YJIT: Side-exit on String#dup when it's not leaf #13921
Conversation
shape.h
Outdated
| @@ -48,7 +48,8 @@ enum shape_id_fl_type { | |||
|
|
|||
| // This masks allows to check if a shape_id contains any ivar. | |||
| // It rely on ROOT_SHAPE_WITH_OBJ_ID==1. | |||
| #define SHAPE_ID_HAS_IVAR_MASK (SHAPE_ID_FL_TOO_COMPLEX | (SHAPE_ID_OFFSET_MASK - 1)) | |||
| #define SHAPE_ID_HAS_IVAR_MASK 0x807fffe | |||
There was a problem hiding this comment.
Why do you need this constant inline?
There was a problem hiding this comment.
This is a bindgen limitation where it can't tell whether a macro is a constant. An alternative is to put the expression in an anonymous enum, which bindgen can figure out since it's a language level thing to force a value to be compile time known and not allocate storage for it.
Could you try that? If it doesn't work for some C code or whatever, as is isn't too horrible.
There was a problem hiding this comment.
A secret third thing is to make a uint32_t abc = SHAPE_ID_HAS_IVAR_MASK in jit.c and then bindgen that
There was a problem hiding this comment.
(If we care about signedness, which it looks like we do, we can't really rely on enum signedness. and I think enum width, too)
There was a problem hiding this comment.
No need to worry about signedness because the operands in the expression come from enum themselves 😁 Go integer promotion.
This comment has been minimized.
This comment has been minimized.
This fixes ccbbe06.
rb_obj_dupis not leaf, so it gives a wrong backtrace if you skip pushing a frame when the string has ivars and falls back torb_obj_dup. This PR changes it to side-exit, reversing the simplification we made in #13606.