void prepareForNextHHBC(IRGS& env,
const NormalizedInstruction* ni,
SrcKey newSk,
bool lastBcInst) {
FTRACE(1, "------------------- prepareForNextHHBC ------------------\n");
env.currentNormalizedInstruction = ni;
always_assert_flog(
IMPLIES(isInlining(env), !env.lastBcInst),
"Tried to end trace while inlining."
);
always_assert_flog(
IMPLIES(isInlining(env), !env.firstBcInst),
"Inlining while still at the first region instruction."
);
always_assert(env.bcStateStack.size() >= env.inlineLevel + 1);
auto pops = env.bcStateStack.size() - 1 - env.inlineLevel;
while (pops--) env.bcStateStack.pop_back();
always_assert_flog(env.bcStateStack.back().func() == newSk.func(),
"Tried to update current SrcKey with a different func");
env.bcStateStack.back().setOffset(newSk.offset());
updateMarker(env);
env.lastBcInst = lastBcInst;
env.catchCreator = nullptr;
env.irb->prepareForNextHHBC();
}
Block* findMainExitBlock(const IRUnit& unit, SrcKey lastSk) {
Block* mainExit = nullptr;
FTRACE(5, "findMainExitBlock: starting on unit:\n{}\n", show(unit));
for (auto block : rpoSortCfg(unit)) {
if (endsUnitAtSrcKey(block, lastSk)) {
if (mainExit == nullptr) {
mainExit = block;
continue;
}
always_assert_flog(
mainExit->hint() == Block::Hint::Unlikely ||
block->hint() == Block::Hint::Unlikely,
"findMainExit: 2 likely exits found: B{} and B{}\nlastSk = {}",
mainExit->id(), block->id(), showShort(lastSk));
if (mainExit->hint() == Block::Hint::Unlikely) mainExit = block;
}
}
always_assert_flog(mainExit, "findMainExit: no exit found for lastSk = {}",
showShort(lastSk));
FTRACE(5, "findMainExitBlock: mainExit = B{}\n", mainExit->id());
return mainExit;
}
bool check_invariants(const FrameState& state) {
for (auto id = uint32_t{0}; id < state.locals.size(); ++id) {
auto const& local = state.locals[id];
always_assert_flog(
local.predictedType <= local.type,
"local {} failed prediction invariants; pred = {}, type = {}\n",
id,
local.predictedType,
local.type
);
always_assert_flog(
local.value == nullptr || local.value->type() == local.type,
"local {} had type {}, but value {}\n",
id,
local.type,
local.value->toString()
);
if (state.curFunc->isPseudoMain()) {
always_assert_flog(
local.value == nullptr,
"We should never be tracking values for locals in a pseudomain "
"right now. Local {} had value {}",
id,
local.value->toString()
);
always_assert_flog(
local.type == TGen,
"We should never be tracking non-predicted types for locals in "
"a pseudomain right now. Local {} had type {}",
id,
local.type.toString()
);
}
}
// We require the memory stack is always at least as big as the irSPOff,
// unless irSPOff went negative (because we're returning and have freed the
// ActRec). Note that there are some "wasted" slots where locals/iterators
// would be in the vector right now.
always_assert_flog(
state.irSPOff < FPInvOffset{0} ||
state.stack.size() >= state.irSPOff.offset,
"stack was smaller than possible"
);
return true;
}
Type relaxType(Type t, DataTypeCategory cat) {
always_assert_flog(t <= TGen && t != TBottom, "t = {}", t);
if (cat == DataTypeGeneric) return TGen;
auto const relaxed =
(t & TCell) <= TBottom ? TBottom : relaxCell(t & TCell, cat);
return t <= TCell ? relaxed : relaxed | TBoxedInitCell;
}
/*
* Lookup a catch trace for the given TCA, returning nullptr if none was
* found. Will abort if a nullptr catch trace was registered, meaning this call
* isn't allowed to throw.
*/
TCA lookup_catch_trace(TCA rip, _Unwind_Exception* exn) {
if (auto catchTraceOpt = mcg->getCatchTrace(rip)) {
if (auto catchTrace = *catchTraceOpt) return catchTrace;
// A few of our optimization passes must be aware of every path out of
// the trace, so throwing through jitted code without a catch block is
// very bad. This is indicated with a present but nullptr entry in the
// catch trace map.
const size_t kCallSize = 5;
const uint8_t kCallOpcode = 0xe8;
auto callAddr = rip - kCallSize;
TCA helperAddr = nullptr;
if (*callAddr == kCallOpcode) {
helperAddr = rip + *reinterpret_cast<int32_t*>(callAddr + 1);
}
always_assert_flog(false,
"Translated call to {} threw '{}' without "
"catch block, return address: {}\n",
getNativeFunctionName(helperAddr),
typeInfoFromUnwindException(exn).name(),
rip);
}
return nullptr;
}
MemEffects memory_effects(const IRInstruction& inst) {
auto const ret = memory_effects_impl(inst);
if (debug) {
// In debug let's do some type checking in case people move instruction
// argument numbers.
auto const fp = match<SSATmp*>(
ret,
[&] (UnknownEffects) { return nullptr; },
[&] (IrrelevantEffects) { return nullptr; },
[&] (ReadAllLocals) { return nullptr; },
[&] (KillFrameLocals l) { return l.fp; },
[&] (ReadLocal l) { return l.fp; },
[&] (ReadLocal2 l) { return l.fp; },
[&] (StoreLocal l) { return l.fp; },
[&] (StoreLocalNT l) { return l.fp; }
);
if (fp != nullptr) {
always_assert_flog(
fp->type() <= Type::FramePtr,
"Non frame pointer in memory effects:\n inst: {}\n effects: {}",
inst.toString(),
show(ret)
);
}
}
return ret;
}
/*
* MOpFlags is a bitmask so it doesn't fit into the [0,n) pattern of the other
* subops above.
*/
const char* subopToName(MOpFlags f) {
switch (f) {
#define FLAG(name, val) case MOpFlags::name: return #name;
M_OP_FLAGS
#undef FLAG
}
always_assert_flog(false, "Invalid MOpFlags: {}", uint8_t(f));
}
void safe_cast_failure(const std::string& valStr,
const char* fn,
const char* what) {
always_assert_flog(
false,
"conversion of {} failed in {} : {}\n",
valStr, fn, what
);
}
Vptr lookupDestructor(Vout& v, Vreg type) {
auto const table = reinterpret_cast<intptr_t>(g_destructors);
always_assert_flog(deltaFits(table, sz::dword),
"Destructor function table is expected to be in the data "
"segment, with addresses less than 2^31"
);
auto index = v.makeReg();
v << shrli{kShiftDataTypeToDestrIndex, type, index, v.makeReg()};
return baseless(index * 8 + safe_cast<int>(table));
}
TCA emitFuncPrologue(Func* func, int argc, TransKind kind) {
try {
return emitFuncPrologueImpl(func, argc, kind);
} catch (const DataBlockFull& dbFull) {
// Fail hard if the block isn't code.hot.
always_assert_flog(dbFull.name == "hot",
"data block = {}\nmessage: {}\n",
dbFull.name, dbFull.what());
// Otherwise, fall back to code.main and retry.
code().disableHot();
try {
return emitFuncPrologueImpl(func, argc, kind);
} catch (const DataBlockFull& dbFull) {
always_assert_flog(0, "data block = {}\nmessage: {}\n",
dbFull.name, dbFull.what());
}
}
}
void FrameState::refineLocalType(uint32_t id, Type type, SSATmp* typeSource) {
always_assert(id < m_locals.size());
auto& local = m_locals[id];
Type newType = refineType(local.type, type);
ITRACE(2, "updating local {}'s type: {} -> {}\n",
id, local.type, newType);
always_assert_flog(newType != Type::Bottom,
"Bad new type for local {}: {} & {} = {}",
id, local.type, type, newType);
local.type = newType;
local.typeSource = typeSource;
}
bool install_catch_trace(_Unwind_Context* ctx, _Unwind_Exception* exn,
bool do_side_exit, TypedValue unwinder_tv) {
auto const rip = (TCA)_Unwind_GetIP(ctx);
auto catchTraceOpt = mcg->getCatchTrace(rip);
FTRACE(1, "No catch trace entry for ip {}; bailing\n", rip);
if (!catchTraceOpt) return false;
auto catchTrace = *catchTraceOpt;
if (!catchTrace) {
// A few of our optimization passes must be aware of every path out of the
// trace, so throwing through jitted code without a catch block is very
// bad. This is indicated with a present but nullptr entry in the catch
// trace map.
const size_t kCallSize = 5;
const uint8_t kCallOpcode = 0xe8;
auto callAddr = rip - kCallSize;
TCA helperAddr = nullptr;
std::string helperName;
if (*callAddr == kCallOpcode) {
helperAddr = rip + *reinterpret_cast<int32_t*>(callAddr + 1);
}
always_assert_flog(false,
"Translated call to {} threw '{}' without "
"catch block, return address: {}\n",
getNativeFunctionName(helperAddr),
exceptionFromUnwindException(exn)->what(),
rip);
return false;
}
FTRACE(1, "installing catch trace {} for call {} with tv {}, "
"returning _URC_INSTALL_CONTEXT\n",
catchTrace, rip, unwinder_tv.pretty());
// In theory the unwind api will let us set registers in the frame
// before executing our landing pad. In practice, trying to use
// their recommended scratch registers results in a SEGV inside
// _Unwind_SetGR, so we pass things to the handler using the
// RDS. This also simplifies the handler code because it doesn't
// have to worry about saving its arguments somewhere while
// executing the exit trace.
unwindRdsInfo->unwinderScratch = (int64_t)exn;
unwindRdsInfo->doSideExit = do_side_exit;
if (do_side_exit) {
unwindRdsInfo->unwinderTv = unwinder_tv;
}
_Unwind_SetIP(ctx, (uint64_t)catchTrace);
tl_regState = VMRegState::DIRTY;
return true;
}
pid_t LightProcess::proc_open(const char *cmd, const std::vector<int> &created,
const std::vector<int> &desired,
const char *cwd,
const std::vector<std::string> &env) {
int id = GetId();
Lock lock(g_procs[id].m_procMutex);
always_assert(Available());
always_assert(created.size() == desired.size());
auto fout = g_procs[id].m_afdt_fd;
lwp_write(fout, "proc_open");
lwp_write(fout, cmd);
lwp_write(fout, cwd ? cwd : "");
lwp_write_int32(fout, (int)env.size());
for (unsigned int i = 0; i < env.size(); i++) {
lwp_write(fout, env[i]);
}
lwp_write_int32(fout, (int)created.size());
for (unsigned int i = 0; i < desired.size(); i++) {
lwp_write_int32(fout, desired[i]);
}
bool error_send = false;
int save_errno = 0;
for (unsigned int i = 0; i < created.size(); i++) {
if (!send_fd(g_procs[id].m_afdt_fd, created[i])) {
error_send = true;
save_errno = errno;
break;
}
}
std::string buf;
auto fin = g_procs[id].m_afdt_fd;
lwp_read(fin, buf);
if (buf == "error") {
lwp_read_int32(fin, errno);
if (error_send) {
// On this error, the receiver side returns dummy errno,
// use the sender side errno here.
errno = save_errno;
}
return -1;
}
always_assert_flog(buf == "success",
"Unexpected message from light process: `{}'", buf);
int64_t pid = -1;
lwp_read_int64(fin, pid);
always_assert(pid);
return (pid_t)pid;
}
GuardConstraint relaxConstraint(GuardConstraint origGc,
Type knownType, Type toRelax) {
ITRACE(4, "relaxConstraint({}, knownType = {}, toRelax = {})\n",
origGc, knownType, toRelax);
Trace::Indent _i;
// AssertType can be given TCtx, which should never relax.
if (toRelax.maybe(TCctx)) {
always_assert(toRelax <= TCtx);
return origGc;
}
auto const dstType = knownType & toRelax;
always_assert_flog(typeFitsConstraint(dstType, origGc),
"refine({}, {}) doesn't fit {}",
knownType, toRelax, origGc);
// Preserve origGc's weak property.
GuardConstraint newGc{DataTypeGeneric};
newGc.weak = origGc.weak;
while (true) {
if (newGc.isSpecialized()) {
// We need to ask for the right kind of specialization, so grab it from
// origGc.
if (origGc.wantArrayKind()) newGc.setWantArrayKind();
if (origGc.wantClass()) newGc.setDesiredClass(origGc.desiredClass());
}
auto const relaxed = relaxType(toRelax, newGc.category);
auto const newDstType = relaxed & knownType;
if (typeFitsConstraint(newDstType, origGc)) break;
ITRACE(5, "newDstType = {}, newGc = {}; incrementing constraint\n",
newDstType, newGc);
incCategory(newGc.category);
}
// DataTypeCountness can be relaxed to DataTypeGeneric in
// optimizeProfiledGuards, so we can't rely on this category to give type
// information through guards. Since relaxConstraint is used to relax the
// DataTypeCategory for guards, we cannot return DataTypeCountness unless we
// already had it to start with. Instead, we return DataTypeBoxCountness,
// which won't be further relaxed by optimizeProfiledGuards.
if (newGc.category == DataTypeCountness && origGc != DataTypeCountness) {
newGc.category = DataTypeBoxAndCountness;
}
ITRACE(4, "Returning {}\n", newGc);
// newGc shouldn't be any more specific than origGc.
always_assert(newGc.category <= origGc.category);
return newGc;
}
/*
* Build the CFG, then the dominator tree, then use it to validate SSA.
* 1. Each src must be defined by some other instruction, and each dst must
* be defined by the current instruction.
* 2. Each src must be defined earlier in the same block or in a dominator.
* 3. Each dst must not be previously defined.
* 4. Treat tmps defined by DefConst as always defined.
* 5. Each predecessor of a reachable block must be reachable (deleted
* blocks must not have out-edges to reachable blocks).
* 6. The entry block must not have any predecessors.
* 7. The entry block starts with a DefFP instruction.
*/
bool checkCfg(const IRUnit& unit) {
auto const blocksIds = rpoSortCfgWithIds(unit);
auto const& blocks = blocksIds.blocks;
jit::hash_set<const Edge*> edges;
// Entry block can't have predecessors.
always_assert(unit.entry()->numPreds() == 0);
// Entry block starts with DefFP
always_assert(!unit.entry()->empty() &&
unit.entry()->begin()->op() == DefFP);
// Check valid successor/predecessor edges.
for (Block* b : blocks) {
auto checkEdge = [&] (const Edge* e) {
always_assert(e->from() == b);
edges.insert(e);
for (auto& p : e->to()->preds()) if (&p == e) return;
always_assert(false); // did not find edge.
};
checkBlock(b);
if (auto *e = b->nextEdge()) checkEdge(e);
if (auto *e = b->takenEdge()) checkEdge(e);
}
for (Block* b : blocks) {
for (auto const &e : b->preds()) {
always_assert(&e == e.inst()->takenEdge() || &e == e.inst()->nextEdge());
always_assert(e.to() == b);
}
}
// Visit every instruction and make sure their sources are defined in a block
// that dominates the block containing the instruction.
auto const idoms = findDominators(unit, blocksIds);
forEachInst(blocks, [&] (const IRInstruction* inst) {
for (auto src : inst->srcs()) {
if (src->inst()->is(DefConst)) continue;
auto const dom = findDefiningBlock(src);
always_assert_flog(
dom && dominates(dom, inst->block(), idoms),
"src '{}' in '{}' came from '{}', which is not a "
"DefConst and is not defined at this use site",
src->toString(), inst->toString(),
src->inst()->toString()
);
}
});
return true;
}
AStack::AStack(SSATmp* base, int32_t o, int32_t s)
: offset(o), size(s)
{
// Always canonicalize to the outermost frame pointer.
if (base->isA(TStkPtr)) {
auto const defSP = base->inst();
always_assert_flog(defSP->is(DefSP),
"unexpected StkPtr: {}\n", base->toString());
offset -= defSP->extra<DefSP>()->offset.offset;
return;
}
assertx(base->isA(TFramePtr));
auto const defInlineFP = base->inst();
if (defInlineFP->is(DefInlineFP)) {
auto const sp = defInlineFP->src(0)->inst();
offset += defInlineFP->extra<DefInlineFP>()->spOffset.offset;
offset -= sp->extra<DefSP>()->offset.offset;
always_assert_flog(sp->src(0)->inst()->is(DefFP),
"failed to canonicalize to outermost FramePtr: {}\n",
sp->src(0)->toString());
}
}
Block* findMainExitBlock(const IRUnit& unit, SrcKey lastSk) {
bool unreachable = false;
Block* mainExit = nullptr;
FTRACE(5, "findMainExitBlock: looking for exit at {} in unit:\n{}\n",
showShort(lastSk), show(unit));
for (auto block : rpoSortCfg(unit)) {
if (block->back().is(Unreachable)) unreachable = true;
if (endsUnitAtSrcKey(block, lastSk)) {
if (mainExit == nullptr) {
mainExit = block;
continue;
}
always_assert_flog(
mainExit->hint() == Block::Hint::Unlikely ||
block->hint() == Block::Hint::Unlikely,
"findMainExit: 2 likely exits found: B{} and B{}\nlastSk = {}",
mainExit->id(), block->id(), showShort(lastSk)
);
if (mainExit->hint() == Block::Hint::Unlikely) mainExit = block;
}
}
always_assert_flog(
mainExit || unreachable,
"findMainExit: no exit found for lastSk = {}",
showShort(lastSk)
);
FTRACE(5, "findMainExitBlock: mainExit = B{}\n", mainExit->id());
return mainExit;
}
TCA emit_bindjcc1st_stub(CodeBlock& cb, FPInvOffset spOff, TCA jcc,
SrcKey taken, SrcKey next, ConditionCode cc) {
always_assert_flog(taken.resumed() == next.resumed(),
"bind_jcc_1st was confused about resumables");
return emit_ephemeral(
cb,
mcg->getFreeStub(cb, &mcg->cgFixups()),
taken.resumed() ? folly::none : folly::make_optional(spOff),
REQ_BIND_JCC_FIRST,
jcc,
taken.toAtomicInt(),
next.toAtomicInt(),
cc
);
}
Vptr lookupDestructor(Vout& v, Vreg type) {
auto const table = reinterpret_cast<intptr_t>(g_destructors);
always_assert_flog(deltaFits(table, sz::dword),
"Destructor function table is expected to be in the data "
"segment, with addresses less than 2^31"
);
static_assert((KindOfString >> kShiftDataTypeToDestrIndex == 1) &&
(KindOfArray >> kShiftDataTypeToDestrIndex == 2) &&
(KindOfObject >> kShiftDataTypeToDestrIndex == 3) &&
(KindOfResource >> kShiftDataTypeToDestrIndex == 4) &&
(KindOfRef >> kShiftDataTypeToDestrIndex == 5),
"lookup of destructors depends on KindOf* values");
auto index = v.makeReg();
v << shrli{kShiftDataTypeToDestrIndex, type, index, v.makeReg()};
return baseless(index * 8 + safe_cast<int>(table));
}
Type refineType(Type oldType, Type newType) {
// It's OK for the old and new inner types of boxed values not to
// intersect, since the inner type is really just a prediction.
// But if they do intersect, we keep the intersection. This is
// necessary to keep the type known in situations like:
// oldType: Boxed{Obj}
// newType: Boxed{Obj<C>, InitNull}
if (oldType.isBoxed() && newType.isBoxed() && oldType.not(newType)) {
return oldType < newType ? oldType : newType;
}
auto const result = oldType & newType;
always_assert_flog(result != Type::Bottom,
"refineType({}, {}) failed", oldType, newType);
return result;
}
/*
* relaxConstraint returns the least specific TypeConstraint 'tc' that doesn't
* prevent the intersection of knownType and relaxType(toRelax, tc) from
* satisfying origTc. It is used in IRBuilder::constrainValue and
* IRBuilder::constrainStack to determine how to constrain the typeParam and
* src values of CheckType/CheckStk instructions, and the src values of
* AssertType/AssertStk instructions.
*
* AssertType example:
* t24:Obj<C> = AssertType<{Obj<C>|InitNull}> t4:Obj
*
* If constrainValue is called with (t24, DataTypeSpecialized), relaxConstraint
* will be called with (DataTypeSpecialized, Obj<C>|InitNull, Obj). After a few
* iterations it will determine that constraining Obj with DataTypeCountness
* will still allow the result type of the AssertType instruction to satisfy
* DataTypeSpecialized, because relaxType(Obj, DataTypeCountness) == Obj.
*/
TypeConstraint relaxConstraint(const TypeConstraint origTc,
const Type knownType, const Type toRelax) {
ITRACE(4, "relaxConstraint({}, knownType = {}, toRelax = {})\n",
origTc, knownType, toRelax);
Trace::Indent _i;
auto const dstType = refineType(knownType, toRelax);
always_assert_flog(typeFitsConstraint(dstType, origTc),
"refine({}, {}) doesn't fit {}",
knownType, toRelax, origTc);
// Preserve origTc's weak property.
TypeConstraint newTc{DataTypeGeneric, DataTypeGeneric};
newTc.weak = origTc.weak;
while (true) {
if (newTc.isSpecialized()) {
// We need to ask for the right kind of specialization, so grab it from
// origTc.
if (origTc.wantArrayKind()) newTc.setWantArrayKind();
if (origTc.wantClass()) newTc.setDesiredClass(origTc.desiredClass());
}
auto const relaxed = relaxType(toRelax, newTc);
auto const newDstType = refineType(relaxed, knownType);
if (typeFitsConstraint(newDstType, origTc)) break;
ITRACE(5, "newDstType = {}, newTc = {}; ", newDstType, newTc);
if (newTc.category == DataTypeGeneric ||
!typeFitsOuterConstraint(newDstType, origTc)) {
FTRACE(5, "incrementing outer\n");
incCategory(newTc.category);
} else if (!typeFitsInnerConstraint(newDstType, origTc)) {
FTRACE(5, "incrementing inner\n");
incCategory(newTc.innerCat);
} else {
not_reached();
}
}
ITRACE(4, "Returning {}\n", newTc);
// newTc shouldn't be any more specific than origTc.
always_assert(newTc.category <= origTc.category &&
newTc.innerCat <= origTc.innerCat);
return newTc;
}
/*
* Modify a GeneralEffects to take potential VM re-entry into account. This
* affects may-load, may-store, and kills information for the instruction. The
* GeneralEffects should already contain AHeapAny in both loads and stores if
* it affects those locations for reasons other than re-entry, but does not
* need to if it doesn't.
*
* For loads, we need to take into account EnableArgsInBacktraces: if this flag
* is on, any instruction that could re-enter could call debug_backtrace, which
* could read the argument locals of any activation record in the callstack.
* We don't try to limit the load effects to argument locals here, though, and
* just union in all the locals.
*
* For kills, locations on the eval stack below the re-entry depth should all
* be added.
*
* Important note: because of the `kills' set modifications, an instruction may
* not report that it can re-enter if it actually can't. The reason this can
* go wrong is that if the instruction was in an inlined function, if we've
* removed the DefInlineFP its spOff will not be meaningful (unless it's a
* DecRef instruction, which we explicitly adjust in dce.cpp). In this case
* the `kills' set will refer to the wrong stack locations. In general this
* means instructions that can re-enter must have catch traces---but a few
* other instructions are exceptions, either since they are not allowed in
* inlined functions or because they take the (possibly-inlined) FramePtr as a
* source.
*/
GeneralEffects may_reenter(const IRInstruction& inst, GeneralEffects x) {
auto const may_reenter_is_ok =
(inst.taken() && inst.taken()->isCatch()) ||
inst.is(DecRef,
ReleaseVVAndSkip,
CIterFree,
MIterFree,
MIterNext,
MIterNextK,
IterFree,
ABCUnblock,
GenericRetDecRefs);
always_assert_flog(
may_reenter_is_ok,
"instruction {} claimed may_reenter, but it isn't allowed to say that",
inst
);
/*
* We want to union `killed_stack' into whatever else the instruction already
* said it must kill, but if we end up with an unrepresentable AliasClass we
* can't return a set that's too big (the `kills' set is unlike the other
* AliasClasses in GeneralEffects in that means it kills /everything/ in the
* set, since it's must-information).
*
* If we can't represent the union, just take the stack, in part because we
* have some debugging asserts about this right now---but also nothing
* actually uses may_reenter with a non-AEmpty kills at the time of this
* writing anyway.
*/
auto const killed_stack =
stack_below(inst.marker().fp(), -inst.marker().spOff().offset - 1);
auto const kills_union = x.kills.precise_union(killed_stack);
auto const new_kills = kills_union ? *kills_union : killed_stack;
return GeneralEffects {
x.loads | AHeapAny
| (RuntimeOption::EnableArgsInBacktraces ? AFrameAny : AEmpty),
x.stores | AHeapAny,
x.moves,
new_kills
};
}
DataType Type::toDataType() const {
assertx(!maybe(TPtrToGen) || m_bits == kBottom);
assertx(isKnownDataType());
// Order is important here: types must progress from more specific
// to less specific to return the most specific DataType.
if (*this <= TUninit) return KindOfUninit;
if (*this <= TInitNull) return KindOfNull;
if (*this <= TBool) return KindOfBoolean;
if (*this <= TInt) return KindOfInt64;
if (*this <= TDbl) return KindOfDouble;
if (*this <= TStaticStr) return KindOfStaticString;
if (*this <= TStr) return KindOfString;
if (*this <= TArr) return KindOfArray;
if (*this <= TObj) return KindOfObject;
if (*this <= TRes) return KindOfResource;
if (*this <= TBoxedCell) return KindOfRef;
if (*this <= TCls) return KindOfClass;
always_assert_flog(false,
"Bad Type {} in Type::toDataType()", *this);
}
DataType Type::toDataType() const {
assert(!isPtr());
assert(isKnownDataType());
// Order is important here: types must progress from more specific
// to less specific to return the most specific DataType.
if (subtypeOf(Uninit)) return KindOfUninit;
if (subtypeOf(InitNull)) return KindOfNull;
if (subtypeOf(Bool)) return KindOfBoolean;
if (subtypeOf(Int)) return KindOfInt64;
if (subtypeOf(Dbl)) return KindOfDouble;
if (subtypeOf(StaticStr)) return KindOfStaticString;
if (subtypeOf(Str)) return KindOfString;
if (subtypeOf(Arr)) return KindOfArray;
if (subtypeOf(Obj)) return KindOfObject;
if (subtypeOf(Res)) return KindOfResource;
if (subtypeOf(BoxedCell)) return KindOfRef;
if (subtypeOf(Cls)) return KindOfClass;
always_assert_flog(false,
"Bad Type {} in Type::toDataType()", *this);
}
Block* findDefiningBlock(const SSATmp* t, const IdomVector& idoms) {
assertx(!t->inst()->is(DefConst));
auto const srcInst = t->inst();
if (srcInst->hasEdges()) {
auto const next = srcInst->next();
UNUSED auto const taken = srcInst->taken();
always_assert_flog(
next && taken,
"hasEdges instruction defining a dst had no edges:\n {}\n",
srcInst->toString()
);
for (const auto& arc : next->preds()) {
auto pred = arc.from();
if (pred != srcInst->block() && !dominates(next, pred, idoms)) {
return nullptr;
}
}
return next;
}
return srcInst->block();
}
void Extension::CompileSystemlib(const std::string &slib,
const std::string &name) {
// TODO (t3443556) Bytecode repo compilation expects that any errors
// encountered during systemlib compilation have valid filename pointers
// which won't be the case for now unless these pointers are long-lived.
auto const moduleName = makeStaticString(name.c_str());
auto const unit = compile_systemlib_string(slib.c_str(), slib.size(),
moduleName->data());
always_assert_flog(unit, "No unit created for systemlib `{}'", name);
const StringData* msg;
int line;
if (unit->compileTimeFatal(msg, line) ||
unit->parseFatal(msg, line)) {
std::fprintf(stderr, "Systemlib `%s' contains a fataling unit: %s, %d\n",
name.c_str(),
msg->data(),
line);
_Exit(0);
}
unit->merge();
s_systemlib_units.push_back(unit);
}
bool dontGuardAnyInputs(Op op) {
switch (op) {
case Op::IterBreak:
case Op::DecodeCufIter:
case Op::IterNext:
case Op::IterNextK:
case Op::WIterInit:
case Op::WIterInitK:
case Op::WIterNext:
case Op::WIterNextK:
case Op::MIterInit:
case Op::MIterInitK:
case Op::MIterNext:
case Op::MIterNextK:
case Op::IterInitK:
case Op::IterInit:
case Op::JmpZ:
case Op::JmpNZ:
case Op::Jmp:
case Op::JmpNS:
case Op::FCallArray:
case Op::FCall:
case Op::FCallD:
case Op::FCallAwait:
case Op::ClsCnsD:
case Op::FPassCW:
case Op::FPassCE:
case Op::FPassR:
case Op::FPassV:
case Op::FPassG:
case Op::FPassL:
case Op::FPassS:
case Op::FCallBuiltin:
case Op::NewStructArray:
case Op::Switch:
case Op::SSwitch:
case Op::Lt:
case Op::Lte:
case Op::Gt:
case Op::Gte:
case Op::Cmp:
case Op::SetOpL:
case Op::InitProp:
case Op::BreakTraceHint:
case Op::IsTypeL:
case Op::IsTypeC:
case Op::IncDecL:
case Op::DefCls:
case Op::FPushCuf:
case Op::FPushCufF:
case Op::FPushCufSafe:
case Op::IncStat:
case Op::Eq:
case Op::Neq:
case Op::AssertRATL:
case Op::AssertRATStk:
case Op::SetL:
case Op::BindL:
case Op::EmptyL:
case Op::CastBool:
case Op::Same:
case Op::NSame:
case Op::Yield:
case Op::YieldK:
case Op::ContEnter:
case Op::ContRaise:
case Op::CreateCont:
case Op::Await:
case Op::BitAnd:
case Op::BitOr:
case Op::BitXor:
case Op::Sub:
case Op::Mul:
case Op::SubO:
case Op::MulO:
case Op::Add:
case Op::AddO:
case Op::AGetC:
case Op::AGetL:
case Op::AKExists:
case Op::AddElemC:
case Op::AddNewElemC:
case Op::Array:
case Op::ArrayIdx:
case Op::BareThis:
case Op::BindG:
case Op::BindS:
case Op::BitNot:
case Op::CGetG:
case Op::CGetQuietG:
case Op::CGetL:
case Op::CGetQuietL:
case Op::CGetL2:
case Op::CGetS:
case Op::CUGetL:
case Op::CIterFree:
case Op::CastArray:
case Op::CastDouble:
case Op::CastInt:
case Op::CastObject:
case Op::CastString:
case Op::CheckProp:
case Op::CheckThis:
case Op::Clone:
case Op::Cns:
case Op::CnsE:
case Op::CnsU:
case Op::MapAddElemC:
case Op::ColAddNewElemC:
case Op::ColFromArray:
case Op::ConcatN:
case Op::Concat:
case Op::ContCheck:
case Op::ContCurrent:
case Op::ContKey:
case Op::ContValid:
case Op::ContStarted:
case Op::ContGetReturn:
case Op::CreateCl:
case Op::DefCns:
case Op::DefFunc:
case Op::Dir:
case Op::Div:
case Op::Double:
case Op::Dup:
case Op::EmptyG:
case Op::EmptyS:
case Op::FPushClsMethodD:
case Op::FPushClsMethod:
case Op::FPushClsMethodF:
case Op::FPushCtor:
case Op::FPushCtorD:
case Op::FPushCufIter:
case Op::FPushFunc:
case Op::FPushFuncD:
case Op::FPushFuncU:
case Op::FPushObjMethodD:
case Op::False:
case Op::File:
case Op::GetMemoKey:
case Op::Idx:
case Op::InitThisLoc:
case Op::InstanceOf:
case Op::InstanceOfD:
case Op::Int:
case Op::IssetG:
case Op::IssetL:
case Op::IssetS:
case Op::IterFree:
case Op::LateBoundCls:
case Op::MIterFree:
case Op::Mod:
case Op::Pow:
case Op::NameA:
case Op::NativeImpl:
case Op::NewArray:
case Op::NewCol:
case Op::NewLikeArrayL:
case Op::NewMixedArray:
case Op::NewDictArray:
case Op::NewPackedArray:
case Op::NewVecArray:
case Op::Not:
case Op::Null:
case Op::NullUninit:
case Op::OODeclExists:
case Op::Parent:
case Op::PopA:
case Op::PopC:
case Op::PopR:
case Op::PopV:
case Op::Print:
case Op::PushL:
case Op::RetC:
case Op::RetV:
case Op::Self:
case Op::SetG:
case Op::SetS:
case Op::Shl:
case Op::Shr:
case Op::Silence:
case Op::StaticLoc:
case Op::StaticLocInit:
case Op::String:
case Op::This:
case Op::True:
case Op::Unbox:
case Op::UnboxR:
case Op::UnsetL:
case Op::VGetG:
case Op::VGetL:
case Op::VGetS:
case Op::VerifyParamType:
case Op::VerifyRetTypeC:
case Op::VerifyRetTypeV:
case Op::WHResult:
case Op::Xor:
case Op::BaseNC:
case Op::BaseNL:
case Op::BaseGC:
case Op::BaseGL:
case Op::FPassBaseNC:
case Op::FPassBaseNL:
case Op::FPassBaseGC:
case Op::FPassBaseGL:
case Op::BaseSC:
case Op::BaseSL:
case Op::BaseL:
case Op::FPassBaseL:
case Op::BaseC:
case Op::BaseR:
case Op::BaseH:
case Op::Dim:
case Op::FPassDim:
case Op::QueryM:
case Op::VGetM:
case Op::FPassM:
case Op::SetM:
case Op::IncDecM:
case Op::SetOpM:
case Op::BindM:
case Op::UnsetM:
case Op::SetWithRefLML:
case Op::SetWithRefRML:
return false;
// These are instructions that are always interp-one'd, or are always no-ops.
case Op::LowInvalid:
case Op::Nop:
case Op::Box:
case Op::BoxR:
case Op::BoxRNop:
case Op::UnboxRNop:
case Op::RGetCNop:
case Op::AddElemV:
case Op::AddNewElemV:
case Op::ClsCns:
case Op::Exit:
case Op::Fatal:
case Op::Unwind:
case Op::Throw:
case Op::CGetL3:
case Op::CGetN:
case Op::CGetQuietN:
case Op::VGetN:
case Op::IssetN:
case Op::EmptyN:
case Op::SetN:
case Op::SetOpN:
case Op::SetOpG:
case Op::SetOpS:
case Op::IncDecN:
case Op::IncDecG:
case Op::IncDecS:
case Op::BindN:
case Op::UnsetN:
case Op::UnsetG:
case Op::FPushObjMethod:
case Op::FPassC:
case Op::FPassVNop:
case Op::FPassN:
case Op::FCallUnpack:
case Op::CufSafeArray:
case Op::CufSafeReturn:
case Op::Incl:
case Op::InclOnce:
case Op::Req:
case Op::ReqOnce:
case Op::ReqDoc:
case Op::Eval:
case Op::DefClsNop:
case Op::DefTypeAlias:
case Op::Catch:
case Op::HighInvalid:
case Op::ContAssignDelegate:
case Op::ContEnterDelegate:
case Op::YieldFromDelegate:
case Op::ContUnsetDelegate:
return true;
}
always_assert_flog(0, "invalid opcode {}\n", static_cast<uint32_t>(op));
}
/*
* Attempts to begin inlining, and returns whether or not it successed.
*
* When doing gen-time inlining, we set up a series of IR instructions
* that looks like this:
*
* fp0 = DefFP
* sp = DefSP<offset>
*
* // ... normal stuff happens ...
*
* // FPI region:
* SpillFrame sp, ...
* // ... probably some StStks due to argument expressions
* fp2 = DefInlineFP<func,retBC,retSP,off> sp
*
* // ... callee body ...
*
* InlineReturn fp2
*
* In DCE we attempt to remove the InlineReturn and DefInlineFP instructions if
* they aren't needed.
*/
bool beginInlining(IRGS& env,
unsigned numParams,
const Func* target,
Offset returnBcOffset) {
auto const& fpiStack = env.irb->fpiStack();
assertx(!fpiStack.empty() &&
"Inlining does not support calls with the FPush* in a different Tracelet");
assertx(returnBcOffset >= 0 && "returnBcOffset before beginning of caller");
assertx(curFunc(env)->base() + returnBcOffset < curFunc(env)->past() &&
"returnBcOffset past end of caller");
FTRACE(1, "[[[ begin inlining: {}\n", target->fullName()->data());
SSATmp** params = (SSATmp**)alloca(sizeof(SSATmp*) * numParams);
for (unsigned i = 0; i < numParams; ++i) {
params[numParams - i - 1] = popF(env);
}
auto const prevSP = fpiStack.front().returnSP;
auto const prevSPOff = fpiStack.front().returnSPOff;
spillStack(env);
auto const calleeSP = sp(env);
always_assert_flog(
prevSP == calleeSP,
"FPI stack pointer and callee stack pointer didn't match in beginInlining"
);
auto const& info = fpiStack.front();
always_assert(!isFPushCuf(info.fpushOpc) && !info.interp);
auto ctx = [&] {
if (info.ctx || isFPushFunc(info.fpushOpc)) {
return info.ctx;
}
constexpr int32_t adjust = offsetof(ActRec, m_r) - offsetof(ActRec, m_this);
IRSPOffset ctxOff{invSPOff(env) - info.returnSPOff - adjust};
return gen(env, LdStk, TCtx, IRSPOffsetData{ctxOff}, sp(env));
}();
DefInlineFPData data;
data.target = target;
data.retBCOff = returnBcOffset;
data.fromFPushCtor = isFPushCtor(info.fpushOpc);
data.ctx = ctx;
data.retSPOff = prevSPOff;
data.spOffset = offsetFromIRSP(env, BCSPOffset{0});
// Push state and update the marker before emitting any instructions so
// they're all given markers in the callee.
auto const key = SrcKey {
target,
target->getEntryForNumArgs(numParams),
false
};
env.bcStateStack.emplace_back(key);
env.inlineLevel++;
updateMarker(env);
auto const calleeFP = gen(env, DefInlineFP, data, calleeSP, fp(env));
for (unsigned i = 0; i < numParams; ++i) {
stLocRaw(env, i, calleeFP, params[i]);
}
const bool hasVariadicArg = target->hasVariadicCaptureParam();
for (unsigned i = numParams; i < target->numLocals() - hasVariadicArg; ++i) {
/*
* Here we need to be generating hopefully-dead stores to initialize
* non-parameter locals to KindOfUninit in case we have to leave the trace.
*/
stLocRaw(env, i, calleeFP, cns(env, TUninit));
}
if (hasVariadicArg) {
auto argNum = target->numLocals() - 1;
always_assert(numParams <= argNum);
stLocRaw(env, argNum, calleeFP, cns(env, staticEmptyArray()));
}
return true;
}
void unknownBaseType(const TypedValue* tv) {
always_assert_flog(
false,
"Unknown KindOf: {} in member operation base",
static_cast<uint8_t>(tv->m_type));
}
void Repo::initCentral() {
std::string error;
assert(m_dbc == nullptr);
auto tryPath = [this, &error](const char* path) {
std::string subErr;
if (openCentral(path, subErr) == RepoStatus::error) {
folly::format(&error, " {}\n", subErr.empty() ? path : subErr);
return false;
}
return true;
};
auto fail_no_repo = [&error] {
error = "Failed to initialize central HHBC repository:\n" + error;
// Database initialization failed; this is an unrecoverable state.
Logger::Error("%s", error.c_str());
if (Process::IsInMainThread()) {
exit(1);
}
always_assert_flog(false, "{}", error);
};
// Try Repo.Central.Path
if (!RuntimeOption::RepoCentralPath.empty() &&
tryPath(RuntimeOption::RepoCentralPath.c_str())) {
return;
}
// Try HHVM_REPO_CENTRAL_PATH
const char* HHVM_REPO_CENTRAL_PATH = getenv("HHVM_REPO_CENTRAL_PATH");
if (HHVM_REPO_CENTRAL_PATH != nullptr &&
tryPath(HHVM_REPO_CENTRAL_PATH)) {
return;
}
if (!RuntimeOption::RepoAllowFallbackPath) fail_no_repo();
// Try "$HOME/.hhvm.hhbc".
char* HOME = getenv("HOME");
if (HOME != nullptr) {
std::string centralPath = HOME;
centralPath += "/.hhvm.hhbc";
if (tryPath(centralPath.c_str())) {
return;
}
}
#ifndef _WIN32
// Try the equivalent of "$HOME/.hhvm.hhbc", but look up the home directory
// in the password database.
{
passwd pwbuf;
passwd* pwbufp;
long bufsize = sysconf(_SC_GETPW_R_SIZE_MAX);
if (bufsize != -1) {
auto buf = new char[bufsize];
SCOPE_EXIT { delete[] buf; };
if (!getpwuid_r(getuid(), &pwbuf, buf, size_t(bufsize), &pwbufp)
&& (HOME == nullptr || strcmp(HOME, pwbufp->pw_dir))) {
std::string centralPath = pwbufp->pw_dir;
centralPath += "/.hhvm.hhbc";
if (tryPath(centralPath.c_str())) {
return;
}
}
}
}