Message ID | 20171019173159.21402-6-hjl.tools@gmail.com |
---|---|
State | New |
Headers | show |
Series | x86: Use the generic math implementations | expand |
On Thu, Oct 19, 2017 at 10:31 AM, H.J. Lu <hjl.tools@gmail.com> wrote: > This patch replaces i386 assembly versions of e_logf with generic > e_logf.c. For workload-spec2017.wrf, on Nehalem, it improves > performance by: > > Before After Improvement > reciprocal-throughput 73.3865 40.0454 83% > latency 90.0985 54.4479 65% > > On Skylake, it improves performance by: > > Before After Improvement > reciprocal-throughput 75.1384 22.1452 239% > latency 91.9441 50.7925 81% On IvyBridge with --disable-multi-arch, it improves performance by: Before After Improvement reciprocal-throughput 84.5575 28.7879 193% latency 103.971 57.5231 80% > * sysdeps/i386/fpu/e_logf.S: Removed. > * sysdeps/i386/fpu/e_logf_data.c: Likewise. > * sysdeps/i386/fpu/w_logf.c: Likewise. > * sysdeps/i386/i686/fpu/e_logf.S: Likewise. > * sysdeps/i386/fpu/libm-test-ulps: Updated for generic e_logf.c. > * sysdeps/i386/i686/fpu/multiarch/libm-test-ulps: Likewise. > * sysdeps/i386/i686/fpu/multiarch/Makefile (libm-sysdep_routines): > Add e_logf-sse2. > (CFLAGS-e_logf-sse2.c): New. > * sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c: New file. > * sysdeps/i386/i686/fpu/multiarch/e_logf.c: Likewise. > --- > sysdeps/i386/fpu/e_logf.S | 93 -------------------------- > sysdeps/i386/fpu/e_logf_data.c | 1 - > sysdeps/i386/fpu/libm-test-ulps | 20 +++--- > sysdeps/i386/fpu/w_logf.c | 1 - > sysdeps/i386/i686/fpu/e_logf.S | 30 --------- > sysdeps/i386/i686/fpu/multiarch/Makefile | 5 +- > sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c | 3 + > sysdeps/i386/i686/fpu/multiarch/e_logf.c | 40 +++++++++++ > sysdeps/i386/i686/fpu/multiarch/libm-test-ulps | 12 ++-- > 9 files changed, 62 insertions(+), 143 deletions(-) > delete mode 100644 sysdeps/i386/fpu/e_logf.S > delete mode 100644 sysdeps/i386/fpu/e_logf_data.c > delete mode 100644 sysdeps/i386/fpu/w_logf.c > delete mode 100644 sysdeps/i386/i686/fpu/e_logf.S > create mode 100644 sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c > create mode 100644 sysdeps/i386/i686/fpu/multiarch/e_logf.c > > diff --git a/sysdeps/i386/fpu/e_logf.S b/sysdeps/i386/fpu/e_logf.S > deleted file mode 100644 > index de967a31f5..0000000000 > --- a/sysdeps/i386/fpu/e_logf.S > +++ /dev/null > @@ -1,93 +0,0 @@ > -/* > - * Written by J.T. Conklin <jtc@netbsd.org>. > - * Public domain. > - * Adapted for float by Ulrich Drepper <drepper@cygnus.com>. > - * > - * Changed to use fyl2xp1 for values near 1, <drepper@cygnus.com>. > - */ > - > -#include <machine/asm.h> > - > - .section .rodata.cst8,"aM",@progbits,8 > - > - .p2align 3 > - .type one,@object > -one: .double 1.0 > - ASM_SIZE_DIRECTIVE(one) > - /* It is not important that this constant is precise. It is only > - a value which is known to be on the safe side for using the > - fyl2xp1 instruction. */ > - .type limit,@object > -limit: .double 0.29 > - ASM_SIZE_DIRECTIVE(limit) > - > - > -#ifdef PIC > -# define MO(op) op##@GOTOFF(%edx) > -#else > -# define MO(op) op > -#endif > - > - .text > -ENTRY(__ieee754_logf) > - fldln2 // log(2) > - flds 4(%esp) // x : log(2) > - fxam > - fnstsw > -#ifdef PIC > - LOAD_PIC_REG (dx) > -#endif > - fld %st // x : x : log(2) > - sahf > - jc 3f // in case x is NaN or +-Inf > -4: fsubl MO(one) // x-1 : x : log(2) > - fld %st // x-1 : x-1 : x : log(2) > - fabs // |x-1| : x-1 : x : log(2) > - fcompl MO(limit) // x-1 : x : log(2) > - fnstsw // x-1 : x : log(2) > - andb $0x45, %ah > - jz 2f > - fxam > - fnstsw > - andb $0x45, %ah > - cmpb $0x40, %ah > - jne 5f > - fabs // log(1) is +0 in all rounding modes. > -5: fstp %st(1) // x-1 : log(2) > - fyl2xp1 // log(x) > - ret > - > -2: fstp %st(0) // x : log(2) > - fyl2x // log(x) > - ret > - > -3: jp 4b // in case x is +-Inf > - fstp %st(1) > - fstp %st(1) > - ret > -END (__ieee754_logf) > - > -ENTRY(__logf_finite) > - fldln2 // log(2) > - flds 4(%esp) // x : log(2) > -#ifdef PIC > - LOAD_PIC_REG (dx) > -#endif > - fld %st // x : x : log(2) > - fsubl MO(one) // x-1 : x : log(2) > - fld %st // x-1 : x-1 : x : log(2) > - fabs // |x-1| : x-1 : x : log(2) > - fcompl MO(limit) // x-1 : x : log(2) > - fnstsw // x-1 : x : log(2) > - andb $0x45, %ah > - jz 2b > - fxam > - fnstsw > - andb $0x45, %ah > - cmpb $0x40, %ah > - jne 6f > - fabs // log(1) is +0 in all rounding modes. > -6: fstp %st(1) // x-1 : log(2) > - fyl2xp1 // log(x) > - ret > -END(__logf_finite) > diff --git a/sysdeps/i386/fpu/e_logf_data.c b/sysdeps/i386/fpu/e_logf_data.c > deleted file mode 100644 > index 1cc8931700..0000000000 > --- a/sysdeps/i386/fpu/e_logf_data.c > +++ /dev/null > @@ -1 +0,0 @@ > -/* Not needed. */ > diff --git a/sysdeps/i386/fpu/libm-test-ulps b/sysdeps/i386/fpu/libm-test-ulps > index 5a319f147b..698d1b3a1d 100644 > --- a/sysdeps/i386/fpu/libm-test-ulps > +++ b/sysdeps/i386/fpu/libm-test-ulps > @@ -2000,17 +2000,17 @@ ldouble: 4 > > Function: "gamma_downward": > double: 4 > -float: 4 > +float: 5 > idouble: 4 > -ifloat: 4 > +ifloat: 5 > ildouble: 7 > ldouble: 7 > > Function: "gamma_towardzero": > double: 4 > -float: 2 > +float: 3 > idouble: 4 > -ifloat: 2 > +ifloat: 3 > ildouble: 7 > ldouble: 7 > > @@ -2186,20 +2186,20 @@ ldouble: 4 > > Function: "lgamma_downward": > double: 4 > -float: 4 > +float: 5 > float128: 8 > idouble: 4 > -ifloat: 4 > +ifloat: 5 > ifloat128: 8 > ildouble: 7 > ldouble: 7 > > Function: "lgamma_towardzero": > double: 4 > -float: 2 > +float: 3 > float128: 5 > idouble: 4 > -ifloat: 2 > +ifloat: 3 > ifloat128: 5 > ildouble: 7 > ldouble: 7 > @@ -2641,10 +2641,10 @@ ldouble: 5 > > Function: "y0_towardzero": > double: 2 > -float: 2 > +float: 3 > float128: 3 > idouble: 2 > -ifloat: 2 > +ifloat: 3 > ifloat128: 3 > ildouble: 5 > ldouble: 5 > diff --git a/sysdeps/i386/fpu/w_logf.c b/sysdeps/i386/fpu/w_logf.c > deleted file mode 100644 > index ea48d1356e..0000000000 > --- a/sysdeps/i386/fpu/w_logf.c > +++ /dev/null > @@ -1 +0,0 @@ > -#include <sysdeps/../math/w_logf.c> > diff --git a/sysdeps/i386/i686/fpu/e_logf.S b/sysdeps/i386/i686/fpu/e_logf.S > deleted file mode 100644 > index 6fd39d50d3..0000000000 > --- a/sysdeps/i386/i686/fpu/e_logf.S > +++ /dev/null > @@ -1,30 +0,0 @@ > -/* > - * Written by J.T. Conklin <jtc@netbsd.org>. > - * Public domain. > - * Adapted for float by Ulrich Drepper <drepper@cygnus.com>. > - * > - * Adapted for i686 instructions. > - */ > - > -#include <machine/asm.h> > - > - > - .text > -ENTRY(__ieee754_logf) > - fldln2 // log(2) > - flds 4(%esp) // x : log(2) > - fucomi %st > - jp 3f > - fyl2x // log(x) > - ret > - > -3: fstp %st(1) > - ret > -END (__ieee754_logf) > - > -ENTRY(__logf_finite) > - fldln2 // log(2) > - flds 4(%esp) // x : log(2) > - fyl2x // log(x) > - ret > -END(__logf_finite) > diff --git a/sysdeps/i386/i686/fpu/multiarch/Makefile b/sysdeps/i386/i686/fpu/multiarch/Makefile > index 2723f570af..ee4c3f39e7 100644 > --- a/sysdeps/i386/i686/fpu/multiarch/Makefile > +++ b/sysdeps/i386/i686/fpu/multiarch/Makefile > @@ -1,7 +1,8 @@ > ifeq ($(subdir),math) > -libm-sysdep_routines += e_exp2f-sse2 e_expf-sse2 s_sinf-sse2 s_cosf-sse2 \ > - s_sincosf-sse2 > +libm-sysdep_routines += e_exp2f-sse2 e_expf-sse2 e_logf-sse2 \ > + s_sinf-sse2 s_cosf-sse2 s_sincosf-sse2 > > CFLAGS-e_exp2f-sse2.c = -msse2 -mfpmath=sse > CFLAGS-e_expf-sse2.c = -msse2 -mfpmath=sse > +CFLAGS-e_logf-sse2.c = -msse2 -mfpmath=sse > endif > diff --git a/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c b/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c > new file mode 100644 > index 0000000000..11621fc122 > --- /dev/null > +++ b/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c > @@ -0,0 +1,3 @@ > +#define __logf __logf_sse2 > + > +#include <sysdeps/ieee754/flt-32/e_logf.c> > diff --git a/sysdeps/i386/i686/fpu/multiarch/e_logf.c b/sysdeps/i386/i686/fpu/multiarch/e_logf.c > new file mode 100644 > index 0000000000..1414d17d25 > --- /dev/null > +++ b/sysdeps/i386/i686/fpu/multiarch/e_logf.c > @@ -0,0 +1,40 @@ > +/* Multiple versions of logf. > + Copyright (C) 2017 Free Software Foundation, Inc. > + This file is part of the GNU C Library. > + > + The GNU C Library is free software; you can redistribute it and/or > + modify it under the terms of the GNU Lesser General Public > + License as published by the Free Software Foundation; either > + version 2.1 of the License, or (at your option) any later version. > + > + The GNU C Library is distributed in the hope that it will be useful, > + but WITHOUT ANY WARRANTY; without even the implied warranty of > + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU > + Lesser General Public License for more details. > + > + You should have received a copy of the GNU Lesser General Public > + License along with the GNU C Library; if not, see > + <http://www.gnu.org/licenses/>. */ > + > +extern float __redirect_logf (float); > + > +#define SYMBOL_NAME logf > +#include "ifunc-sse2.h" > + > +libc_ifunc_redirected (__redirect_logf, __logf, IFUNC_SELECTOR ()); > + > +#ifdef SHARED > +__hidden_ver1 (__logf_ia32, __GI___logf, __redirect_logf) > + __attribute__ ((visibility ("hidden"))); > + > +# include <shlib-compat.h> > +versioned_symbol (libm, __logf, logf, GLIBC_2_27); > +#else > +weak_alias (__logf, logf) > +#endif > + > +strong_alias (__logf, __ieee754_logf) > +strong_alias (__logf, __logf_finite) > + > +#define __logf __logf_ia32 > +#include <sysdeps/ieee754/flt-32/e_logf.c> > diff --git a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > index 2d056a9260..c86294b778 100644 > --- a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > +++ b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps > @@ -2000,9 +2000,9 @@ ldouble: 4 > > Function: "gamma_downward": > double: 4 > -float: 4 > +float: 5 > idouble: 4 > -ifloat: 4 > +ifloat: 5 > ildouble: 7 > ldouble: 7 > > @@ -2186,10 +2186,10 @@ ldouble: 4 > > Function: "lgamma_downward": > double: 4 > -float: 4 > +float: 5 > float128: 8 > idouble: 4 > -ifloat: 4 > +ifloat: 5 > ifloat128: 8 > ildouble: 7 > ldouble: 7 > @@ -2625,10 +2625,10 @@ ldouble: 5 > > Function: "y0_towardzero": > double: 2 > -float: 2 > +float: 3 > float128: 3 > idouble: 2 > -ifloat: 2 > +ifloat: 3 > ifloat128: 3 > ildouble: 5 > ldouble: 5 > -- > 2.13.6 >
diff --git a/sysdeps/i386/fpu/e_logf.S b/sysdeps/i386/fpu/e_logf.S deleted file mode 100644 index de967a31f5..0000000000 --- a/sysdeps/i386/fpu/e_logf.S +++ /dev/null @@ -1,93 +0,0 @@ -/* - * Written by J.T. Conklin <jtc@netbsd.org>. - * Public domain. - * Adapted for float by Ulrich Drepper <drepper@cygnus.com>. - * - * Changed to use fyl2xp1 for values near 1, <drepper@cygnus.com>. - */ - -#include <machine/asm.h> - - .section .rodata.cst8,"aM",@progbits,8 - - .p2align 3 - .type one,@object -one: .double 1.0 - ASM_SIZE_DIRECTIVE(one) - /* It is not important that this constant is precise. It is only - a value which is known to be on the safe side for using the - fyl2xp1 instruction. */ - .type limit,@object -limit: .double 0.29 - ASM_SIZE_DIRECTIVE(limit) - - -#ifdef PIC -# define MO(op) op##@GOTOFF(%edx) -#else -# define MO(op) op -#endif - - .text -ENTRY(__ieee754_logf) - fldln2 // log(2) - flds 4(%esp) // x : log(2) - fxam - fnstsw -#ifdef PIC - LOAD_PIC_REG (dx) -#endif - fld %st // x : x : log(2) - sahf - jc 3f // in case x is NaN or +-Inf -4: fsubl MO(one) // x-1 : x : log(2) - fld %st // x-1 : x-1 : x : log(2) - fabs // |x-1| : x-1 : x : log(2) - fcompl MO(limit) // x-1 : x : log(2) - fnstsw // x-1 : x : log(2) - andb $0x45, %ah - jz 2f - fxam - fnstsw - andb $0x45, %ah - cmpb $0x40, %ah - jne 5f - fabs // log(1) is +0 in all rounding modes. -5: fstp %st(1) // x-1 : log(2) - fyl2xp1 // log(x) - ret - -2: fstp %st(0) // x : log(2) - fyl2x // log(x) - ret - -3: jp 4b // in case x is +-Inf - fstp %st(1) - fstp %st(1) - ret -END (__ieee754_logf) - -ENTRY(__logf_finite) - fldln2 // log(2) - flds 4(%esp) // x : log(2) -#ifdef PIC - LOAD_PIC_REG (dx) -#endif - fld %st // x : x : log(2) - fsubl MO(one) // x-1 : x : log(2) - fld %st // x-1 : x-1 : x : log(2) - fabs // |x-1| : x-1 : x : log(2) - fcompl MO(limit) // x-1 : x : log(2) - fnstsw // x-1 : x : log(2) - andb $0x45, %ah - jz 2b - fxam - fnstsw - andb $0x45, %ah - cmpb $0x40, %ah - jne 6f - fabs // log(1) is +0 in all rounding modes. -6: fstp %st(1) // x-1 : log(2) - fyl2xp1 // log(x) - ret -END(__logf_finite) diff --git a/sysdeps/i386/fpu/e_logf_data.c b/sysdeps/i386/fpu/e_logf_data.c deleted file mode 100644 index 1cc8931700..0000000000 --- a/sysdeps/i386/fpu/e_logf_data.c +++ /dev/null @@ -1 +0,0 @@ -/* Not needed. */ diff --git a/sysdeps/i386/fpu/libm-test-ulps b/sysdeps/i386/fpu/libm-test-ulps index 5a319f147b..698d1b3a1d 100644 --- a/sysdeps/i386/fpu/libm-test-ulps +++ b/sysdeps/i386/fpu/libm-test-ulps @@ -2000,17 +2000,17 @@ ldouble: 4 Function: "gamma_downward": double: 4 -float: 4 +float: 5 idouble: 4 -ifloat: 4 +ifloat: 5 ildouble: 7 ldouble: 7 Function: "gamma_towardzero": double: 4 -float: 2 +float: 3 idouble: 4 -ifloat: 2 +ifloat: 3 ildouble: 7 ldouble: 7 @@ -2186,20 +2186,20 @@ ldouble: 4 Function: "lgamma_downward": double: 4 -float: 4 +float: 5 float128: 8 idouble: 4 -ifloat: 4 +ifloat: 5 ifloat128: 8 ildouble: 7 ldouble: 7 Function: "lgamma_towardzero": double: 4 -float: 2 +float: 3 float128: 5 idouble: 4 -ifloat: 2 +ifloat: 3 ifloat128: 5 ildouble: 7 ldouble: 7 @@ -2641,10 +2641,10 @@ ldouble: 5 Function: "y0_towardzero": double: 2 -float: 2 +float: 3 float128: 3 idouble: 2 -ifloat: 2 +ifloat: 3 ifloat128: 3 ildouble: 5 ldouble: 5 diff --git a/sysdeps/i386/fpu/w_logf.c b/sysdeps/i386/fpu/w_logf.c deleted file mode 100644 index ea48d1356e..0000000000 --- a/sysdeps/i386/fpu/w_logf.c +++ /dev/null @@ -1 +0,0 @@ -#include <sysdeps/../math/w_logf.c> diff --git a/sysdeps/i386/i686/fpu/e_logf.S b/sysdeps/i386/i686/fpu/e_logf.S deleted file mode 100644 index 6fd39d50d3..0000000000 --- a/sysdeps/i386/i686/fpu/e_logf.S +++ /dev/null @@ -1,30 +0,0 @@ -/* - * Written by J.T. Conklin <jtc@netbsd.org>. - * Public domain. - * Adapted for float by Ulrich Drepper <drepper@cygnus.com>. - * - * Adapted for i686 instructions. - */ - -#include <machine/asm.h> - - - .text -ENTRY(__ieee754_logf) - fldln2 // log(2) - flds 4(%esp) // x : log(2) - fucomi %st - jp 3f - fyl2x // log(x) - ret - -3: fstp %st(1) - ret -END (__ieee754_logf) - -ENTRY(__logf_finite) - fldln2 // log(2) - flds 4(%esp) // x : log(2) - fyl2x // log(x) - ret -END(__logf_finite) diff --git a/sysdeps/i386/i686/fpu/multiarch/Makefile b/sysdeps/i386/i686/fpu/multiarch/Makefile index 2723f570af..ee4c3f39e7 100644 --- a/sysdeps/i386/i686/fpu/multiarch/Makefile +++ b/sysdeps/i386/i686/fpu/multiarch/Makefile @@ -1,7 +1,8 @@ ifeq ($(subdir),math) -libm-sysdep_routines += e_exp2f-sse2 e_expf-sse2 s_sinf-sse2 s_cosf-sse2 \ - s_sincosf-sse2 +libm-sysdep_routines += e_exp2f-sse2 e_expf-sse2 e_logf-sse2 \ + s_sinf-sse2 s_cosf-sse2 s_sincosf-sse2 CFLAGS-e_exp2f-sse2.c = -msse2 -mfpmath=sse CFLAGS-e_expf-sse2.c = -msse2 -mfpmath=sse +CFLAGS-e_logf-sse2.c = -msse2 -mfpmath=sse endif diff --git a/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c b/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c new file mode 100644 index 0000000000..11621fc122 --- /dev/null +++ b/sysdeps/i386/i686/fpu/multiarch/e_logf-sse2.c @@ -0,0 +1,3 @@ +#define __logf __logf_sse2 + +#include <sysdeps/ieee754/flt-32/e_logf.c> diff --git a/sysdeps/i386/i686/fpu/multiarch/e_logf.c b/sysdeps/i386/i686/fpu/multiarch/e_logf.c new file mode 100644 index 0000000000..1414d17d25 --- /dev/null +++ b/sysdeps/i386/i686/fpu/multiarch/e_logf.c @@ -0,0 +1,40 @@ +/* Multiple versions of logf. + Copyright (C) 2017 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <http://www.gnu.org/licenses/>. */ + +extern float __redirect_logf (float); + +#define SYMBOL_NAME logf +#include "ifunc-sse2.h" + +libc_ifunc_redirected (__redirect_logf, __logf, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (__logf_ia32, __GI___logf, __redirect_logf) + __attribute__ ((visibility ("hidden"))); + +# include <shlib-compat.h> +versioned_symbol (libm, __logf, logf, GLIBC_2_27); +#else +weak_alias (__logf, logf) +#endif + +strong_alias (__logf, __ieee754_logf) +strong_alias (__logf, __logf_finite) + +#define __logf __logf_ia32 +#include <sysdeps/ieee754/flt-32/e_logf.c> diff --git a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps index 2d056a9260..c86294b778 100644 --- a/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps +++ b/sysdeps/i386/i686/fpu/multiarch/libm-test-ulps @@ -2000,9 +2000,9 @@ ldouble: 4 Function: "gamma_downward": double: 4 -float: 4 +float: 5 idouble: 4 -ifloat: 4 +ifloat: 5 ildouble: 7 ldouble: 7 @@ -2186,10 +2186,10 @@ ldouble: 4 Function: "lgamma_downward": double: 4 -float: 4 +float: 5 float128: 8 idouble: 4 -ifloat: 4 +ifloat: 5 ifloat128: 8 ildouble: 7 ldouble: 7 @@ -2625,10 +2625,10 @@ ldouble: 5 Function: "y0_towardzero": double: 2 -float: 2 +float: 3 float128: 3 idouble: 2 -ifloat: 2 +ifloat: 3 ifloat128: 3 ildouble: 5 ldouble: 5