[v3,1/3] powerpc: Add optimized ilogb* for POWER9
Commit Message
Changes since v2:
- Moved the GCC version test to math_private.h and start using
__has_builtin().
- Removed the optimization from long double as it was converting
ibm128 to float128.
---8<---
The instructions xsxexpdp and xsxexpqp introduced on POWER9 extract
the exponent from a double-precision and quad-precision floating-point
respectively, thus they can be used to improve ilogb, ilogbf and ilogbf128.
---
sysdeps/powerpc/fpu/math_private.h | 26 +++++++++++++++-
.../powerpc64/le/fpu/w_ilogb_template.c | 30 +++++++++++++++++++
sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c | 3 ++
3 files changed, 58 insertions(+), 1 deletion(-)
create mode 100644 sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
create mode 100644 sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
Comments
On 3/4/21 9:00 AM, Raphael Moreira Zinsly wrote:
> Changes since v2:
> - Moved the GCC version test to math_private.h and start using
> __has_builtin().
> - Removed the optimization from long double as it was converting
> ibm128 to float128.
>
> ---8<---
>
> The instructions xsxexpdp and xsxexpqp introduced on POWER9 extract
> the exponent from a double-precision and quad-precision floating-point
> respectively, thus they can be used to improve ilogb, ilogbf and ilogbf128.
> ---
> sysdeps/powerpc/fpu/math_private.h | 26 +++++++++++++++-
> .../powerpc64/le/fpu/w_ilogb_template.c | 30 +++++++++++++++++++
> sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c | 3 ++
> 3 files changed, 58 insertions(+), 1 deletion(-)
> create mode 100644 sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
> create mode 100644 sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
>
> diff --git a/sysdeps/powerpc/fpu/math_private.h b/sysdeps/powerpc/fpu/math_private.h
> index 91b1361749..21628f3bda 100644
> --- a/sysdeps/powerpc/fpu/math_private.h
> +++ b/sysdeps/powerpc/fpu/math_private.h
> @@ -25,7 +25,28 @@
>
> #include_next <math_private.h>
>
> -#if defined _ARCH_PWR9 && __HAVE_DISTINCT_FLOAT128
> +#ifdef _ARCH_PWR9
> +
> +#if __GNUC_PREREQ (8, 0)
> +# define _GL_HAS_BUILTIN_ILOGB 1
> +#elif defined __has_builtin
> +# define _GL_HAS_BUILTIN_ILOGB __has_builtin (__builtin_vsx_scalar_extract_exp)
> +#else
> +# define _GL_HAS_BUILTIN_ILOGB 0
> +#endif
> +
> +#define __builtin_test_dc_ilogbf __builtin_test_dc_ilogb
> +#define __builtin_ilogbf __builtin_ilogb
> +
> +#define __builtin_test_dc_ilogb(x, y) \
> + __builtin_vsx_scalar_test_data_class_dp(x, y)
> +#define __builtin_ilogb(x) __builtin_vsx_scalar_extract_exp(x) - 0x3ff
> +
> +#define __builtin_test_dc_ilogbf128(x, y) \
> + __builtin_vsx_scalar_test_data_class_qp(x, y)
> +#define __builtin_ilogbf128(x) __builtin_vsx_scalar_extract_expq(x) - 0x3fff
> +
> +#if __HAVE_DISTINCT_FLOAT128
> extern __always_inline _Float128
> __ieee754_sqrtf128 (_Float128 __x)
> {
> @@ -34,5 +55,8 @@ __ieee754_sqrtf128 (_Float128 __x)
> return __z;
> }
> #endif
> +#else /* !_ARCH_PWR9 */
> +#define _GL_HAS_BUILTIN_ILOGB 0
> +#endif
>
> #endif /* _PPC_MATH_PRIVATE_H_ */
> diff --git a/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
> new file mode 100644
> index 0000000000..b5c1c0aa9d
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
> @@ -0,0 +1,30 @@
> +#include <math.h>
> +#include <errno.h>
> +#include <limits.h>
> +#include <math_private.h>
> +#include <fenv.h>
> +
> +#if _GL_HAS_BUILTIN_ILOGB
> +int
> +M_DECL_FUNC (__ilogb) (FLOAT x)
> +{
> + int r;
> + /* Check for exceptional cases. */
> + if (! M_SUF(__builtin_test_dc_ilogb) (x, 0x7f))
> + r = M_SUF (__builtin_ilogb) (x);
> + else
> + /* Fallback to the generic ilogb if x is NaN, Inf or subnormal. */
> + r = M_SUF (__ieee754_ilogb) (x);
> + if (__builtin_expect (r == FP_ILOGB0, 0)
> + || __builtin_expect (r == FP_ILOGBNAN, 0)
> + || __builtin_expect (r == INT_MAX, 0))
> + {
> + __set_errno (EDOM);
> + __feraiseexcept (FE_INVALID);
> + }
> + return r;
> +}
> +declare_mgen_alias (__ilogb, ilogb)
> +#else
> +#include <math/w_ilogb_template.c>
> +#endif
> diff --git a/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
> new file mode 100644
> index 0000000000..215a00141d
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
> @@ -0,0 +1,3 @@
> +/* Skip the optimization for long double as it uses ibm128. */
I would recommend rewording this as "... as ibm128 does not provide an
optimized builtin".
Though, I suspect you could use the double version of these built-ins by
extracting the significant double of the ibm128, and testing/extracting
it in a similar manner.
Anyhow, this patch is OK with a minor rewording of this comment.
> +#include <math-type-macros-ldouble.h>
> +#include <math/w_ilogb_template.c>
>
On 04/03/2021 12:30, Paul E Murphy wrote:
>
>> diff --git a/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
>> b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
>> new file mode 100644
>> index 0000000000..215a00141d
>> --- /dev/null
>> +++ b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
>> @@ -0,0 +1,3 @@
>> +/* Skip the optimization for long double as it uses ibm128. */
>
> I would recommend rewording this as "... as ibm128 does not provide an
> optimized builtin".
>
> Though, I suspect you could use the double version of these built-ins by
> extracting the significant double of the ibm128, and testing/extracting
> it in a similar manner.
>
> Anyhow, this patch is OK with a minor rewording of this comment.
>
I fixed that and pushed as 56c81132ccc6f468fa4fc29c536db060e18e9d87, thanks!
Best Regards,
@@ -25,7 +25,28 @@
#include_next <math_private.h>
-#if defined _ARCH_PWR9 && __HAVE_DISTINCT_FLOAT128
+#ifdef _ARCH_PWR9
+
+#if __GNUC_PREREQ (8, 0)
+# define _GL_HAS_BUILTIN_ILOGB 1
+#elif defined __has_builtin
+# define _GL_HAS_BUILTIN_ILOGB __has_builtin (__builtin_vsx_scalar_extract_exp)
+#else
+# define _GL_HAS_BUILTIN_ILOGB 0
+#endif
+
+#define __builtin_test_dc_ilogbf __builtin_test_dc_ilogb
+#define __builtin_ilogbf __builtin_ilogb
+
+#define __builtin_test_dc_ilogb(x, y) \
+ __builtin_vsx_scalar_test_data_class_dp(x, y)
+#define __builtin_ilogb(x) __builtin_vsx_scalar_extract_exp(x) - 0x3ff
+
+#define __builtin_test_dc_ilogbf128(x, y) \
+ __builtin_vsx_scalar_test_data_class_qp(x, y)
+#define __builtin_ilogbf128(x) __builtin_vsx_scalar_extract_expq(x) - 0x3fff
+
+#if __HAVE_DISTINCT_FLOAT128
extern __always_inline _Float128
__ieee754_sqrtf128 (_Float128 __x)
{
@@ -34,5 +55,8 @@ __ieee754_sqrtf128 (_Float128 __x)
return __z;
}
#endif
+#else /* !_ARCH_PWR9 */
+#define _GL_HAS_BUILTIN_ILOGB 0
+#endif
#endif /* _PPC_MATH_PRIVATE_H_ */
new file mode 100644
@@ -0,0 +1,30 @@
+#include <math.h>
+#include <errno.h>
+#include <limits.h>
+#include <math_private.h>
+#include <fenv.h>
+
+#if _GL_HAS_BUILTIN_ILOGB
+int
+M_DECL_FUNC (__ilogb) (FLOAT x)
+{
+ int r;
+ /* Check for exceptional cases. */
+ if (! M_SUF(__builtin_test_dc_ilogb) (x, 0x7f))
+ r = M_SUF (__builtin_ilogb) (x);
+ else
+ /* Fallback to the generic ilogb if x is NaN, Inf or subnormal. */
+ r = M_SUF (__ieee754_ilogb) (x);
+ if (__builtin_expect (r == FP_ILOGB0, 0)
+ || __builtin_expect (r == FP_ILOGBNAN, 0)
+ || __builtin_expect (r == INT_MAX, 0))
+ {
+ __set_errno (EDOM);
+ __feraiseexcept (FE_INVALID);
+ }
+ return r;
+}
+declare_mgen_alias (__ilogb, ilogb)
+#else
+#include <math/w_ilogb_template.c>
+#endif
new file mode 100644
@@ -0,0 +1,3 @@
+/* Skip the optimization for long double as it uses ibm128. */
+#include <math-type-macros-ldouble.h>
+#include <math/w_ilogb_template.c>