void
StupidAllocator::syncForBlockEnd(LBlock *block, LInstruction *ins)
{
// Sync any dirty registers, and update the synced state for phi nodes at
// each successor of a block. We cannot conflate the storage for phis with
// that of their inputs, as we cannot prove the live ranges of the phi and
// its input do not overlap. The values for the two may additionally be
// different, as the phi could be for the value of the input in a previous
// loop iteration.
for (size_t i = 0; i < registerCount; i++)
syncRegister(ins, i);
LMoveGroup *group = nullptr;
MBasicBlock *successor = block->mir()->successorWithPhis();
if (successor) {
uint32_t position = block->mir()->positionInPhiSuccessor();
LBlock *lirsuccessor = graph.getBlock(successor->id());
for (size_t i = 0; i < lirsuccessor->numPhis(); i++) {
LPhi *phi = lirsuccessor->getPhi(i);
uint32_t sourcevreg = phi->getOperand(position)->toUse()->virtualRegister();
uint32_t destvreg = phi->getDef(0)->virtualRegister();
if (sourcevreg == destvreg)
continue;
LAllocation *source = stackLocation(sourcevreg);
LAllocation *dest = stackLocation(destvreg);
if (!group) {
// The moves we insert here need to happen simultaneously with
// each other, yet after any existing moves before the instruction.
LMoveGroup *input = getInputMoveGroup(ins->id());
if (input->numMoves() == 0) {
group = input;
} else {
group = new LMoveGroup(alloc());
block->insertAfter(input, group);
}
}
group->add(source, dest);
}
}
}
bool
GreedyAllocator::buildPhiMoves(LBlock *block)
{
IonSpew(IonSpew_RegAlloc, " Merging phi state.");
phiMoves = Mover();
MBasicBlock *mblock = block->mir();
if (!mblock->successorWithPhis())
return true;
// Insert moves from our state into our successor's phi.
uint32 pos = mblock->positionInPhiSuccessor();
LBlock *successor = mblock->successorWithPhis()->lir();
for (size_t i = 0; i < successor->numPhis(); i++) {
LPhi *phi = successor->getPhi(i);
JS_ASSERT(phi->numDefs() == 1);
VirtualRegister *phiReg = getVirtualRegister(phi->getDef(0));
allocateStack(phiReg);
LAllocation *in = phi->getOperand(pos);
VirtualRegister *inReg = getVirtualRegister(in->toUse());
allocateStack(inReg);
// Try to get a register for the input.
if (!inReg->hasRegister() && !allocatableRegs().empty(inReg->isDouble())) {
if (!allocateReg(inReg))
return false;
}
// Add a move from the input to the phi.
if (inReg->hasRegister()) {
if (!phiMoves.move(inReg->reg(), phiReg->backingStack()))
return false;
} else {
if (!phiMoves.move(inReg->backingStack(), phiReg->backingStack()))
return false;
}
}
return true;
}
void
RegisterAllocator::dumpInstructions()
{
#ifdef DEBUG
fprintf(stderr, "Instructions:\n");
for (size_t blockIndex = 0; blockIndex < graph.numBlocks(); blockIndex++) {
LBlock* block = graph.getBlock(blockIndex);
MBasicBlock* mir = block->mir();
fprintf(stderr, "\nBlock %lu", static_cast<unsigned long>(blockIndex));
for (size_t i = 0; i < mir->numSuccessors(); i++)
fprintf(stderr, " [successor %u]", mir->getSuccessor(i)->id());
fprintf(stderr, "\n");
for (size_t i = 0; i < block->numPhis(); i++) {
LPhi* phi = block->getPhi(i);
fprintf(stderr, "[%u,%u Phi] [def %s]",
inputOf(phi).bits(),
outputOf(phi).bits(),
phi->getDef(0)->toString());
for (size_t j = 0; j < phi->numOperands(); j++)
fprintf(stderr, " [use %s]", phi->getOperand(j)->toString());
fprintf(stderr, "\n");
}
for (LInstructionIterator iter = block->begin(); iter != block->end(); iter++) {
LInstruction* ins = *iter;
fprintf(stderr, "[");
if (ins->id() != 0)
fprintf(stderr, "%u,%u ", inputOf(ins).bits(), outputOf(ins).bits());
fprintf(stderr, "%s]", ins->opName());
if (ins->isMoveGroup()) {
LMoveGroup* group = ins->toMoveGroup();
for (int i = group->numMoves() - 1; i >= 0; i--) {
// Use two printfs, as LAllocation::toString is not reentant.
fprintf(stderr, " [%s", group->getMove(i).from()->toString());
fprintf(stderr, " -> %s]", group->getMove(i).to()->toString());
}
fprintf(stderr, "\n");
continue;
}
for (size_t i = 0; i < ins->numDefs(); i++)
fprintf(stderr, " [def %s]", ins->getDef(i)->toString());
for (size_t i = 0; i < ins->numTemps(); i++) {
LDefinition* temp = ins->getTemp(i);
if (!temp->isBogusTemp())
fprintf(stderr, " [temp %s]", temp->toString());
}
for (LInstruction::InputIterator alloc(*ins); alloc.more(); alloc.next()) {
if (!alloc->isBogus())
fprintf(stderr, " [use %s]", alloc->toString());
}
fprintf(stderr, "\n");
}
}
fprintf(stderr, "\n");
#endif // DEBUG
}
bool
AllocationIntegrityState::record()
{
// Ignore repeated record() calls.
if (!instructions.empty())
return true;
if (!instructions.appendN(InstructionInfo(), graph.numInstructions()))
return false;
if (!virtualRegisters.appendN((LDefinition*)nullptr, graph.numVirtualRegisters()))
return false;
if (!blocks.reserve(graph.numBlocks()))
return false;
for (size_t i = 0; i < graph.numBlocks(); i++) {
blocks.infallibleAppend(BlockInfo());
LBlock* block = graph.getBlock(i);
MOZ_ASSERT(block->mir()->id() == i);
BlockInfo& blockInfo = blocks[i];
if (!blockInfo.phis.reserve(block->numPhis()))
return false;
for (size_t j = 0; j < block->numPhis(); j++) {
blockInfo.phis.infallibleAppend(InstructionInfo());
InstructionInfo& info = blockInfo.phis[j];
LPhi* phi = block->getPhi(j);
MOZ_ASSERT(phi->numDefs() == 1);
uint32_t vreg = phi->getDef(0)->virtualRegister();
virtualRegisters[vreg] = phi->getDef(0);
if (!info.outputs.append(*phi->getDef(0)))
return false;
for (size_t k = 0, kend = phi->numOperands(); k < kend; k++) {
if (!info.inputs.append(*phi->getOperand(k)))
return false;
}
}
for (LInstructionIterator iter = block->begin(); iter != block->end(); iter++) {
LInstruction* ins = *iter;
InstructionInfo& info = instructions[ins->id()];
for (size_t k = 0; k < ins->numTemps(); k++) {
if (!ins->getTemp(k)->isBogusTemp()) {
uint32_t vreg = ins->getTemp(k)->virtualRegister();
virtualRegisters[vreg] = ins->getTemp(k);
}
if (!info.temps.append(*ins->getTemp(k)))
return false;
}
for (size_t k = 0; k < ins->numDefs(); k++) {
if (!ins->getDef(k)->isBogusTemp()) {
uint32_t vreg = ins->getDef(k)->virtualRegister();
virtualRegisters[vreg] = ins->getDef(k);
}
if (!info.outputs.append(*ins->getDef(k)))
return false;
}
for (LInstruction::InputIterator alloc(*ins); alloc.more(); alloc.next()) {
if (!info.inputs.append(**alloc))
return false;
}
}
}
return seen.init();
}
bool
LiveRangeAllocator<VREG>::buildLivenessInfo()
{
if (!init())
return false;
Vector<MBasicBlock *, 1, SystemAllocPolicy> loopWorkList;
BitSet *loopDone = BitSet::New(alloc(), graph.numBlockIds());
if (!loopDone)
return false;
for (size_t i = graph.numBlocks(); i > 0; i--) {
if (mir->shouldCancel("Build Liveness Info (main loop)"))
return false;
LBlock *block = graph.getBlock(i - 1);
MBasicBlock *mblock = block->mir();
BitSet *live = BitSet::New(alloc(), graph.numVirtualRegisters());
if (!live)
return false;
liveIn[mblock->id()] = live;
// Propagate liveIn from our successors to us
for (size_t i = 0; i < mblock->lastIns()->numSuccessors(); i++) {
MBasicBlock *successor = mblock->lastIns()->getSuccessor(i);
// Skip backedges, as we fix them up at the loop header.
if (mblock->id() < successor->id())
live->insertAll(liveIn[successor->id()]);
}
// Add successor phis
if (mblock->successorWithPhis()) {
LBlock *phiSuccessor = mblock->successorWithPhis()->lir();
for (unsigned int j = 0; j < phiSuccessor->numPhis(); j++) {
LPhi *phi = phiSuccessor->getPhi(j);
LAllocation *use = phi->getOperand(mblock->positionInPhiSuccessor());
uint32_t reg = use->toUse()->virtualRegister();
live->insert(reg);
}
}
// Variables are assumed alive for the entire block, a define shortens
// the interval to the point of definition.
for (BitSet::Iterator liveRegId(*live); liveRegId; liveRegId++) {
if (!vregs[*liveRegId].getInterval(0)->addRangeAtHead(inputOf(block->firstId()),
outputOf(block->lastId()).next()))
{
return false;
}
}
// Shorten the front end of live intervals for live variables to their
// point of definition, if found.
for (LInstructionReverseIterator ins = block->rbegin(); ins != block->rend(); ins++) {
// Calls may clobber registers, so force a spill and reload around the callsite.
if (ins->isCall()) {
for (AnyRegisterIterator iter(allRegisters_); iter.more(); iter++) {
if (forLSRA) {
if (!addFixedRangeAtHead(*iter, inputOf(*ins), outputOf(*ins)))
return false;
} else {
bool found = false;
for (size_t i = 0; i < ins->numDefs(); i++) {
if (ins->getDef(i)->isPreset() &&
*ins->getDef(i)->output() == LAllocation(*iter)) {
found = true;
break;
}
}
if (!found && !addFixedRangeAtHead(*iter, outputOf(*ins), outputOf(*ins).next()))
return false;
}
}
}
for (size_t i = 0; i < ins->numDefs(); i++) {
if (ins->getDef(i)->policy() != LDefinition::PASSTHROUGH) {
LDefinition *def = ins->getDef(i);
CodePosition from;
if (def->policy() == LDefinition::PRESET && def->output()->isRegister() && forLSRA) {
// The fixed range covers the current instruction so the
// interval for the virtual register starts at the next
// instruction. If the next instruction has a fixed use,
// this can lead to unnecessary register moves. To avoid
// special handling for this, assert the next instruction
// has no fixed uses. defineFixed guarantees this by inserting
// an LNop.
JS_ASSERT(!NextInstructionHasFixedUses(block, *ins));
AnyRegister reg = def->output()->toRegister();
if (!addFixedRangeAtHead(reg, inputOf(*ins), outputOf(*ins).next()))
return false;
from = outputOf(*ins).next();
} else {
from = forLSRA ? inputOf(*ins) : outputOf(*ins);
}
if (def->policy() == LDefinition::MUST_REUSE_INPUT) {
// MUST_REUSE_INPUT is implemented by allocating an output
// register and moving the input to it. Register hints are
// used to avoid unnecessary moves. We give the input an
// LUse::ANY policy to avoid allocating a register for the
// input.
LUse *inputUse = ins->getOperand(def->getReusedInput())->toUse();
JS_ASSERT(inputUse->policy() == LUse::REGISTER);
JS_ASSERT(inputUse->usedAtStart());
*inputUse = LUse(inputUse->virtualRegister(), LUse::ANY, /* usedAtStart = */ true);
}
LiveInterval *interval = vregs[def].getInterval(0);
interval->setFrom(from);
// Ensure that if there aren't any uses, there's at least
// some interval for the output to go into.
if (interval->numRanges() == 0) {
if (!interval->addRangeAtHead(from, from.next()))
return false;
}
live->remove(def->virtualRegister());
}
}
for (size_t i = 0; i < ins->numTemps(); i++) {
LDefinition *temp = ins->getTemp(i);
if (temp->isBogusTemp())
continue;
if (forLSRA) {
if (temp->policy() == LDefinition::PRESET) {
if (ins->isCall())
continue;
AnyRegister reg = temp->output()->toRegister();
if (!addFixedRangeAtHead(reg, inputOf(*ins), outputOf(*ins)))
return false;
// Fixed intervals are not added to safepoints, so do it
// here.
if (LSafepoint *safepoint = ins->safepoint())
AddRegisterToSafepoint(safepoint, reg, *temp);
} else {
JS_ASSERT(!ins->isCall());
if (!vregs[temp].getInterval(0)->addRangeAtHead(inputOf(*ins), outputOf(*ins)))
return false;
}
} else {
// Normally temps are considered to cover both the input
// and output of the associated instruction. In some cases
// though we want to use a fixed register as both an input
// and clobbered register in the instruction, so watch for
// this and shorten the temp to cover only the output.
CodePosition from = inputOf(*ins);
if (temp->policy() == LDefinition::PRESET) {
AnyRegister reg = temp->output()->toRegister();
for (LInstruction::InputIterator alloc(**ins); alloc.more(); alloc.next()) {
if (alloc->isUse()) {
LUse *use = alloc->toUse();
if (use->isFixedRegister()) {
if (GetFixedRegister(vregs[use].def(), use) == reg)
from = outputOf(*ins);
}
}
}
}
CodePosition to =
ins->isCall() ? outputOf(*ins) : outputOf(*ins).next();
if (!vregs[temp].getInterval(0)->addRangeAtHead(from, to))
return false;
}
}
DebugOnly<bool> hasUseRegister = false;
DebugOnly<bool> hasUseRegisterAtStart = false;
for (LInstruction::InputIterator inputAlloc(**ins); inputAlloc.more(); inputAlloc.next()) {
if (inputAlloc->isUse()) {
LUse *use = inputAlloc->toUse();
// The first instruction, LLabel, has no uses.
JS_ASSERT(inputOf(*ins) > outputOf(block->firstId()));
// Call uses should always be at-start or fixed, since the fixed intervals
// use all registers.
JS_ASSERT_IF(ins->isCall() && !inputAlloc.isSnapshotInput(),
use->isFixedRegister() || use->usedAtStart());
#ifdef DEBUG
// Don't allow at-start call uses if there are temps of the same kind,
// so that we don't assign the same register.
if (ins->isCall() && use->usedAtStart()) {
for (size_t i = 0; i < ins->numTemps(); i++)
JS_ASSERT(vregs[ins->getTemp(i)].isDouble() != vregs[use].isDouble());
}
// If there are both useRegisterAtStart(x) and useRegister(y)
// uses, we may assign the same register to both operands due to
// interval splitting (bug 772830). Don't allow this for now.
if (use->policy() == LUse::REGISTER) {
if (use->usedAtStart()) {
if (!IsInputReused(*ins, use))
hasUseRegisterAtStart = true;
} else {
hasUseRegister = true;
}
}
JS_ASSERT(!(hasUseRegister && hasUseRegisterAtStart));
#endif
// Don't treat RECOVERED_INPUT uses as keeping the vreg alive.
if (use->policy() == LUse::RECOVERED_INPUT)
continue;
CodePosition to;
if (forLSRA) {
if (use->isFixedRegister()) {
AnyRegister reg = GetFixedRegister(vregs[use].def(), use);
if (!addFixedRangeAtHead(reg, inputOf(*ins), outputOf(*ins)))
return false;
to = inputOf(*ins);
// Fixed intervals are not added to safepoints, so do it
// here.
LSafepoint *safepoint = ins->safepoint();
if (!ins->isCall() && safepoint)
AddRegisterToSafepoint(safepoint, reg, *vregs[use].def());
} else {
to = use->usedAtStart() ? inputOf(*ins) : outputOf(*ins);
}
} else {
to = (use->usedAtStart() || ins->isCall())
? inputOf(*ins) : outputOf(*ins);
if (use->isFixedRegister()) {
LAllocation reg(AnyRegister::FromCode(use->registerCode()));
for (size_t i = 0; i < ins->numDefs(); i++) {
LDefinition *def = ins->getDef(i);
if (def->policy() == LDefinition::PRESET && *def->output() == reg)
to = inputOf(*ins);
}
}
}
LiveInterval *interval = vregs[use].getInterval(0);
if (!interval->addRangeAtHead(inputOf(block->firstId()), forLSRA ? to : to.next()))
return false;
interval->addUse(new(alloc()) UsePosition(use, to));
live->insert(use->virtualRegister());
}
}
}
// Phis have simultaneous assignment semantics at block begin, so at
// the beginning of the block we can be sure that liveIn does not
// contain any phi outputs.
for (unsigned int i = 0; i < block->numPhis(); i++) {
LDefinition *def = block->getPhi(i)->getDef(0);
if (live->contains(def->virtualRegister())) {
live->remove(def->virtualRegister());
} else {
// This is a dead phi, so add a dummy range over all phis. This
// can go away if we have an earlier dead code elimination pass.
if (!vregs[def].getInterval(0)->addRangeAtHead(inputOf(block->firstId()),
outputOf(block->firstId())))
{
return false;
}
}
}
if (mblock->isLoopHeader()) {
// A divergence from the published algorithm is required here, as
// our block order does not guarantee that blocks of a loop are
// contiguous. As a result, a single live interval spanning the
// loop is not possible. Additionally, we require liveIn in a later
// pass for resolution, so that must also be fixed up here.
MBasicBlock *loopBlock = mblock->backedge();
while (true) {
// Blocks must already have been visited to have a liveIn set.
JS_ASSERT(loopBlock->id() >= mblock->id());
// Add an interval for this entire loop block
CodePosition from = inputOf(loopBlock->lir()->firstId());
CodePosition to = outputOf(loopBlock->lir()->lastId()).next();
for (BitSet::Iterator liveRegId(*live); liveRegId; liveRegId++) {
if (!vregs[*liveRegId].getInterval(0)->addRange(from, to))
return false;
}
// Fix up the liveIn set to account for the new interval
liveIn[loopBlock->id()]->insertAll(live);
// Make sure we don't visit this node again
loopDone->insert(loopBlock->id());
// If this is the loop header, any predecessors are either the
// backedge or out of the loop, so skip any predecessors of
// this block
if (loopBlock != mblock) {
for (size_t i = 0; i < loopBlock->numPredecessors(); i++) {
MBasicBlock *pred = loopBlock->getPredecessor(i);
if (loopDone->contains(pred->id()))
continue;
if (!loopWorkList.append(pred))
return false;
}
}
// Terminate loop if out of work.
if (loopWorkList.empty())
break;
// Grab the next block off the work list, skipping any OSR block.
while (!loopWorkList.empty()) {
loopBlock = loopWorkList.popCopy();
if (loopBlock->lir() != graph.osrBlock())
break;
}
// If end is reached without finding a non-OSR block, then no more work items were found.
if (loopBlock->lir() == graph.osrBlock()) {
JS_ASSERT(loopWorkList.empty());
break;
}
}
// Clear the done set for other loops
loopDone->clear();
}
JS_ASSERT_IF(!mblock->numPredecessors(), live->empty());
}
validateVirtualRegisters();
// If the script has an infinite loop, there may be no MReturn and therefore
// no fixed intervals. Add a small range to fixedIntervalsUnion so that the
// rest of the allocator can assume it has at least one range.
if (fixedIntervalsUnion->numRanges() == 0) {
if (!fixedIntervalsUnion->addRangeAtHead(CodePosition(0, CodePosition::INPUT),
CodePosition(0, CodePosition::OUTPUT)))
{
return false;
}
}
return true;
}
bool
GreedyAllocator::allocateRegisters()
{
// Allocate registers bottom-up, such that we see all uses before their
// definitions.
for (size_t i = graph.numBlocks() - 1; i < graph.numBlocks(); i--) {
LBlock *block = graph.getBlock(i);
IonSpew(IonSpew_RegAlloc, "Allocating block %d", (uint32)i);
// All registers should be free.
JS_ASSERT(state.free == RegisterSet::All());
// Allocate stack for any phis.
for (size_t j = 0; j < block->numPhis(); j++) {
LPhi *phi = block->getPhi(j);
VirtualRegister *vreg = getVirtualRegister(phi->getDef(0));
allocateStack(vreg);
}
// Allocate registers.
if (!allocateRegistersInBlock(block))
return false;
LMoveGroup *entrySpills = block->getEntryMoveGroup();
// We've reached the top of the block. Spill all registers by inserting
// moves from their stack locations.
for (AnyRegisterIterator iter(RegisterSet::All()); iter.more(); iter++) {
VirtualRegister *vreg = state[*iter];
if (!vreg) {
JS_ASSERT(state.free.has(*iter));
continue;
}
JS_ASSERT(vreg->reg() == *iter);
JS_ASSERT(!state.free.has(vreg->reg()));
allocateStack(vreg);
LAllocation *from = LAllocation::New(vreg->backingStack());
LAllocation *to = LAllocation::New(vreg->reg());
if (!entrySpills->add(from, to))
return false;
killReg(vreg);
vreg->unsetRegister();
}
// Before killing phis, ensure that each phi input has its own stack
// allocation. This ensures we won't allocate the same slot for any phi
// as its input, which technically may be legal (since the phi becomes
// the last use of the slot), but we avoid for sanity.
for (size_t i = 0; i < block->numPhis(); i++) {
LPhi *phi = block->getPhi(i);
for (size_t j = 0; j < phi->numOperands(); j++) {
VirtualRegister *in = getVirtualRegister(phi->getOperand(j)->toUse());
allocateStack(in);
}
}
// Kill phis.
for (size_t i = 0; i < block->numPhis(); i++) {
LPhi *phi = block->getPhi(i);
VirtualRegister *vr = getVirtualRegister(phi->getDef(0));
JS_ASSERT(!vr->hasRegister());
killStack(vr);
}
}
return true;
}
