Description
Feature or enhancement
Proposal:
Partially discussed here: faster-cpython/ideas#728
We want to teach the JIT types optimizer the types of its stack variables. This, when combined with pure operations, will allow us to shift more things to the DSL. There are numerous benefits to moving things to the DSL:
- Less error prone.
- The DSL can work with more information. Right now the
pureoperations don't check for their types before executing them. In a simple uop format, this is safe, but it's not robust if we change the uop format or project more advanced uops.
The implementation will only affect optimizer_bytecodes.c. There has been some discussion about the growing complexity of bytecodes.c. Since this is also specific to the types specializer, we should keep it out of bytecodes.c.
Consider the following:
Current:
op(_GUARD_NOS_UNICODE, (nos, unused -- nos, unused)) {
if (sym_matches_type(nos, &PyUnicode_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
sym_set_type(nos, &PyUnicode_Type);
}
New:
op(_GUARD_NOS_UNICODE, (nos, unused -- type(&PyUnicode_Type) nos, unused)) {
if (sym_matches_type(nos, &PyUnicode_Type)) {
REPLACE_OP(this_instr, _NOP, 0, 0);
}
}
This isn't very powerful on it's own, but when combined with the pure optimization, it will allow us to validate properly advanced pure uops before evaluating them.
Consider the following code right now:
op(_BINARY_OP_ADD_INT, (left, right -- res)) {
if (sym_is_const(ctx, left) && sym_is_const(ctx, right)) {
assert(PyLong_CheckExact(sym_get_const(ctx, left)));
assert(PyLong_CheckExact(sym_get_const(ctx, right)));
PyObject *temp = _PyLong_Add((PyLongObject *)sym_get_const(ctx, left),
(PyLongObject *)sym_get_const(ctx, right));
if (temp == NULL) {
goto error;
}
res = sym_new_const(ctx, temp);
Py_DECREF(temp);
// TODO gh-115506:
// replace opcode with constant propagated one and add tests!
}
else {
res = sym_new_type(ctx, &PyLong_Type);
}
}
The constant eval part doesn't check for matching types first. This is currently safe as _BINARY_OP_ADD_INT always follows a guard in tier 1. However, it's not safe once we do things like combining specialization passes (#128939), reorganizing them, or introducing more advanced pure uops. it would be tedious and error-prone to handwrite these checks ourselves. When the pure optimization is combined with the type attributes, or with const attributes, they will be able to automatically validate themselves.
Has this already been discussed elsewhere?
No response given
Links to previous discussion of this feature:
No response