static void registerClobberCheck(AssemblyHelpers& jit, RegisterSet dontClobber)
{
if (!Options::clobberAllRegsInFTLICSlowPath())
return;
RegisterSet clobber = RegisterSet::allRegisters();
clobber.exclude(RegisterSet::reservedHardwareRegisters());
clobber.exclude(RegisterSet::stackRegisters());
clobber.exclude(RegisterSet::calleeSaveRegisters());
clobber.exclude(dontClobber);
GPRReg someGPR;
for (Reg reg = Reg::first(); reg <= Reg::last(); reg = reg.next()) {
if (!clobber.get(reg) || !reg.isGPR())
continue;
jit.move(AssemblyHelpers::TrustedImm32(0x1337beef), reg.gpr());
someGPR = reg.gpr();
}
for (Reg reg = Reg::first(); reg <= Reg::last(); reg = reg.next()) {
if (!clobber.get(reg) || !reg.isFPR())
continue;
jit.move64ToDouble(someGPR, reg.fpr());
}
}
void ValueRep::emitRestore(AssemblyHelpers& jit, Reg reg) const
{
if (reg.isGPR()) {
switch (kind()) {
case LateRegister:
case Register:
if (isGPR())
jit.move(gpr(), reg.gpr());
else
jit.moveDoubleTo64(fpr(), reg.gpr());
break;
case Stack:
jit.load64(AssemblyHelpers::Address(GPRInfo::callFrameRegister, offsetFromFP()), reg.gpr());
break;
case Constant:
jit.move(AssemblyHelpers::TrustedImm64(value()), reg.gpr());
break;
default:
RELEASE_ASSERT_NOT_REACHED();
break;
}
return;
}
switch (kind()) {
case LateRegister:
case Register:
if (isGPR())
jit.move64ToDouble(gpr(), reg.fpr());
else
jit.moveDouble(fpr(), reg.fpr());
break;
case Stack:
jit.loadDouble(AssemblyHelpers::Address(GPRInfo::callFrameRegister, offsetFromFP()), reg.fpr());
break;
case Constant:
jit.move(AssemblyHelpers::TrustedImm64(value()), jit.scratchRegister());
jit.move64ToDouble(jit.scratchRegister(), reg.fpr());
break;
default:
RELEASE_ASSERT_NOT_REACHED();
break;
}
}
void CallFrameShuffler::emitDisplace(CachedRecovery& cachedRecovery)
{
Reg wantedReg;
if (!(wantedReg = Reg { cachedRecovery.wantedJSValueRegs().gpr() }))
wantedReg = Reg { cachedRecovery.wantedFPR() };
ASSERT(wantedReg);
ASSERT(!m_lockedRegisters.get(wantedReg));
if (CachedRecovery* current = m_registers[wantedReg]) {
if (current == &cachedRecovery) {
if (verbose)
dataLog(" + ", wantedReg, " is OK\n");
return;
}
// We could do a more complex thing by finding cycles
// etc. in that case.
// However, ending up in this situation will be super
// rare, and should actually be outright impossible for
// non-FTL tiers, since:
// (a) All doubles have been converted into JSValues with
// ValueRep nodes, so FPRs are initially free
//
// (b) The only recoveries with wanted registers are the
// callee (which always starts out in a register) and
// the callee-save registers
//
// (c) The callee-save registers are the first things we
// load (after the return PC), and they are loaded as JSValues
//
// (d) We prefer loading JSValues into FPRs if their
// wanted GPR is not available
//
// (e) If we end up spilling some registers with a
// target, we won't load them again before the very
// end of the algorithm
//
// Combined, this means that we will never load a recovery
// with a wanted GPR into any GPR other than its wanted
// GPR. The callee could however have been initially in
// one of the callee-save registers - but since the wanted
// GPR for the callee is always regT0, it will be the
// first one to be displaced, and we won't see it when
// handling any of the callee-save registers.
//
// Thus, the only way we could ever reach this path is in
// the FTL, when there is so much pressure that we
// absolutely need to load the callee-save registers into
// different GPRs initially but not enough pressure to
// then have to spill all of them. And even in that case,
// depending on the order in which B3 saves the
// callee-saves, we will probably still be safe. Anyway,
// the couple extra move instructions compared to an
// efficient cycle-based algorithm are not going to hurt
// us.
if (wantedReg.isFPR()) {
FPRReg tempFPR = getFreeFPR();
if (verbose)
dataLog(" * Moving ", wantedReg, " into ", tempFPR, "\n");
m_jit.moveDouble(wantedReg.fpr(), tempFPR);
updateRecovery(*current,
ValueRecovery::inFPR(tempFPR, current->recovery().dataFormat()));
} else {
GPRReg tempGPR = getFreeGPR();
if (verbose)
dataLog(" * Moving ", wantedReg.gpr(), " into ", tempGPR, "\n");
m_jit.move(wantedReg.gpr(), tempGPR);
updateRecovery(*current,
ValueRecovery::inGPR(tempGPR, current->recovery().dataFormat()));
}
}
ASSERT(!m_registers[wantedReg]);
if (cachedRecovery.recovery().isConstant()) {
// We only care about callee saves for wanted FPRs, and those are never constants
ASSERT(wantedReg.isGPR());
if (verbose)
dataLog(" * Loading ", cachedRecovery.recovery().constant(), " into ", wantedReg, "\n");
m_jit.moveTrustedValue(cachedRecovery.recovery().constant(), JSValueRegs { wantedReg.gpr() });
updateRecovery(
cachedRecovery,
ValueRecovery::inRegister(wantedReg, DataFormatJS));
} else if (cachedRecovery.recovery().isInGPR()) {
if (verbose)
dataLog(" * Moving ", cachedRecovery.recovery(), " into ", wantedReg, "\n");
if (wantedReg.isGPR())
m_jit.move(cachedRecovery.recovery().gpr(), wantedReg.gpr());
else
m_jit.move64ToDouble(cachedRecovery.recovery().gpr(), wantedReg.fpr());
RELEASE_ASSERT(cachedRecovery.recovery().dataFormat() == DataFormatJS);
updateRecovery(cachedRecovery,
ValueRecovery::inRegister(wantedReg, DataFormatJS));
} else {
ASSERT(cachedRecovery.recovery().isInFPR());
if (cachedRecovery.recovery().dataFormat() == DataFormatDouble) {
// We only care about callee saves for wanted FPRs, and those are always DataFormatJS
ASSERT(wantedReg.isGPR());
// This will automatically pick the wanted GPR
emitBox(cachedRecovery);
} else {
if (verbose)
dataLog(" * Moving ", cachedRecovery.recovery().fpr(), " into ", wantedReg, "\n");
if (wantedReg.isGPR())
m_jit.moveDoubleTo64(cachedRecovery.recovery().fpr(), wantedReg.gpr());
else
m_jit.moveDouble(cachedRecovery.recovery().fpr(), wantedReg.fpr());
RELEASE_ASSERT(cachedRecovery.recovery().dataFormat() == DataFormatJS);
updateRecovery(cachedRecovery,
ValueRecovery::inRegister(wantedReg, DataFormatJS));
}
}
ASSERT(m_registers[wantedReg] == &cachedRecovery);
}
static void compileStub(
unsigned exitID, JITCode* jitCode, OSRExit& exit, VM* vm, CodeBlock* codeBlock)
{
StackMaps::Record* record = nullptr;
for (unsigned i = jitCode->stackmaps.records.size(); i--;) {
record = &jitCode->stackmaps.records[i];
if (record->patchpointID == exit.m_stackmapID)
break;
}
RELEASE_ASSERT(record->patchpointID == exit.m_stackmapID);
// This code requires framePointerRegister is the same as callFrameRegister
static_assert(MacroAssembler::framePointerRegister == GPRInfo::callFrameRegister, "MacroAssembler::framePointerRegister and GPRInfo::callFrameRegister must be the same");
CCallHelpers jit(vm, codeBlock);
// We need scratch space to save all registers, to build up the JS stack, to deal with unwind
// fixup, pointers to all of the objects we materialize, and the elements inside those objects
// that we materialize.
// Figure out how much space we need for those object allocations.
unsigned numMaterializations = 0;
size_t maxMaterializationNumArguments = 0;
for (ExitTimeObjectMaterialization* materialization : exit.m_materializations) {
numMaterializations++;
maxMaterializationNumArguments = std::max(
maxMaterializationNumArguments,
materialization->properties().size());
}
ScratchBuffer* scratchBuffer = vm->scratchBufferForSize(
sizeof(EncodedJSValue) * (
exit.m_values.size() + numMaterializations + maxMaterializationNumArguments) +
requiredScratchMemorySizeInBytes() +
codeBlock->calleeSaveRegisters()->size() * sizeof(uint64_t));
EncodedJSValue* scratch = scratchBuffer ? static_cast<EncodedJSValue*>(scratchBuffer->dataBuffer()) : 0;
EncodedJSValue* materializationPointers = scratch + exit.m_values.size();
EncodedJSValue* materializationArguments = materializationPointers + numMaterializations;
char* registerScratch = bitwise_cast<char*>(materializationArguments + maxMaterializationNumArguments);
uint64_t* unwindScratch = bitwise_cast<uint64_t*>(registerScratch + requiredScratchMemorySizeInBytes());
HashMap<ExitTimeObjectMaterialization*, EncodedJSValue*> materializationToPointer;
unsigned materializationCount = 0;
for (ExitTimeObjectMaterialization* materialization : exit.m_materializations) {
materializationToPointer.add(
materialization, materializationPointers + materializationCount++);
}
// Note that we come in here, the stack used to be as LLVM left it except that someone called pushToSave().
// We don't care about the value they saved. But, we do appreciate the fact that they did it, because we use
// that slot for saveAllRegisters().
saveAllRegisters(jit, registerScratch);
// Bring the stack back into a sane form and assert that it's sane.
jit.popToRestore(GPRInfo::regT0);
jit.checkStackPointerAlignment();
if (vm->m_perBytecodeProfiler && codeBlock->jitCode()->dfgCommon()->compilation) {
Profiler::Database& database = *vm->m_perBytecodeProfiler;
Profiler::Compilation* compilation = codeBlock->jitCode()->dfgCommon()->compilation.get();
Profiler::OSRExit* profilerExit = compilation->addOSRExit(
exitID, Profiler::OriginStack(database, codeBlock, exit.m_codeOrigin),
exit.m_kind, exit.m_kind == UncountableInvalidation);
jit.add64(CCallHelpers::TrustedImm32(1), CCallHelpers::AbsoluteAddress(profilerExit->counterAddress()));
}
// The remaining code assumes that SP/FP are in the same state that they were in the FTL's
// call frame.
// Get the call frame and tag thingies.
// Restore the exiting function's callFrame value into a regT4
jit.move(MacroAssembler::TrustedImm64(TagTypeNumber), GPRInfo::tagTypeNumberRegister);
jit.move(MacroAssembler::TrustedImm64(TagMask), GPRInfo::tagMaskRegister);
// Do some value profiling.
if (exit.m_profileDataFormat != DataFormatNone) {
record->locations[0].restoreInto(jit, jitCode->stackmaps, registerScratch, GPRInfo::regT0);
reboxAccordingToFormat(
exit.m_profileDataFormat, jit, GPRInfo::regT0, GPRInfo::regT1, GPRInfo::regT2);
if (exit.m_kind == BadCache || exit.m_kind == BadIndexingType) {
CodeOrigin codeOrigin = exit.m_codeOriginForExitProfile;
if (ArrayProfile* arrayProfile = jit.baselineCodeBlockFor(codeOrigin)->getArrayProfile(codeOrigin.bytecodeIndex)) {
jit.load32(MacroAssembler::Address(GPRInfo::regT0, JSCell::structureIDOffset()), GPRInfo::regT1);
jit.store32(GPRInfo::regT1, arrayProfile->addressOfLastSeenStructureID());
jit.load8(MacroAssembler::Address(GPRInfo::regT0, JSCell::indexingTypeOffset()), GPRInfo::regT1);
jit.move(MacroAssembler::TrustedImm32(1), GPRInfo::regT2);
jit.lshift32(GPRInfo::regT1, GPRInfo::regT2);
jit.or32(GPRInfo::regT2, MacroAssembler::AbsoluteAddress(arrayProfile->addressOfArrayModes()));
}
}
if (!!exit.m_valueProfile)
jit.store64(GPRInfo::regT0, exit.m_valueProfile.getSpecFailBucket(0));
}
// Materialize all objects. Don't materialize an object until all
// of the objects it needs have been materialized. We break cycles
// by populating objects late - we only consider an object as
// needing another object if the later is needed for the
// allocation of the former.
HashSet<ExitTimeObjectMaterialization*> toMaterialize;
for (ExitTimeObjectMaterialization* materialization : exit.m_materializations)
toMaterialize.add(materialization);
while (!toMaterialize.isEmpty()) {
unsigned previousToMaterializeSize = toMaterialize.size();
Vector<ExitTimeObjectMaterialization*> worklist;
worklist.appendRange(toMaterialize.begin(), toMaterialize.end());
for (ExitTimeObjectMaterialization* materialization : worklist) {
// Check if we can do anything about this right now.
bool allGood = true;
for (ExitPropertyValue value : materialization->properties()) {
if (!value.value().isObjectMaterialization())
continue;
if (!value.location().neededForMaterialization())
continue;
if (toMaterialize.contains(value.value().objectMaterialization())) {
// Gotta skip this one, since it needs a
// materialization that hasn't been materialized.
allGood = false;
break;
}
}
if (!allGood)
continue;
// All systems go for materializing the object. First we
// recover the values of all of its fields and then we
// call a function to actually allocate the beast.
// We only recover the fields that are needed for the allocation.
for (unsigned propertyIndex = materialization->properties().size(); propertyIndex--;) {
const ExitPropertyValue& property = materialization->properties()[propertyIndex];
const ExitValue& value = property.value();
if (!property.location().neededForMaterialization())
continue;
compileRecovery(
jit, value, record, jitCode->stackmaps, registerScratch,
materializationToPointer);
jit.storePtr(GPRInfo::regT0, materializationArguments + propertyIndex);
}
// This call assumes that we don't pass arguments on the stack.
jit.setupArgumentsWithExecState(
CCallHelpers::TrustedImmPtr(materialization),
CCallHelpers::TrustedImmPtr(materializationArguments));
jit.move(CCallHelpers::TrustedImmPtr(bitwise_cast<void*>(operationMaterializeObjectInOSR)), GPRInfo::nonArgGPR0);
jit.call(GPRInfo::nonArgGPR0);
jit.storePtr(GPRInfo::returnValueGPR, materializationToPointer.get(materialization));
// Let everyone know that we're done.
toMaterialize.remove(materialization);
}
// We expect progress! This ensures that we crash rather than looping infinitely if there
// is something broken about this fixpoint. Or, this could happen if we ever violate the
// "materializations form a DAG" rule.
RELEASE_ASSERT(toMaterialize.size() < previousToMaterializeSize);
}
// Now that all the objects have been allocated, we populate them
// with the correct values. This time we can recover all the
// fields, including those that are only needed for the allocation.
for (ExitTimeObjectMaterialization* materialization : exit.m_materializations) {
for (unsigned propertyIndex = materialization->properties().size(); propertyIndex--;) {
const ExitValue& value = materialization->properties()[propertyIndex].value();
compileRecovery(
jit, value, record, jitCode->stackmaps, registerScratch,
materializationToPointer);
jit.storePtr(GPRInfo::regT0, materializationArguments + propertyIndex);
}
// This call assumes that we don't pass arguments on the stack
jit.setupArgumentsWithExecState(
CCallHelpers::TrustedImmPtr(materialization),
CCallHelpers::TrustedImmPtr(materializationToPointer.get(materialization)),
CCallHelpers::TrustedImmPtr(materializationArguments));
jit.move(CCallHelpers::TrustedImmPtr(bitwise_cast<void*>(operationPopulateObjectInOSR)), GPRInfo::nonArgGPR0);
jit.call(GPRInfo::nonArgGPR0);
}
// Save all state from wherever the exit data tells us it was, into the appropriate place in
// the scratch buffer. This also does the reboxing.
for (unsigned index = exit.m_values.size(); index--;) {
compileRecovery(
jit, exit.m_values[index], record, jitCode->stackmaps, registerScratch,
materializationToPointer);
jit.store64(GPRInfo::regT0, scratch + index);
}
// Henceforth we make it look like the exiting function was called through a register
// preservation wrapper. This implies that FP must be nudged down by a certain amount. Then
// we restore the various things according to either exit.m_values or by copying from the
// old frame, and finally we save the various callee-save registers into where the
// restoration thunk would restore them from.
// Before we start messing with the frame, we need to set aside any registers that the
// FTL code was preserving.
for (unsigned i = codeBlock->calleeSaveRegisters()->size(); i--;) {
RegisterAtOffset entry = codeBlock->calleeSaveRegisters()->at(i);
jit.load64(
MacroAssembler::Address(MacroAssembler::framePointerRegister, entry.offset()),
GPRInfo::regT0);
jit.store64(GPRInfo::regT0, unwindScratch + i);
}
jit.load32(CCallHelpers::payloadFor(JSStack::ArgumentCount), GPRInfo::regT2);
// Let's say that the FTL function had failed its arity check. In that case, the stack will
// contain some extra stuff.
//
// We compute the padded stack space:
//
// paddedStackSpace = roundUp(codeBlock->numParameters - regT2 + 1)
//
// The stack will have regT2 + CallFrameHeaderSize stuff.
// We want to make the stack look like this, from higher addresses down:
//
// - argument padding
// - actual arguments
// - call frame header
// This code assumes that we're dealing with FunctionCode.
RELEASE_ASSERT(codeBlock->codeType() == FunctionCode);
jit.add32(
MacroAssembler::TrustedImm32(-codeBlock->numParameters()), GPRInfo::regT2,
GPRInfo::regT3);
MacroAssembler::Jump arityIntact = jit.branch32(
MacroAssembler::GreaterThanOrEqual, GPRInfo::regT3, MacroAssembler::TrustedImm32(0));
jit.neg32(GPRInfo::regT3);
jit.add32(MacroAssembler::TrustedImm32(1 + stackAlignmentRegisters() - 1), GPRInfo::regT3);
jit.and32(MacroAssembler::TrustedImm32(-stackAlignmentRegisters()), GPRInfo::regT3);
jit.add32(GPRInfo::regT3, GPRInfo::regT2);
arityIntact.link(&jit);
CodeBlock* baselineCodeBlock = jit.baselineCodeBlockFor(exit.m_codeOrigin);
// First set up SP so that our data doesn't get clobbered by signals.
unsigned conservativeStackDelta =
(exit.m_values.numberOfLocals() + baselineCodeBlock->calleeSaveSpaceAsVirtualRegisters()) * sizeof(Register) +
maxFrameExtentForSlowPathCall;
conservativeStackDelta = WTF::roundUpToMultipleOf(
stackAlignmentBytes(), conservativeStackDelta);
jit.addPtr(
MacroAssembler::TrustedImm32(-conservativeStackDelta),
MacroAssembler::framePointerRegister, MacroAssembler::stackPointerRegister);
jit.checkStackPointerAlignment();
RegisterSet allFTLCalleeSaves = RegisterSet::ftlCalleeSaveRegisters();
RegisterAtOffsetList* baselineCalleeSaves = baselineCodeBlock->calleeSaveRegisters();
for (Reg reg = Reg::first(); reg <= Reg::last(); reg = reg.next()) {
if (!allFTLCalleeSaves.get(reg))
continue;
unsigned unwindIndex = codeBlock->calleeSaveRegisters()->indexOf(reg);
RegisterAtOffset* baselineRegisterOffset = baselineCalleeSaves->find(reg);
if (reg.isGPR()) {
GPRReg regToLoad = baselineRegisterOffset ? GPRInfo::regT0 : reg.gpr();
if (unwindIndex == UINT_MAX) {
// The FTL compilation didn't preserve this register. This means that it also
// didn't use the register. So its value at the beginning of OSR exit should be
// preserved by the thunk. Luckily, we saved all registers into the register
// scratch buffer, so we can restore them from there.
jit.load64(registerScratch + offsetOfReg(reg), regToLoad);
} else {
// The FTL compilation preserved the register. Its new value is therefore
// irrelevant, but we can get the value that was preserved by using the unwind
// data. We've already copied all unwind-able preserved registers into the unwind
// scratch buffer, so we can get it from there.
jit.load64(unwindScratch + unwindIndex, regToLoad);
}
if (baselineRegisterOffset)
jit.store64(regToLoad, MacroAssembler::Address(MacroAssembler::framePointerRegister, baselineRegisterOffset->offset()));
} else {
FPRReg fpRegToLoad = baselineRegisterOffset ? FPRInfo::fpRegT0 : reg.fpr();
if (unwindIndex == UINT_MAX)
jit.loadDouble(MacroAssembler::TrustedImmPtr(registerScratch + offsetOfReg(reg)), fpRegToLoad);
else
jit.loadDouble(MacroAssembler::TrustedImmPtr(unwindScratch + unwindIndex), fpRegToLoad);
if (baselineRegisterOffset)
jit.storeDouble(fpRegToLoad, MacroAssembler::Address(MacroAssembler::framePointerRegister, baselineRegisterOffset->offset()));
}
}
size_t baselineVirtualRegistersForCalleeSaves = baselineCodeBlock->calleeSaveSpaceAsVirtualRegisters();
// Now get state out of the scratch buffer and place it back into the stack. The values are
// already reboxed so we just move them.
for (unsigned index = exit.m_values.size(); index--;) {
VirtualRegister reg = exit.m_values.virtualRegisterForIndex(index);
if (reg.isLocal() && reg.toLocal() < static_cast<int>(baselineVirtualRegistersForCalleeSaves))
continue;
jit.load64(scratch + index, GPRInfo::regT0);
jit.store64(GPRInfo::regT0, AssemblyHelpers::addressFor(reg));
}
handleExitCounts(jit, exit);
reifyInlinedCallFrames(jit, exit);
adjustAndJumpToTarget(jit, exit, false);
LinkBuffer patchBuffer(*vm, jit, codeBlock);
exit.m_code = FINALIZE_CODE_IF(
shouldDumpDisassembly() || Options::verboseOSR() || Options::verboseFTLOSRExit(),
patchBuffer,
("FTL OSR exit #%u (%s, %s) from %s, with operands = %s, and record = %s",
exitID, toCString(exit.m_codeOrigin).data(),
exitKindToString(exit.m_kind), toCString(*codeBlock).data(),
toCString(ignoringContext<DumpContext>(exit.m_values)).data(),
toCString(*record).data()));
}