AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-12-22 10:50:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Optimize exp

Browse Source 
Add __exp*_finite optimizations and rewrite some wrappers.
This commit is contained in:
Ulrich Drepper 2011-10-15 20:22:59 -04:00
parent ba1a0d5938
commit bcf01e6d80
14 changed files with 169 additions and 146 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
ChangeLog
View File

@ -1,5 +1,20 @@
2011-10-15 Ulrich Drepper <drepper@gmail.com> 2011-10-15 Ulrich Drepper <drepper@gmail.com>
* math/Versions [libm] (GLIBC_2.15): Add __exp_finite, __expf_finite,
__expl_finite.
* math/bits/math-finite.h: Add entries for exp.
* math/e_expl.c: Add __*_finite alias.
* sysdeps/i386/fpu/e_exp.S: Likewise.
* sysdeps/i386/fpu/e_expf.S: Likewise.
* sysdeps/i386/fpu/e_expl.c: Likewise.
* sysdeps/ieee754/dbl-64/e_exp.c: Likewise.
* sysdeps/ieee754/flt-32/e_expf.c: Likewise.
* sysdeps/ieee754/ldbl-128/e_expl.c: Likewise.
* sysdeps/ieee754/ldbl-128ibm/e_expl.c: Likewise.
* sysdeps/ieee754/dbl-64/w_exp.c: Complete rewrite.
* sysdeps/ieee754/flt-32/w_expf.c: Likewise.
* sysdeps/ieee754/ldbl-96/w_expl.c: Likewise.
* sysdeps/i386/i686/fpu/e_logf.S: No need for the fyl2xp1 use, fyl2x * sysdeps/i386/i686/fpu/e_logf.S: No need for the fyl2xp1 use, fyl2x
is sufficient, at least on modern CPUs. is sufficient, at least on modern CPUs.

1
math/Versions
View File

@ -196,5 +196,6 @@ libm {
__sinh_finite; __sinhf_finite; __sinhl_finite; __sinh_finite; __sinhf_finite; __sinhl_finite;
__sqrt_finite; __sqrtf_finite; __sqrtl_finite; __sqrt_finite; __sqrtf_finite; __sqrtl_finite;
__gamma_r_finite; __gammaf_r_finite; __gammal_r_finite; __gamma_r_finite; __gammaf_r_finite; __gammal_r_finite;
__exp_finite; __expf_finite; __expl_finite;
} }
} }

7
math/bits/math-finite.h
View File

@ -67,6 +67,13 @@ extern float coshf (float) __asm__ ("__coshf_finite");
extern long double coshl (long double) __asm__ ("__coshl_finite"); extern long double coshl (long double) __asm__ ("__coshl_finite");
#endif #endif
/* exp. */
extern double exp (double) __asm__ ("__exp_finite");
extern float expf (float) __asm__ ("__expf_finite");
#ifdef __MATH_DECLARE_LDOUBLE
extern long double expl (long double) __asm__ ("__expl_finite");
#endif
#ifdef __USE_GNU #ifdef __USE_GNU
/* exp10. */ /* exp10. */
extern double exp10 (double) __asm__ ("__exp10_finite"); extern double exp10 (double) __asm__ ("__exp10_finite");

1
math/e_expl.c
View File

@ -9,6 +9,7 @@ __ieee754_expl (long double x)
__set_errno (ENOSYS); __set_errno (ENOSYS);
return 0.0; return 0.0;
} }
strong_alias (__ieee754_expl, __expl_finite)
stub_warning (expl) stub_warning (expl)
#include <stub-tag.h> #include <stub-tag.h>

17
sysdeps/i386/fpu/e_exp.S
View File

@ -5,7 +5,6 @@
#include <machine/asm.h> #include <machine/asm.h>
RCSID("$NetBSD: e_exp.S,v 1.7 1996/07/03 17:31:28 jtc Exp $")
/* e^x = 2^(x * log2(e)) */ /* e^x = 2^(x * log2(e)) */
ENTRY(__ieee754_exp) ENTRY(__ieee754_exp)
@ -39,3 +38,19 @@ ENTRY(__ieee754_exp)
fldz /* Set result to 0. */ fldz /* Set result to 0. */
2: ret 2: ret
END (__ieee754_exp) END (__ieee754_exp)
ENTRY(__exp_finite)
fldl2e
fmull 4(%esp) /* x * log2(e) */
fld %st
frndint /* int(x * log2(e)) */
fsubr %st,%st(1) /* fract(x * log2(e)) */
fxch
f2xm1 /* 2^(fract(x * log2(e))) - 1 */
fld1
faddp /* 2^(fract(x * log2(e))) */
fscale /* e^x */
fstp %st(1)
ret
END(__exp_finite)

17
sysdeps/i386/fpu/e_expf.S
View File

@ -6,7 +6,6 @@
#include <machine/asm.h> #include <machine/asm.h>
RCSID("$NetBSD: $")
/* e^x = 2^(x * log2(e)) */ /* e^x = 2^(x * log2(e)) */
ENTRY(__ieee754_expf) ENTRY(__ieee754_expf)
@ -40,3 +39,19 @@ ENTRY(__ieee754_expf)
fldz /* Set result to 0. */ fldz /* Set result to 0. */
2: ret 2: ret
END (__ieee754_expf) END (__ieee754_expf)
ENTRY(__expf_finite)
fldl2e
fmuls 4(%esp) /* x * log2(e) */
fld %st
frndint /* int(x * log2(e)) */
fsubr %st,%st(1) /* fract(x * log2(e)) */
fxch
f2xm1 /* 2^(fract(x * log2(e))) - 1 */
fld1
faddp /* 2^(fract(x * log2(e))) */
fscale /* e^x */
fstp %st(1)
ret
END(__expf_finite)

3
sysdeps/i386/fpu/e_expl.c
View File

@ -63,7 +63,7 @@ __ieee754_expl (long double x)
"fld1\n\t" /* 4 1.0 */ "fld1\n\t" /* 4 1.0 */
"faddp\n\t" /* 3 2^(fract(x * log2(e))) */ "faddp\n\t" /* 3 2^(fract(x * log2(e))) */
"fstp %%st(1)\n\t" /* 2 */ "fstp %%st(1)\n\t" /* 2 */
"fscale\n\t" /* 2 scale factor is st(1); e^x */ "fscale\n\t" /* 2 scale factor is st(1); e^x */
"fstp %%st(1)\n\t" /* 1 */ "fstp %%st(1)\n\t" /* 1 */
"fstp %%st(1)\n\t" /* 0 */ "fstp %%st(1)\n\t" /* 0 */
"jmp 2f\n\t" "jmp 2f\n\t"
@ -75,3 +75,4 @@ __ieee754_expl (long double x)
: "=t" (res) : "0" (x), "m" (c0), "m" (c1) : "ax", "dx"); : "=t" (res) : "0" (x), "m" (c0), "m" (c1) : "ax", "dx");
return res; return res;
} }
strong_alias (__ieee754_expl, __expl_finite)

3
sysdeps/ieee754/dbl-64/e_exp.c
View File

@ -1,7 +1,7 @@
/* /*
* IBM Accurate Mathematical Library * IBM Accurate Mathematical Library
* written by International Business Machines Corp. * written by International Business Machines Corp.
* Copyright (C) 2001 Free Software Foundation * Copyright (C) 2001, 2011 Free Software Foundation
* *
* This program is free software; you can redistribute it and/or modify * This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU Lesser General Public License as published by * it under the terms of the GNU Lesser General Public License as published by
@ -145,6 +145,7 @@ double __ieee754_exp(double x) {
else return __slowexp(x); else return __slowexp(x);
} }
} }
strong_alias (__ieee754_exp, __exp_finite)
/************************************************************************/ /************************************************************************/
/* Compute e^(x+xx)(Double-Length number) .The routine also receive */ /* Compute e^(x+xx)(Double-Length number) .The routine also receive */

77
sysdeps/ieee754/dbl-64/w_exp.c
View File

@ -1,55 +1,46 @@
/* @(#)w_exp.c 5.1 93/09/24 */ /* Copyright (C) 2011 Free Software Foundation, Inc.
/* This file is part of the GNU C Library.
* ==================================================== Contributed by Ulrich Drepper <drepper@gmail.com>, 2011.
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
*
* Developed at SunPro, a Sun Microsystems, Inc. business.
* Permission to use, copy, modify, and distribute this
* software is freely granted, provided that this notice
* is preserved.
* ====================================================
*/
#if defined(LIBM_SCCS) && !defined(lint) The GNU C Library is free software; you can redistribute it and/or
static char rcsid[] = "$NetBSD: w_exp.c,v 1.6 1995/05/10 20:48:51 jtc Exp $"; modify it under the terms of the GNU Lesser General Public
#endif 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,
* wrapper exp(x) 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.
#include "math.h" You should have received a copy of the GNU Lesser General Public
#include "math_private.h" License along with the GNU C Library; if not, write to the Free
Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include <math.h>
#include <math_private.h>
#ifdef __STDC__
static const double static const double
#else
static double
#endif
o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */ o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */
u_threshold= -7.45133219101941108420e+02; /* 0xc0874910, 0xD52D3051 */ u_threshold= -7.45133219101941108420e+02; /* 0xc0874910, 0xD52D3051 */
#ifdef __STDC__
double __exp(double x) /* wrapper exp */ /* wrapper exp */
#else double
double __exp(x) /* wrapper exp */ __exp (double x)
double x;
#endif
{ {
#ifdef _IEEE_LIBM if (__builtin_expect (x > o_threshold, 0))
return __ieee754_exp(x); {
#else if (_LIB_VERSION != _IEEE_)
double z; return __kernel_standard_f (x, x, 6);
z = __ieee754_exp(x); }
if(_LIB_VERSION == _IEEE_) return z; else if (__builtin_expect (x < u_threshold, 0))
if(__finite(x)) { {
if(x>o_threshold) if (_LIB_VERSION != _IEEE_)
return __kernel_standard(x,x,6); /* exp overflow */ return __kernel_standard_f (x, x, 7);
else if(x<u_threshold) }
return __kernel_standard(x,x,7); /* exp underflow */
} return __ieee754_exp (x);
return z;
#endif
} }
hidden_def (__exp) hidden_def (__exp)
weak_alias (__exp, exp) weak_alias (__exp, exp)

7
sysdeps/ieee754/flt-32/e_expf.c
View File

@ -1,5 +1,5 @@
/* Single-precision floating point e^x. /* Single-precision floating point e^x.
Copyright (C) 1997, 1998, 2005, 2006 Free Software Foundation, Inc. Copyright (C) 1997, 1998, 2005, 2006, 2011 Free Software Foundation, Inc.
This file is part of the GNU C Library. This file is part of the GNU C Library.
Contributed by Geoffrey Keating <geoffk@ozemail.com.au> Contributed by Geoffrey Keating <geoffk@ozemail.com.au>
@ -33,8 +33,8 @@
Then e^x is approximated as Then e^x is approximated as
e^x = 2^n ( e^(t/512 + delta[t]) e^x = 2^n ( e^(t/512 + delta[t])
+ ( e^(t/512 + delta[t]) + ( e^(t/512 + delta[t])
* ( p(x + delta[t] + n * ln(2)) - delta ) ) ) * ( p(x + delta[t] + n * ln(2)) - delta ) ) )
where where
- p(x) is a polynomial approximating e(x)-1; - p(x) is a polynomial approximating e(x)-1;
@ -138,3 +138,4 @@ __ieee754_expf (float x)
/* Return x, if x is a NaN or Inf; or overflow, otherwise. */ /* Return x, if x is a NaN or Inf; or overflow, otherwise. */
return TWO127*x; return TWO127*x;
} }
strong_alias (__ieee754_expf, __expf_finite)

80
sysdeps/ieee754/flt-32/w_expf.c
View File

@ -1,60 +1,46 @@
/* w_expf.c -- float version of w_exp.c. /* Copyright (C) 2011 Free Software Foundation, Inc.
* Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com. This file is part of the GNU C Library.
*/ Contributed by Ulrich Drepper <drepper@gmail.com>, 2011.
/* The GNU C Library is free software; you can redistribute it and/or
* ==================================================== modify it under the terms of the GNU Lesser General Public
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
* Developed at SunPro, a Sun Microsystems, Inc. business.
* Permission to use, copy, modify, and distribute this
* software is freely granted, provided that this notice
* is preserved.
* ====================================================
*/
#if defined(LIBM_SCCS) && !defined(lint) The GNU C Library is distributed in the hope that it will be useful,
static char rcsid[] = "$NetBSD: w_expf.c,v 1.3 1995/05/10 20:48:53 jtc Exp $"; but WITHOUT ANY WARRANTY; without even the implied warranty of
#endif 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
* wrapper expf(x) License along with the GNU C Library; if not, write to the Free
*/ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include "math.h" #include <math.h>
#include "math_private.h" #include <math_private.h>
#ifdef __STDC__
static const float static const float
#else
static float
#endif
o_threshold= 8.8722831726e+01, /* 0x42b17217 */ o_threshold= 8.8722831726e+01, /* 0x42b17217 */
u_threshold= -1.0397208405e+02; /* 0xc2cff1b5 */ u_threshold= -1.0397208405e+02; /* 0xc2cff1b5 */
#ifdef __STDC__
float __expf(float x) /* wrapper expf */ /* wrapper expf */
#else float
float __expf(x) /* wrapper expf */ __expf (float x)
float x;
#endif
{ {
#ifdef _IEEE_LIBM if (__builtin_expect (x > o_threshold, 0))
return __ieee754_expf(x); {
#else if (_LIB_VERSION != _IEEE_)
float z; return __kernel_standard_f (x, x, 106);
z = __ieee754_expf(x); }
if(_LIB_VERSION == _IEEE_) return z; else if (__builtin_expect (x < u_threshold, 0))
if(__finitef(x)) { {
if(x>o_threshold) if (_LIB_VERSION != _IEEE_)
/* exp overflow */ return __kernel_standard_f (x, x, 107);
return (float)__kernel_standard((double)x,(double)x,106); }
else if(x<u_threshold)
/* exp underflow */ return __ieee754_expf (x);
return (float)__kernel_standard((double)x,(double)x,107);
}
return z;
#endif
} }
hidden_def (__expf) hidden_def (__expf)
weak_alias (__expf, expf) weak_alias (__expf, expf)

5
sysdeps/ieee754/ldbl-128/e_expl.c
View File

@ -1,5 +1,5 @@
/* Quad-precision floating point e^x. /* Quad-precision floating point e^x.
Copyright (C) 1999 Free Software Foundation, Inc. Copyright (C) 1999, 2011 Free Software Foundation, Inc.
This file is part of the GNU C Library. This file is part of the GNU C Library.
Contributed by Jakub Jelinek <jj@ultra.linux.cz> Contributed by Jakub Jelinek <jj@ultra.linux.cz>
Partly based on double-precision code Partly based on double-precision code
@ -73,7 +73,7 @@ static const long double C[] = {
/* Smallest integer x for which e^x overflows. */ /* Smallest integer x for which e^x overflows. */
#define himark C[0] #define himark C[0]
11356.523406294143949491931077970765L, 11356.523406294143949491931077970765L,
/* Largest integer x for which e^x underflows. */ /* Largest integer x for which e^x underflows. */
#define lomark C[1] #define lomark C[1]
-11433.4627433362978788372438434526231L, -11433.4627433362978788372438434526231L,
@ -247,3 +247,4 @@ __ieee754_expl (long double x)
/* Return x, if x is a NaN or Inf; or overflow, otherwise. */ /* Return x, if x is a NaN or Inf; or overflow, otherwise. */
return TWO16383*x; return TWO16383*x;
} }
strong_alias (__ieee754_expl, __expl_finite)

3
sysdeps/ieee754/ldbl-128ibm/e_expl.c
View File

@ -1,5 +1,5 @@
/* Quad-precision floating point e^x. /* Quad-precision floating point e^x.
Copyright (C) 1999,2004,2006, 2008 Free Software Foundation, Inc. Copyright (C) 1999,2004,2006, 2008, 2011 Free Software Foundation, Inc.
This file is part of the GNU C Library. This file is part of the GNU C Library.
Contributed by Jakub Jelinek <jj@ultra.linux.cz> Contributed by Jakub Jelinek <jj@ultra.linux.cz>
Partly based on double-precision code Partly based on double-precision code
@ -255,3 +255,4 @@ __ieee754_expl (long double x)
/* Return x, if x is a NaN or Inf; or overflow, otherwise. */ /* Return x, if x is a NaN or Inf; or overflow, otherwise. */
return TWO1023*x; return TWO1023*x;
} }
strong_alias (__ieee754_expl, __expl_finite)

79
sysdeps/ieee754/ldbl-96/w_expl.c
View File

@ -1,61 +1,48 @@
/* w_expl.c -- long double version of w_exp.c. /* Copyright (C) 2011 Free Software Foundation, Inc.
* Conversion to long double by Ulrich Drepper, This file is part of the GNU C Library.
* Cygnus Support, drepper@cygnus.com. Contributed by Ulrich Drepper <drepper@gmail.com>, 2011.
*/
/* The GNU C Library is free software; you can redistribute it and/or
* ==================================================== modify it under the terms of the GNU Lesser General Public
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
* Developed at SunPro, a Sun Microsystems, Inc. business.
* Permission to use, copy, modify, and distribute this
* software is freely granted, provided that this notice
* is preserved.
* ====================================================
*/
#if defined(LIBM_SCCS) && !defined(lint) The GNU C Library is distributed in the hope that it will be useful,
static char rcsid[] = "$NetBSD: $"; but WITHOUT ANY WARRANTY; without even the implied warranty of
#endif 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
* wrapper expl(x) License along with the GNU C Library; if not, write to the Free
*/ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include "math.h" #include <math.h>
#include "math_private.h" #include <math_private.h>
#ifdef __STDC__
static const long double static const long double
#else
static long double
#endif
o_threshold= 1.135652340629414394949193107797076489134e4, o_threshold= 1.135652340629414394949193107797076489134e4,
/* 0x400C, 0xB17217F7, 0xD1CF79AC */ /* 0x400C, 0xB17217F7, 0xD1CF79AC */
u_threshold= -1.140019167866942050398521670162263001513e4; u_threshold= -1.140019167866942050398521670162263001513e4;
/* 0x400C, 0xB220C447, 0x69C201E8 */ /* 0x400C, 0xB220C447, 0x69C201E8 */
#ifdef __STDC__
long double __expl(long double x) /* wrapper exp */ /* wrapper expl */
#else long double
long double __expl(x) /* wrapper exp */ __expl (long double x)
long double x;
#endif
{ {
#ifdef _IEEE_LIBM if (__builtin_expect (x > o_threshold, 0))
return __ieee754_expl(x); {
#else if (_LIB_VERSION != _IEEE_)
long double z; return __kernel_standard (x, x, 206);
z = __ieee754_expl(x); }
if(_LIB_VERSION == _IEEE_) return z; else if (__builtin_expect (x < u_threshold, 0))
if(__finitel(x)) { {
if(x>o_threshold) if (_LIB_VERSION != _IEEE_)
return __kernel_standard(x,x,206); /* exp overflow */ return __kernel_standard (x, x, 207);
else if(x<u_threshold) }
return __kernel_standard(x,x,207); /* exp underflow */
} return __ieee754_expl (x);
return z;
#endif
} }
hidden_def (__expl) hidden_def (__expl)
weak_alias (__expl, expl) weak_alias (__expl, expl)