diff mbox series

[25/28] powerpc: Refactor modf{f}

Message ID 20190329133529.22523-26-adhemerval.zanella@linaro.org
State New
Headers show
Series powerpc floating-point optimization refactor | expand

Commit Message

Adhemerval Zanella Netto March 29, 2019, 1:35 p.m. UTC
The modf{f} optimization is not an optimization for ISA 2.07+.  This
patch move the IFUNC for powerpc64 only, move the power5+ to generic
location, and include the generic implementation for ISA 2.07+.

The performance changes are based on modf benchtests:

  * POWER9 - generic code
  "modf": {
   "": {
    "duration": 4.97057e+09,
    "iterations": 1.00688e+09,
    "max": 28.76,
    "min": 4.912,
    "mean": 4.9366
   }
  }
  * POWER9 - power5+ optimization
  "modf": {
   "": {
    "duration": 4.98291e+09,
    "iterations": 9.32818e+08,
    "max": 15.058,
    "min": 5.107,
    "mean": 5.34178
   }
  }

  * POWER8 - generic code
   "modf": {
   "": {
    "duration": 5.05329e+09,
    "iterations": 8.38814e+08,
    "max": 518.051,
    "min": 5.79,
    "mean": 6.02433
   }
  }
  * POWER8 - power5+ optimization
  "modf": {
   "": {
    "duration": 5.05573e+09,
    "iterations": 8.35254e+08,
    "max": 63.141,
    "min": 5.873,
    "mean": 6.05293
   }
  }

  * POWER7 - generic code
  "modf": {
   "": {
    "duration": 4.89818e+09,
    "iterations": 1.08408e+09,
    "max": 57.556,
    "min": 3.953,
    "mean": 4.51827
   }
  }
  * POWER7 - power5+ optimization
  "modf": {
   "": {
    "duration": 4.83789e+09,
    "iterations": 1.33409e+09,
    "max": 46.608,
    "min": 2.224,
    "mean": 3.62636
   }
  }

Checked on powerpc-linux-gnu (built without --with-cpu, with
--with-cpu=power4 and with --with-cpu=power5+ and --disable-multi-arch),
powerpc64-linux-gnu (built without --with-cp and with --with-cpu=power5+
and --disable-multi-arch).

	* sysdeps/powerpc/power5+/fpu/s_modf.c: Move to ...
	* sysdeps/powerpc/fpu/s_modf.c: ... here.  Add ISA 2.07 optimization.
	* sysdeps/powerpc/power5+/fpu/s_modff.c: Move to ...
	* sysdeps/powerpc/fpu/s_modff.c: ... here.  Add ISA 2.07 optimization.
	* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c:
	Adjust include.
	* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c:
	Likewise.
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/Makefile (sysdep_calls,
	sysdep_routines): Add s_modf* objects.
	(CFLAGS-s_modf-power5+.c, CFLAGS-s_modff-power5+.c,
	CFLAGS-s_modf-ppc64.c, CFLAGS-s_modff-ppc64.c): New rule.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-power5+.c: Move
	to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-power5+.c:
	... here.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-power5+.c: Movo
	to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-power5+.c: Move
	... here.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf.c: Move to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf.c: ... here.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-power5+.c: Move
	to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-power5+.c:
	... here.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-ppc64.c: Move to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-ppc64.c:
	... here.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff.c: Move to ...
	* sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff.c: ... here.
---
 sysdeps/powerpc/{power5+ => }/fpu/s_modf.c    | 17 +++++++++++++----
 sysdeps/powerpc/{power5+ => }/fpu/s_modff.c   | 13 +++++++++++--
 .../power4/fpu/multiarch/s_modf-power5+.c     | 13 +------------
 .../power4/fpu/multiarch/s_modff-power5+.c    |  9 +--------
 .../powerpc64/be/fpu/multiarch/Makefile       | 19 ++++++++++++++++++-
 .../{ => be}/fpu/multiarch/s_modf-power5+.c   |  3 ++-
 .../{ => be}/fpu/multiarch/s_modf-ppc64.c     |  0
 .../powerpc64/{ => be}/fpu/multiarch/s_modf.c |  0
 .../{ => be}/fpu/multiarch/s_modff-power5+.c  |  3 ++-
 .../{ => be}/fpu/multiarch/s_modff-ppc64.c    |  0
 .../{ => be}/fpu/multiarch/s_modff.c          |  0
 .../powerpc/powerpc64/fpu/multiarch/Makefile  | 13 -------------
 12 files changed, 48 insertions(+), 42 deletions(-)
 rename sysdeps/powerpc/{power5+ => }/fpu/s_modf.c (74%)
 rename sysdeps/powerpc/{power5+ => }/fpu/s_modff.c (77%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modf-power5+.c (91%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modf-ppc64.c (100%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modf.c (100%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modff-power5+.c (91%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modff-ppc64.c (100%)
 rename sysdeps/powerpc/powerpc64/{ => be}/fpu/multiarch/s_modff.c (100%)

-- 
2.17.1

Comments

Gabriel F. T. Gomes July 5, 2019, 1:23 a.m. UTC | #1
On Fri, Mar 29 2019, Adhemerval Zanella wrote:
> The modf{f} optimization is not an optimization for ISA 2.07+.  This

> patch move the IFUNC for powerpc64 only, move the power5+ to generic

> location, and include the generic implementation for ISA 2.07+.


Looks good to me. Thanks.

Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>
Adhemerval Zanella Netto July 8, 2019, 6:30 p.m. UTC | #2
On 04/07/2019 22:23, Gabriel F. T. Gomes wrote:
> On Fri, Mar 29 2019, Adhemerval Zanella wrote:

>> The modf{f} optimization is not an optimization for ISA 2.07+.  This

>> patch move the IFUNC for powerpc64 only, move the power5+ to generic

>> location, and include the generic implementation for ISA 2.07+.

> 

> Looks good to me. Thanks.

> 

> Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>

> 


If no one opposes it, I will commit this shortly.
diff mbox series

Patch

diff --git a/sysdeps/powerpc/power5+/fpu/s_modf.c b/sysdeps/powerpc/fpu/s_modf.c
similarity index 74%
rename from sysdeps/powerpc/power5+/fpu/s_modf.c
rename to sysdeps/powerpc/fpu/s_modf.c
index dbb11652e1..2304fc48ed 100644
--- a/sysdeps/powerpc/power5+/fpu/s_modf.c
+++ b/sysdeps/powerpc/fpu/s_modf.c
@@ -15,9 +15,15 @@ 
    License along with the GNU C Library; see the file COPYING.LIB.  If
    not, see <http://www.gnu.org/licenses/>.  */
 
-#include <math.h>
-#include <math_ldbl_opt.h>
-#include <libm-alias-double.h>
+/* ISA 2.07 provides fast GPR to FP instruction (mfvsr{d,wz}) which make
+   generic implementation faster.  Also disables for old ISAs that do not
+   have ceil/floor instructions.  */
+#if defined(_ARCH_PWR8) || !defined(_ARCH_PWR5X)
+# include <sysdeps/ieee754/ldbl-opt/s_modf.c>
+#else
+# include <math.h>
+# include <math_ldbl_opt.h>
+# include <libm-alias-double.h>
 
 double
 __modf (double x, double *iptr)
@@ -44,7 +50,10 @@  __modf (double x, double *iptr)
       return copysign (x - *iptr, x);
     }
 }
+# ifndef __modf
 libm_alias_double (__modf, modf)
-#if LONG_DOUBLE_COMPAT (libc, GLIBC_2_0)
+#  if LONG_DOUBLE_COMPAT (libc, GLIBC_2_0)
 compat_symbol (libc, __modf, modfl, GLIBC_2_0);
+#  endif
+# endif
 #endif
diff --git a/sysdeps/powerpc/power5+/fpu/s_modff.c b/sysdeps/powerpc/fpu/s_modff.c
similarity index 77%
rename from sysdeps/powerpc/power5+/fpu/s_modff.c
rename to sysdeps/powerpc/fpu/s_modff.c
index 87c9f020f7..2a0f114b20 100644
--- a/sysdeps/powerpc/power5+/fpu/s_modff.c
+++ b/sysdeps/powerpc/fpu/s_modff.c
@@ -15,8 +15,14 @@ 
    License along with the GNU C Library; see the file COPYING.LIB.  If
    not, see <http://www.gnu.org/licenses/>.  */
 
-#include <math.h>
-#include <libm-alias-float.h>
+/* ISA 2.07 provides fast GPR to FP instruction (mfvsr{d,wz}) which make
+   generic implementation faster.  Also disables for old ISAs that do not
+   have ceil/floor instructions.  */
+#if defined(_ARCH_PWR8) || !defined(_ARCH_PWR5X)
+# include <sysdeps/ieee754/flt-32/s_modff.c>
+#else
+# include <math.h>
+# include <libm-alias-float.h>
 
 float
 __modff (float x, float *iptr)
@@ -43,4 +49,7 @@  __modff (float x, float *iptr)
       return copysignf (x - *iptr, x);
     }
 }
+# ifndef __modff
 libm_alias_float (__modf, modf)
+# endif
+#endif
diff --git a/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c b/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c
index b1d0540b31..6f93c2b652 100644
--- a/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c
+++ b/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c
@@ -16,16 +16,5 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-#include <math.h>
-#include <math_ldbl_opt.h>
-
-#undef weak_alias
-#define weak_alias(a,b)
-#undef strong_alias
-#define strong_alias(a,b)
-#undef compat_symbol
-#define compat_symbol(a,b,c,d)
-
 #define __modf __modf_power5plus
-
-#include <sysdeps/powerpc/power5+/fpu/s_modf.c>
+#include <sysdeps/powerpc/fpu/s_modf.c>
diff --git a/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c b/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c
index 8b333eae0d..2e701881e8 100644
--- a/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c
+++ b/sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c
@@ -16,12 +16,5 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-#include <math.h>
-#include <math_ldbl_opt.h>
-
-#undef weak_alias
-#define weak_alias(a,b)
-
 #define __modff __modff_power5plus
-
-#include <sysdeps/powerpc/power5+/fpu/s_modff.c>
+#include <sysdeps/powerpc/fpu/s_modff.c>
diff --git a/sysdeps/powerpc/powerpc64/be/fpu/multiarch/Makefile b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/Makefile
index 9917cd15d5..63de39bbf1 100644
--- a/sysdeps/powerpc/powerpc64/be/fpu/multiarch/Makefile
+++ b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/Makefile
@@ -1,4 +1,13 @@ 
 ifeq ($(subdir),math)
+# These functions are built both for libc and libm because they're required
+# by printf.  While the libc objects have the prefix s_, the libm ones are
+# prefixed with  m_.
+sysdep_calls := s_modf-power5+ \
+		s_modf-ppc64 \
+		s_modff-power5+ \
+		s_modff-ppc64
+
+sysdep_routines += $(sysdep_calls)
 libm-sysdep_routines += s_ceil-power5+ \
 			s_ceil-ppc64 \
 			s_ceilf-power5+ \
@@ -22,7 +31,8 @@  libm-sysdep_routines += s_ceil-power5+ \
 			s_llround-power6x \
 			s_llround-power5+ \
 			s_llround-ppc64 \
-			s_llroundf-ppc64
+			s_llroundf-ppc64 \
+			$(sysdep_calls:s_%=m_%)
 
 CFLAGS-s_ceil-power5+.c = -mcpu=power5+
 CFLAGS-s_ceilf-power5+.c = -mcpu=power5+
@@ -35,4 +45,11 @@  CFLAGS-s_truncf-power5+.c = -mcpu=power5+
 CFLAGS-s_llround-power8.c += -mcpu=power8
 CFLAGS-s_llround-power6x.c += -mcpu=power6x
 CFLAGS-s_llround-power5+.c += -mcpu=power5+
+
+CFLAGS-s_modf-power5+.c += -mcpu=power5+
+CFLAGS-s_modff-power5+.c += -mcpu=power5+
+# These files quiet sNaNs in a way that is optimized away without
+# -fsignaling-nans.
+CFLAGS-s_modf-ppc64.c += -fsignaling-nans
+CFLAGS-s_modff-ppc64.c += -fsignaling-nans
 endif
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-power5+.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-power5+.c
similarity index 91%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-power5+.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-power5+.c
index 1a958de178..6f93c2b652 100644
--- a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-power5+.c
+++ b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-power5+.c
@@ -16,4 +16,5 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-#include <sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modf-power5+.c>
+#define __modf __modf_power5plus
+#include <sysdeps/powerpc/fpu/s_modf.c>
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-ppc64.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-ppc64.c
similarity index 100%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf-ppc64.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf-ppc64.c
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf.c
similarity index 100%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modf.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modf.c
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-power5+.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-power5+.c
similarity index 91%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-power5+.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-power5+.c
index 4939d4bc2b..2e701881e8 100644
--- a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-power5+.c
+++ b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-power5+.c
@@ -16,4 +16,5 @@ 
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-#include <sysdeps/powerpc/powerpc32/power4/fpu/multiarch/s_modff-power5+.c>
+#define __modff __modff_power5plus
+#include <sysdeps/powerpc/fpu/s_modff.c>
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-ppc64.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-ppc64.c
similarity index 100%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff-ppc64.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff-ppc64.c
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff.c b/sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff.c
similarity index 100%
rename from sysdeps/powerpc/powerpc64/fpu/multiarch/s_modff.c
rename to sysdeps/powerpc/powerpc64/be/fpu/multiarch/s_modff.c
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile b/sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile
index 534d5a7133..d7ad1e2724 100644
--- a/sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile
+++ b/sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile
@@ -1,10 +1,4 @@ 
 ifeq ($(subdir),math)
-# These functions are built both for libc and libm because they're required
-# by printf.  While the libc objects have the prefix s_, the libm ones are
-# prefixed with  m_.
-sysdep_calls := s_modf-power5+ s_modf-ppc64 \
-		s_modff-power5+ s_modff-ppc64
-
 sysdep_routines += $(sysdep_calls)
 libm-sysdep_routines += s_logb-power7 s_logbf-power7 \
 			s_logbl-power7 s_logb-ppc64 s_logbf-ppc64 \
@@ -14,11 +8,4 @@  libm-sysdep_routines += s_logb-power7 s_logbf-power7 \
 CFLAGS-s_logbf-power7.c = -mcpu=power7
 CFLAGS-s_logbl-power7.c = -mcpu=power7
 CFLAGS-s_logb-power7.c = -mcpu=power7
-CFLAGS-s_modf-power5+.c = -mcpu=power5+
-CFLAGS-s_modff-power5+.c = -mcpu=power5+
-
-# These files quiet sNaNs in a way that is optimized away without
-# -fsignaling-nans.
-CFLAGS-s_modf-ppc64.c += -fsignaling-nans
-CFLAGS-s_modff-ppc64.c += -fsignaling-nans
 endif