diff options
Diffstat (limited to 'REORG.TODO/sysdeps/i386/strrchr.S')
| -rw-r--r-- | REORG.TODO/sysdeps/i386/strrchr.S | 334 |
1 files changed, 334 insertions, 0 deletions
diff --git a/REORG.TODO/sysdeps/i386/strrchr.S b/REORG.TODO/sysdeps/i386/strrchr.S new file mode 100644 index 0000000000..95b304dc0b --- /dev/null +++ b/REORG.TODO/sysdeps/i386/strrchr.S @@ -0,0 +1,334 @@ +/* strrchr (str, ch) -- Return pointer to last occurrence of CH in STR. + For Intel 80x86, x>=3. + Copyright (C) 1994-2017 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Ulrich Drepper <drepper@gnu.ai.mit.edu> + Some optimisations by Alan Modra <Alan@SPRI.Levels.UniSA.Edu.Au> + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 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 + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <http://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "asm-syntax.h" + +#define PARMS 4+8 /* space for 2 saved regs */ +#define RTN PARMS +#define STR RTN +#define CHR STR+4 + + .text +ENTRY (strrchr) + + pushl %edi /* Save callee-safe registers used here. */ + cfi_adjust_cfa_offset (4) + cfi_rel_offset (edi, 0) + pushl %esi + cfi_adjust_cfa_offset (4) + + xorl %eax, %eax + movl STR(%esp), %esi + cfi_rel_offset (esi, 0) + movl CHR(%esp), %ecx + + /* 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 inherent + boundaries are multiples of 4. */ + + testl $3, %esi /* correctly aligned ? */ + jz L(19) /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L(11) /* target found => return */ + movl %esi, %eax /* remember pointer as possible result */ +L(11): orb %dl, %dl /* is NUL? */ + jz L(2) /* yes => return NULL */ + incl %esi /* increment pointer */ + + testl $3, %esi /* correctly aligned ? */ + jz L(19) /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L(12) /* target found => return */ + movl %esi, %eax /* remember pointer as result */ +L(12): orb %dl, %dl /* is NUL? */ + jz L(2) /* yes => return NULL */ + incl %esi /* increment pointer */ + + testl $3, %esi /* correctly aligned ? */ + jz L(19) /* yes => begin loop */ + movb (%esi), %dl /* load byte in question (we need it twice) */ + cmpb %dl, %cl /* compare byte */ + jne L(13) /* target found => return */ + movl %esi, %eax /* remember pointer as result */ +L(13): orb %dl, %dl /* is NUL? */ + jz L(2) /* yes => return NULL */ + incl %esi /* increment pointer */ + + /* No we have reached alignment. */ + jmp L(19) /* 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 +#ifndef PROF + /* Profiling adds some code and so changes the alignment. */ + .byte 0 +#endif + +L(4): subl $4, %esi /* adjust pointer */ +L(41): subl $4, %esi +L(42): subl $4, %esi +L(43): testl $0xff000000, %edx /* is highest byte == C? */ + jnz L(33) /* no => try other bytes */ + leal 15(%esi), %eax /* store address as result */ + jmp L(1) /* and start loop again */ + +L(3): subl $4, %esi /* adjust pointer */ +L(31): subl $4, %esi +L(32): subl $4, %esi +L(33): testl $0xff0000, %edx /* is C in third byte? */ + jnz L(51) /* no => try other bytes */ + leal 14(%esi), %eax /* store address as result */ + jmp L(1) /* and start loop again */ + +L(51): + /* 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 L(1) /* yes => start loop */ + leal 13(%esi), %eax /* correct guess to second byte */ + +L(1): addl $16, %esi /* increment pointer for full round */ + +L(19): 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 occurred in the last byte => it was 0. */ + + jnc L(20) /* 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 L(20) /* 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 L(4) /* 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 L(3) /* 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 L(21) /* 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 L(21) /* 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 L(41) /* 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 L(31) /* 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 L(22) /* 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 L(22) /* 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 L(42) /* 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 L(32) /* 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 L(23) /* 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 L(23) /* 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 L(43) /* 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 L(1) /* C is not detected => restart loop */ + jmp L(33) /* examine word */ + +L(23): addl $4, %esi /* adjust pointer */ +L(22): addl $4, %esi +L(21): 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. */ + +L(20): cmpb %cl, %dl /* is first byte == C? */ + jne L(24) /* no => skip */ + movl %esi, %eax /* store address as result */ +L(24): testb %dl, %dl /* is first byte == NUL? */ + jz L(2) /* yes => return */ + + cmpb %cl, %dh /* is second byte == C? */ + jne L(25) /* no => skip */ + leal 1(%esi), %eax /* store address as result */ +L(25): testb %dh, %dh /* is second byte == NUL? */ + jz L(2) /* yes => return */ + + shrl $16,%edx /* make upper bytes accessible */ + cmpb %cl, %dl /* is third byte == C */ + jne L(26) /* no => skip */ + leal 2(%esi), %eax /* store address as result */ +L(26): testb %dl, %dl /* is third byte == NUL */ + jz L(2) /* yes => return */ + + cmpb %cl, %dh /* is fourth byte == C */ + jne L(2) /* no => skip */ + leal 3(%esi), %eax /* store address as result */ + +L(2): popl %esi /* restore saved register content */ + cfi_adjust_cfa_offset (-4) + cfi_restore (esi) + popl %edi + cfi_adjust_cfa_offset (-4) + cfi_restore (edi) + + ret +END (strrchr) + +weak_alias (strrchr, rindex) +libc_hidden_builtin_def (strrchr) |