[13/13] Use GCC builtins for copysign functions if desired.
Commit Message
On 12/3/19 5:50 PM, Joseph Myers wrote:
> On Tue, 3 Dec 2019, Stefan Liebler wrote:
>
>> On 12/2/19 10:00 PM, Joseph Myers wrote:
>>> On Mon, 2 Dec 2019, Stefan Liebler wrote:
>>>
>>>> This patch is using the corresponding GCC builtin for copysignf, copysign,
>>>> copysignl and copysignf128 if the USE_FUNCTION_BUILTIN macros are defined
>>>> to one
>>>> in math-use-builtins.h.
>>>
>>> I believe this is always safe for these implementations (the only case
>>> where GCC might not expand copysign functions inline is copysignl for IBM
>>> long double, in the soft-float case).
>>>
>> Thus you mean we can do the following preset in
>> sysdeps/generic/math-use-builtins.h?
>> #define USE_COPYSIGN_BUILTIN 1
>> #define USE_COPYSIGNF_BUILTIN 1
>> #define USE_COPYSIGNL_BUILTIN 0
>> #define USE_COPYSIGNF128_BUILTIN 0
>>
>> Or even also set USE_COPYSIGNL_BUILTIN to one as IBM long double has its own
>> implementation in ./sysdeps/ieee754/ldbl-128ibm/s_copysignl.c.
>
> I think they can all be 1, given that IBM long double has its own
> implementation. Except that you shouldn't need the indirection through
> these macros at all. Just use __builtin_copysign etc. directly in the
> implementations (other than IBM long double), unconditionally, the same
> way that sysdeps/ieee754/dbl-64/s_fabs.c uses __builtin_fabs directly
> without such macros being needed.
>
Okay. I've set them to 1 and also used build-many-glibcs.py and had a
look into the s_copysign*.os files. There is no function call to
copysign* itself.
If one architecture encounters issues with those builtins, an
architecture specific math-use-builtins.sh file can set the macros to zero.
Bye,
Stefan
Comments
On Wed, 4 Dec 2019, Stefan Liebler wrote:
> > I think they can all be 1, given that IBM long double has its own
> > implementation. Except that you shouldn't need the indirection through
> > these macros at all. Just use __builtin_copysign etc. directly in the
> > implementations (other than IBM long double), unconditionally, the same
> > way that sysdeps/ieee754/dbl-64/s_fabs.c uses __builtin_fabs directly
> > without such macros being needed.
> >
>
> Okay. I've set them to 1 and also used build-many-glibcs.py and had a look
> into the s_copysign*.os files. There is no function call to copysign* itself.
I think we should handle this like fabs and not have the macros at all,
just call __builtin_copysign etc. unconditionally.
commit 91e337064c1684bd10c63ecc559e8340656f0cf5
Author: Stefan Liebler <stli@linux.ibm.com>
Date: Mon Dec 2 15:32:44 2019 +0100
Use GCC builtins for copysign functions if desired.
This patch is using the corresponding GCC builtin for copysignf, copysign,
copysignl and copysignf128 if the USE_FUNCTION_BUILTIN macros are defined to one
in math-use-builtins.h.
By default the float, double and long double macros are set to one.
According to Joseph Myers it is safe to just use those builtins.
Also the objdumps of s_copysign*.os files generated by build-many-glibcs.py
do not show copysign* function calls.
If one architecture encounters issues with those builtins, an architecture
specific math-use-builtins.sh file can set the macros to zero.
@@ -51,4 +51,9 @@
#define USE_ROUNDL_BUILTIN 0
#define USE_ROUNDF128_BUILTIN 0
+#define USE_COPYSIGN_BUILTIN 1
+#define USE_COPYSIGNF_BUILTIN 1
+#define USE_COPYSIGNL_BUILTIN 1
+#define USE_COPYSIGNF128_BUILTIN 0
+
#endif /* math-use-builtins.h */
@@ -10,7 +10,9 @@
* ====================================================
*/
-#if defined(LIBM_SCCS) && !defined(lint)
+#include <math-use-builtins.h>
+
+#if ! USE_COPYSIGN_BUILTIN && defined (LIBM_SCCS) && ! defined (lint)
static char rcsid[] = "$NetBSD: s_copysign.c,v 1.8 1995/05/10 20:46:57 jtc Exp $";
#endif
@@ -28,10 +30,14 @@ static char rcsid[] = "$NetBSD: s_copysign.c,v 1.8 1995/05/10 20:46:57 jtc Exp $
double
__copysign (double x, double y)
{
+#if USE_COPYSIGN_BUILTIN
+ return __builtin_copysign (x, y);
+#else
uint32_t hx, hy;
GET_HIGH_WORD (hx, x);
GET_HIGH_WORD (hy, y);
SET_HIGH_WORD (x, (hx & 0x7fffffff) | (hy & 0x80000000));
return x;
+#endif
}
libm_alias_double (__copysign, copysign)
@@ -152,6 +152,8 @@
#define USE_TRUNCL_BUILTIN USE_TRUNCF128_BUILTIN
#undef USE_ROUNDL_BUILTIN
#define USE_ROUNDL_BUILTIN USE_ROUNDF128_BUILTIN
+#undef USE_COPYSIGNL_BUILTIN
+#define USE_COPYSIGNL_BUILTIN USE_COPYSIGNF128_BUILTIN
/* IEEE function renames. */
#define __ieee754_acoshl __ieee754_acoshf128
@@ -361,6 +363,7 @@
#define __builtin_ceill __builtin_ceilf128
#define __builtin_truncl __builtin_truncf128
#define __builtin_roundl __builtin_roundf128
+#define __builtin_copysignl __builtin_copysignf128
/* Get the constant suffix from bits/floatn-compat.h. */
#define L(x) __f128 (x)
@@ -13,7 +13,9 @@
* ====================================================
*/
-#if defined(LIBM_SCCS) && !defined(lint)
+#include <math-use-builtins.h>
+
+#if ! USE_COPYSIGNF_BUILTIN && defined (LIBM_SCCS) && ! defined (lint)
static char rcsid[] = "$NetBSD: s_copysignf.c,v 1.4 1995/05/10 20:46:59 jtc Exp $";
#endif
@@ -28,12 +30,17 @@ static char rcsid[] = "$NetBSD: s_copysignf.c,v 1.4 1995/05/10 20:46:59 jtc Exp
#include <math_private.h>
#include <libm-alias-float.h>
-float __copysignf(float x, float y)
+float
+__copysignf (float x, float y)
{
- uint32_t ix,iy;
- GET_FLOAT_WORD(ix,x);
- GET_FLOAT_WORD(iy,y);
- SET_FLOAT_WORD(x,(ix&0x7fffffff)|(iy&0x80000000));
- return x;
+#if USE_COPYSIGNF_BUILTIN
+ return __builtin_copysignf (x, y);
+#else
+ uint32_t ix, iy;
+ GET_FLOAT_WORD (ix, x);
+ GET_FLOAT_WORD (iy, y);
+ SET_FLOAT_WORD (x, (ix & 0x7fffffff) | (iy & 0x80000000));
+ return x;
+#endif
}
libm_alias_float (__copysign, copysign)
@@ -13,7 +13,9 @@
* ====================================================
*/
-#if defined(LIBM_SCCS) && !defined(lint)
+#include <math-use-builtins.h>
+
+#if ! USE_COPYSIGNL_BUILTIN && defined (LIBM_SCCS) && ! defined (lint)
static char rcsid[] = "$NetBSD: $";
#endif
@@ -28,13 +30,18 @@ static char rcsid[] = "$NetBSD: $";
#include <math_private.h>
#include <libm-alias-ldouble.h>
-_Float128 __copysignl(_Float128 x, _Float128 y)
+_Float128
+__copysignl (_Float128 x, _Float128 y)
{
- uint64_t hx,hy;
- GET_LDOUBLE_MSW64(hx,x);
- GET_LDOUBLE_MSW64(hy,y);
- SET_LDOUBLE_MSW64(x,(hx&0x7fffffffffffffffULL)
- |(hy&0x8000000000000000ULL));
- return x;
+#if USE_COPYSIGNL_BUILTIN
+ return __builtin_copysignl (x, y);
+#else
+ uint64_t hx, hy;
+ GET_LDOUBLE_MSW64 (hx, x);
+ GET_LDOUBLE_MSW64 (hy, y);
+ SET_LDOUBLE_MSW64 (x, (hx & 0x7fffffffffffffffULL)
+ | (hy & 0x8000000000000000ULL));
+ return x;
+#endif
}
libm_alias_ldouble (__copysign, copysign)
@@ -50,6 +50,13 @@
# define USE_ROUNDF_BUILTIN 1
# define USE_ROUNDL_BUILTIN 1
+/* GCC emits the z9-ec zarch "copy sign" instruction for these
+ builtins if build with at least --march=z196 -mzarch.
+ Note: z196 is no typo but a bug in GCC! */
+# define USE_COPYSIGN_BUILTIN 1
+# define USE_COPYSIGNF_BUILTIN 1
+# define USE_COPYSIGNL_BUILTIN 1
+
# if __GNUC_PREREQ (8, 1)
# define USE_NEARBYINTF128_BUILTIN 1
# define USE_RINTF128_BUILTIN 1
@@ -57,6 +64,7 @@
# define USE_CEILF128_BUILTIN 1
# define USE_TRUNCF128_BUILTIN 1
# define USE_ROUNDF128_BUILTIN 1
+# define USE_COPYSIGNF128_BUILTIN 1
# else
# define USE_NEARBYINTF128_BUILTIN 0
# define USE_RINTF128_BUILTIN 0
@@ -64,6 +72,7 @@
# define USE_CEILF128_BUILTIN 0
# define USE_TRUNCF128_BUILTIN 0
# define USE_ROUNDF128_BUILTIN 0
+# define USE_COPYSIGNF128_BUILTIN 0
# endif
#else