aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/powerpc/strlen.s
diff options
context:
space:
mode:
authorUlrich Drepper <drepper@redhat.com>1997-04-02 14:47:34 +0000
committerUlrich Drepper <drepper@redhat.com>1997-04-02 14:47:34 +0000
commit22d57dd3690a0fe623de1a56036306a93fa9a945 (patch)
tree72d15a44833db599574dc62e78ea02e8d30567e3 /sysdeps/powerpc/strlen.s
parent5929563ff20245bbaa1b3ca2f4bfcee623ac6fb5 (diff)
downloadglibc-22d57dd3690a0fe623de1a56036306a93fa9a945.tar
glibc-22d57dd3690a0fe623de1a56036306a93fa9a945.tar.gz
glibc-22d57dd3690a0fe623de1a56036306a93fa9a945.tar.bz2
glibc-22d57dd3690a0fe623de1a56036306a93fa9a945.zip
Update.
1997-04-02 16:28 Ulrich Drepper <drepper@cygnus.com> * grp/fgetgrent.c: Don't use fixed buffer length. Allow dynamic adjustment if the underlying function say it is too small. * pwd/fgetpwent.c: Likewise. * shadow/fgetspent.c: Likewise. * shadow/sgetspent.c: Likewise. * grp/fgetgrent_r.c (__fgetgrent_r): If `parse_result' returns -1 this means there was not enough room for the gr_mem array. * nss/nss_files/files-XXX.c: Leave loop and return error if `parse_result' returns -1. * nss/nss_files/files-parse.c (parse_line): Return -1 if parsing of trailing list fails because of low memory. (parse_list): Rewrite. The logic was completly incorrect since the checks for overflowing the array were not executed when needed. * shadow/sgetspent_r.c: Only a return value > 0 from parse_line signals success. * nss/getXXbyYY.c: Make sure NULL is returned if we ran out of memory. * nss/getXXent.c: Likewise. * stdlib/strtod.c (SET_MANTISSA): Make sure that at least one bit is set in mantissa. * stdlib/strtof.c: Likewise. * stdlib/strtold.c: Likewise. * wcsmbs/wcstof.c: Likewise. * wcsmbs/wcstold.c: Likewise. * sysdeps/libm-ieee754/s_nan.c: Use strtod is parameter is not empty string. * sysdeps/libm-ieee754/s_nanf.c: Likewise. * sysdeps/libm-ieee754/s_nanld.c: Likewise. 1997-04-02 13:56 Ulrich Drepper <drepper@cygnus.com> * manual/errno.texi: Add description of strerror_r. 1997-04-01 19:07 H.J. Lu (hjl@gnu.ai.mit.edu) * manual/errno.texi: Add ENOMEDIUM and EMEDIUMTYPE. 1997-04-02 13:55 Ulrich Drepper <drepper@cygnus.com> * sysdeps/gnu/errlist.awk: The Texinfo file is named errno.texi. 1997-03-30 10:59 Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> * sysdeps/m68k/fpu/s_cexp.c: New file. * sysdeps/m68k/fpu/s_cexpf.c: New file. * sysdeps/m68k/fpu/s_cexpl.c: New file. * sysdeps/m68k/fpu/s_nearbyint.c: New file. * sysdeps/m68k/fpu/s_nearbyintf.c: New file. * sysdeps/m68k/fpu/s_nearbyintl.c: New file. * sysdeps/m68k/fpu/s_trunc.c: New file. * sysdeps/m68k/fpu/s_truncf.c: New file. * sysdeps/m68k/fpu/s_truncl.c: New file. * sysdeps/m68k/fpu/s_modf.c: Rewritten. * sysdeps/m68k/fpu/s_modff.c: Adapted. * sysdeps/m68k/fpu/s_modfl.c: Likewise. * sysdeps/m68k/fpu/e_atan2.c: Use __signbit inline. * sysdeps/m68k/fpu/e_pow.c: Likewise. Write 0.0/0.0 for NaN. Fix typo in test for `y = 0.5'. * sysdeps/m68k/fpu/__math.h: Define __trunc, __signbit, __nearbyint. Remove __modf. * sysdeps/m68k/Makefile (CFLAGS-s_copysignl.c) [$(subdir)=math]: Define for gcc bug workaround. * math/libm-test.c (log2_test): Add epsilon for long double to "log2 (e)" and "log2 (16)". (modf_test): Fix test name. (sqrt_test): Add missing FUNC. 1997-03-29 16:24 Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> * sunrpc/Makefile (CFLAGS-xrquota.c): Fix typo in variable name. 1997-03-28 20:56 Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> * stdlib/stdlib.h: Declare __clearenv. 1997-03-28 18:03 Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> * sysdeps/libm-ieee754/s_asinhl.c: Fix Ulrich Dreppers mangling of the last change. 1997-03-28 17:55 Andreas Schwab <schwab@issan.informatik.uni-dortmund.de> * configure.in: Call AC_CANONICAL_BUILD before referencing $build. Fix test for BUILD_CC. 1997-04-02 01:29 Philip Blundell <pjb27@cam.ac.uk> * sysdeps/libm-ieee754/s_log2f.c (__log2f): Fix typo. 1997-03-31 18:53 Philip Blundell <pjb27@cam.ac.uk> * sysdeps/unix/sysv/linux/net/route.h: Add new IPv6 definitions. 1997-04-01 14:08 Ulrich Drepper <drepper@cygnus.com> * sysdeps/i386/memset.c: Fix typo. * sysdeps/stub/fchmod.c: Fix typo. * sysdeps/stub/fchdir.c: Fix typo. Reported by Fila Kolodny <fila@ibi.com>.
Diffstat (limited to 'sysdeps/powerpc/strlen.s')
-rw-r--r--sysdeps/powerpc/strlen.s145
1 files changed, 145 insertions, 0 deletions
diff --git a/sysdeps/powerpc/strlen.s b/sysdeps/powerpc/strlen.s
new file mode 100644
index 0000000000..eb6a88e2b8
--- /dev/null
+++ b/sysdeps/powerpc/strlen.s
@@ -0,0 +1,145 @@
+ # Optimized strlen implementation for PowerPC.
+ # 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.
+
+ # The algorithm here uses the following techniques:
+ #
+ # 1) Given a word 'x', we can test to see if it contains any 0 bytes
+ # by subtracting 0x01010101, and seeing if any of the high bits of each
+ # byte changed from 0 to 1. This works because the least significant
+ # 0 byte must have had no incoming carry (otherwise it's not the least
+ # significant), so it is 0x00 - 0x01 == 0xff. For all other
+ # byte values, either they have the high bit set initially, or when
+ # 1 is subtracted you get a value in the range 0x00-0x7f, none of which
+ # have their high bit set. The expression here is
+ # (x + 0xfefefeff) & ~(x | 0x7f7f7f7f), which gives 0x00000000 when
+ # there were no 0x00 bytes in the word.
+ #
+ # 2) Given a word 'x', we can test to see _which_ byte was zero by
+ # calculating ~(((x & 0x7f7f7f7f) + 0x7f7f7f7f) | x | 0x7f7f7f7f).
+ # This produces 0x80 in each byte that was zero, and 0x00 in all
+ # the other bytes. The '| 0x7f7f7f7f' clears the low 7 bits in each
+ # byte, and the '| x' part ensures that bytes with the high bit set
+ # produce 0x00. The addition will carry into the high bit of each byte
+ # iff that byte had one of its low 7 bits set. We can then just see
+ # which was the most significant bit set and divide by 8 to find how
+ # many to add to the index.
+ # This is from the book 'The PowerPC Compiler Writer's Guide',
+ # by Steve Hoxey, Faraydon Karim, Bill Hay and Hank Warren.
+ #
+ # We deal with strings not aligned to a word boundary by taking the
+ # first word and ensuring that bytes not part of the string
+ # are treated as nonzero. To allow for memory latency, we unroll the
+ # loop a few times, being careful to ensure that we do not read ahead
+ # across cache line boundaries.
+ #
+ # Questions to answer:
+ # 1) How long are strings passed to strlen? If they're often really long,
+ # we should probably use cache management instructions and/or unroll the
+ # loop more. If they're often quite short, it might be better to use
+ # fact (2) in the inner loop than have to recalculate it.
+ # 2) How popular are bytes with the high bit set? If they are very rare,
+ # on some processors it might be useful to use the simpler expression
+ # ~((x - 0x01010101) | 0x7f7f7f7f) (that is, on processors with only one
+ # ALU), but this fails when any character has its high bit set.
+
+ # Some notes on register usage: Under the SVR4 ABI, we can use registers
+ # 0 and 3 through 12 (so long as we don't call any procedures) without
+ # saving them. We can also use registers 14 through 31 if we save them.
+ # We can't use r1 (it's the stack pointer), nor r2 or r13 because the user
+ # program may expect them to be hold their usual value if we get sent
+ # a signal. Integer parameters are passed in r3 through r10.
+ # We can use condition registers cr0, cr1, cr5, cr6, and cr7 without saving
+ # them, the others we must save.
+
+ .section ".text"
+ .align 3
+ .globl strlen
+ .type strlen,@function
+strlen:
+ # On entry, r3 points to the string, and it's left that way.
+ # We use r6 to store 0x01010101, and r7 to store 0x7f7f7f7f.
+ # r4 is used to keep the current index into the string; r5 holds
+ # the number of padding bits we prepend to the string to make it
+ # start at a word boundary. r8 holds the 'current' word.
+ # r9-12 are temporaries. r0 is used as a temporary and for discarded
+ # results.
+ clrrwi 4,3,2
+ lis 6,0xfeff
+ lis 7,0x7f7f
+ rlwinm 10,3,0,29,29
+ lwz 8,0(4)
+ addi 7,7,0x7f7f
+ rlwinm 5,3,3,27,28
+ cmpwi 1,10,0
+ li 9,-1
+ # That's the setup done, now do the first pair of words.
+ # We make an exception and use method (2) on the first two words, to reduce
+ # overhead.
+ srw 9,9,5
+ and 0,7,8
+ or 10,7,8
+ add 0,0,7
+ nor 0,10,0
+ and. 8,0,9
+ bne done0
+ # Handle second word of pair. Put addi between branches to avoid hurting
+ # branch prediction.
+ addi 6,6,0xfffffeff
+
+ bne 1,loop
+ lwzu 8,4(4)
+ and 0,7,8
+ or 10,7,8
+ add 0,0,7
+ nor. 0,10,0
+ bne done0
+
+ # The loop.
+
+loop: lwz 8,4(4)
+ lwzu 9,8(4)
+ add 0,6,8
+ nor 10,7,8
+ and. 0,0,10
+ add 11,6,9
+ nor 12,7,9
+ bne done1
+ and. 0,11,12
+ beq loop
+
+ and 0,7,9
+ or 10,7,9
+ b done2
+
+done1: addi 4,4,-4
+ and 0,7,9
+ or 10,7,9
+done2: add 0,0,7
+ nor 0,10,0
+
+ # When we get to here, r4 points to the first word in the string that
+ # contains a zero byte, and the most significant set bit in r8 is in that
+ # byte.
+done0: cntlzw 11,8
+ subf 0,3,4
+ srwi 11,11,3
+ add 3,0,11
+ blr
+0:
+ .size strlen,0b-strlen