Specialization for longs should check for compactness #138302
Description
Bug report
Bug description:
As part of #135479, I made the specializing interpreter only specialize compact longs.
I forgot to check for the long's compactness in the specializer. The downside is that there's a bit of perf lost when adding large ints now due to repeated failed specialization and deopt.
For example for the following code:
def foo():
for _ in range(100000000):
y = 1000000000000000000000000000000000
y + y
foo()
y + y specializes to _BINARY_OP_ADD_INT, when it should be _BINARY_OP/_BINARY_OP_EXTEND
The fix is simple:
Change this line
https://github.com/python/cpython/blob/main/Python/specialize.c#L2598
from
if (PyLong_CheckExact(lhs)) {
specialize(instr, BINARY_OP_ADD_INT);
return;
}
to
if (_PyLong_CheckExactAndCompact(lhs) && _PyLong_CheckExactAndCompact(rhs)) {
specialize(instr, BINARY_OP_ADD_INT);
return;
}
Then, add a test in Lib/test/test_opcache.py, checking for _BINARY_OP/_BINARY_OP_EXTEND.
CPython versions tested on:
CPython main branch, 3.15, 3.14
Operating systems tested on:
No response