From daa891c0e8900d0f03c16fde9c09f8d6f9a26a7d Mon Sep 17 00:00:00 2001 From: Ulrich Drepper Date: Tue, 10 Jan 2012 22:26:22 -0500 Subject: [PATCH] Optimize s_modf --- ChangeLog | 6 ++ NEWS | 5 +- sysdeps/ieee754/dbl-64/s_modf.c | 23 ++------ sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c | 65 +++++++++++++++++++++ sysdeps/ieee754/flt-32/s_modff.c | 20 ++----- sysdeps/ieee754/ldbl-96/s_modfl.c | 20 ++----- 6 files changed, 88 insertions(+), 51 deletions(-) create mode 100644 sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c diff --git a/ChangeLog b/ChangeLog index c25ea20c7e..9ebbdf9902 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,5 +1,11 @@ 2012-01-10 Ulrich Drepper + * sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c: New file. + + * sysdeps/ieee754/dbl-64/s_modf.c: Add branch prediction. + * sysdeps/ieee754/flt-32/s_modff.c: Likewise. + * sysdeps/ieee754/ldbl-96/s_modfl.c: Likewise. + * math/bits/mathcalls.h: Add const attribute to fmin and fmax. * sysdeps/ieee754/dbl-64/s_scalbn.c: Add branch prediction. diff --git a/NEWS b/NEWS index be6a37a65d..5e062afdaa 100644 --- a/NEWS +++ b/NEWS @@ -1,4 +1,4 @@ -GNU C Library NEWS -- history of user-visible changes. 2012-1-8 +GNU C Library NEWS -- history of user-visible changes. 2012-1-10 Copyright (C) 1992-2009, 2010, 2011, 2012 Free Software Foundation, Inc. See the end for copying conditions. @@ -39,6 +39,9 @@ Version 2.16 * Checking versions of poll, ppoll added. Implemented by Ulrich Drepper. + +* More generic and 64-bit performance optimizations to math functions. + Implemented by Ulrich Drepper. Version 2.15 diff --git a/sysdeps/ieee754/dbl-64/s_modf.c b/sysdeps/ieee754/dbl-64/s_modf.c index 7851f675a4..a884de5d5f 100644 --- a/sysdeps/ieee754/dbl-64/s_modf.c +++ b/sysdeps/ieee754/dbl-64/s_modf.c @@ -1,4 +1,3 @@ -/* @(#)s_modf.c 5.1 93/09/24 */ /* * ==================================================== * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. @@ -10,10 +9,6 @@ * ==================================================== */ -#if defined(LIBM_SCCS) && !defined(lint) -static char rcsid[] = "$NetBSD: s_modf.c,v 1.8 1995/05/10 20:47:55 jtc Exp $"; -#endif - /* * modf(double x, double *iptr) * return fraction part of x, and return x's integral part in *iptr. @@ -27,18 +22,10 @@ static char rcsid[] = "$NetBSD: s_modf.c,v 1.8 1995/05/10 20:47:55 jtc Exp $"; #include "math.h" #include "math_private.h" -#ifdef __STDC__ static const double one = 1.0; -#else -static double one = 1.0; -#endif -#ifdef __STDC__ - double __modf(double x, double *iptr) -#else - double __modf(x, iptr) - double x,*iptr; -#endif +double +__modf(double x, double *iptr) { int32_t i0,i1,j0; u_int32_t i; @@ -46,7 +33,7 @@ static double one = 1.0; j0 = ((i0>>20)&0x7ff)-0x3ff; /* exponent of x */ if(j0<20) { /* integer part in high x */ if(j0<0) { /* |x|<1 */ - INSERT_WORDS(*iptr,i0&0x80000000,0); /* *iptr = +-0 */ + INSERT_WORDS(*iptr,i0&0x80000000,0); /* *iptr = +-0 */ return x; } else { i = (0x000fffff)>>j0; @@ -59,7 +46,7 @@ static double one = 1.0; return x - *iptr; } } - } else if (j0>51) { /* no fraction part */ + } else if (__builtin_expect(j0>51, 0)) { /* no fraction part */ *iptr = x*one; /* We must handle NaNs separately. */ if (j0 == 0x400 && ((i0 & 0xfffff) | i1)) @@ -73,7 +60,7 @@ static double one = 1.0; INSERT_WORDS(x,i0&0x80000000,0); /* return +-0 */ return x; } else { - INSERT_WORDS(*iptr,i0,i1&(~i)); + INSERT_WORDS(*iptr,i0,i1&(~i)); return x - *iptr; } } diff --git a/sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c b/sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c new file mode 100644 index 0000000000..0cba548abd --- /dev/null +++ b/sysdeps/ieee754/dbl-64/wordsize-64/s_modf.c @@ -0,0 +1,65 @@ +/* Rewritten for 64-bit machines by Ulrich Drepper . */ +/* + * ==================================================== + * 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. + * ==================================================== + */ + +/* + * modf(double x, double *iptr) + * return fraction part of x, and return x's integral part in *iptr. + * Method: + * Bit twiddling. + * + * Exception: + * No exception. + */ + +#include "math.h" +#include "math_private.h" + +static const double one = 1.0; + +double +__modf(double x, double *iptr) +{ + int64_t i0; + int32_t j0; + EXTRACT_WORDS64(i0,x); + j0 = ((i0>>52)&0x7ff)-0x3ff; /* exponent of x */ + if(j0<52) { /* integer part in x */ + if(j0<0) { /* |x|<1 */ + /* *iptr = +-0 */ + INSERT_WORDS64(*iptr,i0&UINT64_C(0x8000000000000000)); + return x; + } else { + uint64_t i = UINT64_C(0x000fffffffffffff)>>j0; + if((i0&i)==0) { /* x is integral */ + *iptr = x; + /* return +-0 */ + INSERT_WORDS64(x,i0&UINT64_C(0x8000000000000000)); + return x; + } else { + INSERT_WORDS64(*iptr,i0&(~i)); + return x - *iptr; + } + } + } else { /* no fraction part */ + *iptr = x*one; + /* We must handle NaNs separately. */ + if (j0 == 0x400 && (i0 & UINT64_C(0xfffffffffffff))) + return x*one; + INSERT_WORDS64(x,i0&UINT64_C(0x8000000000000000)); /* return +-0 */ + return x; + } +} +weak_alias (__modf, modf) +#ifdef NO_LONG_DOUBLE +strong_alias (__modf, __modfl) +weak_alias (__modf, modfl) +#endif diff --git a/sysdeps/ieee754/flt-32/s_modff.c b/sysdeps/ieee754/flt-32/s_modff.c index e6c22b2add..741b1c9b16 100644 --- a/sysdeps/ieee754/flt-32/s_modff.c +++ b/sysdeps/ieee754/flt-32/s_modff.c @@ -13,33 +13,21 @@ * ==================================================== */ -#if defined(LIBM_SCCS) && !defined(lint) -static char rcsid[] = "$NetBSD: s_modff.c,v 1.4 1995/05/10 20:47:56 jtc Exp $"; -#endif - #include "math.h" #include "math_private.h" -#ifdef __STDC__ static const float one = 1.0; -#else -static float one = 1.0; -#endif -#ifdef __STDC__ - float __modff(float x, float *iptr) -#else - float __modff(x, iptr) - float x,*iptr; -#endif +float +__modff(float x, float *iptr) { int32_t i0,j0; u_int32_t i; GET_FLOAT_WORD(i0,x); j0 = ((i0>>23)&0xff)-0x7f; /* exponent of x */ - if(j0<23) { /* integer part in x */ + if(__builtin_expect(j0<23, 1)) { /* integer part in x */ if(j0<0) { /* |x|<1 */ - SET_FLOAT_WORD(*iptr,i0&0x80000000); /* *iptr = +-0 */ + SET_FLOAT_WORD(*iptr,i0&0x80000000); /* *iptr = +-0 */ return x; } else { i = (0x007fffff)>>j0; diff --git a/sysdeps/ieee754/ldbl-96/s_modfl.c b/sysdeps/ieee754/ldbl-96/s_modfl.c index 99c0752efa..c7659119f3 100644 --- a/sysdeps/ieee754/ldbl-96/s_modfl.c +++ b/sysdeps/ieee754/ldbl-96/s_modfl.c @@ -14,10 +14,6 @@ * ==================================================== */ -#if defined(LIBM_SCCS) && !defined(lint) -static char rcsid[] = "$NetBSD: $"; -#endif - /* * modfl(long double x, long double *iptr) * return fraction part of x, and return x's integral part in *iptr. @@ -31,18 +27,10 @@ static char rcsid[] = "$NetBSD: $"; #include "math.h" #include "math_private.h" -#ifdef __STDC__ static const long double one = 1.0; -#else -static long double one = 1.0; -#endif -#ifdef __STDC__ - long double __modfl(long double x, long double *iptr) -#else - long double __modfl(x, iptr) - long double x,*iptr; -#endif +long double +__modfl(long double x, long double *iptr) { int32_t i0,i1,j0; u_int32_t i,se; @@ -50,7 +38,7 @@ static long double one = 1.0; j0 = (se&0x7fff)-0x3fff; /* exponent of x */ if(j0<32) { /* integer part in high x */ if(j0<0) { /* |x|<1 */ - SET_LDOUBLE_WORDS(*iptr,se&0x8000,0,0); /* *iptr = +-0 */ + SET_LDOUBLE_WORDS(*iptr,se&0x8000,0,0); /* *iptr = +-0 */ return x; } else { i = (0x7fffffff)>>j0; @@ -63,7 +51,7 @@ static long double one = 1.0; return x - *iptr; } } - } else if (j0>63) { /* no fraction part */ + } else if (__builtin_expect(j0>63, 0)) { /* no fraction part */ *iptr = x*one; /* We must handle NaNs separately. */ if (j0 == 0x4000 && ((i0 & 0x7fffffff) | i1))