|
Packit |
6c4009 |
/* Optimized memchr implementation for PowerPC64/POWER7 using cmpb insn.
|
|
Packit |
6c4009 |
Copyright (C) 2010-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
Contributed by Luis Machado <luisgpm@br.ibm.com>.
|
|
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 |
/* int [r3] memchr (char *s [r3], int byte [r4], int size [r5]) */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#ifndef MEMCHR
|
|
Packit |
6c4009 |
# define MEMCHR __memchr
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
.machine power7
|
|
Packit |
6c4009 |
ENTRY_TOCLESS (MEMCHR)
|
|
Packit |
6c4009 |
CALL_MCOUNT 3
|
|
Packit |
6c4009 |
dcbt 0,r3
|
|
Packit |
6c4009 |
clrrdi r8,r3,3
|
|
Packit |
6c4009 |
insrdi r4,r4,8,48
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Calculate the last acceptable address and check for possible
|
|
Packit |
6c4009 |
addition overflow by using satured math:
|
|
Packit |
6c4009 |
r7 = r3 + r5
|
|
Packit |
6c4009 |
r7 |= -(r7 < x) */
|
|
Packit |
6c4009 |
add r7,r3,r5
|
|
Packit |
6c4009 |
subfc r6,r3,r7
|
|
Packit |
6c4009 |
subfe r9,r9,r9
|
|
Packit |
6c4009 |
extsw r6,r9
|
|
Packit |
6c4009 |
or r7,r7,r6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
insrdi r4,r4,16,32
|
|
Packit |
6c4009 |
cmpldi r5,32
|
|
Packit |
6c4009 |
li r9, -1
|
|
Packit |
6c4009 |
rlwinm r6,r3,3,26,28 /* Calculate padding. */
|
|
Packit |
6c4009 |
insrdi r4,r4,32,0
|
|
Packit |
6c4009 |
addi r7,r7,-1
|
|
Packit |
6c4009 |
#ifdef __LITTLE_ENDIAN__
|
|
Packit |
6c4009 |
sld r9,r9,r6
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
srd r9,r9,r6
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
ble L(small_range)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ld r12,0(r8) /* Load doubleword from memory. */
|
|
Packit |
6c4009 |
cmpb r3,r12,r4 /* Check for BYTEs in DWORD1. */
|
|
Packit |
6c4009 |
and r3,r3,r9
|
|
Packit |
6c4009 |
clrldi r5,r7,61 /* Byte count - 1 in last dword. */
|
|
Packit |
6c4009 |
clrrdi r7,r7,3 /* Address of last doubleword. */
|
|
Packit |
6c4009 |
cmpldi cr7,r3,0 /* Does r3 indicate we got a hit? */
|
|
Packit |
6c4009 |
bne cr7,L(done)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mtcrf 0x01,r8
|
|
Packit |
6c4009 |
/* Are we now aligned to a quadword boundary? If so, skip to
|
|
Packit |
6c4009 |
the main loop. Otherwise, go through the alignment code. */
|
|
Packit |
6c4009 |
bt 28,L(loop_setup)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Handle DWORD2 of pair. */
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr7,r3,0
|
|
Packit |
6c4009 |
bne cr7,L(done)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(loop_setup):
|
|
Packit |
6c4009 |
/* The last dword we want to read in the loop below is the one
|
|
Packit |
6c4009 |
containing the last byte of the string, ie. the dword at
|
|
Packit |
6c4009 |
(s + size - 1) & ~7, or r7. The first dword read is at
|
|
Packit |
6c4009 |
r8 + 8, we read 2 * cnt dwords, so the last dword read will
|
|
Packit |
6c4009 |
be at r8 + 8 + 16 * cnt - 8. Solving for cnt gives
|
|
Packit |
6c4009 |
cnt = (r7 - r8) / 16 */
|
|
Packit |
6c4009 |
sub r6,r7,r8
|
|
Packit |
6c4009 |
srdi r6,r6,4 /* Number of loop iterations. */
|
|
Packit |
6c4009 |
mtctr r6 /* Setup the counter. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Main loop to look for BYTE in the string. Since
|
|
Packit |
6c4009 |
it's a small loop (8 instructions), align it to 32-bytes. */
|
|
Packit |
6c4009 |
.align 5
|
|
Packit |
6c4009 |
L(loop):
|
|
Packit |
6c4009 |
/* Load two doublewords, compare and merge in a
|
|
Packit |
6c4009 |
single register for speed. This is an attempt
|
|
Packit |
6c4009 |
to speed up the byte-checking process for bigger strings. */
|
|
Packit |
6c4009 |
ld r12,8(r8)
|
|
Packit |
6c4009 |
ldu r11,16(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpb r9,r11,r4
|
|
Packit |
6c4009 |
or r6,r9,r3 /* Merge everything in one doubleword. */
|
|
Packit |
6c4009 |
cmpldi cr7,r6,0
|
|
Packit |
6c4009 |
bne cr7,L(found)
|
|
Packit |
6c4009 |
bdnz L(loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* We may have one more dword to read. */
|
|
Packit |
6c4009 |
cmpld r8,r7
|
|
Packit |
6c4009 |
beqlr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
bne cr6,L(done)
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.align 4
|
|
Packit |
6c4009 |
L(found):
|
|
Packit |
6c4009 |
/* OK, one (or both) of the doublewords contains BYTE. Check
|
|
Packit |
6c4009 |
the first doubleword and decrement the address in case the first
|
|
Packit |
6c4009 |
doubleword really contains BYTE. */
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
addi r8,r8,-8
|
|
Packit |
6c4009 |
bne cr6,L(done)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* BYTE must be in the second doubleword. Adjust the address
|
|
Packit |
6c4009 |
again and move the result of cmpb to r3 so we can calculate the
|
|
Packit |
6c4009 |
pointer. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mr r3,r9
|
|
Packit |
6c4009 |
addi r8,r8,8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* r3 has the output of the cmpb instruction, that is, it contains
|
|
Packit |
6c4009 |
0xff in the same position as BYTE in the original
|
|
Packit |
6c4009 |
doubleword from the string. Use that to calculate the pointer.
|
|
Packit |
6c4009 |
We need to make sure BYTE is *before* the end of the range. */
|
|
Packit |
6c4009 |
L(done):
|
|
Packit |
6c4009 |
#ifdef __LITTLE_ENDIAN__
|
|
Packit |
6c4009 |
addi r0,r3,-1
|
|
Packit |
6c4009 |
andc r0,r0,r3
|
|
Packit |
6c4009 |
popcntd r0,r0 /* Count trailing zeros. */
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
cntlzd r0,r3 /* Count leading zeros before the match. */
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
cmpld r8,r7 /* Are we on the last dword? */
|
|
Packit |
6c4009 |
srdi r0,r0,3 /* Convert leading/trailing zeros to bytes. */
|
|
Packit |
6c4009 |
add r3,r8,r0
|
|
Packit |
6c4009 |
cmpld cr7,r0,r5 /* If on the last dword, check byte offset. */
|
|
Packit |
6c4009 |
bnelr
|
|
Packit |
6c4009 |
blelr cr7
|
|
Packit |
6c4009 |
li r3,0
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.align 4
|
|
Packit |
6c4009 |
L(null):
|
|
Packit |
6c4009 |
li r3,0
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Deals with size <= 32. */
|
|
Packit |
6c4009 |
.align 4
|
|
Packit |
6c4009 |
L(small_range):
|
|
Packit |
6c4009 |
cmpldi r5,0
|
|
Packit |
6c4009 |
beq L(null)
|
|
Packit |
6c4009 |
ld r12,0(r8) /* Load word from memory. */
|
|
Packit |
6c4009 |
cmpb r3,r12,r4 /* Check for BYTE in DWORD1. */
|
|
Packit |
6c4009 |
and r3,r3,r9
|
|
Packit |
6c4009 |
cmpldi cr7,r3,0
|
|
Packit |
6c4009 |
clrldi r5,r7,61 /* Byte count - 1 in last dword. */
|
|
Packit |
6c4009 |
clrrdi r7,r7,3 /* Address of last doubleword. */
|
|
Packit |
6c4009 |
cmpld r8,r7 /* Are we done already? */
|
|
Packit |
6c4009 |
bne cr7,L(done)
|
|
Packit |
6c4009 |
beqlr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
cmpld r8,r7
|
|
Packit |
6c4009 |
bne cr6,L(done) /* Found something. */
|
|
Packit |
6c4009 |
beqlr /* Hit end of string (length). */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
cmpld r8,r7
|
|
Packit |
6c4009 |
bne cr6,L(done)
|
|
Packit |
6c4009 |
beqlr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
cmpld r8,r7
|
|
Packit |
6c4009 |
bne cr6,L(done)
|
|
Packit |
6c4009 |
beqlr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ldu r12,8(r8)
|
|
Packit |
6c4009 |
cmpb r3,r12,r4
|
|
Packit |
6c4009 |
cmpldi cr6,r3,0
|
|
Packit |
6c4009 |
bne cr6,L(done)
|
|
Packit |
6c4009 |
blr
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (MEMCHR)
|
|
Packit |
6c4009 |
weak_alias (__memchr, memchr)
|
|
Packit |
6c4009 |
libc_hidden_builtin_def (memchr)
|