From ffdd5e50e18b0cb212acad135e421d932cf3d3a2 Mon Sep 17 00:00:00 2001 From: Ulrich Drepper Date: Wed, 26 May 2004 04:47:00 +0000 Subject: Update. 2004-05-25 Steven Munroe * sysdeps/powerpc/fpu/Makefile: Make ld.so a dependency of libm.so. * sysdeps/powerpc/fpu/bits/mathinline.h [__LIBC_INERNAL_MATH_INLINES] (__ieee754_sqrt): Define as __MATH_INLINE using fsqrt instruction. (__ieee754_sqrtf): Define as __MATH_INLINE using fsqrts instruction. * sysdeps/powerpc/fpu/e_sqrt.c (__slow_ieee754_sqrt): Moved implementation from w_sqrt.c. * sysdeps/powerpc/fpu/e_sqrtf.c (__slow_ieee754_sqrtf): Moved implementation from w_sqrtf.c. * sysdeps/powerpc/fpu/w_sqrt.c (__sqrt): Wrapper implementation using inline __ieee754_sqrt(). * sysdeps/powerpc/fpu/w_sqrtf.c (__sqrtf): Wrapper implementation using inline __ieee754_sqrtf(). * sysdeps/powerpc/powerpc32/sysdep.h [__ASSEMBLER__]: Include independent of __ASSEMBLER__. * sysdeps/powerpc/sysdep.h [__ASSEMBLER__] (PPC_FEATURE_*): Define PPC_FEATURE_* independent of __ASSEMBLER__. 2004-05-25 Jakub Jelinek * sysdeps/pthread/aio_notify.c: Use <> instead of "" for aio_misc.h include. (aio_start_notify_thread): Define if not defined. (notify_func_wrapper): Use it. * sysdeps/pthread/aio_misc.c: Use <> instead of "" for aio_misc.h include. (aio_create_helper_thread): Define if not defined. (__aio_create_helper_thread): New function. (__aio_enqueue_request): Use aio_create_helper_thread. * nis/ypclnt.c (ypall_data, ypall_foreach): Remove. (struct ypresp_all_data): New type. (__xdr_ypresp_all): Change second argument to struct ypresp_all_data *. Replace ypall_foreach and ypall_data with objp->foreach and objp->data. (yp_all): Remove status variable, add data. Replace all uses of status with data.status. Initialize data.foreach and data.data instead of ypall_foreach and ypall_data. 2004-05-24 Jakub Jelinek * elf/dl-lookup.c (add_dependency): Set DF_1_NODELETE bit in l_flags_1, not in l_flags. --- ChangeLog | 45 ++++++++ nis/ypclnt.c | 38 ++++--- nptl/ChangeLog | 4 + nptl/sysdeps/unix/sysv/linux/aio_misc.h | 67 ++++++++++++ sysdeps/powerpc/fpu/Makefile | 3 + sysdeps/powerpc/fpu/bits/mathinline.h | 52 +++++++++ sysdeps/powerpc/fpu/e_sqrt.c | 186 +++++++++++++++++++++++++++++++- sysdeps/powerpc/fpu/e_sqrtf.c | 163 +++++++++++++++++++++++++++- sysdeps/powerpc/fpu/w_sqrt.c | 143 +++++------------------- sysdeps/powerpc/fpu/w_sqrtf.c | 138 +++++------------------- sysdeps/powerpc/powerpc32/sysdep.h | 4 +- sysdeps/powerpc/sysdep.h | 30 +++--- sysdeps/pthread/aio_misc.c | 28 +++-- sysdeps/pthread/aio_notify.c | 9 +- 14 files changed, 636 insertions(+), 274 deletions(-) create mode 100644 nptl/sysdeps/unix/sysv/linux/aio_misc.h diff --git a/ChangeLog b/ChangeLog index 9c1166efe8..4d7a509a55 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,3 +1,48 @@ +2004-05-25 Steven Munroe + + * sysdeps/powerpc/fpu/Makefile: Make ld.so a dependency of libm.so. + * sysdeps/powerpc/fpu/bits/mathinline.h [__LIBC_INERNAL_MATH_INLINES] + (__ieee754_sqrt): Define as __MATH_INLINE using fsqrt instruction. + (__ieee754_sqrtf): Define as __MATH_INLINE using fsqrts instruction. + * sysdeps/powerpc/fpu/e_sqrt.c (__slow_ieee754_sqrt): Moved + implementation from w_sqrt.c. + * sysdeps/powerpc/fpu/e_sqrtf.c (__slow_ieee754_sqrtf): Moved + implementation from w_sqrtf.c. + * sysdeps/powerpc/fpu/w_sqrt.c (__sqrt): Wrapper implementation + using inline __ieee754_sqrt(). + * sysdeps/powerpc/fpu/w_sqrtf.c (__sqrtf): Wrapper implementation + using inline __ieee754_sqrtf(). + * sysdeps/powerpc/powerpc32/sysdep.h [__ASSEMBLER__]: Include + independent of __ASSEMBLER__. + * sysdeps/powerpc/sysdep.h [__ASSEMBLER__] (PPC_FEATURE_*): Define + PPC_FEATURE_* independent of __ASSEMBLER__. + +2004-05-25 Jakub Jelinek + + * sysdeps/pthread/aio_notify.c: Use <> instead of "" for aio_misc.h + include. + (aio_start_notify_thread): Define if not defined. + (notify_func_wrapper): Use it. + * sysdeps/pthread/aio_misc.c: Use <> instead of "" for aio_misc.h + include. + (aio_create_helper_thread): Define if not defined. + (__aio_create_helper_thread): New function. + (__aio_enqueue_request): Use aio_create_helper_thread. + + * nis/ypclnt.c (ypall_data, ypall_foreach): Remove. + (struct ypresp_all_data): New type. + (__xdr_ypresp_all): Change second argument to + struct ypresp_all_data *. Replace ypall_foreach and + ypall_data with objp->foreach and objp->data. + (yp_all): Remove status variable, add data. Replace + all uses of status with data.status. Initialize data.foreach + and data.data instead of ypall_foreach and ypall_data. + +2004-05-24 Jakub Jelinek + + * elf/dl-lookup.c (add_dependency): Set DF_1_NODELETE bit + in l_flags_1, not in l_flags. + 2004-04-10 Robert Millan * sysdeps/unix/sysv/linux/bits/in.h: Cosmetic fixes to get in sync diff --git a/nis/ypclnt.c b/nis/ypclnt.c index d6e2205fd9..ab28e6002b 100644 --- a/nis/ypclnt.c +++ b/nis/ypclnt.c @@ -618,12 +618,16 @@ yp_order (const char *indomain, const char *inmap, unsigned int *outorder) return YPERR_SUCCESS; } -static void *ypall_data; -static int (*ypall_foreach) (int status, char *key, int keylen, - char *val, int vallen, char *data); +struct ypresp_all_data +{ + unsigned long status; + void *data; + int (*foreach) (int status, char *key, int keylen, + char *val, int vallen, char *data); +}; static bool_t -__xdr_ypresp_all (XDR *xdrs, u_long *objp) +__xdr_ypresp_all (XDR *xdrs, struct ypresp_all_data *objp) { while (1) { @@ -633,13 +637,13 @@ __xdr_ypresp_all (XDR *xdrs, u_long *objp) if (!xdr_ypresp_all (xdrs, &resp)) { xdr_free ((xdrproc_t) xdr_ypresp_all, (char *) &resp); - *objp = YP_YPERR; + objp->status = YP_YPERR; return FALSE; } if (resp.more == 0) { xdr_free ((xdrproc_t) xdr_ypresp_all, (char *) &resp); - *objp = YP_NOMORE; + objp->status = YP_NOMORE; return TRUE; } @@ -656,24 +660,24 @@ __xdr_ypresp_all (XDR *xdrs, u_long *objp) But we are allowed to add data behind the buffer, if we don't modify the length. So add an extra NUL character to avoid trouble with broken code. */ - *objp = YP_TRUE; + objp->status = YP_TRUE; memcpy (key, resp.ypresp_all_u.val.key.keydat_val, keylen); key[keylen] = '\0'; memcpy (val, resp.ypresp_all_u.val.val.valdat_val, vallen); val[vallen] = '\0'; xdr_free ((xdrproc_t) xdr_ypresp_all, (char *) &resp); - if ((*ypall_foreach) (*objp, key, keylen, - val, vallen, ypall_data)) + if ((*objp->foreach) (objp->status, key, keylen, + val, vallen, objp->data)) return TRUE; } break; default: - *objp = resp.ypresp_all_u.val.stat; + objp->status = resp.ypresp_all_u.val.stat; xdr_free ((xdrproc_t) xdr_ypresp_all, (char *) &resp); /* Sun says we don't need to make this call, but must return immediatly. Since Solaris makes this call, we will call the callback function, too. */ - (*ypall_foreach) (*objp, NULL, 0, NULL, 0, ypall_data); + (*objp->foreach) (objp->status, NULL, 0, NULL, 0, objp->data); return TRUE; } } @@ -689,7 +693,7 @@ yp_all (const char *indomain, const char *inmap, enum clnt_stat result; struct sockaddr_in clnt_sin; CLIENT *clnt; - unsigned long status; + struct ypresp_all_data data; int clnt_sock; int saved_errno = errno; @@ -725,12 +729,12 @@ yp_all (const char *indomain, const char *inmap, req.domain = (char *) indomain; req.map = (char *) inmap; - ypall_foreach = incallback->foreach; - ypall_data = (void *) incallback->data; + data.foreach = incallback->foreach; + data.data = (void *) incallback->data; result = clnt_call (clnt, YPPROC_ALL, (xdrproc_t) xdr_ypreq_nokey, (caddr_t) &req, (xdrproc_t) __xdr_ypresp_all, - (caddr_t) &status, RPCTIMEOUT); + (caddr_t) &data, RPCTIMEOUT); if (result != RPC_SUCCESS) { @@ -744,10 +748,10 @@ yp_all (const char *indomain, const char *inmap, clnt_destroy (clnt); - if (res == YPERR_SUCCESS && status != YP_NOMORE) + if (res == YPERR_SUCCESS && data.status != YP_NOMORE) { __set_errno (saved_errno); - return ypprot_err (status); + return ypprot_err (data.status); } ++try; } diff --git a/nptl/ChangeLog b/nptl/ChangeLog index 5f1f61c088..9164affc9d 100644 --- a/nptl/ChangeLog +++ b/nptl/ChangeLog @@ -1,3 +1,7 @@ +2004-05-25 Jakub Jelinek + + * sysdeps/unix/sysv/linux/aio_misc.h: New file. + 2004-05-21 Jakub Jelinek * sysdeps/pthread/pthread_cond_wait.c (__pthread_cond_wait): Compare diff --git a/nptl/sysdeps/unix/sysv/linux/aio_misc.h b/nptl/sysdeps/unix/sysv/linux/aio_misc.h new file mode 100644 index 0000000000..7b0bac75f2 --- /dev/null +++ b/nptl/sysdeps/unix/sysv/linux/aio_misc.h @@ -0,0 +1,67 @@ +/* Copyright (C) 2004 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek , 2004. + + 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; see the file COPYING.LIB. If not, + write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, + Boston, MA 02111-1307, USA. */ + +#ifndef _AIO_MISC_H +# include_next +# include +# include +# include + +# define aio_start_notify_thread __aio_start_notify_thread +# define aio_create_helper_thread __aio_create_helper_thread + +extern inline void +__aio_start_notify_thread (void) +{ + sigset_t ss; + sigemptyset (&ss); + INTERNAL_SYSCALL_DECL (err); + INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_SETMASK, &ss, NULL, _NSIG / 8); +} + +extern inline int +__aio_create_helper_thread (pthread_t *threadp, void *(*tf) (void *), void *arg) +{ + pthread_attr_t attr; + + /* Make sure the thread is created detached. */ + pthread_attr_init (&attr); + pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); + + /* The helper thread needs only very little resources. */ + (void) pthread_attr_setstacksize (&attr, PTHREAD_STACK_MIN); + + /* Block all signals in the helper thread. To do this thoroughly we + temporarily have to block all signals here. */ + sigset_t ss; + sigset_t oss; + sigfillset (&ss); + INTERNAL_SYSCALL_DECL (err); + INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_SETMASK, &ss, &oss, _NSIG / 8); + + int ret = pthread_create (threadp, &attr, tf, arg); + + /* Restore the signal mask. */ + INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_SETMASK, &oss, NULL, + _NSIG / 8); + + (void) pthread_attr_destroy (&attr); + return ret; +} +#endif diff --git a/sysdeps/powerpc/fpu/Makefile b/sysdeps/powerpc/fpu/Makefile index bf2ed92e7b..d0fe4a8ba1 100644 --- a/sysdeps/powerpc/fpu/Makefile +++ b/sysdeps/powerpc/fpu/Makefile @@ -1,3 +1,6 @@ ifeq ($(subdir),math) libm-support += fenv_const fe_nomask t_sqrt + +# libm needs ld.so to access dl_hwcap +$(objpfx)libm.so: $(elfobjdir)/ld.so endif diff --git a/sysdeps/powerpc/fpu/bits/mathinline.h b/sysdeps/powerpc/fpu/bits/mathinline.h index e692df9b1a..d9206d4fac 100644 --- a/sysdeps/powerpc/fpu/bits/mathinline.h +++ b/sysdeps/powerpc/fpu/bits/mathinline.h @@ -121,4 +121,56 @@ fdimf (float __x, float __y) __THROW #endif /* __USE_ISOC99 */ #endif /* !__NO_MATH_INLINES && __OPTIMIZE__ */ + +/* This code is used internally in the GNU libc. */ +# ifdef __LIBC_INTERNAL_MATH_INLINES + +#include +#include +#include + +extern double __slow_ieee754_sqrt (double); +__MATH_INLINE double +__ieee754_sqrt (double __x) +{ + double __z; + + /* If the CPU is 64-bit we can use the optional FP instructions we. */ + if ((GLRO(dl_hwcap) & PPC_FEATURE_64) != 0) + { + /* Volatile is required to prevent the compiler from moving the + fsqrt instruction above the branch. */ + __asm __volatile ( + " fsqrt %0,%1\n" + : "=f" (__z) + : "f" (__x)); + } + else + __z = __slow_ieee754_sqrt(__x); + + return __z; +} + +extern float __slow_ieee754_sqrtf (float); +__MATH_INLINE float +__ieee754_sqrtf (float __x) +{ + float __z; + + /* If the CPU is 64-bit we can use the optional FP instructions we. */ + if ((GLRO(dl_hwcap) & PPC_FEATURE_64) != 0) + { + /* Volatile is required to prevent the compiler from moving the + fsqrts instruction above the branch. */ + __asm __volatile ( + " fsqrts %0,%1\n" + : "=f" (__z) + : "f" (__x)); + } + else + __z = __slow_ieee754_sqrtf(__x); + + return __z; +} +# endif /* __LIBC_INTERNAL_MATH_INLINES */ #endif /* __GNUC__ && !_SOFT_FLOAT */ diff --git a/sysdeps/powerpc/fpu/e_sqrt.c b/sysdeps/powerpc/fpu/e_sqrt.c index 9416ea60c8..eb9984d0a1 100644 --- a/sysdeps/powerpc/fpu/e_sqrt.c +++ b/sysdeps/powerpc/fpu/e_sqrt.c @@ -1 +1,185 @@ -/* __ieee754_sqrt is in w_sqrt.c */ +/* Double-precision floating point square root. + Copyright (C) 1997, 2002, 2003, 2004 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, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include +#include +#include +#include + +#include +#include +#include + +static const double almost_half = 0.5000000000000001; /* 0.5 + 2^-53 */ +static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; +static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; +static const float two108 = 3.245185536584267269e+32; +static const float twom54 = 5.551115123125782702e-17; +extern const float __t_sqrt[1024]; + +/* The method is based on a description in + Computation of elementary functions on the IBM RISC System/6000 processor, + P. W. Markstein, IBM J. Res. Develop, 34(1) 1990. + Basically, it consists of two interleaved Newton-Rhapson approximations, + one to find the actual square root, and one to find its reciprocal + without the expense of a division operation. The tricky bit here + is the use of the POWER/PowerPC multiply-add operation to get the + required accuracy with high speed. + + The argument reduction works by a combination of table lookup to + obtain the initial guesses, and some careful modification of the + generated guesses (which mostly runs on the integer unit, while the + Newton-Rhapson is running on the FPU). */ + +#ifdef __STDC__ +double +__slow_ieee754_sqrt (double x) +#else +double +__slow_ieee754_sqrt (x) + double x; +#endif +{ + const float inf = a_inf.value; + + if (x > 0) + { + /* schedule the EXTRACT_WORDS to get separation between the store + and the load. */ + ieee_double_shape_type ew_u; + ieee_double_shape_type iw_u; + ew_u.value = (x); + if (x != inf) + { + /* Variables named starting with 's' exist in the + argument-reduced space, so that 2 > sx >= 0.5, + 1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... . + Variables named ending with 'i' are integer versions of + floating-point values. */ + double sx; /* The value of which we're trying to find the + square root. */ + double sg, g; /* Guess of the square root of x. */ + double sd, d; /* Difference between the square of the guess and x. */ + double sy; /* Estimate of 1/2g (overestimated by 1ulp). */ + double sy2; /* 2*sy */ + double e; /* Difference between y*g and 1/2 (se = e * fsy). */ + double shx; /* == sx * fsg */ + double fsg; /* sg*fsg == g. */ + fenv_t fe; /* Saved floating-point environment (stores rounding + mode and whether the inexact exception is + enabled). */ + uint32_t xi0, xi1, sxi, fsgi; + const float *t_sqrt; + + fe = fegetenv_register (); + /* complete the EXTRACT_WORDS (xi0,xi1,x) operation. */ + xi0 = ew_u.parts.msw; + xi1 = ew_u.parts.lsw; + relax_fenv_state (); + sxi = (xi0 & 0x3fffffff) | 0x3fe00000; + /* schedule the INSERT_WORDS (sx, sxi, xi1) to get separation + between the store and the load. */ + iw_u.parts.msw = sxi; + iw_u.parts.lsw = xi1; + t_sqrt = __t_sqrt + (xi0 >> (52 - 32 - 8 - 1) & 0x3fe); + sg = t_sqrt[0]; + sy = t_sqrt[1]; + /* complete the INSERT_WORDS (sx, sxi, xi1) operation. */ + sx = iw_u.value; + + /* Here we have three Newton-Rhapson iterations each of a + division and a square root and the remainder of the + argument reduction, all interleaved. */ + sd = -(sg * sg - sx); + fsgi = (xi0 + 0x40000000) >> 1 & 0x7ff00000; + sy2 = sy + sy; + sg = sy * sd + sg; /* 16-bit approximation to sqrt(sx). */ + + /* schedule the INSERT_WORDS (fsg, fsgi, 0) to get separation + between the store and the load. */ + INSERT_WORDS (fsg, fsgi, 0); + iw_u.parts.msw = fsgi; + iw_u.parts.lsw = (0); + e = -(sy * sg - almost_half); + sd = -(sg * sg - sx); + if ((xi0 & 0x7ff00000) == 0) + goto denorm; + sy = sy + e * sy2; + sg = sg + sy * sd; /* 32-bit approximation to sqrt(sx). */ + sy2 = sy + sy; + /* complete the INSERT_WORDS (fsg, fsgi, 0) operation. */ + fsg = iw_u.value; + e = -(sy * sg - almost_half); + sd = -(sg * sg - sx); + sy = sy + e * sy2; + shx = sx * fsg; + sg = sg + sy * sd; /* 64-bit approximation to sqrt(sx), + but perhaps rounded incorrectly. */ + sy2 = sy + sy; + g = sg * fsg; + e = -(sy * sg - almost_half); + d = -(g * sg - shx); + sy = sy + e * sy2; + fesetenv_register (fe); + return g + sy * d; + denorm: + /* For denormalised numbers, we normalise, calculate the + square root, and return an adjusted result. */ + fesetenv_register (fe); + return __slow_ieee754_sqrt (x * two108) * twom54; + } + } + else if (x < 0) + { + /* For some reason, some PowerPC32 processors don't implement + FE_INVALID_SQRT. */ +#ifdef FE_INVALID_SQRT + feraiseexcept (FE_INVALID_SQRT); + if (!fetestexcept (FE_INVALID)) +#endif + feraiseexcept (FE_INVALID); + x = a_nan.value; + } + return f_wash (x); +} + +#ifdef __STDC__ +double +__ieee754_sqrt (double x) +#else +double +__ieee754_sqrt (x) + double x; +#endif +{ + double z; + + /* If the CPU is 64-bit we can use the optional FP instructions we. */ + if ((GLRO (dl_hwcap) & PPC_FEATURE_64) != 0) + { + /* Volatile is required to prevent the compiler from moving the + fsqrt instruction above the branch. */ + __asm __volatile (" fsqrt %0,%1\n" + :"=f" (z):"f" (x)); + } + else + z = __slow_ieee754_sqrt (x); + + return z; +} diff --git a/sysdeps/powerpc/fpu/e_sqrtf.c b/sysdeps/powerpc/fpu/e_sqrtf.c index 01c76d6757..9b701012af 100644 --- a/sysdeps/powerpc/fpu/e_sqrtf.c +++ b/sysdeps/powerpc/fpu/e_sqrtf.c @@ -1 +1,162 @@ -/* __ieee754_sqrtf is in w_sqrtf.c */ +/* Single-precision floating point square root. + Copyright (C) 1997, 2003, 2004 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, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include +#include +#include +#include + +#include +#include +#include + +static const float almost_half = 0.50000006; /* 0.5 + 2^-24 */ +static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; +static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; +static const float two48 = 281474976710656.0; +static const float twom24 = 5.9604644775390625e-8; +extern const float __t_sqrt[1024]; + +/* The method is based on a description in + Computation of elementary functions on the IBM RISC System/6000 processor, + P. W. Markstein, IBM J. Res. Develop, 34(1) 1990. + Basically, it consists of two interleaved Newton-Rhapson approximations, + one to find the actual square root, and one to find its reciprocal + without the expense of a division operation. The tricky bit here + is the use of the POWER/PowerPC multiply-add operation to get the + required accuracy with high speed. + + The argument reduction works by a combination of table lookup to + obtain the initial guesses, and some careful modification of the + generated guesses (which mostly runs on the integer unit, while the + Newton-Rhapson is running on the FPU). */ + +#ifdef __STDC__ +float +__slow_ieee754_sqrtf (float x) +#else +float +__slow_ieee754_sqrtf (x) + float x; +#endif +{ + const float inf = a_inf.value; + + if (x > 0) + { + if (x != inf) + { + /* Variables named starting with 's' exist in the + argument-reduced space, so that 2 > sx >= 0.5, + 1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... . + Variables named ending with 'i' are integer versions of + floating-point values. */ + float sx; /* The value of which we're trying to find the square + root. */ + float sg, g; /* Guess of the square root of x. */ + float sd, d; /* Difference between the square of the guess and x. */ + float sy; /* Estimate of 1/2g (overestimated by 1ulp). */ + float sy2; /* 2*sy */ + float e; /* Difference between y*g and 1/2 (note that e==se). */ + float shx; /* == sx * fsg */ + float fsg; /* sg*fsg == g. */ + fenv_t fe; /* Saved floating-point environment (stores rounding + mode and whether the inexact exception is + enabled). */ + uint32_t xi, sxi, fsgi; + const float *t_sqrt; + + GET_FLOAT_WORD (xi, x); + fe = fegetenv_register (); + relax_fenv_state (); + sxi = (xi & 0x3fffffff) | 0x3f000000; + SET_FLOAT_WORD (sx, sxi); + t_sqrt = __t_sqrt + (xi >> (23 - 8 - 1) & 0x3fe); + sg = t_sqrt[0]; + sy = t_sqrt[1]; + + /* Here we have three Newton-Rhapson iterations each of a + division and a square root and the remainder of the + argument reduction, all interleaved. */ + sd = -(sg * sg - sx); + fsgi = (xi + 0x40000000) >> 1 & 0x7f800000; + sy2 = sy + sy; + sg = sy * sd + sg; /* 16-bit approximation to sqrt(sx). */ + e = -(sy * sg - almost_half); + SET_FLOAT_WORD (fsg, fsgi); + sd = -(sg * sg - sx); + sy = sy + e * sy2; + if ((xi & 0x7f800000) == 0) + goto denorm; + shx = sx * fsg; + sg = sg + sy * sd; /* 32-bit approximation to sqrt(sx), + but perhaps rounded incorrectly. */ + sy2 = sy + sy; + g = sg * fsg; + e = -(sy * sg - almost_half); + d = -(g * sg - shx); + sy = sy + e * sy2; + fesetenv_register (fe); + return g + sy * d; + denorm: + /* For denormalised numbers, we normalise, calculate the + square root, and return an adjusted result. */ + fesetenv_register (fe); + return __slow_ieee754_sqrtf (x * two48) * twom24; + } + } + else if (x < 0) + { + /* For some reason, some PowerPC32 processors don't implement + FE_INVALID_SQRT. */ +#ifdef FE_INVALID_SQRT + feraiseexcept (FE_INVALID_SQRT); + if (!fetestexcept (FE_INVALID)) +#endif + feraiseexcept (FE_INVALID); + x = a_nan.value; + } + return f_washf (x); +} + + +#ifdef __STDC__ +float +__ieee754_sqrtf (float x) +#else +float +__ieee754_sqrtf (x) + float x; +#endif +{ + double z; + + /* If the CPU is 64-bit we can use the optional FP instructions we. */ + if ((GLRO (dl_hwcap) & PPC_FEATURE_64) != 0) + { + /* Volatile is required to prevent the compiler from moving the + fsqrt instruction above the branch. */ + __asm __volatile (" fsqrts %0,%1\n" + :"=f" (z):"f" (x)); + } + else + z = __slow_ieee754_sqrtf (x); + + return z; +} diff --git a/sysdeps/powerpc/fpu/w_sqrt.c b/sysdeps/powerpc/fpu/w_sqrt.c index ff0331725c..806d8e4907 100644 --- a/sysdeps/powerpc/fpu/w_sqrt.c +++ b/sysdeps/powerpc/fpu/w_sqrt.c @@ -1,5 +1,5 @@ -/* Double-precision floating point square root. - Copyright (C) 1997, 2002, 2003 Free Software Foundation, Inc. +/* Double-precision floating point square root wrapper. + Copyright (C) 2004 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 @@ -17,130 +17,35 @@ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. */ -#include -#include +#include "math.h" +#include "math_private.h" #include -#include -static const double almost_half = 0.5000000000000001; /* 0.5 + 2^-53 */ -static const ieee_float_shape_type a_nan = { .word = 0x7fc00000 }; -static const ieee_float_shape_type a_inf = { .word = 0x7f800000 }; -static const float two108 = 3.245185536584267269e+32; -static const float twom54 = 5.551115123125782702e-17; -extern const float __t_sqrt[1024]; - -/* The method is based on a description in - Computation of elementary functions on the IBM RISC System/6000 processor, - P. W. Markstein, IBM J. Res. Develop, 34(1) 1990. - Basically, it consists of two interleaved Newton-Rhapson approximations, - one to find the actual square root, and one to find its reciprocal - without the expense of a division operation. The tricky bit here - is the use of the POWER/PowerPC multiply-add operation to get the - required accuracy with high speed. - - The argument reduction works by a combination of table lookup to - obtain the initial guesses, and some careful modification of the - generated guesses (which mostly runs on the integer unit, while the - Newton-Rhapson is running on the FPU). */ +#ifdef __STDC__ double -__sqrt(double x) -{ - const float inf = a_inf.value; - /* x = f_wash(x); *//* This ensures only one exception for SNaN. */ - if (x > 0) - { - if (x != inf) - { - /* Variables named starting with 's' exist in the - argument-reduced space, so that 2 > sx >= 0.5, - 1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... . - Variables named ending with 'i' are integer versions of - floating-point values. */ - double sx; /* The value of which we're trying to find the - square root. */ - double sg,g; /* Guess of the square root of x. */ - double sd,d; /* Difference between the square of the guess and x. */ - double sy; /* Estimate of 1/2g (overestimated by 1ulp). */ - double sy2; /* 2*sy */ - double e; /* Difference between y*g and 1/2 (se = e * fsy). */ - double shx; /* == sx * fsg */ - double fsg; /* sg*fsg == g. */ - fenv_t fe; /* Saved floating-point environment (stores rounding - mode and whether the inexact exception is - enabled). */ - uint32_t xi0, xi1, sxi, fsgi; - const float *t_sqrt; - - fe = fegetenv_register(); - EXTRACT_WORDS (xi0,xi1,x); - relax_fenv_state(); - sxi = (xi0 & 0x3fffffff) | 0x3fe00000; - INSERT_WORDS (sx, sxi, xi1); - t_sqrt = __t_sqrt + (xi0 >> (52-32-8-1) & 0x3fe); - sg = t_sqrt[0]; - sy = t_sqrt[1]; - - /* Here we have three Newton-Rhapson iterations each of a - division and a square root and the remainder of the - argument reduction, all interleaved. */ - sd = -(sg*sg - sx); - fsgi = (xi0 + 0x40000000) >> 1 & 0x7ff00000; - sy2 = sy + sy; - sg = sy*sd + sg; /* 16-bit approximation to sqrt(sx). */ - INSERT_WORDS (fsg, fsgi, 0); - e = -(sy*sg - almost_half); - sd = -(sg*sg - sx); - if ((xi0 & 0x7ff00000) == 0) - goto denorm; - sy = sy + e*sy2; - sg = sg + sy*sd; /* 32-bit approximation to sqrt(sx). */ - sy2 = sy + sy; - e = -(sy*sg - almost_half); - sd = -(sg*sg - sx); - sy = sy + e*sy2; - shx = sx * fsg; - sg = sg + sy*sd; /* 64-bit approximation to sqrt(sx), - but perhaps rounded incorrectly. */ - sy2 = sy + sy; - g = sg * fsg; - e = -(sy*sg - almost_half); - d = -(g*sg - shx); - sy = sy + e*sy2; - fesetenv_register (fe); - return g + sy*d; - denorm: - /* For denormalised numbers, we normalise, calculate the - square root, and return an adjusted result. */ - fesetenv_register (fe); - return __sqrt(x * two108) * twom54; - } - } - else if (x < 0) - { -#ifdef FE_INVALID_SQRT - feraiseexcept (FE_INVALID_SQRT); - /* For some reason, some PowerPC processors don't implement - FE_INVALID_SQRT. I guess no-one ever thought they'd be - used for square roots... :-) */ - if (!fetestexcept (FE_INVALID)) +__sqrt (double x) /* wrapper sqrt */ +#else +double +__sqrt (x) /* wrapper sqrt */ + double x; #endif - feraiseexcept (FE_INVALID); -#ifndef _IEEE_LIBM - if (_LIB_VERSION != _IEEE_) - x = __kernel_standard(x,x,26); - else +{ +#ifdef _IEEE_LIBM + return __ieee754_sqrt (x); +#else + double z; + z = __ieee754_sqrt (x); + if (_LIB_VERSION == _IEEE_ || (x != x)) + return z; + + if (x < 0.0) + return __kernel_standard (x, x, 26); /* sqrt(negative) */ + else + return z; #endif - x = a_nan.value; - } - return f_wash(x); } weak_alias (__sqrt, sqrt) -/* Strictly, this is wrong, but the only places where _ieee754_sqrt is - used will not pass in a negative result. */ -strong_alias(__sqrt,__ieee754_sqrt) - #ifdef NO_LONG_DOUBLE -weak_alias (__sqrt, __sqrtl) -weak_alias (__sqrt, sqrtl) + strong_alias (__sqrt, __sqrtl) weak_alias (__sqrt, sqrtl) #endif diff --git a/sysdeps/powerpc/fpu/w_sqrtf.c b/sysdeps/powerpc/fpu/w_sqrtf.c index 8eb94d8e3b..e3f3c995e8 100644 --- a/sysdeps/powerpc/fpu/w_sqrtf.c +++ b/sysdeps/powerpc/fpu/w_sqrtf.c @@ -1,5 +1,5 @@ -/* Single-precision floating point square root. - Copyright (C) 1997, 2003 Free Software Foundation, Inc. +/* Single-precision floating point square root wrapper. + Copyright (C) 2004 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 @@ -17,120 +17,38 @@ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. */ -#include -#include +#include "math.h" +#include "math_private.h" #include -#include -static const float almost_half = 0.50000006; /* 0.5 + 2^-24 */ -static const ieee_float_shape_type a_nan = { .word = 0x7fc00000 }; -static const ieee_float_shape_type a_inf = { .word = 0x7f800000 }; -static const float two48 = 281474976710656.0; -static const float twom24 = 5.9604644775390625e-8; -extern const float __t_sqrt[1024]; +#include +#include +#include -/* The method is based on a description in - Computation of elementary functions on the IBM RISC System/6000 processor, - P. W. Markstein, IBM J. Res. Develop, 34(1) 1990. - Basically, it consists of two interleaved Newton-Rhapson approximations, - one to find the actual square root, and one to find its reciprocal - without the expense of a division operation. The tricky bit here - is the use of the POWER/PowerPC multiply-add operation to get the - required accuracy with high speed. - - The argument reduction works by a combination of table lookup to - obtain the initial guesses, and some careful modification of the - generated guesses (which mostly runs on the integer unit, while the - Newton-Rhapson is running on the FPU). */ +#ifdef __STDC__ float -__sqrtf(float x) -{ - const float inf = a_inf.value; - /* x = f_washf(x); *//* This ensures only one exception for SNaN. */ - if (x > 0) - { - if (x != inf) - { - /* Variables named starting with 's' exist in the - argument-reduced space, so that 2 > sx >= 0.5, - 1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... . - Variables named ending with 'i' are integer versions of - floating-point values. */ - float sx; /* The value of which we're trying to find the square - root. */ - float sg,g; /* Guess of the square root of x. */ - float sd,d; /* Difference between the square of the guess and x. */ - float sy; /* Estimate of 1/2g (overestimated by 1ulp). */ - float sy2; /* 2*sy */ - float e; /* Difference between y*g and 1/2 (note that e==se). */ - float shx; /* == sx * fsg */ - float fsg; /* sg*fsg == g. */ - fenv_t fe; /* Saved floating-point environment (stores rounding - mode and whether the inexact exception is - enabled). */ - uint32_t xi, sxi, fsgi; - const float *t_sqrt; - - GET_FLOAT_WORD (xi, x); - fe = fegetenv_register (); - relax_fenv_state (); - sxi = (xi & 0x3fffffff) | 0x3f000000; - SET_FLOAT_WORD (sx, sxi); - t_sqrt = __t_sqrt + (xi >> (23-8-1) & 0x3fe); - sg = t_sqrt[0]; - sy = t_sqrt[1]; - - /* Here we have three Newton-Rhapson iterations each of a - division and a square root and the remainder of the - argument reduction, all interleaved. */ - sd = -(sg*sg - sx); - fsgi = (xi + 0x40000000) >> 1 & 0x7f800000; - sy2 = sy + sy; - sg = sy*sd + sg; /* 16-bit approximation to sqrt(sx). */ - e = -(sy*sg - almost_half); - SET_FLOAT_WORD (fsg, fsgi); - sd = -(sg*sg - sx); - sy = sy + e*sy2; - if ((xi & 0x7f800000) == 0) - goto denorm; - shx = sx * fsg; - sg = sg + sy*sd; /* 32-bit approximation to sqrt(sx), - but perhaps rounded incorrectly. */ - sy2 = sy + sy; - g = sg * fsg; - e = -(sy*sg - almost_half); - d = -(g*sg - shx); - sy = sy + e*sy2; - fesetenv_register (fe); - return g + sy*d; - denorm: - /* For denormalised numbers, we normalise, calculate the - square root, and return an adjusted result. */ - fesetenv_register (fe); - return __sqrtf(x * two48) * twom24; - } - } - else if (x < 0) - { -#ifdef FE_INVALID_SQRT - feraiseexcept (FE_INVALID_SQRT); - /* For some reason, some PowerPC processors don't implement - FE_INVALID_SQRT. I guess no-one ever thought they'd be - used for square roots... :-) */ - if (!fetestexcept (FE_INVALID)) +__sqrtf (float x) /* wrapper sqrtf */ +#else +float +__sqrtf (x) /* wrapper sqrtf */ + float x; #endif - feraiseexcept (FE_INVALID); -#ifndef _IEEE_LIBM - if (_LIB_VERSION != _IEEE_) - x = __kernel_standard(x,x,126); - else +{ +#ifdef _IEEE_LIBM + return __ieee754_sqrtf (x); +#else + float z; + z = __ieee754_sqrtf (x); + + if (_LIB_VERSION == _IEEE_ || (x != x)) + return z; + + if (x < (float) 0.0) + /* sqrtf(negative) */ + return (float) __kernel_standard ((double) x, (double) x, 126); + else + return z; #endif - x = a_nan.value; - } - return f_washf(x); } weak_alias (__sqrtf, sqrtf) -/* Strictly, this is wrong, but the only places where _ieee754_sqrt is - used will not pass in a negative result. */ -strong_alias(__sqrtf,__ieee754_sqrtf) diff --git a/sysdeps/powerpc/powerpc32/sysdep.h b/sysdeps/powerpc/powerpc32/sysdep.h index 43179d14f0..59761f75b9 100644 --- a/sysdeps/powerpc/powerpc32/sysdep.h +++ b/sysdeps/powerpc/powerpc32/sysdep.h @@ -17,10 +17,10 @@ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. */ -#ifdef __ASSEMBLER__ - #include +#ifdef __ASSEMBLER__ + #ifdef __ELF__ /* If compiled for profiling, call `_mcount' at the start of each diff --git a/sysdeps/powerpc/sysdep.h b/sysdeps/powerpc/sysdep.h index 214ac6a2ed..fb668ad51f 100644 --- a/sysdeps/powerpc/sysdep.h +++ b/sysdeps/powerpc/sysdep.h @@ -16,6 +16,20 @@ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. */ +/* + * Powerpc Feature masks for the Aux Vector Hardware Capabilities (AT_HWCAP). + * This entry is copied to _dl_hwcap or rtld_global._dl_hwcap during startup. + * The following must match the kernels linux/asm/cputable.h. + */ +#define PPC_FEATURE_32 0x80000000 /* 32-bit mode. */ +#define PPC_FEATURE_64 0x40000000 /* 64-bit mode. */ +#define PPC_FEATURE_601_INSTR 0x20000000 /* 601 chip, Old POWER ISA. */ +#define PPC_FEATURE_HAS_ALTIVEC 0x10000000 /* SIMD/Vector Unit. */ +#define PPC_FEATURE_HAS_FPU 0x08000000 /* Floating Point Unit. */ +#define PPC_FEATURE_HAS_MMU 0x04000000 /* Memory Management Unit. */ +#define PPC_FEATURE_HAS_4xxMAC 0x02000000 /* 4xx Multiply Accumulator. */ +#define PPC_FEATURE_UNIFIED_CACHE 0x01000000 /* Unified I/D cache. */ + #ifdef __ASSEMBLER__ /* Symbolic names for the registers. The only portable way to write asm @@ -146,19 +160,5 @@ #define ASM_SIZE_DIRECTIVE(name) .size name,.-name #endif /* __ELF__ */ - -/* - * Powerpc Feature masks for the Aux Vector Hardware Capabilities (AT_HWCAP). - * This entry is copied to _dl_hwcap or rtld_global._dl_hwcap during startup. - * The following must match the kernels linux/asm/cputable.h. - */ -#define PPC_FEATURE_32 0x80000000 /* 32-bit mode. */ -#define PPC_FEATURE_64 0x40000000 /* 64-bit mode. */ -#define PPC_FEATURE_601_INSTR 0x20000000 /* 601 chip, Old POWER ISA. */ -#define PPC_FEATURE_HAS_ALTIVEC 0x10000000 /* SIMD/Vector Unit. */ -#define PPC_FEATURE_HAS_FPU 0x08000000 /* Floating Point Unit. */ -#define PPC_FEATURE_HAS_MMU 0x04000000 /* Memory Management Unit. */ -#define PPC_FEATURE_HAS_4xxMAC 0x02000000 /* 4xx Multiply Accumulator. */ -#define PPC_FEATURE_UNIFIED_CACHE 0x01000000 /* Unified I/D cache. */ - #endif /* __ASSEMBLER__ */ + diff --git a/sysdeps/pthread/aio_misc.c b/sysdeps/pthread/aio_misc.c index c2eb674212..1da3ad2c73 100644 --- a/sysdeps/pthread/aio_misc.c +++ b/sysdeps/pthread/aio_misc.c @@ -27,8 +27,27 @@ #include #include #include +#include -#include "aio_misc.h" +#ifndef aio_create_helper_thread +# define aio_create_helper_thread __aio_create_helper_thread + +extern inline int +__aio_create_helper_thread (pthread_t *threadp, void *(*tf) (void *), void *arg) +{ + pthread_attr_t attr; + + /* Make sure the thread is created detached. */ + pthread_attr_init (&attr); + pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); + + int ret = pthread_create (threadp, &attr, tf, arg); + + (void) pthread_attr_destroy (&attr); + return ret; +} + +#endif static void add_request_to_runlist (struct requestlist *newrequest); @@ -400,16 +419,11 @@ __aio_enqueue_request (aiocb_union *aiocbp, int operation) if (nthreads < optim.aio_threads && idle_thread_count == 0) { pthread_t thid; - pthread_attr_t attr; - - /* Make sure the thread is created detached. */ - pthread_attr_init (&attr); - pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED); running = newp->running = allocated; /* Now try to start a thread. */ - if (pthread_create (&thid, &attr, handle_fildes_io, newp) == 0) + if (aio_create_helper_thread (&thid, handle_fildes_io, newp) == 0) /* We managed to enqueue the request. All errors which can happen now can be recognized by calls to `aio_return' and `aio_error'. */ diff --git a/sysdeps/pthread/aio_notify.c b/sysdeps/pthread/aio_notify.c index d1c0a63fbf..877e8d9363 100644 --- a/sysdeps/pthread/aio_notify.c +++ b/sysdeps/pthread/aio_notify.c @@ -1,5 +1,6 @@ /* Notify initiator of AIO request. - Copyright (C) 1997,98,99,2000,2001,2003 Free Software Foundation, Inc. + Copyright (C) 1997, 1998, 1999, 2000, 2001, 2003, 2004 + Free Software Foundation, Inc. This file is part of the GNU C Library. Contributed by Ulrich Drepper , 1997. @@ -22,8 +23,11 @@ #include #include #include -#include "aio_misc.h" +#include +#ifndef aio_start_notify_thread +# define aio_start_notify_thread() do { } while (0) +#endif struct notify_func { @@ -34,6 +38,7 @@ struct notify_func static void * notify_func_wrapper (void *arg) { + aio_start_notify_thread (); struct notify_func *const n = arg; void (*func) (sigval_t) = n->func; sigval_t value = n->value; -- cgit 1.4.1