[Cfp-interest 3354] Re: JTC1/SC22/WG14/CFP 2025/01/08 Meeting minutes

[Joshua Cranmer]

On 1/9/2025 0:10, Damian McGuckin wrote:
>
> Joshua,
>
> On Wed, 8 Jan 2025, Joshua Cranmer wrote:
>
>> On 1/8/2025 14:13, Jerome Coonen wrote:
>>       Joshua: C++ doesn't want pragmas, but wants templates instead. 
>> Have
>>       tried to propose, but have found resistance.
>
> Can you share an example of the template approach, specifically the 
> one with the 11 template arguments?
One hasn't been proposed yet, but the general thinking is that something 
like this would exist:
template <typename T, round mode, exception behavior, ...>
class float;

which starts to get obscene if you represent the fast-math-related 
behavior as individual parameters.

>
>>         Joshua: C++ does all FENV_ACCESS on a per-function basis.
>>
>> C++ implementations allow something like FENV_ACCESS on a 
>> per-function basis.
>
> Do you have an example please?

MSVC implements a few FP programs, like #pragma fp_control 
<https://learn.microsoft.com/en-us/cpp/preprocessor/float-control?view=msvc-170> 
(the documentation doesn't say it, but tests on compiler explorer 
confirm that it's function-level).

Clang has a rich set of pragmas 
(https://clang.llvm.org/docs/LanguageExtensions.html#extensions-to-specify-floating-point-flags), 
in addition to implementing the MSVC pragmas, that work on a per-block 
basis. It also implements #pragma STDC FENV_ACCESS directly.

GCC has the -frounding-math parameter that controls the same thing as 
FENV_ACCESS, and I think (though I haven't tested) it's possible to 
specify this with the per-function attribute allowing custom 
optimizations to be applied to functions.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20250109/75292207/attachment.htm>