S390: Allow "v" constraint for long double math_opt_barrier and math_force_eval with GCC 11.
Commit Message
Starting with GCC 11, long double values can also be processed in vector
registers if build with -march >= z14. Then GCC defines the
__LONG_DOUBLE_VX__ macro.
FYI: GCC commit "IBM Z: Introduce __LONG_DOUBLE_VX__ macro"
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=f47df2af313d2ce7f9149149010a142c2237beda
---
sysdeps/s390/fpu/math-barriers.h | 21 +++++++++++++++++++--
1 file changed, 19 insertions(+), 2 deletions(-)
Comments
* Stefan Liebler via Libc-alpha:
with GCC 11, long double values can also be processed in vector
> registers if build with -march >= z14. Then GCC defines the
> __LONG_DOUBLE_VX__ macro.
>
> FYI: GCC commit "IBM Z: Introduce __LONG_DOUBLE_VX__ macro"
> https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=f47df2af313d2ce7f9149149010a142c2237beda
> ---
> sysdeps/s390/fpu/math-barriers.h | 21 +++++++++++++++++++--
> 1 file changed, 19 insertions(+), 2 deletions(-)
>
> diff --git a/sysdeps/s390/fpu/math-barriers.h b/sysdeps/s390/fpu/math-barriers.h
> index 68a3e55098..a47cda8b24 100644
> --- a/sysdeps/s390/fpu/math-barriers.h
> +++ b/sysdeps/s390/fpu/math-barriers.h
> @@ -25,9 +25,22 @@
> # define ASM_CONSTRAINT_VR
> #endif
>
> +/* Starting with gcc 11, long double values can also be processed in vector
> + registers if build with -march >= z14. Then GCC defines the
> + __LONG_DOUBLE_VX__ macro. */
> +#ifdef __LONG_DOUBLE_VX__
> +# define ASM_LONG_DOUBLE_IN_VR 1
> +#else
> +# define ASM_LONG_DOUBLE_IN_VR 0
> +#endif
Above that there is:
#ifdef HAVE_S390_VX_GCC_SUPPORT
# define ASM_CONSTRAINT_VR "v"
#else
# define ASM_CONSTRAINT_VR
#endif
I think conceptually, ASM_LONG_DOUBLE_IN_VR should be set under
HAVE_S390_VX_GCC_SUPPORT, right?
Or maybe enforce this with #error?
On 31/03/2021 11:34, Florian Weimer wrote:
> * Stefan Liebler via Libc-alpha:
>> +/* Starting with gcc 11, long double values can also be processed in vector
>> + registers if build with -march >= z14. Then GCC defines the
>> + __LONG_DOUBLE_VX__ macro. */
>> +#ifdef __LONG_DOUBLE_VX__
>> +# define ASM_LONG_DOUBLE_IN_VR 1
>> +#else
>> +# define ASM_LONG_DOUBLE_IN_VR 0
>> +#endif
>
> Above that there is:
>
> #ifdef HAVE_S390_VX_GCC_SUPPORT
> # define ASM_CONSTRAINT_VR "v"
> #else
> # define ASM_CONSTRAINT_VR
> #endif
>
> I think conceptually, ASM_LONG_DOUBLE_IN_VR should be set under
> HAVE_S390_VX_GCC_SUPPORT, right?
Hi Florian,
yes, you are right. GCC will only define __LONG_DOUBLE_VX__ if it knows
about vector registers and HAVE_S390_VX_GCC_SUPPORT will always be
defined in this case.
If HAVE_S390_VX_GCC_SUPPORT would not be defined, but
__LONG_DOUBLE_VX__, then the constraint "fm" will be used in all asm
statements.
But nevertheless, setting ASM_LONG_DOUBLE_IN_VR under
HAVE_S390_VX_GCC_SUPPORT makes it more cleanly. I've just posted a v2:
https://sourceware.org/pipermail/libc-alpha/2021-March/124606.html
Thanks,
Stefan
@@ -25,9 +25,22 @@
# define ASM_CONSTRAINT_VR
#endif
+/* Starting with gcc 11, long double values can also be processed in vector
+ registers if build with -march >= z14. Then GCC defines the
+ __LONG_DOUBLE_VX__ macro. */
+#ifdef __LONG_DOUBLE_VX__
+# define ASM_LONG_DOUBLE_IN_VR 1
+#else
+# define ASM_LONG_DOUBLE_IN_VR 0
+#endif
+
#define math_opt_barrier(x) \
({ __typeof (x) __x = (x); \
- if (__builtin_types_compatible_p (__typeof (x), _Float128)) \
+ if (! ASM_LONG_DOUBLE_IN_VR \
+ && (__builtin_types_compatible_p (__typeof (x), _Float128) \
+ || __builtin_types_compatible_p (__typeof (x), long double) \
+ ) \
+ ) \
__asm__ ("# math_opt_barrier_f128 %0" : "+fm" (__x)); \
else \
__asm__ ("# math_opt_barrier %0" \
@@ -35,7 +48,11 @@
__x; })
#define math_force_eval(x) \
({ __typeof (x) __x = (x); \
- if (__builtin_types_compatible_p (__typeof (x), _Float128)) \
+ if (! ASM_LONG_DOUBLE_IN_VR \
+ && (__builtin_types_compatible_p (__typeof (x), _Float128) \
+ || __builtin_types_compatible_p (__typeof (x), long double) \
+ ) \
+ ) \
__asm__ __volatile__ ("# math_force_eval_f128 %0" \
: : "fm" (__x)); \
else \