On Sun, Jun 12, 2016 at 11:50:11PM +0200, Uros Bizjak wrote:
> Attached (mostly mechanical) patch uses equivalent built-in functions
> for fabsq, copysignq and nanq. The patch allows more aggressive
> compiler optimizations, where for fabsq and copysignq, the compiler
> will emit 128bit SSE bitops, and a 128bit constant load instead of
> nanq function call.
>
> The patch also improves detection of required builtins, so for ia64,
> additional patch [1] is needed.
>
> 2016-06-12 Uros Bizjak <ubiz...@gmail.com>
>
> * configure.ac (__float128 support): Also test _builtin_fabsq,
> __builtin_copysignq, __builtin_infq and __builtin_nanq.
> * configure: Regenerate.
> * math/*.c: Use __builtin_fabsq instead of fabsq, __builtin_copysignq
> instead of copysignq, __builtin_infq instead of __builtin_inf and
> __builtin_nanq instead of nanq.
>
> Patch was bootstrapped and regression tested on x86_64-linux-gnu
> {,-m32}. I have also checked, that no fabsq, copysignq or nanq calls
> remain in the library.
Couldn't you instead add into a header inline functions or macros
that map fabsq to __builtin_fabsq etc.? Then you could keep the *.c
files as is.
What I don't really like on the patch is that it diverges too much from the
original libc sources (other than replacing suffixes of functions and
changing types).
Jakub
