From 8f5ca04bc7fd53741d80117df992995ace8f6d2d Mon Sep 17 00:00:00 2001 From: Roland McGrath Date: Mon, 16 Oct 1995 01:37:51 +0000 Subject: Sat Oct 14 02:52:36 1995 Ulrich Drepper * 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 . * 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 [$(subdir)=sunrpc] (headers): Likewise. * sysdeps/unix/sysv/linux/local_lim.h: Rewrite. Instead of defining ourself we use a kernel header file. * sysdeps/unix/sysv/linux/i386/sysdep.h (DO_CALL): Optimize system call handler for i586. * sysdeps/unix/sysv/linux/sys/param.h: Add copyright and clean up. Sat Oct 14 02:52:36 1995 Ulrich Drepper * malloc/malloc.c (_malloc_internal): Performance fix. Move if statement out of loop. * stdio/_itoa.c, stdio/_itoa.h: Complete rewrite. Much faster implementation using GMP functions. Contributed by Torbjorn Granlund and Ulrich Drepper. * stdio/test_rdwr.c: Include . * sysdeps/i386/i586/Implies: New file. New highly optimized string functions for i[345]86. * sysdeps/i386/memchr.S, sysdeps/i386/memcmp.S: New files. * sysdeps/i386/stpcpy.S, sysdeps/i386/stpncpy.S: New files. * sysdeps/i386/strchr.S, sysdeps/i386/strcspn.S: New files. * sysdeps/i386/strpbrk.S, sysdeps/i386/strrchr.S: New files. * sysdeps/i386/strspn.S, sysdeps/i386/i486/strcat.S: New files. * sysdeps/i386/i486/strlen.S, sysdeps/i386/i586/strchr.S: New files. * sysdeps/i386/i586/strlen.S: New file. * sysdeps/i386/memchr.c: Removed. There is now an assembler version. * sysdeps/i386/i586/memcopy.h (WORD_COPY_BWD): Parameters did not correspond to used values. * sysdeps/unix/sysv/linux/nfs/nfs.h: New file. Simply a wrapper around a kernel header file. * sysdeps/unix/sysv/linux/Dist: Add it. * sysdeps/unix/sysv/linux/Makefile [$(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. --- sysdeps/i386/strrchr.S | 321 +++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 321 insertions(+) create mode 100644 sysdeps/i386/strrchr.S (limited to 'sysdeps/i386/strrchr.S') diff --git a/sysdeps/i386/strrchr.S b/sysdeps/i386/strrchr.S new file mode 100644 index 0000000000..468a940d74 --- /dev/null +++ b/sysdeps/i386/strrchr.S @@ -0,0 +1,321 @@ +/* strchr (str, ch) -- Return pointer to last occurrence of CH in STR. +For Intel 80x86, x>=3. +Copyright (C) 1994, 1995 Free Software Foundation, Inc. +Contributed by Ulrich Drepper +Some optimisations by Alan Modra +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 +#include "asm-syntax.h" + +/* + INPUT PARAMETERS: + str (sp + 4) + ch (sp + 8) +*/ + + .text +ENTRY (strrchr) + pushl %edi /* Save callee-safe registers used here. */ + pushl %esi + + xorl %eax, %eax + movl 12(%esp), %esi /* get string pointer */ + movl 16(%esp), %ecx /* get character we are looking for */ + + /* At the moment %ecx 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 %cl, %ch /* now it is 0|0|c|c */ + movl %ecx, %edx + shll $16, %ecx /* now it is c|c|0|0 */ + movw %dx, %cx /* and finally c|c|c|c */ + + /* Before we start with the main loop we process single bytes + until the source pointer is aligned. This has two reasons: + 1. aligned 32-bit memory access is faster + and (more important) + 2. we process in the main loop 32 bit in one step although + we don't know the end of the string. But accessing at + 4-byte alignment guarantees that we never access illegal + memory if this would not also be done by the trivial + implementation (this is because all processor inherant + boundaries are multiples of 4. */ + + testb $3, %esi /* correctly aligned ? */ + jz L19 /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L11 /* target found => return */ + movl %esi, %eax /* remember pointer as possible result */ +L11: orb %dl, %dl /* is NUL? */ + jz L2 /* yes => return NULL */ + incl %esi /* increment pointer */ + + testb $3, %esi /* correctly aligned ? */ + jz L19 /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L12 /* target found => return */ + movl %esi, %eax /* remember pointer as result */ +L12: orb %dl, %dl /* is NUL? */ + jz L2 /* yes => return NULL */ + incl %esi /* increment pointer */ + + testb $3, %esi /* correctly aligned ? */ + jz L19 /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L13 /* target found => return */ + movl %esi, %eax /* remember pointer as result */ +L13: orb %cl, %cl /* is NUL? */ + jz L2 /* yes => return NULL */ + incl %esi /* increment pointer */ + + /* No we have reached alignment. */ + jmp L19 /* begin loop */ + + /* 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. */ + + /* Jump to here when the character is detected. We chose this + way around because the character one is looking for is not + as frequent as the rest and taking a conditional jump is more + expensive than ignoring it. + + Some more words to the code below: it might not be obvious why + we decrement the source pointer here. In the loop the pointer + is not pre-incremented and so it still points before the word + we are looking at. But you should take a look at the instruction + which gets executed before we get into the loop: `addl $16, %esi'. + This makes the following subs into adds. */ + + /* These fill bytes make the main loop be correctly aligned. + We cannot use align because it is not the following instruction + which should be aligned. */ + .byte 0, 0, 0, 0, 0, 0, 0, 0 + +L4: subl $4, %esi /* adjust pointer */ +L41: subl $4, %esi +L42: subl $4, %esi +L43: testl $0xff000000, %edx /* is highest byte == C? */ + jnz L33 /* no => try other bytes */ + leal 15(%esi), %eax /* store address as result */ + jmp L1 /* and start loop again */ + +L3: subl $4, %esi /* adjust pointer */ +L31: subl $4, %esi +L32: subl $4, %esi +L33: testl $0xff0000, %edx /* is C in third byte? */ + jnz L51 /* no => try other bytes */ + leal 14(%esi), %eax /* store address as result */ + jmp L1 /* and start loop again */ + +L51: + /* At this point we know that the byte is in one of the lower bytes. + We make a guess and correct it if necessary. This reduces the + number of necessary jumps. */ + leal 12(%esi), %eax /* guess address of lowest byte as result */ + testb %dh, %dh /* is guess correct? */ + jnz L1 /* yes => start loop */ + leal 13(%esi), %eax /* correct guess to second byte */ + +L1: addl $16, %esi /* increment pointer for full round */ + +L19: movl (%esi), %edx /* get word (= 4 bytes) in question */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %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 L20 /* found NUL => check last word */ + + /* We are only interested in carry bits that change due to the + previous add, so remove original bits */ + xorl %edx, %edi /* (word+magic)^word */ + + /* 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 L20 /* found NUL => check last word */ + + /* Now we made sure the dword does not contain the character we are + looking for. But because we deal with strings we have to check + for the end of string before testing the next dword. */ + + xorl %ecx, %edx /* XOR with word c|c|c|c => bytes of str == c + are now 0 */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L4 /* highest byte is C => examine dword */ + xorl %edx, %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 L3 /* C is detected in the word => examine it */ + + movl 4(%esi), %edx /* get word (= 4 bytes) in question */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L21 /* found NUL => check last word */ + xorl %edx, %edi /* (word+magic)^word */ + 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 L21 /* found NUL => check last word */ + xorl %ecx, %edx /* XOR with word c|c|c|c => bytes of str == c + are now 0 */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L41 /* highest byte is C => examine dword */ + xorl %edx, %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 L31 /* C is detected in the word => examine it */ + + movl 8(%esi), %edx /* get word (= 4 bytes) in question */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L22 /* found NUL => check last word */ + xorl %edx, %edi /* (word+magic)^word */ + 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 L22 /* found NUL => check last word */ + xorl %ecx, %edx /* XOR with word c|c|c|c => bytes of str == c + are now 0 */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L42 /* highest byte is C => examine dword */ + xorl %edx, %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 L32 /* C is detected in the word => examine it */ + + movl 12(%esi), %edx /* get word (= 4 bytes) in question */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L23 /* found NUL => check last word */ + xorl %edx, %edi /* (word+magic)^word */ + 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 L23 /* found NUL => check last word */ + xorl %ecx, %edx /* XOR with word c|c|c|c => bytes of str == c + are now 0 */ + movl $0xfefefeff, %edi /* magic value */ + addl %edx, %edi /* add the magic value to the word. We get + carry bits reported for each byte which + is *not* 0 */ + jnc L43 /* highest byte is C => examine dword */ + xorl %edx, %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. */ + jz L1 /* C is not detected => restart loop */ + jmp L33 /* examine word */ + +L23: addl $4, %esi /* adjust pointer */ +L22: addl $4, %esi +L21: addl $4, %esi + + /* What remains to do is to test which byte the NUL char is and + whether the searched character appears in one of the bytes + before. A special case is that the searched byte maybe NUL. + In this case a pointer to the terminating NUL char has to be + returned. */ + +L20: cmpb %cl, %dl /* is first byte == C? */ + jne L24 /* no => skip */ + movl %esi, %eax /* store address as result */ +L24: testb %dl, %dl /* is first byte == NUL? */ + jz L2 /* yes => return */ + + cmpb %cl, %dh /* is second byte == C? */ + jne L25 /* no => skip */ + leal 1(%esi), %eax /* store address as result */ +L25: testb %dh, %dh /* is second byte == NUL? */ + jz L2 /* yes => return */ + + shrl $16,%edx /* make upper bytes accessible */ + cmpb %cl, %dl /* is third byte == C */ + jne L26 /* no => skip */ + leal 2(%esi), %eax /* store address as result */ +L26: testb %dl, %dl /* is third byte == NUL */ + jz L2 /* yes => return */ + + cmpb %cl, %dh /* is fourth byte == C */ + jne L2 /* no => skip */ + leal 3(%esi), %eax /* store address as result */ + +L2: popl %esi /* restore saved register content */ + popl %edi + + ret + +weak_alias (strrchr, rindex) -- cgit v1.2.3