aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/i386/memchr.S
diff options
context:
space:
mode:
authorRoland McGrath <roland@gnu.org>1995-10-16 01:37:51 +0000
committerRoland McGrath <roland@gnu.org>1995-10-16 01:37:51 +0000
commit8f5ca04bc7fd53741d80117df992995ace8f6d2d (patch)
treee39c13fc198b22ec55647259a8080051988e8c69 /sysdeps/i386/memchr.S
parent5d82cf5c55f56ae10d3b0a205d1fcc7de1cf56a0 (diff)
downloadglibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.tar
glibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.tar.gz
glibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.tar.bz2
glibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.zip
Sat Oct 14 02:52:36 1995 Ulrich Drepper <drepper@ipd.info.uni-karlsruhe.de>
* malloc/malloc.c (_malloc_internal): Performance fix. Move if statement out of loop. * stdio/_itoa.c, stdio/_itoa.h: Complete rewrite. Much faster implementation using GMP functions. Contributed by Torbjorn Granlund and Ulrich Drepper. * stdio/test_rdwr.c: Include <errno.h>. * sysdeps/i386/i586/Implies: New file. New highly optimized string functions for i[345]86. * sysdeps/i386/memchr.S, sysdeps/i386/memcmp.S: New files. * sysdeps/i386/stpcpy.S, sysdeps/i386/stpncpy.S: New files. * sysdeps/i386/strchr.S, sysdeps/i386/strcspn.S: New files. * sysdeps/i386/strpbrk.S, sysdeps/i386/strrchr.S: New files. * sysdeps/i386/strspn.S, sysdeps/i386/i486/strcat.S: New files. * sysdeps/i386/i486/strlen.S, sysdeps/i386/i586/strchr.S: New files. * sysdeps/i386/i586/strlen.S: New file. * sysdeps/i386/memchr.c: Removed. There is now an assembler version. * sysdeps/i386/i586/memcopy.h (WORD_COPY_BWD): Parameters did not correspond to used values. * sysdeps/unix/sysv/linux/nfs/nfs.h: New file. Simply a wrapper around a kernel header file. * sysdeps/unix/sysv/linux/Dist: Add it. * sysdeps/unix/sysv/linux/Makefile [$(subdir)=sunrpc] (headers): Likewise. * sysdeps/unix/sysv/linux/local_lim.h: Rewrite. Instead of defining ourself we use a kernel header file. * sysdeps/unix/sysv/linux/i386/sysdep.h (DO_CALL): Optimize system call handler for i586. * sysdeps/unix/sysv/linux/sys/param.h: Add copyright and clean up. Sat Oct 14 02:52:36 1995 Ulrich Drepper <drepper@ipd.info.uni-karlsruhe.de> * malloc/malloc.c (_malloc_internal): Performance fix. Move if statement out of loop. * stdio/_itoa.c, stdio/_itoa.h: Complete rewrite. Much faster implementation using GMP functions. Contributed by Torbjorn Granlund and Ulrich Drepper. * stdio/test_rdwr.c: Include <errno.h>. * sysdeps/i386/i586/Implies: New file. New highly optimized string functions for i[345]86. * sysdeps/i386/memchr.S, sysdeps/i386/memcmp.S: New files. * sysdeps/i386/stpcpy.S, sysdeps/i386/stpncpy.S: New files. * sysdeps/i386/strchr.S, sysdeps/i386/strcspn.S: New files. * sysdeps/i386/strpbrk.S, sysdeps/i386/strrchr.S: New files. * sysdeps/i386/strspn.S, sysdeps/i386/i486/strcat.S: New files. * sysdeps/i386/i486/strlen.S, sysdeps/i386/i586/strchr.S: New files. * sysdeps/i386/i586/strlen.S: New file. * sysdeps/i386/memchr.c: Removed. There is now an assembler version. * sysdeps/i386/i586/memcopy.h (WORD_COPY_BWD): Parameters did not correspond to used values. * sysdeps/unix/sysv/linux/nfs/nfs.h: New file. Simply a wrapper around a kernel header file. * sysdeps/unix/sysv/linux/Dist: Add it. * sysdeps/unix/sysv/linux/Makefile [$(subdir)=sunrpc] (headers): Likewise. * sysdeps/unix/sysv/linux/local_lim.h: Rewrite. Instead of defining ourself we use a kernel header file. * sysdeps/unix/sysv/linux/i386/sysdep.h (DO_CALL): Optimize system call handler for i586. * sysdeps/unix/sysv/linux/sys/param.h: Add copyright and clean up.
Diffstat (limited to 'sysdeps/i386/memchr.S')
-rw-r--r--sysdeps/i386/memchr.S315
1 files changed, 315 insertions, 0 deletions
diff --git a/sysdeps/i386/memchr.S b/sysdeps/i386/memchr.S
new file mode 100644
index 0000000000..9931f97241
--- /dev/null
+++ b/sysdeps/i386/memchr.S
@@ -0,0 +1,315 @@
+/* memchr (str, ch, n) -- Return pointer to first occurrence of CH in STR less
+ than N.
+For Intel 80x86, x>=3.
+Copyright (C) 1994, 1995 Free Software Foundation, Inc.
+Contributed by Ulrich Drepper <drepper@gnu.ai.mit.edu>
+Optimised a little by Alan Modra <Alan@SPRI.Levels.UniSA.Edu.Au>
+This file is part of the GNU C Library.
+
+This version is developed using the same algorithm as the fast C
+version which carries the following introduction:
+
+Based on strlen implemention by Torbjorn Granlund (tege@sics.se),
+with help from Dan Sahlin (dan@sics.se) and
+commentary by Jim Blandy (jimb@ai.mit.edu);
+adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu),
+and implemented by Roland McGrath (roland@ai.mit.edu).
+
+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 <sysdep.h>
+#include "asm-syntax.h"
+
+/*
+ INPUT PARAMETERS:
+ str (sp + 4)
+ c (sp + 8)
+ len (sp + 12)
+*/
+
+ .text
+ENTRY (memchr)
+ /* Save callee-safe registers used in this function. */
+ pushl %esi
+ pushl %edi
+
+ /* Load parameters into registers. */
+ movl 12(%esp), %eax /* str: pointer to memory block. */
+ movl 16(%esp), %edx /* c: byte we are looking for. */
+ movl 20(%esp), %esi /* len: length of memory block. */
+
+ /* If my must not test more than three characters test
+ them one by one. This is especially true for 0. */
+ cmpl $4, %esi
+ jb L3
+
+ /* At the moment %edx contains C. What we need for the
+ algorithm is C in all bytes of the dword. Avoid
+ operations on 16 bit words because these require an
+ prefix byte (and one more cycle). */
+ movb %dl, %dh /* Now it is 0|0|c|c */
+ movl %edx, %ecx
+ shll $16, %edx /* Now c|c|0|0 */
+ movw %cx, %dx /* And finally c|c|c|c */
+
+ /* Better performance can be achieved if the word (32
+ bit) memory access is aligned on a four-byte-boundary.
+ So process first bytes one by one until boundary is
+ reached. Don't use a loop for better performance. */
+
+ testb $3, %eax /* correctly aligned ? */
+ je L2 /* yes => begin loop */
+ cmpb %dl, (%eax) /* compare byte */
+ je L9 /* target found => return */
+ incl %eax /* increment source pointer */
+ decl %esi /* decrement length counter */
+ je L4 /* len==0 => return NULL */
+
+ testb $3, %eax /* correctly aligned ? */
+ je L2 /* yes => begin loop */
+ cmpb %dl, (%eax) /* compare byte */
+ je L9 /* target found => return */
+ incl %eax /* increment source pointer */
+ decl %esi /* decrement length counter */
+ je L4 /* len==0 => return NULL */
+
+ testb $3, %eax /* correctly aligned ? */
+ je L2 /* yes => begin loop */
+ cmpb %dl, (%eax) /* compare byte */
+ je L9 /* target found => return */
+ incl %eax /* increment source pointer */
+ decl %esi /* decrement length counter */
+ /* no test for len==0 here, because this is done in the
+ loop head */
+ jmp L2
+
+ /* We exit the loop if adding MAGIC_BITS to LONGWORD fails to
+ change any of the hole bits of LONGWORD.
+
+ 1) Is this safe? Will it catch all the zero bytes?
+ Suppose there is a byte with all zeros. Any carry bits
+ propagating from its left will fall into the hole at its
+ least significant bit and stop. Since there will be no
+ carry from its most significant bit, the LSB of the
+ byte to the left will be unchanged, and the zero will be
+ detected.
+
+ 2) Is this worthwhile? Will it ignore everything except
+ zero bytes? Suppose every byte of LONGWORD has a bit set
+ somewhere. There will be a carry into bit 8. If bit 8
+ is set, this will carry into bit 16. If bit 8 is clear,
+ one of bits 9-15 must be set, so there will be a carry
+ into bit 16. Similarly, there will be a carry into bit
+ 24. If one of bits 24-31 is set, there will be a carry
+ into bit 32 (=carry flag), so all of the hole bits will
+ be changed.
+
+ 3) But wait! Aren't we looking for C, not zero?
+ Good point. So what we do is XOR LONGWORD with a longword,
+ each of whose bytes is C. This turns each byte that is C
+ into a zero. */
+
+
+ /* Each round the main loop processes 16 bytes. */
+
+ ALIGN (4)
+
+L1: movl (%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+
+ /* According to the algorithm we had to reverse the effect of the
+ XOR first and then test the overflow bits. But because the
+ following XOR would destroy the carry flag and it would (in a
+ representation with more than 32 bits) not alter then last
+ overflow, we can now test this condition. If no carry is signaled
+ no overflow must have occured in the last byte => it was 0. */
+ jnc L8
+
+ /* We are only interested in carry bits that change due to the
+ previous add, so remove original bits */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+
+ /* Now test for the other three overflow bits. */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+
+ /* If at least one byte of the word is C we don't get 0 in %edi. */
+ jnz L8 /* found it => return pointer */
+
+ /* This process is unfolded four times for better performance.
+ we don't increment the source pointer each time. Instead we
+ use offsets and increment by 16 in each run of the loop. But
+ before probing for the matching byte we need some extra code
+ (following LL(13) below). Even the len can be compared with
+ constants instead of decrementing each time. */
+
+ movl 4(%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L7 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jnz L7 /* found it => return pointer */
+
+ movl 8(%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L6 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jnz L6 /* found it => return pointer */
+
+ movl 12(%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L5 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jnz L5 /* found it => return pointer */
+
+ /* Adjust both counters for a full round, i.e. 16 bytes. */
+ addl $16, %eax
+L2: subl $16, %esi
+ jae L1 /* Still more than 16 bytes remaining */
+
+ /* Process remaining bytes separately. */
+ cmpl $4-16, %esi /* rest < 4 bytes? */
+ jb L3 /* yes, than test byte by byte */
+
+ movl (%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L8 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jne L8 /* found it => return pointer */
+ addl $4, %eax /* adjust source pointer */
+
+ cmpl $8-16, %esi /* rest < 8 bytes? */
+ jb L3 /* yes, than test byte by byte */
+
+ movl (%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L8 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jne L8 /* found it => return pointer */
+ addl $4, %eax /* adjust source pointer */
+
+ cmpl $12-16, %esi /* rest < 12 bytes? */
+ jb L3 /* yes, than test byte by byte */
+
+ movl (%eax), %ecx /* get word (= 4 bytes) in question */
+ movl $0xfefefeff, %edi /* magic value */
+ xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
+ are now 0 */
+ addl %ecx, %edi /* add the magic value to the word. We get
+ carry bits reported for each byte which
+ is *not* 0 */
+ jnc L8 /* highest byte is C => return pointer */
+ xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
+ orl $0xfefefeff, %edi /* set all non-carry bits */
+ incl %edi /* add 1: if one carry bit was *not* set
+ the addition will not result in 0. */
+ jne L8 /* found it => return pointer */
+ addl $4, %eax /* adjust source pointer */
+
+ /* Check the remaining bytes one by one. */
+L3: andl $3, %esi /* mask out uninteresting bytes */
+ jz L4 /* no remaining bytes => return NULL */
+
+ cmpb %dl, (%eax) /* compare byte with C */
+ je L9 /* equal, than return pointer */
+ incl %eax /* increment source pointer */
+ decl %esi /* decrement length */
+ jz L4 /* no remaining bytes => return NULL */
+
+ cmpb %dl, (%eax) /* compare byte with C */
+ je L9 /* equal, than return pointer */
+ incl %eax /* increment source pointer */
+ decl %esi /* decrement length */
+ jz L4 /* no remaining bytes => return NULL */
+
+ cmpb %dl, (%eax) /* compare byte with C */
+ je L9 /* equal, than return pointer */
+
+L4: /* no byte found => return NULL */
+ xorl %eax, %eax
+ jmp L9
+
+ /* add missing source pointer increments */
+L5: addl $4, %eax
+L6: addl $4, %eax
+L7: addl $4, %eax
+
+ /* Test for the matching byte in the word. %ecx contains a NUL
+ char in the byte which originally was the byte we are looking
+ at. */
+L8: testb %cl, %cl /* test first byte in dword */
+ jz L9 /* if zero => return pointer */
+ incl %eax /* increment source pointer */
+
+ testb %ch, %ch /* test second byte in dword */
+ jz L9 /* if zero => return pointer */
+ incl %eax /* increment source pointer */
+
+ testl $0xff0000, %ecx /* test third byte in dword */
+ jz L9 /* if zero => return pointer */
+ incl %eax /* increment source pointer */
+
+ /* No further test needed we we known it is one of the four byytes. */
+
+L9: popl %edi /* pop saved registers */
+ popl %esi
+
+ ret