aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/powerpc
diff options
context:
space:
mode:
authorUlrich Drepper <drepper@redhat.com>1997-08-29 01:19:12 +0000
committerUlrich Drepper <drepper@redhat.com>1997-08-29 01:19:12 +0000
commit39e16978c3b4ac8eaf2201fac56316623910d9da (patch)
tree054fff18119b31464b3133ad91050694130c7d2a /sysdeps/powerpc
parent92f1da4da04a7a86ddee91be5eaf0b10c333ac64 (diff)
downloadglibc-39e16978c3b4ac8eaf2201fac56316623910d9da.tar
glibc-39e16978c3b4ac8eaf2201fac56316623910d9da.tar.gz
glibc-39e16978c3b4ac8eaf2201fac56316623910d9da.tar.bz2
glibc-39e16978c3b4ac8eaf2201fac56316623910d9da.zip
1997-08-29 02:36 Ulrich Drepper <drepper@cygnus.com> * Makefile (version-info.h): Use ISO form for the date. * catgets/catgetsinfo.h: Include <bits/libc-lock.h>. (struct catalog_obj): Add lock field. (__open_catalog): Remove second parameter from prototype. * catgets/catgets.c (catopen): Initialize lock field. (catgets): Don't pass second parameter to __open_catalog. * catgets/gencat.c: Initialize lock field and don't pass second parameter to __open_catalog. * catgets/open_catalog.c (__open_catalog): Decide about use of path by examining path in struct, not based on extra argument. Acquire a the lock before trying to load the catalog and release it before returning. * csu/Makefile (abi-tag.h): Make sure target directory exists. * io/Makefile (headers): Add bits/poll.h. * io/sys/poll.h: Remove definitions of POLL* constants. Include <bits/poll.h>. * sysdeps/generic/bits/poll.h: New file. * sysdeps/unix/sysv/linux/bits/poll.h: New file. * sysdeps/unix/sysv/linux/m68k/bits/poll.h: New file. * sysdeps/unix/sysv/linux/mips/bits/poll.h: New file. * sysdeps/unix/sysv/linux/sparc/bits/poll.h: New file. * libio/fileops.c (_IO_file_read, _IO_file_write): Remove dead code. * malloc/obstack.c: Add casts to keep very verbose compilers on 64bit machine quiet. * nss/Makefile (libnss_db.so): Find libdb.so in db2 directory. 1997-08-28 17:30 Ulrich Drepper <drepper@cygnus.com> * catgets/catgets.c (catopen): Correctly determine length of string in NLSPATH evironment variable. Patch by HJ Lu <hjl@gnu.ai.mit.edu>. 1997-08-27 23:19 Richard Henderson <rth@cygnus.com> * sysdeps/generic/dl-sysdep.c (DL_FIND_ARG_COMPONENTS): Provide default macro to track down arguments from stack start. (_dl_sysdep_start): Use it. * sysdeps/unix/sysv/linux/powerpc/dl-sysdep.c: Truncate to simply providing a special DL_FIND_ARG_COMPONENTS and including the next file up the line. * sysdeps/powerpc/e_sqrt.c: Move contents to w_sqrt.c and provide stub. * sysdeps/powerpc/e_sqrtf.c: Likewise. * sysdeps/powerpc/s_copysignf.S: Provide empty file; symbol is with the double precision version. * sysdeps/powerpc/s_fabsf.S: Likewise. * sysdeps/powerpc/s_isnanf.S: Likewise.
Diffstat (limited to 'sysdeps/powerpc')
-rw-r--r--sysdeps/powerpc/e_sqrt.c142
-rw-r--r--sysdeps/powerpc/e_sqrtf.c137
-rw-r--r--sysdeps/powerpc/s_copysignf.S1
-rw-r--r--sysdeps/powerpc/s_fabsf.S1
-rw-r--r--sysdeps/powerpc/s_isnanf.S1
-rw-r--r--sysdeps/powerpc/w_sqrt.c141
-rw-r--r--sysdeps/powerpc/w_sqrtf.c136
7 files changed, 282 insertions, 277 deletions
diff --git a/sysdeps/powerpc/e_sqrt.c b/sysdeps/powerpc/e_sqrt.c
index df80973f58..9416ea60c8 100644
--- a/sysdeps/powerpc/e_sqrt.c
+++ b/sysdeps/powerpc/e_sqrt.c
@@ -1,141 +1 @@
-/* Single-precision floating point square root.
- Copyright (C) 1997 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 Library General Public License as
- published by the Free Software Foundation; either version 2 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
- Library General Public License for more details.
-
- You should have received a copy of the GNU Library 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. */
-
-#include <math.h>
-#include <math_private.h>
-#include <fenv_libc.h>
-#include <inttypes.h>
-
-static const double almost_half = 0.5000000000000001; /* 0.5 + 2^-53 */
-static const uint32_t a_nan = 0x7fc00000;
-static const uint32_t a_inf = 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). */
-double
-__sqrt(double x)
-{
- const float inf = *(const float *)&a_inf;
- /* 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))
-#endif
- feraiseexcept (FE_INVALID);
-#ifndef _IEEE_LIBM
- if (_LIB_VERSION != _IEEE_)
- x = __kernel_standard(x,x,26);
- else
-#endif
- x = *(const float*)&a_nan;
- }
- 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)
+/* __ieee754_sqrt is in w_sqrt.c */
diff --git a/sysdeps/powerpc/e_sqrtf.c b/sysdeps/powerpc/e_sqrtf.c
index 804dff3c44..01c76d6757 100644
--- a/sysdeps/powerpc/e_sqrtf.c
+++ b/sysdeps/powerpc/e_sqrtf.c
@@ -1,136 +1 @@
-/* Single-precision floating point square root.
- Copyright (C) 1997 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 Library General Public License as
- published by the Free Software Foundation; either version 2 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
- Library General Public License for more details.
-
- You should have received a copy of the GNU Library 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. */
-
-#include <math.h>
-#include <math_private.h>
-#include <fenv_libc.h>
-#include <inttypes.h>
-
-static const float almost_half = 0.50000006; /* 0.5 + 2^-24 */
-static const uint32_t a_nan = 0x7fc00000;
-static const uint32_t a_inf = 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). */
-float
-__sqrtf(float x)
-{
- const float inf = *(const float *)&a_inf;
- /* 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))
-#endif
- feraiseexcept (FE_INVALID);
-#ifndef _IEEE_LIBM
- if (_LIB_VERSION != _IEEE_)
- x = __kernel_standard(x,x,126);
- else
-#endif
- x = *(const float*)&a_nan;
- }
- 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)
+/* __ieee754_sqrtf is in w_sqrtf.c */
diff --git a/sysdeps/powerpc/s_copysignf.S b/sysdeps/powerpc/s_copysignf.S
new file mode 100644
index 0000000000..e05438ae7d
--- /dev/null
+++ b/sysdeps/powerpc/s_copysignf.S
@@ -0,0 +1 @@
+/* __copysignf is in s_copysign.S */
diff --git a/sysdeps/powerpc/s_fabsf.S b/sysdeps/powerpc/s_fabsf.S
new file mode 100644
index 0000000000..877c710ce8
--- /dev/null
+++ b/sysdeps/powerpc/s_fabsf.S
@@ -0,0 +1 @@
+/* __fabsf is in s_fabs.S */
diff --git a/sysdeps/powerpc/s_isnanf.S b/sysdeps/powerpc/s_isnanf.S
new file mode 100644
index 0000000000..fc22f678a1
--- /dev/null
+++ b/sysdeps/powerpc/s_isnanf.S
@@ -0,0 +1 @@
+/* __isnanf is in s_isnan.c */
diff --git a/sysdeps/powerpc/w_sqrt.c b/sysdeps/powerpc/w_sqrt.c
new file mode 100644
index 0000000000..df80973f58
--- /dev/null
+++ b/sysdeps/powerpc/w_sqrt.c
@@ -0,0 +1,141 @@
+/* Single-precision floating point square root.
+ Copyright (C) 1997 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 Library General Public License as
+ published by the Free Software Foundation; either version 2 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
+ Library General Public License for more details.
+
+ You should have received a copy of the GNU Library 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. */
+
+#include <math.h>
+#include <math_private.h>
+#include <fenv_libc.h>
+#include <inttypes.h>
+
+static const double almost_half = 0.5000000000000001; /* 0.5 + 2^-53 */
+static const uint32_t a_nan = 0x7fc00000;
+static const uint32_t a_inf = 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). */
+double
+__sqrt(double x)
+{
+ const float inf = *(const float *)&a_inf;
+ /* 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))
+#endif
+ feraiseexcept (FE_INVALID);
+#ifndef _IEEE_LIBM
+ if (_LIB_VERSION != _IEEE_)
+ x = __kernel_standard(x,x,26);
+ else
+#endif
+ x = *(const float*)&a_nan;
+ }
+ 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)
diff --git a/sysdeps/powerpc/w_sqrtf.c b/sysdeps/powerpc/w_sqrtf.c
new file mode 100644
index 0000000000..804dff3c44
--- /dev/null
+++ b/sysdeps/powerpc/w_sqrtf.c
@@ -0,0 +1,136 @@
+/* Single-precision floating point square root.
+ Copyright (C) 1997 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 Library General Public License as
+ published by the Free Software Foundation; either version 2 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
+ Library General Public License for more details.
+
+ You should have received a copy of the GNU Library 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. */
+
+#include <math.h>
+#include <math_private.h>
+#include <fenv_libc.h>
+#include <inttypes.h>
+
+static const float almost_half = 0.50000006; /* 0.5 + 2^-24 */
+static const uint32_t a_nan = 0x7fc00000;
+static const uint32_t a_inf = 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). */
+float
+__sqrtf(float x)
+{
+ const float inf = *(const float *)&a_inf;
+ /* 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))
+#endif
+ feraiseexcept (FE_INVALID);
+#ifndef _IEEE_LIBM
+ if (_LIB_VERSION != _IEEE_)
+ x = __kernel_standard(x,x,126);
+ else
+#endif
+ x = *(const float*)&a_nan;
+ }
+ 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)