Re: msvcrt: improve accuracy of exp/pow/cosh and sinh

[Ruslan Kabatsayev]

Hi,

Are you sure this will be of any use if one selects SSE instead of x87
instruction set to handle doubles (like -mfpmath=sse)? No one
guarantees the functions like sinh will be evaluated on x87 FPU and
even that this legacy instruction set will used at all.

Regards,
Ruslan

On Sat, Apr 4, 2015 at 7:40 PM, Zheng Chen  wrote:
> The default fpu_cw = 0x27f makes the exp/pow/cosh/sinh functions producing
> less accurate results. This patch temporally changes fpu_cw to 0x37f in
> these functions for more accurate results.
>
> Thanks to Qian for his ideas.
>
> This patch fixes parts of these two bugs: 37149, 37150

msvcrt: improve accuracy of exp/pow/cosh and sinh

[Zheng Chen]

Hi Ruslan,

Sorry for the late reply.

Thanks for pointing out the problem, I havenot thought of the questions
before, and I found that this patch might also cause problems on other
platforms... However, here
https://github.com/wine-compholio/wine-staging/tree/master/patches/msvcrt-Math_Precision
is a patch created by Sebastian and it should work if your environment
supports coshl/sinhl.

Thank you again and wish you a good day! :)

Regards,
Zheng

2015-04-05 2:36 GMT+08:00 Ruslan Kabatsayev :
> Hi,
>
> Are you sure this will be of any use if one selects SSE instead of x87
> instruction set to handle doubles (like -mfpmath=sse)? No one
> guarantees the functions like sinh will be evaluated on x87 FPU and
> even that this legacy instruction set will used at all.
>
> Regards,
> Ruslan
>
> On Sat, Apr 4, 2015 at 7:40 PM, Zheng Chen  wrote:
> > The default fpu_cw = 0x27f makes the exp/pow/cosh/sinh functions
>
> producing
> > less accurate results. This patch temporally changes fpu_cw to 0x37f in
> > these functions for more accurate results.
> >
> > Thanks to Qian for his ideas.
> >
> > This patch fixes parts of these two bugs: 37149, 37150