Op 05-12-2023 om 00:12 schreef Tomas Volf:
> Thank you for the review.
> 
> On 2023-12-04 22:06:42 +0100, Maxime Devos wrote:
>>
>>
>> Op 29-11-2023 om 17:40 schreef Tomas Volf:
>>> +    guix shell -D -f guix.scm --pure -- ./configure --enable-mini-gmp
>>
>> Also -fexcess-precision=standard (see #49368 / #49659 on debbugs) (at least
>> for i*86, should be harmless for other architectures though).
> 
> That explains why I did not catch it (I did run the test suite, but I am on
> x86_64).

I think it might, in theory, also happen on x86_64, and perhaps other 
architectures with ‘excessive’ floats (I don't know any, but they could 
exist).

If you encounter the issue or not depends on how, precisely, the 
compiler optimises things, allocates registers/pushes things on the 
stack ... you might have missed the issue by accident/luck.

> Out of curiosity, since it should be harmless elsewhere, is there a reason this
> is not a flag set by default?

I guess for compatible with non-GCC, non-Clang compilers.
But given how configure.ac now checks for support of '-flto', it should 
be easy to copy/edit things to do the same for -fexcess-precision.

(Also, it is plausible that GCC simply does not recognise the flag on 
non-i*86 and non-x86_64 architectures -- someone would need to actually 
check this.)

Best regards,
Maxime Devos.