critical exits yet");
+ // Get critical predecessors so computeRemainder can deal with them.
+ sa_.getCriticalPreds(Blocks, criticalPreds_);
+
// Create new live interval for the loop.
openIntv();
// Done.
closeIntv();
- return rewrite();
+ finish();
}
// Single Block Splitting
//===----------------------------------------------------------------------===//
+/// getMultiUseBlocks - if curli has more than one use in a basic block, it
+/// may be an advantage to split curli for the duration of the block.
+bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
+ // If curli is local to one block, there is no point to splitting it.
+ if (usingBlocks_.size() <= 1)
+ return false;
+ // Add blocks with multiple uses.
+ for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
+ I != E; ++I)
+ switch (I->second) {
+ case 0:
+ case 1:
+ continue;
+ case 2: {
+ // When there are only two uses and curli is both live in and live out,
+ // we don't really win anything by isolating the block since we would be
+ // inserting two copies.
+ // The remaing register would still have two uses in the block. (Unless it
+ // separates into disconnected components).
+ if (lis_.isLiveInToMBB(*curli_, I->first) &&
+ lis_.isLiveOutOfMBB(*curli_, I->first))
+ continue;
+ } // Fall through.
+ default:
+ Blocks.insert(I->first);
+ }
+ return !Blocks.empty();
+}
+
/// splitSingleBlocks - Split curli into a separate live interval inside each
-/// basic block in Blocks. Return true if curli has been completely replaced,
-/// false if curli is still intact, and needs to be spilled or split further.
-bool SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
+/// basic block in Blocks.
+void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
// Determine the first and last instruction using curli in each block.
typedef std::pair<SlotIndex,SlotIndex> IndexPair;
leaveIntvAfter(IP.second);
closeIntv();
}
- return rewrite();
+ finish();
}
return usingBlocks_.begin()->first;
}
-/// splitInsideBlock - Split curli into multiple intervals inside MBB. Return
-/// true if curli has been completely replaced, false if curli is still
-/// intact, and needs to be spilled or split further.
-bool SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
+/// splitInsideBlock - Split curli into multiple intervals inside MBB.
+void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
SmallVector<SlotIndex, 32> Uses;
Uses.reserve(sa_.usingInstrs_.size());
for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
closeIntv();
}
- return rewrite();
+ finish();
}
projects / oota-llvm.git / blobdiff