If the abstract interpreter ever pushes bottom (i.e., a contradition) to the stack, it might as well stop -- this indicates we're looking at unreachable code (e.g., after a deopt that always deopts, or a branch that always branches).

My original idea was to tweak the code generator to auto-generate such checks for the output stack effects (hand-waving a bit around array outputs, which are special anyways). But when I implemented that, I quickly found that most such pushes are freshly created symbols (e.g. sym_new_const() or sym_new_undefined()) that we already know cannot be bottom.

Also, there's a category of stack writes that evades easy detection, e.g. when _GUARD_BOTH_INT calls sym_set_type(left, &PyLong_Type) (and ditto for right), this is part of an opcode whose input and output stack effects are identical, and for such opcodes, the push operation is actually a write (without changing the stack pointer) that is generated on a different code path in the generator.

So no I'm considering to just hand-write bottom checks where they seem relevant. Perhaps we can change the sym_set_...() functions to return a bool result indicating whether they produced a bottom value, and change the (hand-written) code that calls these to emit something like

if (sym_set_null(sym)) {
    goto hit_bottom;
}

where hit_bottom is a new error label that prints a different debug message.

I'm still weighing my options though...

Linked PRs

• gh-116088: Insert bottom checks after all sym_set_...() calls #116089