`TO_BOOL_INT` repeatedly misses for non-compact exact integers

Feature or enhancement

Proposal:

Versions

CPython 3.15.0b4, Ubuntu 24.04.4 LTS, gcc 13.3.0

Enhancement

I noticed that the specialization function for TO_BOOL and the guard used by TO_BOOL_INT accept different sets of integers.

_Py_Specialize_ToBool() selects TO_BOOL_INT for any exact integer:

if (PyLong_CheckExact(value)) {
    specialized_op = TO_BOOL_INT;
    goto success;
}

However, the TO_BOOL_INT macro uses _GUARD_TOS_INT, and that requires the integer to be compact:

op(_GUARD_TOS_INT, (value -- value)) {
    PyObject *value_o = PyStackRef_AsPyObjectBorrow(value);
    EXIT_IF(!_PyLong_CheckExactAndCompact(value_o));
}

As a result, a non-compact exact integer can cause TO_BOOL_INT to be selected and then immediately miss its guard.

This mismatch appears to have been introduced by GH-143759. Before the refactoring, TO_BOOL_INT had its own PyLong_CheckExact() guard. The refactoring replaced it with a macro using the shared _GUARD_TOS_INT, whose domain is narrower.

I see two alternative ways to fix this.

Option 1: narrow the specialization function

One option would be to change the integer check in _Py_Specialize_ToBool() so that it agrees with the existing guard:

if (_PyLong_CheckExactAndCompact(value)) {
    specialized_op = TO_BOOL_INT;
    goto success;
}

With this change, non-compact integers remain on the generic TO_BOOL path instead of repeatedly entering and missing TO_BOOL_INT.

My main concern with this option was whether the additional compactness check in the specialization function could regress the common case (compact int), so I benchmarked both compact and non-compact integers.

The benchmark target contained 100 TO_BOOL sites and was warmed up before each measurement:

start = time.perf_counter_ns()
for _ in range(10_000):
    target(value)
elapsed = time.perf_counter_ns() - start

Positive values mean that the patched build was faster:

Version	Input	Performance change
CPython 3.15	non-compact exact int	+2.21% (95% CI: +1.64% to +2.86%)
CPython 3.15	compact exact int	+0.31% (95% CI: −0.15% to +0.70%)
CPython main	non-compact exact int	+4.03% (95% CI: +1.43% to +7.10%)
CPython main	compact exact int	−0.23% (95% CI: −0.50% to +0.07%)

The non-compact case improved because it no longer repeatedly enters and misses TO_BOOL_INT. For compact integers, both confidence intervals include zero, so I did not find evidence that the stronger specialization check causes a regression.

Option 2: give TO_BOOL_INT an exact-int guard

The other option is to keep _Py_Specialize_ToBool() unchanged and add a guard that checks exact type without requiring compactness:

op(_GUARD_TOS_EXACT_INT, (value -- value)) {
    PyObject *value_o = PyStackRef_AsPyObjectBorrow(value);
    EXIT_IF(!PyLong_CheckExact(value_o));
}

The TO_BOOL_INT macro would then use the new guard:

macro(TO_BOOL_INT) =
    _GUARD_TOS_EXACT_INT +
    unused/1 +
    unused/2 +
    _TO_BOOL_INT +
    _POP_TOP_INT;

This would restore the specialization domain from before GH-143759.

---

I am not sure which of these two options is preferable, but the current mismatch seems worth fixing, so I am opening this issue to get feedback on which direction would be better.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-155531
• gh-155699
• gh-155700

Start with Python/specialize.c at _Py_Specialize_ToBool() and Python/bytecodes.c at TO_BOOL_INT and _GUARD_TOS_INT. Review linked PRs gh-155531, gh-155699, and gh-155700, then reproduce the stated warm-up benchmark; done means the specialization and guard domains agree without repeated misses.