Re: builtin fenv functions

[Richard Biener <richard dot guenther at gmail dot com>]

On Mon, May 29, 2017 at 12:09 AM, Marc Glisse <marc.glisse@inria.fr> wrote:
> On Fri, 26 May 2017, Richard Biener wrote:
>
>>>>> Similarly, I
>>>>> don't see div as a builtin in that file, only FILE* has special code,
>>>>> but
>>>>> that doesn't seem worth the trouble here. So I am only declaring the 5
>>>>> "simple" functions, with minimal properties: leaf, nothrow, and for
>>>>> fegetround pure (glibc already declares it that way). We can then
>>>>> discuss
>>>>> the safety of future optimizations on a case by case basis.
>>>>
>>>>
>>>>
>>>> +DEF_C99_BUILTIN        (BUILT_IN_FERAISEEXCEPT, "feraiseexcept",
>>>> BT_FN_INT_INT, ATTR_NOTHROW_LEAF_LIST)
>>>>
>>>> I think feraiseexcept shouldn't be nothrow?
>>>
>>>
>>>
>>> glibc marks it as nothrow. I can remove the nothrow flag for now, for
>>> safety. It may trap, but it does not throw a C++ exception AFAIU.
>>
>>
>> Also with -fnon-call-exceptions?
>
>
> Hmm, maybe on windows where trap handlers turn into system exceptions which
> are handled like C++ exceptions... I am happy to remove nothrow.

Ok.  I suppose as glibc has it nothrow differing is somewhat pointless
as as soon
as someone includes the fenv.h header it'll get overridden.

>>>> But it may be pure.
>>>
>>>
>>> It writes to the exception register (aka memory for now), so I would
>>> hardly
>>> call it pure.
>>
>>
>> But it doesn't have to be ordered with control word writes/reads, no?
>
>
> Not sure what you mean here. feraiseexcept(FE_DIVBYZERO) is equivalent to
> 1./0., it writes to the exception status flag. Its order with respect to
> fetestexcept must be preserved.

I see.

>>>> Likewise fetestexcept may be pure?
>>>
>>>
>>>
>>> Too unsafe for now, since any FP operation can write to the memory that
>>> fetestexcept reads.
>>
>>
>> Ah...  but then FP operations are not ordered with the builtins anyway,
>> only FP loads/stores would be.
>
>
> Since gcc doesn't handle fenv properly, people have been using a number of
> workarounds, in particular with pass-through asm, sometimes volatile,
> occasionally with the "memory" clobber.
>
> Some of those versions would still work with pure, but the attribute
> increases the likelyhood of breaking some of those uses, and I don't know if
> it would ever help in practice, so I would rather not add it for now.
> fegetround is very different since it can safely swap position with an
> adjacent float operation.
>
>> After all having builtins is only the first easiest step of properly
>> modeling
>> dependences between FP ops and the FP control/exception registers.
>
>
> Yes, I didn't expect adding those 5 builtins (modulo the nothrow flag) to be
> controversial...

Surely not.  I'm fine with erring on the conservative side for now.

Thanks,
Richard.

> --
> Marc Glisse