Better DLA_FMS

It's better to use __builtin_fma if it works.  Use it for gcc 4.6 and
higher.  Move the x86-64 dla.h to the correct place.
This commit is contained in:
Ulrich Drepper 2011-10-24 22:11:21 -04:00
parent 485683c35f
commit 202c9deb15
2 changed files with 12 additions and 3 deletions

View File

@ -1,7 +1,9 @@
2011-10-24 Ulrich Drepper <drepper@gmail.com>
* sysdeps/x86_64/dla.h (DLA_FMS): Some compilers fail to inline
__builtin_fma in some situations. Have to use an asm.
* sysdeps/x86_64/dla.h: Move to ...
* sysdeps/x86_64/fpu/dla.h: ...here.
(DLA_FMS): Some compilers fail to inline __builtin_fma in some
situations. Use __builtin_fma only for gcc 4.6 and up.
* config.make.in: Add have-mfma4 entry.
* configure.in: Substitute libc_cv_cc_fma4.

View File

@ -1,10 +1,17 @@
#include <features.h>
#ifdef __FMA4__
# define DLA_FMS(x,y,z) \
# if __GNUC_PREREQ (4, 6)
# define DLA_FMS(x,y,z) \
__builtin_fma (x, y, -(z))
# else
# define DLA_FMS(x,y,z) \
({ double __z; \
asm ("vfmsubsd %3, %2, %1, %0" \
: "=x" (__z) \
: "x" ((double) (x)), "xm" ((double) (y)) , "x" ((double) (z))); \
__z; })
# endif
#endif
#include "sysdeps/ieee754/dbl-64/dla.h"