mirrors/cpython
1
0
Fork 0
mirror of https://github.com/python/cpython.git synced 2025-08-04 08:59:19 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 11

gh-132876: workaround broken ldexp() on Windows 10 (#133135)

Browse source 
* gh-132876: workaround broken ldexp() on Windows 10

ldexp() fails to round subnormal results before Windows 11,
so hide their bug.

Co-authored-by: Tim Peters <tim.peters@gmail.com>
This commit is contained in:
Sergey B Kirpichev 2025-05-26 05:44:33 +03:00 • committed by GitHub
parent 0e3bc962c6
commit cf8941c603
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 31 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

21
Modules/mathmodule.c
View file

@ -2161,6 +2161,27 @@ math_ldexp_impl(PyObject *module, double x, PyObject *i)
} else {
errno = 0;
r = ldexp(x, (int)exp);
#ifdef _MSC_VER
if (DBL_MIN > r && r > -DBL_MIN) {
/* Denormal (or zero) results can be incorrectly rounded here (rather,
truncated). Fixed in newer versions of the C runtime, included
with Windows 11. */
int original_exp;
frexp(x, &original_exp);
if (original_exp > DBL_MIN_EXP) {
/* Shift down to the smallest normal binade. No bits lost. */
int shift = DBL_MIN_EXP - original_exp;
x = ldexp(x, shift);
exp -= shift;
}
/* Multiplying by 2**exp finishes the job, and the HW will round as
appropriate. Note: if exp < -DBL_MANT_DIG, all of x is shifted
to be < 0.5ULP of smallest denorm, so should be thrown away. If
exp is so very negative that ldexp underflows to 0, that's fine;
no need to check in advance. */
r = x*ldexp(1.0, (int)exp);
}
#endif
if (isinf(r))
errno = ERANGE;
}