Re: GUB and mpfr/mpc

[Dan Eble]

On Dec 6, 2014, at 09:12 , David Kastrup <address@hidden> wrote:
> 
> Dan Eble <address@hidden> writes:
> 
>> I’ve only skimmed some web pages on this subject, but if the following
>> can be tolerated, it seems like the simplest way to reduce pain in
>> short order:
>> 
>>    On more modern x86 processors that support SSE2, specifying
>>    the compiler options -mfpmath=sse -msse2 ensures all float
>>    and double operations are performed in SSE registers and
>>    correctly rounded.
>>    [https://gcc.gnu.org/wiki/FloatingPointMath]
>> 
>> It worked for my example program.
> 
> "It will tend to fix known test cases and create a new set of previously
> working failures”.
[…]
> I think there is a less architecture-dependent option that keeps gcc
> from using any extra precision for intermediate results but it may be
> somewhat expensive.  On the other hand, numerics are not where LilyPond
> spends most of its time.
> 
> Probably
> 
> '-ffloat-store’
[…]
> would do the trick.  But we really rather should try to become
> independent from exact equalities as algorithm terminators.

I agree that changing the algorithms is preferred; I didn’t mean to suggest 
otherwise.  But if that’s not going to happen overnight, and there is a way to 
mitigate the problem in the meantime without touching the code, the people 
affected would value it.

I saw -ffloat-store too, but I didn’t recommend it because "-ffloat-store will 
not suffice” for specific versions of GCC [1]; and "this option does not force 
unnamed temporaries to be flushed to memory, as shown by experiments” [2].  So, 
maybe it would help, to some extent, with the right compiler; or maybe there 
are no unnamed temporaries that really matter and everything will work.  If you 
don’t want to require SSE2, it's worth a try.

[1] http://christian-seiler.de/projekte/fpmath/
[2] https://hal.archives-ouvertes.fr/hal-00128124/en/
— 
Dan