diff --git a/lib/msun/Makefile b/lib/msun/Makefile index 0be2813bb933..24d740d761d5 100644 --- a/lib/msun/Makefile +++ b/lib/msun/Makefile @@ -1,278 +1,278 @@ # @(#)Makefile 5.1beta 93/09/24 # $FreeBSD$ # # ==================================================== # 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. # ==================================================== # # PACKAGE= clibs ARCH_SUBDIR= ${MACHINE_CPUARCH:S/i386/i387/} .include "${ARCH_SUBDIR}/Makefile.inc" .PATH: ${.CURDIR}/${ARCH_SUBDIR} .if ${MACHINE_CPUARCH} == "i386" || ${MACHINE_CPUARCH} == "amd64" .PATH: ${.CURDIR}/x86 CFLAGS+= -I${.CURDIR}/x86 .endif # long double format .if ${LDBL_PREC} == 64 .PATH: ${.CURDIR}/ld80 CFLAGS+= -I${.CURDIR}/ld80 .elif ${LDBL_PREC} == 113 .PATH: ${.CURDIR}/ld128 CFLAGS+= -I${.CURDIR}/ld128 .endif CFLAGS+= -I${.CURDIR}/${ARCH_SUBDIR} .include .if ${COMPILER_TYPE} == "clang" && ${COMPILER_VERSION} >= 100000 && \ (${MACHINE_CPUARCH} == "amd64" || ${MACHINE_CPUARCH} == "i386") # When using clang with x86_64 CPUs that support AVX, some floating point # transformations may raise exceptions that would not have been raised by the # original code. To avoid this, use the -fp-exception-behavior=maytrap flag, # introduced in clang 10.0.0. # See also: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=254911 CFLAGS+= -ffp-exception-behavior=maytrap .endif # Tell the compiler we don't set errno in any of the math functions. This # stops gcc from trying to generate a function call to set errno when using # a builtin in the implementation CFLAGS+= -fno-math-errno .PATH: ${.CURDIR}/bsdsrc .PATH: ${.CURDIR}/src .PATH: ${.CURDIR}/man LIB= m SHLIBDIR?= /lib SHLIB_MAJOR= 5 WARNS?= 1 IGNORE_PRAGMA= -COMMON_SRCS= b_exp.c b_log.c b_tgamma.c \ +COMMON_SRCS= b_tgamma.c \ e_acos.c e_acosf.c e_acosh.c e_acoshf.c e_asin.c e_asinf.c \ e_atan2.c e_atan2f.c e_atanh.c e_atanhf.c e_cosh.c e_coshf.c e_exp.c \ e_expf.c e_fmod.c e_fmodf.c e_gamma.c e_gamma_r.c e_gammaf.c \ e_gammaf_r.c e_hypot.c e_hypotf.c e_j0.c e_j0f.c e_j1.c e_j1f.c \ e_jn.c e_jnf.c e_lgamma.c e_lgamma_r.c e_lgammaf.c e_lgammaf_r.c \ e_log.c e_log10.c e_log10f.c e_log2.c e_log2f.c e_logf.c \ e_pow.c e_powf.c e_rem_pio2.c \ e_rem_pio2f.c e_remainder.c e_remainderf.c e_scalb.c e_scalbf.c \ e_sinh.c e_sinhf.c e_sqrt.c e_sqrtf.c fenv.c \ imprecise.c \ k_cos.c k_cosf.c k_exp.c k_expf.c k_rem_pio2.c k_sin.c k_sinf.c \ k_tan.c k_tanf.c \ s_asinh.c s_asinhf.c s_atan.c s_atanf.c s_carg.c s_cargf.c s_cargl.c \ s_cbrt.c s_cbrtf.c s_ceil.c s_ceilf.c s_clog.c s_clogf.c \ s_copysign.c s_copysignf.c s_cos.c s_cosf.c \ s_csqrt.c s_csqrtf.c s_erf.c s_erff.c \ s_exp2.c s_exp2f.c s_expm1.c s_expm1f.c s_fabsf.c s_fdim.c \ s_finite.c s_finitef.c \ s_floor.c s_floorf.c s_fma.c s_fmaf.c \ s_fmax.c s_fmaxf.c s_fmin.c \ s_fminf.c s_frexp.c s_frexpf.c s_ilogb.c s_ilogbf.c \ s_ilogbl.c s_isfinite.c s_isnan.c s_isnormal.c \ s_llrint.c s_llrintf.c s_llround.c s_llroundf.c s_llroundl.c \ s_log1p.c s_log1pf.c s_logb.c s_logbf.c s_lrint.c s_lrintf.c \ s_lround.c s_lroundf.c s_lroundl.c s_modff.c \ s_nan.c s_nearbyint.c s_nextafter.c s_nextafterf.c \ s_nexttowardf.c s_remquo.c s_remquof.c \ s_rint.c s_rintf.c s_round.c s_roundf.c \ s_scalbln.c s_scalbn.c s_scalbnf.c s_signbit.c \ s_signgam.c s_significand.c s_significandf.c s_sin.c \ s_sincos.c s_sincosf.c s_sinf.c \ s_tan.c s_tanf.c s_tanh.c s_tanhf.c s_tgammaf.c s_trunc.c s_truncf.c \ w_cabs.c w_cabsf.c w_drem.c w_dremf.c # IEEE-754 2008 and ISO/IEC TS 18661-4 half-cycle trignometric functions COMMON_SRCS+= s_cospi.c s_cospif.c \ s_sinpi.c s_sinpif.c \ s_tanpi.c s_tanpif.c # Location of fpmath.h and _fpmath.h .if exists(${LIBCSRCDIR}/${MACHINE_ARCH}) LIBC_ARCH=${MACHINE_ARCH} .else LIBC_ARCH=${MACHINE_CPUARCH} .endif CFLAGS+= -I${.CURDIR}/src -I${LIBCSRCDIR}/include \ -I${LIBCSRCDIR}/${LIBC_ARCH} SYM_MAPS+= ${.CURDIR}/Symbol.map VERSION_DEF= ${LIBCSRCDIR}/Versions.def SYMBOL_MAPS= ${SYM_MAPS} # C99 long double functions COMMON_SRCS+= s_copysignl.c s_fabsl.c s_llrintl.c s_lrintl.c s_modfl.c .if ${LDBL_PREC} != 53 # If long double != double use these; otherwise, we alias the double versions. COMMON_SRCS+= catrigl.c \ e_acoshl.c e_acosl.c e_asinl.c e_atan2l.c e_atanhl.c \ e_coshl.c e_fmodl.c e_hypotl.c \ e_lgammal.c e_lgammal_r.c e_powl.c \ e_remainderl.c e_sinhl.c e_sqrtl.c \ invtrig.c k_cosl.c k_sinl.c k_tanl.c \ s_asinhl.c s_atanl.c s_cbrtl.c s_ceill.c s_cexpl.c \ s_clogl.c s_cosl.c s_cospil.c s_cprojl.c \ s_csqrtl.c s_erfl.c s_exp2l.c s_expl.c s_floorl.c s_fmal.c \ s_fmaxl.c s_fminl.c s_frexpl.c s_logbl.c s_logl.c s_nanl.c \ s_nextafterl.c s_nexttoward.c s_remquol.c s_rintl.c s_roundl.c \ s_scalbnl.c s_sinl.c s_sincosl.c s_sinpil.c \ s_tanhl.c s_tanl.c s_tanpil.c s_truncl.c w_cabsl.c # Work around this warning from gcc: # lib/msun/ld80/e_powl.c:275:1: error: floating constant exceeds range of # 'long double' [-Werror=overflow] # if( y >= LDBL_MAX ) # See also: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=130067 .if ${COMPILER_TYPE} == "gcc" CFLAGS.e_powl.c+= -Wno-error=overflow .endif .endif # C99 complex functions COMMON_SRCS+= catrig.c catrigf.c \ s_ccosh.c s_ccoshf.c s_cexp.c s_cexpf.c \ s_cimag.c s_cimagf.c s_cimagl.c \ s_conj.c s_conjf.c s_conjl.c \ s_cpow.c s_cpowf.c s_cpowl.c \ s_cproj.c s_cprojf.c s_creal.c s_crealf.c s_creall.c \ s_csinh.c s_csinhf.c s_ctanh.c s_ctanhf.c # FreeBSD's C library supplies these functions: #COMMON_SRCS+= s_fabs.c s_frexp.c s_isnan.c s_ldexp.c s_modf.c # Exclude the generic versions of what we provide in the MD area. .if defined(ARCH_SRCS) && "${MK_MACHDEP_OPTIMIZATIONS}" != no .for i in ${ARCH_SRCS} COMMON_SRCS:= ${COMMON_SRCS:N${i:R}.c} .endfor .endif SRCS= ${COMMON_SRCS} ${ARCH_SRCS} INCS+= fenv.h math.h MAN= acos.3 acosh.3 asin.3 asinh.3 atan.3 atan2.3 atanh.3 \ ceil.3 cacos.3 ccos.3 ccosh.3 cexp.3 \ cimag.3 clog.3 copysign.3 cos.3 cosh.3 cospi.3 \ cpow.3 csqrt.3 erf.3 \ exp.3 fabs.3 fdim.3 \ feclearexcept.3 feenableexcept.3 fegetenv.3 \ fegetround.3 fenv.3 floor.3 \ fma.3 fmax.3 fmod.3 hypot.3 ieee.3 ieee_test.3 ilogb.3 j0.3 \ lgamma.3 log.3 lrint.3 lround.3 math.3 nan.3 \ nextafter.3 remainder.3 rint.3 \ round.3 scalbn.3 signbit.3 sin.3 sincos.3 \ sinh.3 sinpi.3 sqrt.3 tan.3 tanh.3 tanpi.3 trunc.3 \ complex.3 MLINKS+=acos.3 acosf.3 acos.3 acosl.3 MLINKS+=acosh.3 acoshf.3 acosh.3 acoshl.3 MLINKS+=asin.3 asinf.3 asin.3 asinl.3 MLINKS+=asinh.3 asinhf.3 asinh.3 asinhl.3 MLINKS+=atan.3 atanf.3 atan.3 atanl.3 MLINKS+=atanh.3 atanhf.3 atanh.3 atanhl.3 MLINKS+=atan2.3 atan2f.3 atan2.3 atan2l.3 \ atan2.3 carg.3 atan2.3 cargf.3 atan2.3 cargl.3 MLINKS+=cacos.3 cacosf.3 cacos.3 cacosl.3 \ cacos.3 cacosh.3 cacos.3 cacoshf.3 cacos.3 cacoshl.3 \ cacos.3 casin.3 cacos.3 casinf.3 cacos.3 casinl.3 \ cacos.3 casinh.3 cacos.3 casinhf.3 cacos.3 casinhl.3 \ cacos.3 catan.3 cacos.3 catanf.3 cacos.3 catanl.3 \ cacos.3 catanh.3 cacos.3 catanhf.3 cacos.3 catanhl.3 MLINKS+=ccos.3 ccosf.3 ccos.3 csin.3 ccos.3 csinf.3 ccos.3 ctan.3 ccos.3 ctanf.3 MLINKS+=ccosh.3 ccoshf.3 ccosh.3 csinh.3 ccosh.3 csinhf.3 \ ccosh.3 ctanh.3 ccosh.3 ctanhf.3 MLINKS+=ceil.3 ceilf.3 ceil.3 ceill.3 MLINKS+=cexp.3 cexpf.3 cexp.3 cexpl.3 MLINKS+=cimag.3 cimagf.3 cimag.3 cimagl.3 \ cimag.3 conj.3 cimag.3 conjf.3 cimag.3 conjl.3 \ cimag.3 cproj.3 cimag.3 cprojf.3 cimag.3 cprojl.3 \ cimag.3 creal.3 cimag.3 crealf.3 cimag.3 creall.3 MLINKS+=clog.3 clogf.3 clog.3 clogl.3 MLINKS+=copysign.3 copysignf.3 copysign.3 copysignl.3 MLINKS+=cos.3 cosf.3 cos.3 cosl.3 MLINKS+=cosh.3 coshf.3 cosh.3 coshl.3 MLINKS+=cospi.3 cospif.3 cospi.3 cospil.3 MLINKS+=cpow.3 cpowf.3 cpow.3 cpowl.3 MLINKS+=csqrt.3 csqrtf.3 csqrt.3 csqrtl.3 MLINKS+=erf.3 erfc.3 erf.3 erff.3 erf.3 erfcf.3 erf.3 erfl.3 erf.3 erfcl.3 MLINKS+=exp.3 expm1.3 exp.3 expm1f.3 exp.3 expm1l.3 exp.3 pow.3 exp.3 powf.3 \ exp.3 powl.3 exp.3 exp2.3 exp.3 exp2f.3 exp.3 exp2l.3 exp.3 expf.3 \ exp.3 expl.3 MLINKS+=fabs.3 fabsf.3 fabs.3 fabsl.3 MLINKS+=fdim.3 fdimf.3 fdim.3 fdiml.3 MLINKS+=feclearexcept.3 fegetexceptflag.3 feclearexcept.3 feraiseexcept.3 \ feclearexcept.3 fesetexceptflag.3 feclearexcept.3 fetestexcept.3 MLINKS+=feenableexcept.3 fedisableexcept.3 feenableexcept.3 fegetexcept.3 MLINKS+=fegetenv.3 feholdexcept.3 fegetenv.3 fesetenv.3 \ fegetenv.3 feupdateenv.3 MLINKS+=fegetround.3 fesetround.3 MLINKS+=floor.3 floorf.3 floor.3 floorl.3 MLINKS+=fma.3 fmaf.3 fma.3 fmal.3 MLINKS+=fmax.3 fmaxf.3 fmax.3 fmaxl.3 \ fmax.3 fmin.3 fmax.3 fminf.3 fmax.3 fminl.3 MLINKS+=fmod.3 fmodf.3 fmod.3 fmodl.3 MLINKS+=hypot.3 cabs.3 hypot.3 cabsf.3 hypot.3 cabsl.3 \ hypot.3 hypotf.3 hypot.3 hypotl.3 MLINKS+=ieee_test.3 scalb.3 ieee_test.3 scalbf.3 MLINKS+=ieee_test.3 significand.3 ieee_test.3 significandf.3 MLINKS+=ilogb.3 ilogbf.3 ilogb.3 ilogbl.3 \ ilogb.3 logb.3 ilogb.3 logbf.3 ilogb.3 logbl.3 MLINKS+=j0.3 j1.3 j0.3 jn.3 j0.3 y0.3 j0.3 y1.3 j0.3 y1f.3 j0.3 yn.3 MLINKS+=j0.3 j0f.3 j0.3 j1f.3 j0.3 jnf.3 j0.3 y0f.3 j0.3 ynf.3 MLINKS+=lgamma.3 gamma.3 lgamma.3 gammaf.3 \ lgamma.3 lgammaf.3 lgamma.3 lgammal.3 \ lgamma.3 tgamma.3 lgamma.3 tgammaf.3 MLINKS+=log.3 log10.3 log.3 log10f.3 log.3 log10l.3 \ log.3 log1p.3 log.3 log1pf.3 log.3 log1pl.3 \ log.3 logf.3 log.3 logl.3 \ log.3 log2.3 log.3 log2f.3 log.3 log2l.3 MLINKS+=lrint.3 llrint.3 lrint.3 llrintf.3 lrint.3 llrintl.3 \ lrint.3 lrintf.3 lrint.3 lrintl.3 MLINKS+=lround.3 llround.3 lround.3 llroundf.3 lround.3 llroundl.3 \ lround.3 lroundf.3 lround.3 lroundl.3 MLINKS+=nan.3 nanf.3 nan.3 nanl.3 MLINKS+=nextafter.3 nextafterf.3 nextafter.3 nextafterl.3 MLINKS+=nextafter.3 nexttoward.3 nextafter.3 nexttowardf.3 MLINKS+=nextafter.3 nexttowardl.3 MLINKS+=remainder.3 remainderf.3 remainder.3 remainderl.3 \ remainder.3 remquo.3 remainder.3 remquof.3 remainder.3 remquol.3 MLINKS+=rint.3 rintf.3 rint.3 rintl.3 \ rint.3 nearbyint.3 rint.3 nearbyintf.3 rint.3 nearbyintl.3 MLINKS+=round.3 roundf.3 round.3 roundl.3 MLINKS+=scalbn.3 scalbln.3 scalbn.3 scalblnf.3 scalbn.3 scalblnl.3 MLINKS+=scalbn.3 scalbnf.3 scalbn.3 scalbnl.3 MLINKS+=sin.3 sinf.3 sin.3 sinl.3 MLINKS+=sincos.3 sincosf.3 sin.3 sincosl.3 MLINKS+=sinh.3 sinhf.3 sinh.3 sinhl.3 MLINKS+=sinpi.3 sinpif.3 sinpi.3 sinpil.3 MLINKS+=sqrt.3 cbrt.3 sqrt.3 cbrtf.3 sqrt.3 cbrtl.3 sqrt.3 sqrtf.3 \ sqrt.3 sqrtl.3 MLINKS+=tan.3 tanf.3 tan.3 tanl.3 MLINKS+=tanh.3 tanhf.3 tanh.3 tanhl.3 MLINKS+=tanpi.3 tanpif.3 tanpi.3 tanpil.3 MLINKS+=trunc.3 truncf.3 trunc.3 truncl.3 .include HAS_TESTS= SUBDIR.${MK_TESTS}+= tests .include .if ${COMPILER_TYPE} == "clang" && ${COMPILER_VERSION} >= 120000 # Silence '#pragma FENV_ACCESS' is not supported on this target - ignored CWARNFLAGS+= -Wno-error=ignored-pragmas .endif .include diff --git a/lib/msun/bsdsrc/b_exp.c b/lib/msun/bsdsrc/b_exp.c index 89a290529da9..c667293ed7c6 100644 --- a/lib/msun/bsdsrc/b_exp.c +++ b/lib/msun/bsdsrc/b_exp.c @@ -1,173 +1,126 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1985, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* @(#)exp.c 8.1 (Berkeley) 6/4/93 */ #include __FBSDID("$FreeBSD$"); - /* EXP(X) * RETURN THE EXPONENTIAL OF X * DOUBLE PRECISION (IEEE 53 bits, VAX D FORMAT 56 BITS) * CODED IN C BY K.C. NG, 1/19/85; * REVISED BY K.C. NG on 2/6/85, 2/15/85, 3/7/85, 3/24/85, 4/16/85, 6/14/86. * * Required system supported functions: - * scalb(x,n) + * ldexp(x,n) * copysign(x,y) - * finite(x) + * isfinite(x) * * Method: * 1. Argument Reduction: given the input x, find r and integer k such * that - * x = k*ln2 + r, |r| <= 0.5*ln2 . + * x = k*ln2 + r, |r| <= 0.5*ln2. * r will be represented as r := z+c for better accuracy. * * 2. Compute exp(r) by * * exp(r) = 1 + r + r*R1/(2-R1), * where * R1 = x - x^2*(p1+x^2*(p2+x^2*(p3+x^2*(p4+p5*x^2)))). * * 3. exp(x) = 2^k * exp(r) . * * Special cases: * exp(INF) is INF, exp(NaN) is NaN; * exp(-INF)= 0; * for finite argument, only exp(0)=1 is exact. * * Accuracy: * exp(x) returns the exponential of x nearly rounded. In a test run * with 1,156,000 random arguments on a VAX, the maximum observed * error was 0.869 ulps (units in the last place). */ - -#include "mathimpl.h" - -static const double p1 = 0x1.555555555553ep-3; -static const double p2 = -0x1.6c16c16bebd93p-9; -static const double p3 = 0x1.1566aaf25de2cp-14; -static const double p4 = -0x1.bbd41c5d26bf1p-20; -static const double p5 = 0x1.6376972bea4d0p-25; -static const double ln2hi = 0x1.62e42fee00000p-1; -static const double ln2lo = 0x1.a39ef35793c76p-33; -static const double lnhuge = 0x1.6602b15b7ecf2p9; -static const double lntiny = -0x1.77af8ebeae354p9; -static const double invln2 = 0x1.71547652b82fep0; - -#if 0 -double exp(x) -double x; -{ - double z,hi,lo,c; - int k; - -#if !defined(vax)&&!defined(tahoe) - if(x!=x) return(x); /* x is NaN */ -#endif /* !defined(vax)&&!defined(tahoe) */ - if( x <= lnhuge ) { - if( x >= lntiny ) { - - /* argument reduction : x --> x - k*ln2 */ - - k=invln2*x+copysign(0.5,x); /* k=NINT(x/ln2) */ - - /* express x-k*ln2 as hi-lo and let x=hi-lo rounded */ - - hi=x-k*ln2hi; - x=hi-(lo=k*ln2lo); - - /* return 2^k*[1+x+x*c/(2+c)] */ - z=x*x; - c= x - z*(p1+z*(p2+z*(p3+z*(p4+z*p5)))); - return scalb(1.0+(hi-(lo-(x*c)/(2.0-c))),k); - - } - /* end of x > lntiny */ - - else - /* exp(-big#) underflows to zero */ - if(finite(x)) return(scalb(1.0,-5000)); - - /* exp(-INF) is zero */ - else return(0.0); - } - /* end of x < lnhuge */ - - else - /* exp(INF) is INF, exp(+big#) overflows to INF */ - return( finite(x) ? scalb(1.0,5000) : x); -} -#endif +static const double + p1 = 1.6666666666666660e-01, /* 0x3fc55555, 0x55555553 */ + p2 = -2.7777777777564776e-03, /* 0xbf66c16c, 0x16c0ac3c */ + p3 = 6.6137564717940088e-05, /* 0x3f11566a, 0xb5c2ba0d */ + p4 = -1.6534060280704225e-06, /* 0xbebbbd53, 0x273e8fb7 */ + p5 = 4.1437773411069054e-08; /* 0x3e663f2a, 0x09c94b6c */ + +static const double + ln2hi = 0x1.62e42fee00000p-1, /* High 32 bits round-down. */ + ln2lo = 0x1.a39ef35793c76p-33; /* Next 53 bits round-to-nearst. */ + +static const double + lnhuge = 0x1.6602b15b7ecf2p9, /* (DBL_MAX_EXP + 9) * log(2.) */ + lntiny = -0x1.77af8ebeae354p9, /* (DBL_MIN_EXP - 53 - 10) * log(2.) */ + invln2 = 0x1.71547652b82fep0; /* 1 / log(2.) */ /* returns exp(r = x + c) for |c| < |x| with no overlap. */ -double __exp__D(x, c) -double x, c; +static double +__exp__D(double x, double c) { - double z,hi,lo; + double hi, lo, z; int k; - if (x != x) /* x is NaN */ + if (x != x) /* x is NaN. */ return(x); - if ( x <= lnhuge ) { - if ( x >= lntiny ) { - /* argument reduction : x --> x - k*ln2 */ - z = invln2*x; - k = z + copysign(.5, x); - - /* express (x+c)-k*ln2 as hi-lo and let x=hi-lo rounded */ - - hi=(x-k*ln2hi); /* Exact. */ - x= hi - (lo = k*ln2lo-c); - /* return 2^k*[1+x+x*c/(2+c)] */ - z=x*x; - c= x - z*(p1+z*(p2+z*(p3+z*(p4+z*p5)))); - c = (x*c)/(2.0-c); - - return scalb(1.+(hi-(lo - c)), k); + if (x <= lnhuge) { + if (x >= lntiny) { + /* argument reduction: x --> x - k*ln2 */ + z = invln2 * x; + k = z + copysign(0.5, x); + + /* + * Express (x + c) - k * ln2 as hi - lo. + * Let x = hi - lo rounded. + */ + hi = x - k * ln2hi; /* Exact. */ + lo = k * ln2lo - c; + x = hi - lo; + + /* Return 2^k*[1+x+x*c/(2+c)] */ + z = x * x; + c = x - z * (p1 + z * (p2 + z * (p3 + z * (p4 + + z * p5)))); + c = (x * c) / (2 - c); + + return (ldexp(1 + (hi - (lo - c)), k)); + } else { + /* exp(-INF) is 0. exp(-big) underflows to 0. */ + return (isfinite(x) ? ldexp(1., -5000) : 0); } - /* end of x > lntiny */ - - else - /* exp(-big#) underflows to zero */ - if(finite(x)) return(scalb(1.0,-5000)); - - /* exp(-INF) is zero */ - else return(0.0); - } - /* end of x < lnhuge */ - - else + } else /* exp(INF) is INF, exp(+big#) overflows to INF */ - return( finite(x) ? scalb(1.0,5000) : x); + return (isfinite(x) ? ldexp(1., 5000) : x); } diff --git a/lib/msun/bsdsrc/b_log.c b/lib/msun/bsdsrc/b_log.c index c164dfa5014c..9d09ac754706 100644 --- a/lib/msun/bsdsrc/b_log.c +++ b/lib/msun/bsdsrc/b_log.c @@ -1,468 +1,404 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1992, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* @(#)log.c 8.2 (Berkeley) 11/30/93 */ #include __FBSDID("$FreeBSD$"); -#include - -#include "mathimpl.h" - /* Table-driven natural logarithm. * * This code was derived, with minor modifications, from: * Peter Tang, "Table-Driven Implementation of the * Logarithm in IEEE Floating-Point arithmetic." ACM Trans. * Math Software, vol 16. no 4, pp 378-400, Dec 1990). * - * Calculates log(2^m*F*(1+f/F)), |f/j| <= 1/256, + * Calculates log(2^m*F*(1+f/F)), |f/F| <= 1/256, * where F = j/128 for j an integer in [0, 128]. * * log(2^m) = log2_hi*m + log2_tail*m - * since m is an integer, the dominant term is exact. + * The leading term is exact, because m is an integer, * m has at most 10 digits (for subnormal numbers), * and log2_hi has 11 trailing zero bits. * - * log(F) = logF_hi[j] + logF_lo[j] is in tabular form in log_table.h + * log(F) = logF_hi[j] + logF_lo[j] is in table below. * logF_hi[] + 512 is exact. * * log(1+f/F) = 2*f/(2*F + f) + 1/12 * (2*f/(2*F + f))**3 + ... - * the leading term is calculated to extra precision in two + * + * The leading term is calculated to extra precision in two * parts, the larger of which adds exactly to the dominant * m and F terms. + * * There are two cases: - * 1. when m, j are non-zero (m | j), use absolute + * 1. When m and j are non-zero (m | j), use absolute * precision for the leading term. - * 2. when m = j = 0, |1-x| < 1/256, and log(x) ~= (x-1). + * 2. When m = j = 0, |1-x| < 1/256, and log(x) ~= (x-1). * In this case, use a relative precision of 24 bits. * (This is done differently in the original paper) * * Special cases: * 0 return signalling -Inf * neg return signalling NaN * +Inf return +Inf -*/ + */ #define N 128 -/* Table of log(Fj) = logF_head[j] + logF_tail[j], for Fj = 1+j/128. +/* + * Coefficients in the polynomial approximation of log(1+f/F). + * Domain of x is [0,1./256] with 2**(-64.187) precision. + */ +static const double + A1 = 8.3333333333333329e-02, /* 0x3fb55555, 0x55555555 */ + A2 = 1.2499999999943598e-02, /* 0x3f899999, 0x99991a98 */ + A3 = 2.2321527525957776e-03; /* 0x3f624929, 0xe24e70be */ + +/* + * Table of log(Fj) = logF_head[j] + logF_tail[j], for Fj = 1+j/128. * Used for generation of extend precision logarithms. * The constant 35184372088832 is 2^45, so the divide is exact. * It ensures correct reading of logF_head, even for inaccurate * decimal-to-binary conversion routines. (Everybody gets the * right answer for integers less than 2^53.) * Values for log(F) were generated using error < 10^-57 absolute * with the bc -l package. -*/ -static double A1 = .08333333333333178827; -static double A2 = .01250000000377174923; -static double A3 = .002232139987919447809; -static double A4 = .0004348877777076145742; - + */ static double logF_head[N+1] = { 0., .007782140442060381246, .015504186535963526694, .023167059281547608406, .030771658666765233647, .038318864302141264488, .045809536031242714670, .053244514518837604555, .060624621816486978786, .067950661908525944454, .075223421237524235039, .082443669210988446138, .089612158689760690322, .096729626458454731618, .103796793681567578460, .110814366340264314203, .117783035656430001836, .124703478501032805070, .131576357788617315236, .138402322859292326029, .145182009844575077295, .151916042025732167530, .158605030176659056451, .165249572895390883786, .171850256926518341060, .178407657472689606947, .184922338493834104156, .191394852999565046047, .197825743329758552135, .204215541428766300668, .210564769107350002741, .216873938300523150246, .223143551314024080056, .229374101064877322642, .235566071312860003672, .241719936886966024758, .247836163904594286577, .253915209980732470285, .259957524436686071567, .265963548496984003577, .271933715484010463114, .277868451003087102435, .283768173130738432519, .289633292582948342896, .295464212893421063199, .301261330578199704177, .307025035294827830512, .312755710004239517729, .318453731118097493890, .324119468654316733591, .329753286372579168528, .335355541920762334484, .340926586970454081892, .346466767346100823488, .351976423156884266063, .357455888922231679316, .362905493689140712376, .368325561158599157352, .373716409793814818840, .379078352934811846353, .384411698910298582632, .389716751140440464951, .394993808240542421117, .400243164127459749579, .405465108107819105498, .410659924985338875558, .415827895143593195825, .420969294644237379543, .426084395310681429691, .431173464818130014464, .436236766774527495726, .441274560805140936281, .446287102628048160113, .451274644139630254358, .456237433481874177232, .461175715122408291790, .466089729924533457960, .470979715219073113985, .475845904869856894947, .480688529345570714212, .485507815781602403149, .490303988045525329653, .495077266798034543171, .499827869556611403822, .504556010751912253908, .509261901790523552335, .513945751101346104405, .518607764208354637958, .523248143765158602036, .527867089620485785417, .532464798869114019908, .537041465897345915436, .541597282432121573947, .546132437597407260909, .550647117952394182793, .555141507540611200965, .559615787935399566777, .564070138285387656651, .568504735352689749561, .572919753562018740922, .577315365035246941260, .581691739635061821900, .586049045003164792433, .590387446602107957005, .594707107746216934174, .599008189645246602594, .603290851438941899687, .607555250224322662688, .611801541106615331955, .616029877215623855590, .620240409751204424537, .624433288012369303032, .628608659422752680256, .632766669570628437213, .636907462236194987781, .641031179420679109171, .645137961373620782978, .649227946625615004450, .653301272011958644725, .657358072709030238911, .661398482245203922502, .665422632544505177065, .669430653942981734871, .673422675212350441142, .677398823590920073911, .681359224807238206267, .685304003098281100392, .689233281238557538017, .693147180560117703862 }; static double logF_tail[N+1] = { 0., -.00000000000000543229938420049, .00000000000000172745674997061, -.00000000000001323017818229233, -.00000000000001154527628289872, -.00000000000000466529469958300, .00000000000005148849572685810, -.00000000000002532168943117445, -.00000000000005213620639136504, -.00000000000001819506003016881, .00000000000006329065958724544, .00000000000008614512936087814, -.00000000000007355770219435028, .00000000000009638067658552277, .00000000000007598636597194141, .00000000000002579999128306990, -.00000000000004654729747598444, -.00000000000007556920687451336, .00000000000010195735223708472, -.00000000000017319034406422306, -.00000000000007718001336828098, .00000000000010980754099855238, -.00000000000002047235780046195, -.00000000000008372091099235912, .00000000000014088127937111135, .00000000000012869017157588257, .00000000000017788850778198106, .00000000000006440856150696891, .00000000000016132822667240822, -.00000000000007540916511956188, -.00000000000000036507188831790, .00000000000009120937249914984, .00000000000018567570959796010, -.00000000000003149265065191483, -.00000000000009309459495196889, .00000000000017914338601329117, -.00000000000001302979717330866, .00000000000023097385217586939, .00000000000023999540484211737, .00000000000015393776174455408, -.00000000000036870428315837678, .00000000000036920375082080089, -.00000000000009383417223663699, .00000000000009433398189512690, .00000000000041481318704258568, -.00000000000003792316480209314, .00000000000008403156304792424, -.00000000000034262934348285429, .00000000000043712191957429145, -.00000000000010475750058776541, -.00000000000011118671389559323, .00000000000037549577257259853, .00000000000013912841212197565, .00000000000010775743037572640, .00000000000029391859187648000, -.00000000000042790509060060774, .00000000000022774076114039555, .00000000000010849569622967912, -.00000000000023073801945705758, .00000000000015761203773969435, .00000000000003345710269544082, -.00000000000041525158063436123, .00000000000032655698896907146, -.00000000000044704265010452446, .00000000000034527647952039772, -.00000000000007048962392109746, .00000000000011776978751369214, -.00000000000010774341461609578, .00000000000021863343293215910, .00000000000024132639491333131, .00000000000039057462209830700, -.00000000000026570679203560751, .00000000000037135141919592021, -.00000000000017166921336082431, -.00000000000028658285157914353, -.00000000000023812542263446809, .00000000000006576659768580062, -.00000000000028210143846181267, .00000000000010701931762114254, .00000000000018119346366441110, .00000000000009840465278232627, -.00000000000033149150282752542, -.00000000000018302857356041668, -.00000000000016207400156744949, .00000000000048303314949553201, -.00000000000071560553172382115, .00000000000088821239518571855, -.00000000000030900580513238244, -.00000000000061076551972851496, .00000000000035659969663347830, .00000000000035782396591276383, -.00000000000046226087001544578, .00000000000062279762917225156, .00000000000072838947272065741, .00000000000026809646615211673, -.00000000000010960825046059278, .00000000000002311949383800537, -.00000000000058469058005299247, -.00000000000002103748251144494, -.00000000000023323182945587408, -.00000000000042333694288141916, -.00000000000043933937969737844, .00000000000041341647073835565, .00000000000006841763641591466, .00000000000047585534004430641, .00000000000083679678674757695, -.00000000000085763734646658640, .00000000000021913281229340092, -.00000000000062242842536431148, -.00000000000010983594325438430, .00000000000065310431377633651, -.00000000000047580199021710769, -.00000000000037854251265457040, .00000000000040939233218678664, .00000000000087424383914858291, .00000000000025218188456842882, -.00000000000003608131360422557, -.00000000000050518555924280902, .00000000000078699403323355317, -.00000000000067020876961949060, .00000000000016108575753932458, .00000000000058527188436251509, -.00000000000035246757297904791, -.00000000000018372084495629058, .00000000000088606689813494916, .00000000000066486268071468700, .00000000000063831615170646519, .00000000000025144230728376072, -.00000000000017239444525614834 }; - -#if 0 -double -#ifdef _ANSI_SOURCE -log(double x) -#else -log(x) double x; -#endif -{ - int m, j; - double F, f, g, q, u, u2, v, zero = 0.0, one = 1.0; - volatile double u1; - - /* Catch special cases */ - if (x <= 0) - if (x == zero) /* log(0) = -Inf */ - return (-one/zero); - else /* log(neg) = NaN */ - return (zero/zero); - else if (!finite(x)) - return (x+x); /* x = NaN, Inf */ - - /* Argument reduction: 1 <= g < 2; x/2^m = g; */ - /* y = F*(1 + f/F) for |f| <= 2^-8 */ - - m = logb(x); - g = ldexp(x, -m); - if (m == -1022) { - j = logb(g), m += j; - g = ldexp(g, -j); - } - j = N*(g-1) + .5; - F = (1.0/N) * j + 1; /* F*128 is an integer in [128, 512] */ - f = g - F; - - /* Approximate expansion for log(1+f/F) ~= u + q */ - g = 1/(2*F+f); - u = 2*f*g; - v = u*u; - q = u*v*(A1 + v*(A2 + v*(A3 + v*A4))); - - /* case 1: u1 = u rounded to 2^-43 absolute. Since u < 2^-8, - * u1 has at most 35 bits, and F*u1 is exact, as F has < 8 bits. - * It also adds exactly to |m*log2_hi + log_F_head[j] | < 750 - */ - if (m | j) - u1 = u + 513, u1 -= 513; - - /* case 2: |1-x| < 1/256. The m- and j- dependent terms are zero; - * u1 = u to 24 bits. - */ - else - u1 = u, TRUNC(u1); - u2 = (2.0*(f - F*u1) - u1*f) * g; - /* u1 + u2 = 2f/(2F+f) to extra precision. */ - - /* log(x) = log(2^m*F*(1+f/F)) = */ - /* (m*log2_hi+logF_head[j]+u1) + (m*log2_lo+logF_tail[j]+q); */ - /* (exact) + (tiny) */ - - u1 += m*logF_head[N] + logF_head[j]; /* exact */ - u2 = (u2 + logF_tail[j]) + q; /* tiny */ - u2 += logF_tail[N]*m; - return (u1 + u2); -} -#endif - /* * Extra precision variant, returning struct {double a, b;}; - * log(x) = a+b to 63 bits, with a rounded to 26 bits. + * log(x) = a+b to 63 bits, with 'a' rounded to 24 bits. */ -struct Double -#ifdef _ANSI_SOURCE +static struct Double __log__D(double x) -#else -__log__D(x) double x; -#endif { int m, j; - double F, f, g, q, u, v, u2; - volatile double u1; + double F, f, g, q, u, v, u1, u2; struct Double r; - /* Argument reduction: 1 <= g < 2; x/2^m = g; */ - /* y = F*(1 + f/F) for |f| <= 2^-8 */ - - m = logb(x); - g = ldexp(x, -m); + /* + * Argument reduction: 1 <= g < 2; x/2^m = g; + * y = F*(1 + f/F) for |f| <= 2^-8 + */ + g = frexp(x, &m); + g *= 2; + m--; if (m == -1022) { - j = logb(g), m += j; + j = ilogb(g); + m += j; g = ldexp(g, -j); } - j = N*(g-1) + .5; - F = (1.0/N) * j + 1; + j = N * (g - 1) + 0.5; + F = (1. / N) * j + 1; f = g - F; - g = 1/(2*F+f); - u = 2*f*g; - v = u*u; - q = u*v*(A1 + v*(A2 + v*(A3 + v*A4))); - if (m | j) - u1 = u + 513, u1 -= 513; - else - u1 = u, TRUNC(u1); - u2 = (2.0*(f - F*u1) - u1*f) * g; + g = 1 / (2 * F + f); + u = 2 * f * g; + v = u * u; + q = u * v * (A1 + v * (A2 + v * A3)); + if (m | j) { + u1 = u + 513; + u1 -= 513; + } else { + u1 = (float)u; + } + u2 = (2 * (f - F * u1) - u1 * f) * g; - u1 += m*logF_head[N] + logF_head[j]; + u1 += m * logF_head[N] + logF_head[j]; - u2 += logF_tail[j]; u2 += q; - u2 += logF_tail[N]*m; - r.a = u1 + u2; /* Only difference is here */ - TRUNC(r.a); + u2 += logF_tail[j]; + u2 += q; + u2 += logF_tail[N] * m; + r.a = (float)(u1 + u2); /* Only difference is here. */ r.b = (u1 - r.a) + u2; return (r); } diff --git a/lib/msun/bsdsrc/b_tgamma.c b/lib/msun/bsdsrc/b_tgamma.c index 5cb1f93f25ed..493ced3769c7 100644 --- a/lib/msun/bsdsrc/b_tgamma.c +++ b/lib/msun/bsdsrc/b_tgamma.c @@ -1,315 +1,400 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1992, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ +/* + * The original code, FreeBSD's old svn r93211, contained the following + * attribution: + * + * This code by P. McIlroy, Oct 1992; + * + * The financial support of UUNET Communications Services is greatfully + * acknowledged. + * + * The algorithm remains, but the code has been re-arranged to facilitate + * porting to other precisions. + */ + /* @(#)gamma.c 8.1 (Berkeley) 6/4/93 */ #include __FBSDID("$FreeBSD$"); -/* - * This code by P. McIlroy, Oct 1992; - * - * The financial support of UUNET Communications Services is greatfully - * acknowledged. - */ +#include -#include -#include "mathimpl.h" +#include "math.h" +#include "math_private.h" -/* METHOD: - * x < 0: Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x)) - * At negative integers, return NaN and raise invalid. - * - * x < 6.5: - * Use argument reduction G(x+1) = xG(x) to reach the - * range [1.066124,2.066124]. Use a rational - * approximation centered at the minimum (x0+1) to - * ensure monotonicity. - * - * x >= 6.5: Use the asymptotic approximation (Stirling's formula) - * adjusted for equal-ripples: - * - * log(G(x)) ~= (x-.5)*(log(x)-1) + .5(log(2*pi)-1) + 1/x*P(1/(x*x)) +/* Used in b_log.c and below. */ +struct Double { + double a; + double b; +}; + +#include "b_log.c" +#include "b_exp.c" + +/* + * The range is broken into several subranges. Each is handled by its + * helper functions. * - * Keep extra precision in multiplying (x-.5)(log(x)-1), to - * avoid premature round-off. + * x >= 6.0: large_gam(x) + * 6.0 > x >= xleft: small_gam(x) where xleft = 1 + left + x0. + * xleft > x > iota: smaller_gam(x) where iota = 1e-17. + * iota > x > -itoa: Handle x near 0. + * -iota > x : neg_gam * * Special values: * -Inf: return NaN and raise invalid; * negative integer: return NaN and raise invalid; * other x ~< 177.79: return +-0 and raise underflow; * +-0: return +-Inf and raise divide-by-zero; * finite x ~> 171.63: return +Inf and raise overflow; * +Inf: return +Inf; * NaN: return NaN. * * Accuracy: tgamma(x) is accurate to within * x > 0: error provably < 0.9ulp. * Maximum observed in 1,000,000 trials was .87ulp. * x < 0: * Maximum observed error < 4ulp in 1,000,000 trials. */ -static double neg_gam(double); -static double small_gam(double); -static double smaller_gam(double); -static struct Double large_gam(double); -static struct Double ratfun_gam(double, double); - -/* - * Rational approximation, A0 + x*x*P(x)/Q(x), on the interval - * [1.066.., 2.066..] accurate to 4.25e-19. - */ -#define LEFT -.3955078125 /* left boundary for rat. approx */ -#define x0 .461632144968362356785 /* xmin - 1 */ - -#define a0_hi 0.88560319441088874992 -#define a0_lo -.00000000000000004996427036469019695 -#define P0 6.21389571821820863029017800727e-01 -#define P1 2.65757198651533466104979197553e-01 -#define P2 5.53859446429917461063308081748e-03 -#define P3 1.38456698304096573887145282811e-03 -#define P4 2.40659950032711365819348969808e-03 -#define Q0 1.45019531250000000000000000000e+00 -#define Q1 1.06258521948016171343454061571e+00 -#define Q2 -2.07474561943859936441469926649e-01 -#define Q3 -1.46734131782005422506287573015e-01 -#define Q4 3.07878176156175520361557573779e-02 -#define Q5 5.12449347980666221336054633184e-03 -#define Q6 -1.76012741431666995019222898833e-03 -#define Q7 9.35021023573788935372153030556e-05 -#define Q8 6.13275507472443958924745652239e-06 /* * Constants for large x approximation (x in [6, Inf]) * (Accurate to 2.8*10^-19 absolute) */ -#define lns2pi_hi 0.418945312500000 -#define lns2pi_lo -.000006779295327258219670263595 -#define Pa0 8.33333333333333148296162562474e-02 -#define Pa1 -2.77777777774548123579378966497e-03 -#define Pa2 7.93650778754435631476282786423e-04 -#define Pa3 -5.95235082566672847950717262222e-04 -#define Pa4 8.41428560346653702135821806252e-04 -#define Pa5 -1.89773526463879200348872089421e-03 -#define Pa6 5.69394463439411649408050664078e-03 -#define Pa7 -1.44705562421428915453880392761e-02 - -static const double zero = 0., one = 1.0, tiny = 1e-300; -double -tgamma(x) - double x; +static const double zero = 0.; +static const volatile double tiny = 1e-300; +/* + * x >= 6 + * + * Use the asymptotic approximation (Stirling's formula) adjusted fof + * equal-ripples: + * + * log(G(x)) ~= (x-0.5)*(log(x)-1) + 0.5(log(2*pi)-1) + 1/x*P(1/(x*x)) + * + * Keep extra precision in multiplying (x-.5)(log(x)-1), to avoid + * premature round-off. + * + * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error. + */ +static const double + ln2pi_hi = 0.41894531250000000, + ln2pi_lo = -6.7792953272582197e-6, + Pa0 = 8.3333333333333329e-02, /* 0x3fb55555, 0x55555555 */ + Pa1 = -2.7777777777735404e-03, /* 0xbf66c16c, 0x16c145ec */ + Pa2 = 7.9365079044114095e-04, /* 0x3f4a01a0, 0x183de82d */ + Pa3 = -5.9523715464225254e-04, /* 0xbf438136, 0x0e681f62 */ + Pa4 = 8.4161391899445698e-04, /* 0x3f4b93f8, 0x21042a13 */ + Pa5 = -1.9065246069191080e-03, /* 0xbf5f3c8b, 0x357cb64e */ + Pa6 = 5.9047708485785158e-03, /* 0x3f782f99, 0xdaf5d65f */ + Pa7 = -1.6484018705183290e-02; /* 0xbf90e12f, 0xc4fb4df0 */ + +static struct Double +large_gam(double x) { + double p, z, thi, tlo, xhi, xlo; struct Double u; - if (x >= 6) { - if(x > 171.63) - return (x / zero); - u = large_gam(x); - return(__exp__D(u.a, u.b)); - } else if (x >= 1.0 + LEFT + x0) - return (small_gam(x)); - else if (x > 1.e-17) - return (smaller_gam(x)); - else if (x > -1.e-17) { - if (x != 0.0) - u.a = one - tiny; /* raise inexact */ - return (one/x); - } else if (!finite(x)) - return (x - x); /* x is NaN or -Inf */ - else - return (neg_gam(x)); + z = 1 / (x * x); + p = Pa0 + z * (Pa1 + z * (Pa2 + z * (Pa3 + z * (Pa4 + z * (Pa5 + + z * (Pa6 + z * Pa7)))))); + p = p / x; + + u = __log__D(x); + u.a -= 1; + + /* Split (x - 0.5) in high and low parts. */ + x -= 0.5; + xhi = (float)x; + xlo = x - xhi; + + /* Compute t = (x-.5)*(log(x)-1) in extra precision. */ + thi = xhi * u.a; + tlo = xlo * u.a + x * u.b; + + /* Compute thi + tlo + ln2pi_hi + ln2pi_lo + p. */ + tlo += ln2pi_lo; + tlo += p; + u.a = ln2pi_hi + tlo; + u.a += thi; + u.b = thi - u.a; + u.b += ln2pi_hi; + u.b += tlo; + return (u); } /* - * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error. + * Rational approximation, A0 + x * x * P(x) / Q(x), on the interval + * [1.066.., 2.066..] accurate to 4.25e-19. + * + * Returns r.a + r.b = a0 + (z + c)^2 * p / q, with r.a truncated. */ +static const double +#if 0 + a0_hi = 8.8560319441088875e-1, + a0_lo = -4.9964270364690197e-17, +#else + a0_hi = 8.8560319441088875e-01, /* 0x3fec56dc, 0x82a74aef */ + a0_lo = -4.9642368725563397e-17, /* 0xbc8c9deb, 0xaa64afc3 */ +#endif + P0 = 6.2138957182182086e-1, + P1 = 2.6575719865153347e-1, + P2 = 5.5385944642991746e-3, + P3 = 1.3845669830409657e-3, + P4 = 2.4065995003271137e-3, + Q0 = 1.4501953125000000e+0, + Q1 = 1.0625852194801617e+0, + Q2 = -2.0747456194385994e-1, + Q3 = -1.4673413178200542e-1, + Q4 = 3.0787817615617552e-2, + Q5 = 5.1244934798066622e-3, + Q6 = -1.7601274143166700e-3, + Q7 = 9.3502102357378894e-5, + Q8 = 6.1327550747244396e-6; + static struct Double -large_gam(x) - double x; +ratfun_gam(double z, double c) { - double z, p; - struct Double t, u, v; + double p, q, thi, tlo; + struct Double r; - z = one/(x*x); - p = Pa0+z*(Pa1+z*(Pa2+z*(Pa3+z*(Pa4+z*(Pa5+z*(Pa6+z*Pa7)))))); - p = p/x; + q = Q0 + z * (Q1 + z * (Q2 + z * (Q3 + z * (Q4 + z * (Q5 + + z * (Q6 + z * (Q7 + z * Q8))))))); + p = P0 + z * (P1 + z * (P2 + z * (P3 + z * P4))); + p = p / q; - u = __log__D(x); - u.a -= one; - v.a = (x -= .5); - TRUNC(v.a); - v.b = x - v.a; - t.a = v.a*u.a; /* t = (x-.5)*(log(x)-1) */ - t.b = v.b*u.a + x*u.b; - /* return t.a + t.b + lns2pi_hi + lns2pi_lo + p */ - t.b += lns2pi_lo; t.b += p; - u.a = lns2pi_hi + t.b; u.a += t.a; - u.b = t.a - u.a; - u.b += lns2pi_hi; u.b += t.b; - return (u); + /* Split z into high and low parts. */ + thi = (float)z; + tlo = (z - thi) + c; + tlo *= (thi + z); + + /* Split (z+c)^2 into high and low parts. */ + thi *= thi; + q = thi; + thi = (float)thi; + tlo += (q - thi); + + /* Split p/q into high and low parts. */ + r.a = (float)p; + r.b = p - r.a; + + tlo = tlo * p + thi * r.b + a0_lo; + thi *= r.a; /* t = (z+c)^2*(P/Q) */ + r.a = (float)(thi + a0_hi); + r.b = ((a0_hi - r.a) + thi) + tlo; + return (r); /* r = a0 + t */ } /* + * x < 6 + * + * Use argument reduction G(x+1) = xG(x) to reach the range [1.066124, + * 2.066124]. Use a rational approximation centered at the minimum + * (x0+1) to ensure monotonicity. + * * Good to < 1 ulp. (provably .90 ulp; .87 ulp on 1,000,000 runs.) * It also has correct monotonicity. */ +static const double + left = -0.3955078125, /* left boundary for rat. approx */ + x0 = 4.6163214496836236e-1; /* xmin - 1 */ + static double -small_gam(x) - double x; +small_gam(double x) { - double y, ym1, t; + double t, y, ym1; struct Double yy, r; - y = x - one; - ym1 = y - one; - if (y <= 1.0 + (LEFT + x0)) { + + y = x - 1; + if (y <= 1 + (left + x0)) { yy = ratfun_gam(y - x0, 0); return (yy.a + yy.b); } - r.a = y; - TRUNC(r.a); - yy.a = r.a - one; - y = ym1; - yy.b = r.b = y - yy.a; + + r.a = (float)y; + yy.a = r.a - 1; + y = y - 1 ; + r.b = yy.b = y - yy.a; + /* Argument reduction: G(x+1) = x*G(x) */ - for (ym1 = y-one; ym1 > LEFT + x0; y = ym1--, yy.a--) { - t = r.a*yy.a; - r.b = r.a*yy.b + y*r.b; - r.a = t; - TRUNC(r.a); + for (ym1 = y - 1; ym1 > left + x0; y = ym1--, yy.a--) { + t = r.a * yy.a; + r.b = r.a * yy.b + y * r.b; + r.a = (float)t; r.b += (t - r.a); } + /* Return r*tgamma(y). */ yy = ratfun_gam(y - x0, 0); - y = r.b*(yy.a + yy.b) + r.a*yy.b; - y += yy.a*r.a; + y = r.b * (yy.a + yy.b) + r.a * yy.b; + y += yy.a * r.a; return (y); } /* - * Good on (0, 1+x0+LEFT]. Accurate to 1ulp. + * Good on (0, 1+x0+left]. Accurate to 1 ulp. */ static double -smaller_gam(x) - double x; +smaller_gam(double x) { - double t, d; - struct Double r, xx; - if (x < x0 + LEFT) { - t = x, TRUNC(t); - d = (t+x)*(x-t); + double d, rhi, rlo, t, xhi, xlo; + struct Double r; + + if (x < x0 + left) { + t = (float)x; + d = (t + x) * (x - t); t *= t; - xx.a = (t + x), TRUNC(xx.a); - xx.b = x - xx.a; xx.b += t; xx.b += d; - t = (one-x0); t += x; - d = (one-x0); d -= t; d += x; - x = xx.a + xx.b; + xhi = (float)(t + x); + xlo = x - xhi; + xlo += t; + xlo += d; + t = 1 - x0; + t += x; + d = 1 - x0; + d -= t; + d += x; + x = xhi + xlo; } else { - xx.a = x, TRUNC(xx.a); - xx.b = x - xx.a; + xhi = (float)x; + xlo = x - xhi; t = x - x0; - d = (-x0 -t); d += x; + d = - x0 - t; + d += x; } + r = ratfun_gam(t, d); - d = r.a/x, TRUNC(d); - r.a -= d*xx.a; r.a -= d*xx.b; r.a += r.b; - return (d + r.a/x); + d = (float)(r.a / x); + r.a -= d * xhi; + r.a -= d * xlo; + r.a += r.b; + + return (d + r.a / x); } /* - * returns (z+c)^2 * P(z)/Q(z) + a0 + * x < 0 + * + * Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x)). + * At negative integers, return NaN and raise invalid. */ -static struct Double -ratfun_gam(z, c) - double z, c; -{ - double p, q; - struct Double r, t; - - q = Q0 +z*(Q1+z*(Q2+z*(Q3+z*(Q4+z*(Q5+z*(Q6+z*(Q7+z*Q8))))))); - p = P0 + z*(P1 + z*(P2 + z*(P3 + z*P4))); - - /* return r.a + r.b = a0 + (z+c)^2*p/q, with r.a truncated to 26 bits. */ - p = p/q; - t.a = z, TRUNC(t.a); /* t ~= z + c */ - t.b = (z - t.a) + c; - t.b *= (t.a + z); - q = (t.a *= t.a); /* t = (z+c)^2 */ - TRUNC(t.a); - t.b += (q - t.a); - r.a = p, TRUNC(r.a); /* r = P/Q */ - r.b = p - r.a; - t.b = t.b*p + t.a*r.b + a0_lo; - t.a *= r.a; /* t = (z+c)^2*(P/Q) */ - r.a = t.a + a0_hi, TRUNC(r.a); - r.b = ((a0_hi-r.a) + t.a) + t.b; - return (r); /* r = a0 + t */ -} - static double -neg_gam(x) - double x; +neg_gam(double x) { int sgn = 1; struct Double lg, lsine; double y, z; y = ceil(x); if (y == x) /* Negative integer. */ return ((x - x) / zero); + z = y - x; if (z > 0.5) - z = one - z; - y = 0.5 * y; + z = 1 - z; + + y = y / 2; if (y == ceil(y)) sgn = -1; - if (z < .25) - z = sin(M_PI*z); + + if (z < 0.25) + z = sinpi(z); else - z = cos(M_PI*(0.5-z)); + z = cospi(0.5 - z); + /* Special case: G(1-x) = Inf; G(x) may be nonzero. */ if (x < -170) { + if (x < -190) - return ((double)sgn*tiny*tiny); - y = one - x; /* exact: 128 < |x| < 255 */ + return (sgn * tiny * tiny); + + y = 1 - x; /* exact: 128 < |x| < 255 */ lg = large_gam(y); - lsine = __log__D(M_PI/z); /* = TRUNC(log(u)) + small */ + lsine = __log__D(M_PI / z); /* = TRUNC(log(u)) + small */ lg.a -= lsine.a; /* exact (opposite signs) */ lg.b -= lsine.b; y = -(lg.a + lg.b); z = (y + lg.a) + lg.b; y = __exp__D(y, z); if (sgn < 0) y = -y; return (y); } - y = one-x; - if (one-y == x) + + y = 1 - x; + if (1 - y == x) y = tgamma(y); else /* 1-x is inexact */ - y = -x*tgamma(-x); + y = - x * tgamma(-x); + if (sgn < 0) y = -y; - return (M_PI / (y*z)); + return (M_PI / (y * z)); +} +/* + * xmax comes from lgamma(xmax) - emax * log(2) = 0. + * static const float xmax = 35.040095f + * static const double xmax = 171.624376956302725; + * ld80: LD80C(0xdb718c066b352e20, 10, 1.75554834290446291689e+03L), + * ld128: 1.75554834290446291700388921607020320e+03L, + * + * iota is a sloppy threshold to isolate x = 0. + */ +static const double xmax = 171.624376956302725; +static const double iota = 0x1p-56; + +double +tgamma(double x) +{ + struct Double u; + + if (x >= 6) { + if (x > xmax) + return (x / zero); + u = large_gam(x); + return (__exp__D(u.a, u.b)); + } + + if (x >= 1 + left + x0) + return (small_gam(x)); + + if (x > iota) + return (smaller_gam(x)); + + if (x > -iota) { + if (x != 0.) + u.a = 1 - tiny; /* raise inexact */ + return (1 / x); + } + + if (!isfinite(x)) + return (x - x); /* x is NaN or -Inf */ + + return (neg_gam(x)); } + +#if (LDBL_MANT_DIG == 53) +__weak_reference(tgamma, tgammal); +#endif