|
Packit |
6c4009 |
/* Optimized strchr implementation for PowerPC.
|
|
Packit |
6c4009 |
Copyright (C) 1997-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* See strlen.s for comments on how this works. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* char * [r3] strchr (const char *s [r3] , int c [r4] ) */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ENTRY (strchr)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define rTMP1 r0
|
|
Packit |
6c4009 |
#define rRTN r3 /* outgoing result */
|
|
Packit |
6c4009 |
#define rSTR r8 /* current word pointer */
|
|
Packit |
6c4009 |
#define rCHR r4 /* byte we're looking for, spread over the whole word */
|
|
Packit |
6c4009 |
#define rWORD r5 /* the current word */
|
|
Packit |
6c4009 |
#define rCLZB rCHR /* leading zero byte count */
|
|
Packit |
6c4009 |
#define rFEFE r6 /* constant 0xfefefeff (-0x01010101) */
|
|
Packit |
6c4009 |
#define r7F7F r7 /* constant 0x7f7f7f7f */
|
|
Packit |
6c4009 |
#define rTMP2 r9
|
|
Packit |
6c4009 |
#define rIGN r10 /* number of bits we should ignore in the first word */
|
|
Packit |
6c4009 |
#define rMASK r11 /* mask with the bits to ignore set to 0 */
|
|
Packit |
6c4009 |
#define rTMP3 r12
|
|
Packit |
6c4009 |
#define rTMP4 rIGN
|
|
Packit |
6c4009 |
#define rTMP5 rMASK
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
rlwimi rCHR, rCHR, 8, 16, 23
|
|
Packit |
6c4009 |
li rMASK, -1
|
|
Packit |
6c4009 |
rlwimi rCHR, rCHR, 16, 0, 15
|
|
Packit |
6c4009 |
rlwinm rIGN, rRTN, 3, 27, 28
|
|
Packit |
6c4009 |
lis rFEFE, -0x101
|
|
Packit |
6c4009 |
lis r7F7F, 0x7f7f
|
|
Packit |
6c4009 |
clrrwi rSTR, rRTN, 2
|
|
Packit |
6c4009 |
addi rFEFE, rFEFE, -0x101
|
|
Packit |
6c4009 |
addi r7F7F, r7F7F, 0x7f7f
|
|
Packit |
6c4009 |
/* Test the first (partial?) word. */
|
|
Packit |
6c4009 |
lwz rWORD, 0(rSTR)
|
|
Packit |
6c4009 |
#ifdef __LITTLE_ENDIAN__
|
|
Packit |
6c4009 |
slw rMASK, rMASK, rIGN
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
srw rMASK, rMASK, rIGN
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
orc rWORD, rWORD, rMASK
|
|
Packit |
6c4009 |
add rTMP1, rFEFE, rWORD
|
|
Packit |
6c4009 |
nor rTMP2, r7F7F, rWORD
|
|
Packit |
6c4009 |
and. rTMP4, rTMP1, rTMP2
|
|
Packit |
6c4009 |
xor rTMP3, rCHR, rWORD
|
|
Packit |
6c4009 |
orc rTMP3, rTMP3, rMASK
|
|
Packit |
6c4009 |
b L(loopentry)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* The loop. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(loop):
|
|
Packit |
6c4009 |
lwzu rWORD, 4(rSTR)
|
|
Packit |
6c4009 |
and. rTMP5, rTMP1, rTMP2
|
|
Packit |
6c4009 |
/* Test for 0. */
|
|
Packit |
6c4009 |
add rTMP1, rFEFE, rWORD /* x - 0x01010101. */
|
|
Packit |
6c4009 |
nor rTMP2, r7F7F, rWORD /* ~(x | 0x7f7f7f7f) == ~x & 0x80808080. */
|
|
Packit |
6c4009 |
bne L(foundit)
|
|
Packit |
6c4009 |
and. rTMP4, rTMP1, rTMP2 /* (x - 0x01010101) & ~x & 0x80808080. */
|
|
Packit |
6c4009 |
/* Start test for the bytes we're looking for. */
|
|
Packit |
6c4009 |
xor rTMP3, rCHR, rWORD
|
|
Packit |
6c4009 |
L(loopentry):
|
|
Packit |
6c4009 |
add rTMP1, rFEFE, rTMP3
|
|
Packit |
6c4009 |
nor rTMP2, r7F7F, rTMP3
|
|
Packit |
6c4009 |
beq L(loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* There is a zero byte in the word, but may also be a matching byte (either
|
|
Packit |
6c4009 |
before or after the zero byte). In fact, we may be looking for a
|
|
Packit |
6c4009 |
zero byte, in which case we return a match. */
|
|
Packit |
6c4009 |
and. rTMP5, rTMP1, rTMP2
|
|
Packit |
6c4009 |
li rRTN, 0
|
|
Packit |
6c4009 |
beqlr
|
|
Packit |
6c4009 |
/* At this point:
|
|
Packit |
6c4009 |
rTMP5 bytes are 0x80 for each match of c, 0 otherwise.
|
|
Packit |
6c4009 |
rTMP4 bytes are 0x80 for each match of 0, 0 otherwise.
|
|
Packit |
6c4009 |
But there may be false matches in the next most significant byte from
|
|
Packit |
6c4009 |
a true match due to carries. This means we need to recalculate the
|
|
Packit |
6c4009 |
matches using a longer method for big-endian. */
|
|
Packit |
6c4009 |
#ifdef __LITTLE_ENDIAN__
|
|
Packit |
6c4009 |
addi rTMP1, rTMP5, -1
|
|
Packit |
6c4009 |
andc rTMP1, rTMP1, rTMP5
|
|
Packit |
6c4009 |
cntlzw rCLZB, rTMP1
|
|
Packit |
6c4009 |
addi rTMP2, rTMP4, -1
|
|
Packit |
6c4009 |
andc rTMP2, rTMP2, rTMP4
|
|
Packit |
6c4009 |
cmplw rTMP1, rTMP2
|
|
Packit |
6c4009 |
bgtlr
|
|
Packit |
6c4009 |
subfic rCLZB, rCLZB, 32-7
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
/* I think we could reduce this by two instructions by keeping the "nor"
|
|
Packit |
6c4009 |
results from the loop for reuse here. See strlen.S tail. Similarly
|
|
Packit |
6c4009 |
one instruction could be pruned from L(foundit). */
|
|
Packit |
6c4009 |
and rFEFE, r7F7F, rWORD
|
|
Packit |
6c4009 |
or rTMP5, r7F7F, rWORD
|
|
Packit |
6c4009 |
and rTMP1, r7F7F, rTMP3
|
|
Packit |
6c4009 |
or rTMP4, r7F7F, rTMP3
|
|
Packit |
6c4009 |
add rFEFE, rFEFE, r7F7F
|
|
Packit |
6c4009 |
add rTMP1, rTMP1, r7F7F
|
|
Packit |
6c4009 |
nor rWORD, rTMP5, rFEFE
|
|
Packit |
6c4009 |
nor rTMP2, rTMP4, rTMP1
|
|
Packit |
6c4009 |
cntlzw rCLZB, rTMP2
|
|
Packit |
6c4009 |
cmplw rWORD, rTMP2
|
|
Packit |
6c4009 |
bgtlr
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
srwi rCLZB, rCLZB, 3
|
|
Packit |
6c4009 |
add rRTN, rSTR, rCLZB
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(foundit):
|
|
Packit |
6c4009 |
#ifdef __LITTLE_ENDIAN__
|
|
Packit |
6c4009 |
addi rTMP1, rTMP5, -1
|
|
Packit |
6c4009 |
andc rTMP1, rTMP1, rTMP5
|
|
Packit |
6c4009 |
cntlzw rCLZB, rTMP1
|
|
Packit |
6c4009 |
subfic rCLZB, rCLZB, 32-7-32
|
|
Packit |
6c4009 |
srawi rCLZB, rCLZB, 3
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
and rTMP1, r7F7F, rTMP3
|
|
Packit |
6c4009 |
or rTMP4, r7F7F, rTMP3
|
|
Packit |
6c4009 |
add rTMP1, rTMP1, r7F7F
|
|
Packit |
6c4009 |
nor rTMP2, rTMP4, rTMP1
|
|
Packit |
6c4009 |
cntlzw rCLZB, rTMP2
|
|
Packit |
6c4009 |
subi rSTR, rSTR, 4
|
|
Packit |
6c4009 |
srwi rCLZB, rCLZB, 3
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
add rRTN, rSTR, rCLZB
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
END (strchr)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
weak_alias (strchr, index)
|
|
Packit |
6c4009 |
libc_hidden_builtin_def (strchr)
|