glibc/sysdeps/i386/i586/strchr.S
Ulrich Drepper 5929563ff2 Update.
1997-04-01 00:37  Ulrich Drepper  <drepper@cygnus.com>

	* math/Makefile: Prevent inlining for test-{float,double,longdouble}.
	* math/libm-test.c (initialize): Make strings unique and force
	values to memory.

	* posix/regex.c: Debug debugging code.

	* sysdeps/i386/asm-syntax.h: Define `L' macro to emit correct
	syntax for local label.
	* sysdeps/i386/add_n.S: Use correct local label syntax for the
	binary format.
	* sysdeps/i386/addmul_1.S: Likewise.
	* sysdeps/i386/lshift.S: Likewise.
	* sysdeps/i386/memchr.S: Likewise.
	* sysdeps/i386/memset.S: Likewise.
	* sysdeps/i386/mul_1.S: Likewise.
	* sysdeps/i386/rshift.S: Likewise.
	* sysdeps/i386/stpcpy.S: Likewise.
	* sysdeps/i386/stpncpy.S: Likewise.
	* sysdeps/i386/strchr.S: Likewise.
	* sysdeps/i386/strcspn.S: Likewise.
	* sysdeps/i386/strlen.S: Likewise.
	* sysdeps/i386/strpbrk.S: Likewise.
	* sysdeps/i386/strrchr.S: Likewise.
	* sysdeps/i386/strspn.S: Likewise.
	* sysdeps/i386/strtok.S: Likewise.
	* sysdeps/i386/sub_n.S: Likewise.
	* sysdeps/i386/submul_1.S: Likewise.
	* sysdeps/i386/i486/strcat.S: Likewise.
	* sysdeps/i386/i486/strlen.S: Likewise.
	* sysdeps/i386/i586/add_n.S: Likewise.
	* sysdeps/i386/i586/addmul_1.S: Likewise.
	* sysdeps/i386/i586/lshift.S: Likewise.
	* sysdeps/i386/i586/memset.S: Likewise.
	* sysdeps/i386/i586/mul_1.S: Likewise.
	* sysdeps/i386/i586/rshift.S: Likewise.
	* sysdeps/i386/i586/strchr.S: Likewise.
	* sysdeps/i386/i586/strlen.S: Likewise.
	* sysdeps/i386/i586/sub_n.S: Likewise.
	* sysdeps/i386/i586/submul_1.S: Likewise.

	* sysdeps/stub/s_log2l.c: Correct function name.

	* time/africa: Updated from ADO tzdata1997e.
	* time/aantarctica: Likewise.
	* time/asia: Likewise.
	* time/australia: Likewise.
	* time/europe: Likewise.
	* time/northamerica: Likewise.
	* time/southamerica: Likewise.
	* time/private.h: Updated from ADO tzcode1997d.
	* time/scheck.c: Likewise.

1997-03-31 12:42  Ulrich Drepper  <drepper@cygnus.com>

	* sysdeps/unix/sysv/linux/powerpc/Dist: Add kernel_*.h files.

1997-03-23 01:41  Geoff Keating  <geoffk@ozemail.com.au>

	* sysdeps/powerpc/strlen.s: Rewrote for more speed.
	* string/tst-strlen.c: Also test various string alignments.

1997-03-19 06:11  Geoff Keating  <geoffk@ozemail.com.au>

	* sysdeps/unix/sysv/linux/powerpc/statbuf.h: Removed.
	* sysdeps/unix/sysv/linux/powerpc/kernel_stat.h: Added.

	* sysdeps/powerpc/dl-machine.h (ELF_PREFERRED_ADDRESS): Assume
	program base address is 0x01800000 if we can't otherwise tell what
	it is.
	(elf_machine_rela): Replace `assertion failed' with real error
	messages.

	* sysdeps/powerpc/dl-machine.h: More comments, more GNU indenting.
	Put in extra parentheses to make explicit the relative precedence
	of + and >>.

	* sysdeps/powerpc/bsd-setjmp.S: Add `__setjmp' symbol so the tests
	pass.
	* sysdeps/powerpc/clone.S: Rewrite.  Probably still doesn't work.

	* sysdeps/unix/sysv/linux/powerpc/sys/sysdep.h: Use .section "text"
	instead of .text.

1997-03-31 02:28  Ulrich Drepper  <drepper@cygnus.com>

	* misc/tst-tsearch.c: Don't write to stderr, write to stdout
	instead.

	* inet/inet_ntoa.c (inet_ntoa): Complete rewrite.  Now uses
	thread-specific memory in multi-threaded programs.

	* sysdeps/stub/s_exp2.c: New file.  Stub version for exp2 function.
	* sysdeps/stub/s_exp2f.c: New file.
	* sysdeps/stub/s_exp2l.c: New file.
	Reported by a sun <asun@zoology.washington.edu>.

	* sysdeps/unix/sysv/linux/netatalk/atalk.h: Rename to...
	* sysdeps/unix/sysv/linux/netatalk/at.h: ...this.
	Suggested by a sun <asun@zoology.washington.edu>.

1997-03-30 14:53  Ulrich Drepper  <drepper@cygnus.com>

	* sysdeps/unix/sysv/linux/netinet/ip_icmp.h: Add former contents
	of <netinet/icmp.h>.
	Patch by a sun <asun@zoology.washington.edu>.

	* sysdeps/powerpc/strlen.S: Added.
1997-03-31 22:56:24 +00:00

334 lines
8.1 KiB
ArmAsm

/* Find character CH in a NUL terminated string.
Highly optimized version for ix85, x>=5.
Copyright (C) 1995, 1996, 1997 Free Software Foundation, Inc.
This file is part of the GNU C Library.
Contributed by Ulrich Drepper, <drepper@gnu.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"
/* This version is especially optimized for the i586 (and following?)
processors. This is mainly done by using the two pipelines. The
version optimized for i486 is weak in this aspect because to get
as much parallelism we have to execute some *more* instructions.
The code below is structured to reflect the pairing of the instructions
as *I think* it is. I have no processor data book to verify this.
If you find something you think is incorrect let me know. */
/* The magic value which is used throughout in the whole code. */
#define magic 0xfefefeff
/*
INPUT PARAMETERS:
str (sp + 4)
ch (sp + 8)
*/
.text
ENTRY (strchr)
pushl %edi /* Save callee-safe registers. */
pushl %esi
pushl %ebx
pushl %ebp
movl 20(%esp), %eax /* get string pointer */
movl 24(%esp), %edx /* get character we are looking for */
movl %eax, %edi /* duplicate string pointer for later */
xorl %ecx, %ecx /* clear %ecx */
/* 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 */
movb %dl, %cl /* we construct the lower half in %ecx */
shll $16, %edx /* now %edx is c|c|0|0 */
movb %cl, %ch /* now %ecx is 0|0|c|c */
orl %ecx, %edx /* and finally c|c|c|c */
andl $3, %edi /* mask alignment bits */
jz L(11) /* alignment is 0 => start loop */
movb %dl, %cl /* 0 is needed below */
jp L(0) /* exactly two bits set */
xorb (%eax), %cl /* is byte the one we are looking for? */
jz L(2) /* yes => return pointer */
xorb %dl, %cl /* load single byte and test for NUL */
je L(3) /* yes => return NULL */
movb 1(%eax), %cl /* load single byte */
incl %eax
cmpb %cl, %dl /* is byte == C? */
je L(2) /* aligned => return pointer */
cmpb $0, %cl /* is byte NUL? */
je L(3) /* yes => return NULL */
incl %eax
decl %edi
jne L(11)
L(0): movb (%eax), %cl /* load single byte */
cmpb %cl, %dl /* is byte == C? */
je L(2) /* aligned => return pointer */
cmpb $0, %cl /* is byte NUL? */
je L(3) /* yes => return NULL */
incl %eax /* increment pointer */
/* The following code is the preparation for the loop. The
four instruction up to `L1' will not be executed in the loop
because the same code is found at the end of the loop, but
there it is executed in parallel with other instructions. */
L(11): movl (%eax), %ecx
movl $magic, %ebp
movl $magic, %edi
addl %ecx, %ebp
/* The main loop: it looks complex and indeed it is. I would
love to say `it was hard to write, so it should he hard to
read' but I will give some more hints. To fully understand
this code you should first take a look at the i486 version.
The basic algorithm is the same, but here the code organized
in a way which permits to use both pipelines all the time.
I tried to make it a bit more understandable by indenting
the code according to stage in the algorithm. It goes as
follows:
check for 0 in 1st word
check for C in 1st word
check for 0 in 2nd word
check for C in 2nd word
check for 0 in 3rd word
check for C in 3rd word
check for 0 in 4th word
check for C in 4th word
Please note that doing the test for NUL before the test for
C allows us to overlap the test for 0 in the next word with
the test for C. */
L(1): xorl %ecx, %ebp /* (word^magic) */
addl %ecx, %edi /* add magic word */
leal 4(%eax), %eax /* increment pointer */
jnc L(4) /* previous addl caused overflow? */
movl %ecx, %ebx /* duplicate original word */
orl $magic, %ebp /* (word^magic)|magic */
addl $1, %ebp /* (word^magic)|magic == 0xffffffff? */
jne L(4) /* yes => we found word with NUL */
movl $magic, %esi /* load magic value */
xorl %edx, %ebx /* clear words which are C */
movl (%eax), %ecx
addl %ebx, %esi /* (word+magic) */
movl $magic, %edi
jnc L(5) /* previous addl caused overflow? */
movl %edi, %ebp
xorl %ebx, %esi /* (word+magic)^word */
addl %ecx, %ebp
orl $magic, %esi /* ((word+magic)^word)|magic */
addl $1, %esi /* ((word+magic)^word)|magic==0xf..f?*/
jne L(5) /* yes => we found word with C */
xorl %ecx, %ebp
addl %ecx, %edi
leal 4(%eax), %eax
jnc L(4)
movl %ecx, %ebx
orl $magic, %ebp
addl $1, %ebp
jne L(4)
movl $magic, %esi
xorl %edx, %ebx
movl (%eax), %ecx
addl %ebx, %esi
movl $magic, %edi
jnc L(5)
movl %edi, %ebp
xorl %ebx, %esi
addl %ecx, %ebp
orl $magic, %esi
addl $1, %esi
jne L(5)
xorl %ecx, %ebp
addl %ecx, %edi
leal 4(%eax), %eax
jnc L(4)
movl %ecx, %ebx
orl $magic, %ebp
addl $1, %ebp
jne L(4)
movl $magic, %esi
xorl %edx, %ebx
movl (%eax), %ecx
addl %ebx, %esi
movl $magic, %edi
jnc L(5)
movl %edi, %ebp
xorl %ebx, %esi
addl %ecx, %ebp
orl $magic, %esi
addl $1, %esi
jne L(5)
xorl %ecx, %ebp
addl %ecx, %edi
leal 4(%eax), %eax
jnc L(4)
movl %ecx, %ebx
orl $magic, %ebp
addl $1, %ebp
jne L(4)
movl $magic, %esi
xorl %edx, %ebx
movl (%eax), %ecx
addl %ebx, %esi
movl $magic, %edi
jnc L(5)
movl %edi, %ebp
xorl %ebx, %esi
addl %ecx, %ebp
orl $magic, %esi
addl $1, %esi
je L(1)
/* We know there is no NUL byte but a C byte in the word.
%ebx contains NUL in this particular byte. */
L(5): subl $4, %eax /* adjust pointer */
testb %bl, %bl /* first byte == C? */
jz L(2) /* yes => return pointer */
incl %eax /* increment pointer */
testb %bh, %bh /* second byte == C? */
jz L(2) /* yes => return pointer */
shrl $16, %ebx /* make upper bytes accessible */
incl %eax /* increment pointer */
cmp $0, %bl /* third byte == C */
je L(2) /* yes => return pointer */
incl %eax /* increment pointer */
L(2): popl %ebp /* restore saved registers */
popl %ebx
popl %esi
popl %edi
ret
/* We know there is a NUL byte in the word. But we have to test
whether there is an C byte before it in the word. */
L(4): subl $4, %eax /* adjust pointer */
cmpb %dl, %cl /* first byte == C? */
je L(2) /* yes => return pointer */
cmpb $0, %cl /* first byte == NUL? */
je L(3) /* yes => return NULL */
incl %eax /* increment pointer */
cmpb %dl, %ch /* second byte == C? */
je L(2) /* yes => return pointer */
cmpb $0, %ch /* second byte == NUL? */
je L(3) /* yes => return NULL */
shrl $16, %ecx /* make upper bytes accessible */
incl %eax /* increment pointer */
cmpb %dl, %cl /* third byte == C? */
je L(2) /* yes => return pointer */
cmpb $0, %cl /* third byte == NUL? */
je L(3) /* yes => return NULL */
incl %eax /* increment pointer */
/* The test four the fourth byte is necessary! */
cmpb %dl, %ch /* fourth byte == C? */
je L(2) /* yes => return pointer */
L(3): xorl %eax, %eax /* set return value = NULL */
popl %ebp /* restore saved registers */
popl %ebx
popl %esi
popl %edi
ret
END (strchr)
#undef index
weak_alias (strchr, index)