use gcc builtins for sqrt and fma functions
cover use gcc builtins for sqrt and fma functions
Hi,
This was suggested as part of ARC port review.
Build tested for following targets:
aarch64-linux-gnu arm-linux-gnueabi hppa-linux-gnu x86_64-linux-gnu arm-linux-gnueabihf riscv64-linux-gnu-rv64imac-lp64 riscv64-linux-gnu-rv64imafdc-lp64 powerpc-linux-gnu microblaze-linux-gnu nios2-linux-gnu hppa-linux-gnu
Thx,
-Vineet
Vineet Gupta (5):
ARC/dl-runtime helper macros
iee754: prvoide gcc builtins based generic sqrt functions
iee754: prvoide gcc builtins based generic fma functions aarch/fpu: use generic sqrt, fma functions powerpc/fpu: use generic fma functions
sysdeps/aarch64/fpu/e_sqrt.c | 27 -------- sysdeps/aarch64/fpu/e_sqrtf.c | 27 -------- sysdeps/aarch64/fpu/math-use-builtins.h | 70 +++++++++++++++++++++ sysdeps/aarch64/fpu/s_fma.c | 28 --------- sysdeps/aarch64/fpu/s_fmaf.c | 28 --------- sysdeps/arc/dl-runtime.h | 42 +++++++++++++ sysdeps/generic/math-use-builtins.h | 7 +++ sysdeps/ieee754/dbl-64/e_sqrt.c | 6 ++ sysdeps/ieee754/dbl-64/s_fma.c | 6 ++ sysdeps/ieee754/dbl-64/s_fmaf.c | 6 ++ sysdeps/ieee754/float128/float128_private.h | 2 + sysdeps/ieee754/flt-32/e_sqrtf.c | 6 ++ sysdeps/ieee754/ldbl-128/s_fmal.c | 5 ++ sysdeps/powerpc/fpu/math-use-builtins.h | 70 +++++++++++++++++++++ sysdeps/powerpc/fpu/s_fma.c | 27 -------- sysdeps/powerpc/fpu/s_fmaf.c | 27 -------- 16 files changed, 220 insertions(+), 164 deletions(-) delete mode 100644 sysdeps/aarch64/fpu/e_sqrt.c delete mode 100644 sysdeps/aarch64/fpu/e_sqrtf.c create mode 100644 sysdeps/aarch64/fpu/math-use-builtins.h delete mode 100644 sysdeps/aarch64/fpu/s_fma.c delete mode 100644 sysdeps/aarch64/fpu/s_fmaf.c create mode 100644 sysdeps/arc/dl-runtime.h create mode 100644 sysdeps/powerpc/fpu/math-use-builtins.h delete mode 100644 sysdeps/powerpc/fpu/s_fma.c delete mode 100644 sysdeps/powerpc/fpu/s_fmaf.c
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
PATCH 1/5 ARC/dl-runtime helper macros
This is purely for review purposes to attest the interface defined in prior patch
--- sysdeps/arc/dl-runtime.h | 42 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) create mode 100644 sysdeps/arc/dl-runtime.h
sysdeps/arc/dl-runtime.h new file +42 -0
@@ -0,0 +1,42 @@ +/* Helpers for On-demand PLT fixup for shared objects. ARC version. +Copyright (C) 2017-2020 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 +<https://www.gnu.org/licenses/>. */ + +/* PLT jump into resolver passes PC of PLTn, while _dl_fixup expects the +address of corresponding .rela.plt entry. + +- @plt0: runtime pc of first plt entry (DT_PLTGOT) +- @pltn: runtime pc of plt entry being resolved +- @size: size of .plt.rela entry (unused). */ +static inline uintptr_t +reloc_index (uintptr_t plt0,uintptr_t pltn,size_t size) +{ +unsigned long int idx = (unsigned long )pltn - (unsigned long )plt0; + +/* PLT trampoline is 16 bytes. */ + idx /=16 ; + +/* Exclude PLT0 and PLT1. */ +return idx -2 ; +} + +static inline uintptr_t +reloc_offset (uintptr_t plt0,uintptr_t pltn) +{ +size_t sz =sizeof (ElfW (Rela)); +return reloc_index (plt0, pltn, sz) * sz; +}
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 1/5] ARC/dl-runtime helper macros
Quoted 6 lines · Vineet Gupta
This is purely for review purposes to attest the interface defined in prior patch
--- sysdeps/arc/dl-runtime.h | 42 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) create mode 100644 sysdeps/arc/dl-runtime.h
Oops this was from a different series, please ignore. Don't post patches on friday evening :-(
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
PATCH 2/5 iee754: prvoide gcc builtins based generic sqrt functions
In reply to Vineet Gupta--- sysdeps/generic/math-use-builtins.h | 3 +++ sysdeps/ieee754/dbl-64/e_sqrt.c | 6 ++++++ sysdeps/ieee754/flt-32/e_sqrtf.c | 6 ++++++ 3 files changed, 15 insertions(+)
sysdeps/generic/math-use-builtins.h +3 -0
@@ -60,4 +60,7 @@# define USE_COPYSIGNF128_BUILTIN 0#endif +#define USE_SQRT_BUILTIN 0 +#define USE_SQRTF_BUILTIN 0 +#endif /* math-use-builtins.h */
sysdeps/ieee754/dbl-64/e_sqrt.c +6 -0
@@ -41,6 +41,7 @@#include <math_private.h> #include <fenv_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h> /*********************************************************************/ /* An ultimate sqrt routine. Given an IEEE double machine number x */
@@ -50,6 +51,10 @@double __ieee754_sqrt (double x) { +#if USE_SQRT_BUILTIN +return __builtin_sqrt (x); +#else +/* Use generic implementation. */ static const double rt0 =9.99999999859990725855365213134618E-01 , rt1 = 4.99999999495955425917856814202739E-01,
@@ -138,6 +143,7 @@ __ieee754_sqrt (double x)return (x - x) / (x - x);/* sqrt(-ve)=sNaN */ return 0x1p-256 * __ieee754_sqrt (x * 0x1p512); } +#endif /* ! USE_SQRT_BUILTIN */ }#ifndef __ieee754_sqrtlibm_alias_finite (__ieee754_sqrt, __sqrt)
sysdeps/ieee754/flt-32/e_sqrtf.c +6 -0
@@ -16,12 +16,17 @@#include <math.h> #include <math_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h> static const float one =1.0 , tiny=1.0e-30 ;float __ieee754_sqrtf (float x) { +#if USE_SQRTF_BUILTIN +return __builtin_sqrtf (x); +#else +/* Use generic implementation. */ float z;int32_t sign = (int )0x80000000 ;int32_t ix,s,q,m,t,i;
@@ -83,6 +88,7 @@ __ieee754_sqrtf(float x) ix += (m <<23 );SET_FLOAT_WORD (z,ix);return z; +#endif /* ! USE_SQRTF_BUILTIN */ }#ifndef __ieee754_sqrtflibm_alias_finite (__ieee754_sqrtf, __sqrtf)
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
On 29/05/2020 23:00, Vineet Gupta wrote:
LGTM, thanks.
Quoted 11 lines
--- sysdeps/generic/math-use-builtins.h | 3 +++ sysdeps/ieee754/dbl-64/e_sqrt.c | 6 ++++++ sysdeps/ieee754/flt-32/e_sqrtf.c | 6 ++++++ 3 files changed, 15 insertions(+)sysdeps/generic/math-use-builtins.h +0
-0
sysdeps/generic/math-use-builtins.h line 66
Quoted 7 lines
sysdeps/generic/math-use-builtins.h +3
-0@@ -60,4 +60,7 @@# define USE_COPYSIGNF128_BUILTIN 0#endif +#define USE_SQRT_BUILTIN 0 +#define USE_SQRTF_BUILTIN 0 +#endif /* math-use-builtins.h */
Ok.
sysdeps/ieee754/dbl-64/e_sqrt.c line 47
Quoted 12 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +1
-0@@ -41,6 +41,7 @@#include <math_private.h> #include <fenv_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h> /*********************************************************************/ /* An ultimate sqrt routine. Given an IEEE double machine number x */
Ok.
Quoted 12 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +4
-0@@ -50,6 +51,10 @@double __ieee754_sqrt (double x) { +#if USE_SQRT_BUILTIN +return __builtin_sqrt (x); +#else +/* Use generic implementation. */ static const double rt0 =9.99999999859990725855365213134618E-01 , rt1 = 4.99999999495955425917856814202739E-01,
sysdeps/ieee754/dbl-64/e_sqrt.c __ieee754_sqrt (double x) · line 149
Quoted 7 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +1
-0@@ -138,6 +143,7 @@ __ieee754_sqrt (double x)return (x - x) / (x - x);/* sqrt(-ve)=sNaN */ return 0x1p-256 * __ieee754_sqrt (x * 0x1p512); } +#endif /* ! USE_SQRT_BUILTIN */ }#ifndef __ieee754_sqrtlibm_alias_finite (__ieee754_sqrt, __sqrt)
Ok.
Quoted 24 lines
sysdeps/ieee754/flt-32/e_sqrtf.c +5
-0@@ -16,12 +16,17 @@#include <math.h> #include <math_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h> static const float one =1.0 , tiny=1.0e-30 ;float __ieee754_sqrtf (float x) { +#if USE_SQRTF_BUILTIN +return __builtin_sqrtf (x); +#else +/* Use generic implementation. */ float z;int32_t sign = (int )0x80000000 ;int32_t ix,s,q,m,t,i;
sysdeps/ieee754/flt-32/e_sqrtf.c __ieee754_sqrtf(float x) · line 94
Quoted 7 lines
sysdeps/ieee754/flt-32/e_sqrtf.c +1
-0@@ -83,6 +88,7 @@ __ieee754_sqrtf(float x) ix += (m <<23 );SET_FLOAT_WORD (z,ix);return z; +#endif /* ! USE_SQRTF_BUILTIN */ }#ifndef __ieee754_sqrtflibm_alias_finite (__ieee754_sqrtf, __sqrtf)
Ok.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
Quoted 76 lines · Adhemerval Zanella
On 29/05/2020 23:00, Vineet Gupta wrote:
LGTM, thanks.
Reviewed-by Adhemerval Zanella <[email protected]>Quoted 11 lines
--- sysdeps/generic/math-use-builtins.h | 3 +++ sysdeps/ieee754/dbl-64/e_sqrt.c | 6 ++++++ sysdeps/ieee754/flt-32/e_sqrtf.c | 6 ++++++ 3 files changed, 15 insertions(+)sysdeps/generic/math-use-builtins.h +0
-0sysdeps/generic/math-use-builtins.h line 66
Quoted 7 lines
sysdeps/generic/math-use-builtins.h +3
-0@@ -60,4 +60,7 @@# define USE_COPYSIGNF128_BUILTIN 0#endif +#define USE_SQRT_BUILTIN 0 +#define USE_SQRTF_BUILTIN 0 +#endif /* math-use-builtins.h */ Ok.
sysdeps/ieee754/dbl-64/e_sqrt.c line 47
Quoted 12 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +1
-0@@ -41,6 +41,7 @@#include <math_private.h> #include <fenv_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h> /*********************************************************************/ /* An ultimate sqrt routine. Given an IEEE double machine number x */ Ok.
Quoted 12 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +4
-0@@ -50,6 +51,10 @@double __ieee754_sqrt (double x) { +#if USE_SQRT_BUILTIN +return __builtin_sqrt (x); +#else +/* Use generic implementation. */ static const double rt0 =9.99999999859990725855365213134618E-01 , rt1 = 4.99999999495955425917856814202739E-01,sysdeps/ieee754/dbl-64/e_sqrt.c __ieee754_sqrt (double x) · line 149
Quoted 7 lines
sysdeps/ieee754/dbl-64/e_sqrt.c +1
-0@@ -138,6 +143,7 @@ __ieee754_sqrt (double x)return (x - x) / (x - x);/* sqrt(-ve)=sNaN */ return 0x1p-256 * __ieee754_sqrt (x * 0x1p512); } +#endif /* ! USE_SQRT_BUILTIN */ }#ifndef __ieee754_sqrtlibm_alias_finite (__ieee754_sqrt, __sqrt)Ok.
Quoted 10 lines
sysdeps/ieee754/flt-32/e_sqrtf.c +1
-0@@ -16,12 +16,17 @@#include <math.h> #include <math_private.h> #include <libm-alias-finite.h> +#include <math-use-builtins.h>
Quoted 1 line
sysdeps/ieee754/flt-32/e_sqrtf.c line 21
Quoted 1 line
sysdeps/ieee754/flt-32/e_sqrtf.c +0
-0static const float one =1.0 , tiny=1.0e-30 ;
You will need to move this definitions inside the !USE_SQRTF_BUILTIN to avoid defined by not used warnings. Current practice is to just open code the constants and let compiler optimize the constant pool:
sysdeps/ieee754/flt-32/e_sqrtf.c +4 -6
@@ -18,8 +18,6 @@#include <libm-alias-finite.h> #include <math-use-builtins.h> -static const float one =1.0 , tiny=1.0e-30 ; -float __ieee754_sqrtf (float x) {
@@ -75,10 +73,10 @@ __ieee754_sqrtf(float x)/* use floating add to find out rounding direction */ if (ix!=0 ) { - z = one-tiny;/* trigger inexact flag */ -if (z>=one) { - z = one+tiny; -if (z>one) + z =0x1p0 -0x1.4484cp-100 ;/* trigger inexact flag */ +if (z >=0x1p0 ) { + z =0x1p0 +0x1.4484cp-100 ; +if (z >0x1p0 ) q +=2 ;else q += (q&1 );
Quoted 14 lines
Quoted 14 lines
sysdeps/ieee754/flt-32/e_sqrtf.c +4
-0float __ieee754_sqrtf (float x) { +#if USE_SQRTF_BUILTIN +return __builtin_sqrtf (x); +#else +/* Use generic implementation. */ float z;int32_t sign = (int )0x80000000 ;int32_t ix,s,q,m,t,i;@@ -83,6 +88,7 @@ __ieee754_sqrtf(float x) ix += (m <<23 );SET_FLOAT_WORD (z,ix);
Quoted 9 lines
sysdeps/ieee754/flt-32/e_sqrtf.c __ieee754_sqrtf(float x) · line 94
Quoted 6 lines
sysdeps/ieee754/flt-32/e_sqrtf.c +1
-0return z; +#endif /* ! USE_SQRTF_BUILTIN */ }#ifndef __ieee754_sqrtflibm_alias_finite (__ieee754_sqrtf, __sqrtf)Ok.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
Quoted 3 lines · Adhemerval Zanella via Libc-alpha
You will need to move this definitions inside the !USE_SQRTF_BUILTIN to avoid defined by not used warnings. Current practice is to just open code the constants and let compiler optimize the constant pool:
Won't it be better to keep the const variable and trust the compiler to subsume it instead of open coding in multiple places. Makes it more readable ?
-Vineet
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
Quoted 7 lines · Vineet Gupta
Quoted 3 lines · Adhemerval Zanella via Libc-alpha
You will need to move this definitions inside the !USE_SQRTF_BUILTIN to avoid defined by not used warnings. Current practice is to just open code the constants and let compiler optimize the constant pool:
Won't it be better to keep the const variable and trust the compiler to subsume it instead of open coding in multiple places. Makes it more readable ?
I don't have a strong preference, it is just the recent math optimizations use the constant directly in the hex format instead of trying to factoring them out.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
In reply to Vineet GuptaThere's a typo in commit subject: “prvoide”.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 2/5] iee754: prvoide gcc builtins based generic sqrt functions
Quoted 1 line · Florian Weimer
There's a typo in commit subject: “prvoide”.
oops sorry, fixed now.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
PATCH 3/5 iee754: prvoide gcc builtins based generic fma functions
In reply to Vineet Gupta--- sysdeps/generic/math-use-builtins.h | 4 ++++ sysdeps/ieee754/dbl-64/s_fma.c | 6 ++++++ sysdeps/ieee754/dbl-64/s_fmaf.c | 6 ++++++ sysdeps/ieee754/float128/float128_private.h | 2 ++ sysdeps/ieee754/ldbl-128/s_fmal.c | 5 +++++ 5 files changed, 23 insertions(+)
sysdeps/generic/math-use-builtins.h +4 -0
@@ -63,4 +63,8 @@#define USE_SQRT_BUILTIN 0#define USE_SQRTF_BUILTIN 0 +#define USE_FMA_BUILTIN 0 +#define USE_FMAF_BUILTIN 0 +#define USE_FMAL_BUILTIN 0 +#endif /* math-use-builtins.h */
sysdeps/ieee754/dbl-64/s_fma.c +6 -0
@@ -25,6 +25,7 @@#include <fenv_private.h> #include <libm-alias-double.h> #include <tininess.h> +#include <math-use-builtins.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond:
@@ -33,6 +34,10 @@double __fma (double x,double y,double z) { +#if USE_FMA_BUILTIN +return __builtin_fma (x); +#else +/* Use generic implementation. */ union ieee754_double u, v, w;int adjust =0 ; u.d = x;
@@ -292,6 +297,7 @@ __fma (double x, double y, double z) v.ieee.mantissa1 |= j; return v.d * 0x1p-108; } +#endif /* ! USE_FMA_BUILTIN */ }#ifndef __fmalibm_alias_double (__fma, fma)
sysdeps/ieee754/dbl-64/s_fmaf.c +6 -0
@@ -23,6 +23,7 @@#include <math-barriers.h> #include <fenv_private.h> #include <libm-alias-float.h> +#include <math-use-builtins.h> /* This implementation relies on double being more than twice as precise as float and uses rounding to odd in order to avoid problems
@@ -33,6 +34,10 @@float __fmaf (float x,float y,float z) { +#if USE_FMAF_BUILTIN +return __builtin_fmaf (x); +#else +/* Use generic implementation. */ fenv_t env;/* Multiplication is always exact. */
@@ -60,6 +65,7 @@ __fmaf (float x, float y, float z)/* And finally truncation with round to nearest. */ return (float ) u.d; +#endif /* ! USE_FMAF_BUILTIN */ }#ifndef __fmaflibm_alias_float (__fma, fma)
sysdeps/ieee754/float128/float128_private.h +2 -0
@@ -154,6 +154,8 @@#define USE_ROUNDL_BUILTIN USE_ROUNDF128_BUILTIN#undef USE_COPYSIGNL_BUILTIN#define USE_COPYSIGNL_BUILTIN USE_COPYSIGNF128_BUILTIN +#undef USE_FMAL_BUILTIN +#define USE_FMAL_BUILTIN USE_FMA128_BUILTIN/* IEEE function renames. */ #define __ieee754_acoshl __ieee754_acoshf128
sysdeps/ieee754/ldbl-128/s_fmal.c +5 -0
@@ -25,6 +25,7 @@#include <math_private.h> #include <libm-alias-ldouble.h> #include <tininess.h> +#include <math-use-builtins.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond:
@@ -33,6 +34,9 @@_Float128 __fmal (_Float128 x,_Float128 y,_Float128 z) { +#if USE_FMAL_BUILTIN +return __builtin_fmal (x); +#else union ieee854_long_double u, v, w;int adjust =0 ; u.d = x;
@@ -296,5 +300,6 @@ __fmal (_Float128 x, _Float128 y, _Float128 z) v.ieee.mantissa3 |= j; return v.d * L(0x1p-228); } +#endif /* ! USE_FMAL_BUILTIN */ }libm_alias_ldouble (__fma, fma)
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 3/5] iee754: prvoide gcc builtins based generic fma functions
On 29/05/2020 23:00, Vineet Gupta wrote:
Ok with the fixes below.
Quoted 13 lines
--- sysdeps/generic/math-use-builtins.h | 4 ++++ sysdeps/ieee754/dbl-64/s_fma.c | 6 ++++++ sysdeps/ieee754/dbl-64/s_fmaf.c | 6 ++++++ sysdeps/ieee754/float128/float128_private.h | 2 ++ sysdeps/ieee754/ldbl-128/s_fmal.c | 5 +++++ 5 files changed, 23 insertions(+)sysdeps/generic/math-use-builtins.h +0
-0
sysdeps/generic/math-use-builtins.h line 70
Quoted 8 lines
sysdeps/generic/math-use-builtins.h +4
-0@@ -63,4 +63,8 @@#define USE_SQRT_BUILTIN 0#define USE_SQRTF_BUILTIN 0 +#define USE_FMA_BUILTIN 0 +#define USE_FMAF_BUILTIN 0 +#define USE_FMAL_BUILTIN 0 +#endif /* math-use-builtins.h */
For float128 support it should also contain a:
#define USE_FMAF128_BUILTIN
Quoted 13 lines
sysdeps/ieee754/dbl-64/s_fma.c +1
-0@@ -25,6 +25,7 @@#include <fenv_private.h> #include <libm-alias-double.h> #include <tininess.h> +#include <math-use-builtins.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond:
sysdeps/ieee754/dbl-64/s_fma.c line 38
Quoted 5 lines
sysdeps/ieee754/dbl-64/s_fma.c +2
-0@@ -33,6 +34,10 @@double __fma (double x,double y,double z) { +#if USE_FMA_BUILTIN +return __builtin_fma (x);
It should be:
return __builtin_fma (x, y, z);
Same for float and long double variant.
Quoted 47 lines
sysdeps/ieee754/dbl-64/s_fma.c +3
-0+#else +/* Use generic implementation. */ union ieee754_double u, v, w;int adjust =0 ; u.d = x;@@ -292,6 +297,7 @@ __fma (double x, double y, double z) v.ieee.mantissa1 |= j;return v.d *0x1p-108 ; } +#endif /* ! USE_FMA_BUILTIN */ }#ifndef __fma libm_alias_double (__fma, fma)> diff --git a/sysdeps/ieee754/dbl-64/s_fmaf.c b/sysdeps/ieee754/dbl-64/s_fmaf.cindex
57329d0a87fe..f15b18262124100644sysdeps/ieee754/dbl-64/s_fmaf.c +6
-0@@ -23,6 +23,7 @@#include <math-barriers.h> #include <fenv_private.h> #include <libm-alias-float.h> +#include <math-use-builtins.h> /* This implementation relies on double being more than twice as precise as float and uses rounding to odd in order to avoid problems @@ -33,6 +34,10 @@float __fmaf (float x,float y,float z) { +#if USE_FMAF_BUILTIN +return __builtin_fmaf (x); +#else +/* Use generic implementation. */ fenv_t env;/* Multiplication is always exact. */ @@ -60,6 +65,7 @@ __fmaf (float x, float y, float z)/* And finally truncation with round to nearest. */ return (float ) u.d; +#endif /* ! USE_FMAF_BUILTIN */ }#ifndef __fmaf libm_alias_float (__fma, fma)> diff --git a/sysdeps/ieee754/float128/float128_private.h b/sysdeps/ieee754/float128/float128_private.hindex
f97463d9dc1b..a697a7c29038100644
sysdeps/ieee754/float128/float128_private.h line 161
Quoted 8 lines
sysdeps/ieee754/float128/float128_private.h +2
-0@@ -154,6 +154,8 @@#define USE_ROUNDL_BUILTIN USE_ROUNDF128_BUILTIN#undef USE_COPYSIGNL_BUILTIN#define USE_COPYSIGNL_BUILTIN USE_COPYSIGNF128_BUILTIN +#undef USE_FMAL_BUILTIN +#define USE_FMAL_BUILTIN USE_FMA128_BUILTIN/* IEEE function renames. */ #define __ieee754_acoshl __ieee754_acoshf128
Ok.
Quoted 24 lines
sysdeps/ieee754/ldbl-128/s_fmal.c +4
-0@@ -25,6 +25,7 @@#include <math_private.h> #include <libm-alias-ldouble.h> #include <tininess.h> +#include <math-use-builtins.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond: @@ -33,6 +34,9 @@_Float128 __fmal (_Float128 x,_Float128 y,_Float128 z) { +#if USE_FMAL_BUILTIN +return __builtin_fmal (x); +#else union ieee854_long_double u, v, w;int adjust =0 ; u.d = x;
sysdeps/ieee754/ldbl-128/s_fmal.c __fmal (_Float128 x, _Float128 y, _Float128 z) · line 305
Quoted 6 lines
sysdeps/ieee754/ldbl-128/s_fmal.c +1
-0@@ -296,5 +300,6 @@ __fmal (_Float128 x, _Float128 y, _Float128 z) v.ieee.mantissa3 |= j; return v.d * L(0x1p-228); } +#endif /* ! USE_FMAL_BUILTIN */ }libm_alias_ldouble (__fma, fma)
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 3/5] iee754: prvoide gcc builtins based generic fma functions
In reply to Vineet GuptaThe commit subject contains a typo (“prvoide”).
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
PATCH 4/5 aarch/fpu: use generic sqrt, fma functions
In reply to Vineet Gupta--- sysdeps/aarch64/fpu/e_sqrt.c | 27 ---------- sysdeps/aarch64/fpu/e_sqrtf.c | 27 ---------- sysdeps/aarch64/fpu/math-use-builtins.h | 70 +++++++++++++++++++++++++ sysdeps/aarch64/fpu/s_fma.c | 28 ---------- sysdeps/aarch64/fpu/s_fmaf.c | 28 ---------- 5 files changed, 70 insertions(+), 110 deletions(-) delete mode 100644 sysdeps/aarch64/fpu/e_sqrt.c delete mode 100644 sysdeps/aarch64/fpu/e_sqrtf.c create mode 100644 sysdeps/aarch64/fpu/math-use-builtins.h delete mode 100644 sysdeps/aarch64/fpu/s_fma.c delete mode 100644 sysdeps/aarch64/fpu/s_fmaf.c
sysdeps/aarch64/fpu/e_sqrt.c deleted +0 -27
@@ -1,27 +0,0 @@ -/* Square root of floating point number. -Copyright (C) 2015-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <math_private.h> -#include <libm-alias-finite.h> - -double -__ieee754_sqrt (double d) -{ -return __builtin_sqrt (d); -} -libm_alias_finite (__ieee754_sqrt, __sqrt)
sysdeps/aarch64/fpu/e_sqrtf.c deleted +0 -27
@@ -1,27 +0,0 @@ -/* Single-precision floating point square root. -Copyright (C) 2015-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <math_private.h> -#include <libm-alias-finite.h> - -float -__ieee754_sqrtf (float s) -{ -return __builtin_sqrtf (s); -} -libm_alias_finite (__ieee754_sqrtf, __sqrtf)
sysdeps/aarch64/fpu/math-use-builtins.h new file +70 -0
70 lines not on this page · raw message
sysdeps/aarch64/fpu/s_fma.c deleted +0 -28
@@ -1,28 +0,0 @@ -/* Copyright (C) 1996-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <math.h> -#include <libm-alias-double.h> - -double -__fma (double x,double y,double z) -{ -return __builtin_fma (x, y, z); -} - -libm_alias_double (__fma, fma)
sysdeps/aarch64/fpu/s_fmaf.c deleted +0 -28
@@ -1,28 +0,0 @@ -/* Copyright (C) 2011-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <math.h> -#include <libm-alias-float.h> - -float -__fmaf (float x,float y,float z) -{ -return __builtin_fmaf (x, y, z); -} - -libm_alias_float (__fma, fma)
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 4/5] aarch/fpu: use generic sqrt, fma functions
Quoted 1 line · Vineet Gupta
Signed-off-by Vineet Gupta <[email protected]>
LGTM, some comments below.
Quoted 36 lines
--- sysdeps/aarch64/fpu/e_sqrt.c | 27 ---------- sysdeps/aarch64/fpu/e_sqrtf.c | 27 ---------- sysdeps/aarch64/fpu/math-use-builtins.h | 70 +++++++++++++++++++++++++ sysdeps/aarch64/fpu/s_fma.c | 28 ---------- sysdeps/aarch64/fpu/s_fmaf.c | 28 ---------- 5 files changed, 70 insertions(+), 110 deletions(-) delete mode 100644 sysdeps/aarch64/fpu/e_sqrt.c delete mode 100644 sysdeps/aarch64/fpu/e_sqrtf.c create mode 100644 sysdeps/aarch64/fpu/math-use-builtins.h delete mode 100644 sysdeps/aarch64/fpu/s_fma.c delete mode 100644 sysdeps/aarch64/fpu/s_fmaf.csysdeps/aarch64/fpu/e_sqrt.c deleted +0
-17@@ -1,27 +0,0 @@ -/* Square root of floating point number. -Copyright (C) 2015-2020 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 -<https://www.gnu.org/licenses/>. */
sysdeps/aarch64/fpu/e_sqrt.c line 27
Quoted 10 lines
sysdeps/aarch64/fpu/e_sqrt.c deleted +0
-10- -#include <math_private.h> -#include <libm-alias-finite.h> - -double -__ieee754_sqrt (double d) -{ -return __builtin_sqrt (d); -} -libm_alias_finite (__ieee754_sqrt, __sqrt)
Ok.
Quoted 23 lines
sysdeps/aarch64/fpu/e_sqrtf.c deleted +0
-17@@ -1,27 +0,0 @@ -/* Single-precision floating point square root. -Copyright (C) 2015-2020 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 -<https://www.gnu.org/licenses/>. */
sysdeps/aarch64/fpu/e_sqrtf.c line 27
Quoted 10 lines
sysdeps/aarch64/fpu/e_sqrtf.c deleted +0
-10- -#include <math_private.h> -#include <libm-alias-finite.h> - -float -__ieee754_sqrtf (float s) -{ -return __builtin_sqrtf (s); -} -libm_alias_finite (__ieee754_sqrtf, __sqrtf)
Ok.
sysdeps/aarch64/fpu/math-use-builtins.h line 2
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +2
-0@@ -0,0 +1,70 @@ +/* Using math gcc builtins instead of generic implementation. aarch64 version. +Copyright (C) 2019-2020 Free Software Foundation, Inc.
I think it should be just 2020 here.
Quoted 17 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +17
-0+ 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 + <https://www.gnu.org/licenses/>. */ + +#ifndef MATH_USE_BUILTINS_H
sysdeps/aarch64/fpu/math-use-builtins.h line 29
Quoted 10 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +10
-0+#define MATH_USE_BUILTINS_H 1 + +#include <features.h> /* For __GNUC_PREREQ. */ + +/* Define these macros to 1 to use __builtin_xyz instead of the +generic implementation. */ +#define USE_NEARBYINT_BUILTIN 0 +#define USE_NEARBYINTF_BUILTIN 0 +#define USE_NEARBYINTL_BUILTIN 0 +#define USE_NEARBYINTF128_BUILTIN 0
Since we are adding this new file for aarch64, we could also enable it fo nearbyint{f} and remove sysdeps/aarch64/fpu/s_nearbyint{f}.c as well.
sysdeps/aarch64/fpu/math-use-builtins.h line 34
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_RINT_BUILTIN 0 +#define USE_RINTF_BUILTIN 0 +#define USE_RINTL_BUILTIN 0 +#define USE_RINTF128_BUILTIN 0
Ditto.
sysdeps/aarch64/fpu/math-use-builtins.h line 39
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_FLOOR_BUILTIN 0 +#define USE_FLOORF_BUILTIN 0 +#define USE_FLOORL_BUILTIN 0 +#define USE_FLOORF128_BUILTIN 0
Ditto.
sysdeps/aarch64/fpu/math-use-builtins.h line 44
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_CEIL_BUILTIN 0 +#define USE_CEILF_BUILTIN 0 +#define USE_CEILL_BUILTIN 0 +#define USE_CEILF128_BUILTIN 0
Ditto.
sysdeps/aarch64/fpu/math-use-builtins.h line 49
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_TRUNC_BUILTIN 0 +#define USE_TRUNCF_BUILTIN 0 +#define USE_TRUNCL_BUILTIN 0 +#define USE_TRUNCF128_BUILTIN 0
Ditto.
sysdeps/aarch64/fpu/math-use-builtins.h line 54
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_ROUND_BUILTIN 0 +#define USE_ROUNDF_BUILTIN 0 +#define USE_ROUNDL_BUILTIN 0 +#define USE_ROUNDF128_BUILTIN 0
Ditto.
sysdeps/aarch64/fpu/math-use-builtins.h line 62
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +8
-0+ +#define USE_COPYSIGNL_BUILTIN 1 +#if __GNUC_PREREQ (7 ,0 ) +# define USE_COPYSIGNF128_BUILTIN 1 +#else +# define USE_COPYSIGNF128_BUILTIN 0 +#endif +
It should be described in commit message as well (although generated instruction are essentially the same).
sysdeps/aarch64/fpu/math-use-builtins.h line 70
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +8
-0+#define USE_SQRT_BUILTIN 1 +#define USE_SQRTF_BUILTIN 1 + +#define USE_FMA_BUILTIN 1 +#define USE_FMAF_BUILTIN 1 +#define USE_FMAL_BUILTIN 0 + +#endif /* math-use-builtins.h */
Ok.
Quoted 24 lines
sysdeps/aarch64/fpu/s_fma.c deleted +0
-18@@ -1,28 +0,0 @@ -/* Copyright (C) 1996-2020 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 -<https://www.gnu.org/licenses/>. */ -
sysdeps/aarch64/fpu/s_fma.c line 28
Quoted 10 lines
sysdeps/aarch64/fpu/s_fma.c deleted +0
-10-#include <math.h> -#include <libm-alias-double.h> - -double -__fma (double x,double y,double z) -{ -return __builtin_fma (x, y, z); -} - -libm_alias_double (__fma, fma)
Ok.
Quoted 25 lines
sysdeps/aarch64/fpu/s_fmaf.c deleted +0
-18@@ -1,28 +0,0 @@ -/* Copyright (C) 2011-2020 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 -<https://www.gnu.org/licenses/>. */ -
sysdeps/aarch64/fpu/s_fmaf.c line 28
Quoted 10 lines
sysdeps/aarch64/fpu/s_fmaf.c deleted +0
-10-#include <math.h> -#include <libm-alias-float.h> - -float -__fmaf (float x,float y,float z) -{ -return __builtin_fmaf (x, y, z); -} - -libm_alias_float (__fma, fma)
Ok.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 4/5] aarch/fpu: use generic sqrt, fma functions
Quoted 7 lines · Adhemerval Zanella via Libc-alpha
Quoted 1 line · Vineet Gupta
Signed-off-by Vineet Gupta <[email protected]>LGTM, some comments below.
Quoted 11 lines
sysdeps/aarch64/fpu/e_sqrt.c line 27
Quoted 9 lines
sysdeps/aarch64/fpu/e_sqrt.c deleted +0
-9-#include <math_private.h> -#include <libm-alias-finite.h> - -double -__ieee754_sqrt (double d) -{ -return __builtin_sqrt (d); -} -libm_alias_finite (__ieee754_sqrt, __sqrt)Ok.
How is one to test aarch64 port with hard-float. build-many-glibc doesn't have a hf variant and hacking one didn't do the right thing either.
Quoted 10 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 2
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +2
-0@@ -0,0 +1,70 @@ +/* Using math gcc builtins instead of generic implementation. aarch64 version. +Copyright (C) 2019-2020 Free Software Foundation, Inc. I think it should be just 2020 here.
Fixed.
Quoted 10 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 29
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +7
-0+ +/* Define these macros to 1 to use __builtin_xyz instead of the +generic implementation. */ +#define USE_NEARBYINT_BUILTIN 0 +#define USE_NEARBYINTF_BUILTIN 0 +#define USE_NEARBYINTL_BUILTIN 0 +#define USE_NEARBYINTF128_BUILTIN 0Since we are adding this new file for aarch64, we could also enable it fo nearbyint{f} and remove
sysdeps/aarch64/fpu/s_nearbyint{f}.cas well.
OK, but only
+#define USE_NEARBYINT_BUILTIN 1
+#define USE_NEARBYINTF_BUILTIN 1
The other 2 are not defined currently. If they are, then I prefer they be enabled as a separate commit for bisectability.
Quoted 6 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 34
Quoted 4 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +4
-0+#define USE_RINT_BUILTIN 0 +#define USE_RINTF_BUILTIN 0 +#define USE_RINTL_BUILTIN 0 +#define USE_RINTF128_BUILTIN 0Ditto.
OK and again only RINT and RINTF.
Also I don't understand one thing. Both the generic and aarch64 code have this
float__rintf (float x)
...libm_alias_float (__rint, rint)
The alias arg 1 __rint seems to lack suffix 'f' ?
Quoted 3 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 37
Quoted 3 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +3
-0+ +#define USE_FLOOR_BUILTIN 0 +#define USE_FLOORF_BUILTIN 0
Again FLOOR, FLOORF only
Quoted 4 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 39
Quoted 2 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +2
-0+#define USE_FLOORL_BUILTIN 0 +#define USE_FLOORF128_BUILTIN 0Ditto.
Ditto
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 44
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_CEIL_BUILTIN 0 +#define USE_CEILF_BUILTIN 0 +#define USE_CEILL_BUILTIN 0 +#define USE_CEILF128_BUILTIN 0Ditto.
Ditto
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 49
Quoted 6 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_TRUNC_BUILTIN 0 +#define USE_TRUNCF_BUILTIN 0 +#define USE_TRUNCL_BUILTIN 0 +#define USE_TRUNCF128_BUILTIN 0Ditto.
Ditto
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 54
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_ROUND_BUILTIN 0 +#define USE_ROUNDF_BUILTIN 0 +#define USE_ROUNDL_BUILTIN 0 +#define USE_ROUNDF128_BUILTIN 0Ditto.
Ditto
Quoted 11 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 62
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +8
-0+ +#define USE_COPYSIGNL_BUILTIN 1 +#if __GNUC_PREREQ (7 ,0 ) +# define USE_COPYSIGNF128_BUILTIN 1 +#else +# define USE_COPYSIGNF128_BUILTIN 0 +#endif +It should be described in commit message as well (although generated instruction are essentially the same).
Well this one is even more special as it was already using the same code, except that the math-use-builtins.h was generic vs. aarch64 specific one.
Thx
-Vineet
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 4/5] aarch/fpu: use generic sqrt, fma functions
Quoted 21 lines · Vineet Gupta via Libc-alpha
Quoted 7 lines · Adhemerval Zanella via Libc-alpha
Quoted 1 line · Vineet Gupta
Signed-off-by Vineet Gupta <[email protected]>LGTM, some comments below.
Quoted 11 lines
sysdeps/aarch64/fpu/e_sqrt.c line 27
Quoted 9 lines
sysdeps/aarch64/fpu/e_sqrt.c deleted +0
-9-#include <math_private.h> -#include <libm-alias-finite.h> - -double -__ieee754_sqrt (double d) -{ -return __builtin_sqrt (d); -} -libm_alias_finite (__ieee754_sqrt, __sqrt)Ok.
Quoted 2 lines
How is one to test aarch64 port with hard-float. build-many-glibc doesn't have a hf variant and hacking one didn't do the right thing either.
AArch64 is hard-float by default : there is no soft-float variant.
regards
Ramana
Quoted 109 lines
Quoted 11 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 2
Quoted 9 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +2
-0@@ -0,0 +1,70 @@ +/* Using math gcc builtins instead of generic implementation. aarch64 version. +Copyright (C) 2019-2020 Free Software Foundation, Inc. I think it should be just 2020 here.
Fixed.
Quoted 10 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 29
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +7
-0+ +/* Define these macros to 1 to use __builtin_xyz instead of the +generic implementation. */ +#define USE_NEARBYINT_BUILTIN 0 +#define USE_NEARBYINTF_BUILTIN 0 +#define USE_NEARBYINTL_BUILTIN 0 +#define USE_NEARBYINTF128_BUILTIN 0Since we are adding this new file for aarch64, we could also enable it fo nearbyint{f} and remove
sysdeps/aarch64/fpu/s_nearbyint{f}.cas well.OK, but only
(quoted diff) +2
-0+#define USE_NEARBYINT_BUILTIN 1 +#define USE_NEARBYINTF_BUILTIN 1The other 2 are not defined currently. If they are, then I prefer they be enabled as a separate commit for bisectability.
Quoted 6 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 34
Quoted 4 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +4
-0+#define USE_RINT_BUILTIN 0 +#define USE_RINTF_BUILTIN 0 +#define USE_RINTL_BUILTIN 0 +#define USE_RINTF128_BUILTIN 0Ditto.
OK and again only RINT and RINTF.
Also I don't understand one thing. Both the generic and aarch64 code have thisfloat
__rintf(float x)
...libm_alias_float(__rint, rint)The alias arg 1
__rintseems to lack suffix 'f' ?Quoted 3 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 37
Quoted 3 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +3
-0+ +#define USE_FLOOR_BUILTIN 0 +#define USE_FLOORF_BUILTIN 0Again FLOOR, FLOORF only
Quoted 4 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 39
Quoted 2 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +2
-0+#define USE_FLOORL_BUILTIN 0 +#define USE_FLOORF128_BUILTIN 0Ditto.
Ditto
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 44
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_CEIL_BUILTIN 0 +#define USE_CEILF_BUILTIN 0 +#define USE_CEILL_BUILTIN 0 +#define USE_CEILF128_BUILTIN 0Ditto.
Ditto
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 49
Quoted 6 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_TRUNC_BUILTIN 0 +#define USE_TRUNCF_BUILTIN 0 +#define USE_TRUNCL_BUILTIN 0 +#define USE_TRUNCF128_BUILTIN 0Ditto.
Ditto
Quoted 7 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 54
Quoted 5 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +5
-0+ +#define USE_ROUND_BUILTIN 0 +#define USE_ROUNDF_BUILTIN 0 +#define USE_ROUNDL_BUILTIN 0 +#define USE_ROUNDF128_BUILTIN 0Ditto.
Ditto
Quoted 11 lines
sysdeps/aarch64/fpu/math-use-builtins.h line 62
Quoted 8 lines
sysdeps/aarch64/fpu/math-use-builtins.h new file +8
-0+ +#define USE_COPYSIGNL_BUILTIN 1 +#if __GNUC_PREREQ (7 ,0 ) +# define USE_COPYSIGNF128_BUILTIN 1 +#else +# define USE_COPYSIGNF128_BUILTIN 0 +#endif +It should be described in commit message as well (although generated instruction are essentially the same).
Quoted 5 lines
Well this one is even more special as it was already using the same code, except that the
math-use-builtins.hwas generic vs. aarch64 specific one.Thx
-Vineet
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 4/5] aarch/fpu: use generic sqrt, fma functions
In reply to Vineet GuptaQuoted 8 lines · Vineet Gupta via Libc-alpha
Also I don't understand one thing. Both the generic and aarch64 code have this
float
__rintf(float x)
...libm_alias_float(__rint, rint)The alias arg 1
__rintseems to lack suffix 'f' ?
Please see the comments in sysdeps/generic/libm-alias-float.h. The arguments are *unsuffixed* names.
Signature · 7 lines
Joseph S. Myers [email protected] _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
PATCH 5/5 powerpc/fpu: use generic fma functions
In reply to Vineet Gupta--- sysdeps/powerpc/fpu/math-use-builtins.h | 70 +++++++++++++++++++++++++ sysdeps/powerpc/fpu/s_fma.c | 27 ---------- sysdeps/powerpc/fpu/s_fmaf.c | 27 ---------- 3 files changed, 70 insertions(+), 54 deletions(-) create mode 100644 sysdeps/powerpc/fpu/math-use-builtins.h delete mode 100644 sysdeps/powerpc/fpu/s_fma.c delete mode 100644 sysdeps/powerpc/fpu/s_fmaf.c
sysdeps/powerpc/fpu/math-use-builtins.h new file +70 -0
@@ -0,0 +1,70 @@ +/* Using math gcc builtins instead of generic implementation. PowerPC version. +Copyright (C) 2019-2020 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 +<https://www.gnu.org/licenses/>. */ + +#ifndef MATH_USE_BUILTINS_H +#define MATH_USE_BUILTINS_H 1 + +#include <features.h> /* For __GNUC_PREREQ. */ + +/* Define these macros to 1 to use __builtin_xyz instead of the +generic implementation. */ +#define USE_NEARBYINT_BUILTIN 0 +#define USE_NEARBYINTF_BUILTIN 0 +#define USE_NEARBYINTL_BUILTIN 0 +#define USE_NEARBYINTF128_BUILTIN 0 + +#define USE_RINT_BUILTIN 0 +#define USE_RINTF_BUILTIN 0 +#define USE_RINTL_BUILTIN 0 +#define USE_RINTF128_BUILTIN 0 + +#define USE_FLOOR_BUILTIN 0 +#define USE_FLOORF_BUILTIN 0 +#define USE_FLOORL_BUILTIN 0 +#define USE_FLOORF128_BUILTIN 0 + +#define USE_CEIL_BUILTIN 0 +#define USE_CEILF_BUILTIN 0 +#define USE_CEILL_BUILTIN 0 +#define USE_CEILF128_BUILTIN 0 + +#define USE_TRUNC_BUILTIN 0 +#define USE_TRUNCF_BUILTIN 0 +#define USE_TRUNCL_BUILTIN 0 +#define USE_TRUNCF128_BUILTIN 0 + +#define USE_ROUND_BUILTIN 0 +#define USE_ROUNDF_BUILTIN 0 +#define USE_ROUNDL_BUILTIN 0 +#define USE_ROUNDF128_BUILTIN 0 + +#define USE_COPYSIGNL_BUILTIN 1 +#if __GNUC_PREREQ (7 ,0 ) +# define USE_COPYSIGNF128_BUILTIN 1 +#else +# define USE_COPYSIGNF128_BUILTIN 0 +#endif + +#define USE_SQRT_BUILTIN 0 +#define USE_SQRTF_BUILTIN 0 + +#define USE_FMA_BUILTIN 1 +#define USE_FMAF_BUILTIN 1 +#define USE_FMAL_BUILTIN 0 + +#endif /* math-use-builtins.h */
sysdeps/powerpc/fpu/s_fma.c deleted +0 -27
@@ -1,27 +0,0 @@ -/* Compute x * y + z as ternary operation. PowerPC version. -Copyright (C) 2019-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <sysdep.h> -#include <libm-alias-double.h> - -double -__fma (double x,double y,double z) -{ -return __builtin_fma (x, y, z); -} -libm_alias_double (__fma, fma)
sysdeps/powerpc/fpu/s_fmaf.c deleted +0 -27
@@ -1,27 +0,0 @@ -/* Compute x * y + z as ternary operation. PowerPC version. -Copyright (C) 2019-2020 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 -<https://www.gnu.org/licenses/>. */ - -#include <sysdep.h> -#include <libm-alias-float.h> - -float -__fmaf (float x,float y,float z) -{ -return __builtin_fmaf (x, y, z); -} -libm_alias_float (__fma, fma)
Signature · 7 lines
2.20.1 _______________________________________________ linux-snps-arc mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-snps-arc
Re: [PATCH 5/5] powerpc/fpu: use generic fma functions
On 29/05/2020 23:00, Vineet Gupta wrote:
LGTM, thanks.
This new addition, along with the aarch64 one, makes me wonder if it would be better to decompose the USE_* defined in multiple files so the architecture adds only the required ones (instead of copy/paste the whole file with the generic ones).
Also for powerpc, maybe a future cleanup would to move the logic to select the builtin to math-use-builtins.h. Something like:
#ifdef _ARCH_PWR5X # define USE_FLOOR_BUILTIN 1 # define USE_FLOORF_BUILTIN 1 #else # define USE_FLOOR_BUILTIN 0 # define USE_FLOORF_BUILTIN 0 #endif #define USE_FLOORL_BUILTIN 0 #define USE_FLOORF128_BUILTIN 0
Quoted 75 lines
--- sysdeps/powerpc/fpu/math-use-builtins.h | 70 +++++++++++++++++++++++++ sysdeps/powerpc/fpu/s_fma.c | 27 ---------- sysdeps/powerpc/fpu/s_fmaf.c | 27 ---------- 3 files changed, 70 insertions(+), 54 deletions(-) create mode 100644 sysdeps/powerpc/fpu/math-use-builtins.h delete mode 100644 sysdeps/powerpc/fpu/s_fma.c delete mode 100644 sysdeps/powerpc/fpu/s_fmaf.csysdeps/powerpc/fpu/math-use-builtins.h new file +60
-0@@ -0,0 +1,70 @@ +/* Using math gcc builtins instead of generic implementation. PowerPC version. +Copyright (C) 2019-2020 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 +<https://www.gnu.org/licenses/>. */ + +#ifndef MATH_USE_BUILTINS_H +#define MATH_USE_BUILTINS_H 1 + +#include <features.h> /* For __GNUC_PREREQ. */ + +/* Define these macros to 1 to use __builtin_xyz instead of the +generic implementation. */ +#define USE_NEARBYINT_BUILTIN 0 +#define USE_NEARBYINTF_BUILTIN 0 +#define USE_NEARBYINTL_BUILTIN 0 +#define USE_NEARBYINTF128_BUILTIN 0 + +#define USE_RINT_BUILTIN 0 +#define USE_RINTF_BUILTIN 0 +#define USE_RINTL_BUILTIN 0 +#define USE_RINTF128_BUILTIN 0 + +#define USE_FLOOR_BUILTIN 0 +#define USE_FLOORF_BUILTIN 0 +#define USE_FLOORL_BUILTIN 0 +#define USE_FLOORF128_BUILTIN 0 + +#define USE_CEIL_BUILTIN 0 +#define USE_CEILF_BUILTIN 0 +#define USE_CEILL_BUILTIN 0 +#define USE_CEILF128_BUILTIN 0 + +#define USE_TRUNC_BUILTIN 0 +#define USE_TRUNCF_BUILTIN 0 +#define USE_TRUNCL_BUILTIN 0 +#define USE_TRUNCF128_BUILTIN 0 + +#define USE_ROUND_BUILTIN 0 +#define USE_ROUNDF_BUILTIN 0 +#define USE_ROUNDL_BUILTIN 0 +#define USE_ROUNDF128_BUILTIN 0 + +#define USE_COPYSIGNL_BUILTIN 1 +#if __GNUC_PREREQ (7 ,0 ) +# define USE_COPYSIGNF128_BUILTIN 1 +#else +# define USE_COPYSIGNF128_BUILTIN 0
sysdeps/powerpc/fpu/math-use-builtins.h line 70
Quoted 10 lines
sysdeps/powerpc/fpu/math-use-builtins.h new file +10
-0+#endif + +#define USE_SQRT_BUILTIN 0 +#define USE_SQRTF_BUILTIN 0 + +#define USE_FMA_BUILTIN 1 +#define USE_FMAF_BUILTIN 1 +#define USE_FMAL_BUILTIN 0 + +#endif /* math-use-builtins.h */
Ok.
Quoted 23 lines
sysdeps/powerpc/fpu/s_fma.c deleted +0
-17@@ -1,27 +0,0 @@ -/* Compute x * y + z as ternary operation. PowerPC version. -Copyright (C) 2019-2020 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 -<https://www.gnu.org/licenses/>. */
sysdeps/powerpc/fpu/s_fma.c line 27
Quoted 10 lines
sysdeps/powerpc/fpu/s_fma.c deleted +0
-10- -#include <sysdep.h> -#include <libm-alias-double.h> - -double -__fma (double x,double y,double z) -{ -return __builtin_fma (x, y, z); -} -libm_alias_double (__fma, fma)
Ok.
Quoted 24 lines
sysdeps/powerpc/fpu/s_fmaf.c deleted +0
-17@@ -1,27 +0,0 @@ -/* Compute x * y + z as ternary operation. PowerPC version. -Copyright (C) 2019-2020 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 -<https://www.gnu.org/licenses/>. */
sysdeps/powerpc/fpu/s_fmaf.c line 27
Quoted 10 lines
sysdeps/powerpc/fpu/s_fmaf.c deleted +0
-10- -#include <sysdep.h> -#include <libm-alias-float.h> - -float -__fmaf (float x,float y,float z) -{ -return __builtin_fmaf (x, y, z); -} -libm_alias_float (__fma, fma)
Ok.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc