alpha/lib/strlen.S

0001 /* SPDX-License-Identifier: GPL-2.0 */
0002 /*
0003  * strlen.S (c) 1995 David Mosberger (davidm@cs.arizona.edu)
0004  *
0005  * Finds length of a 0-terminated string.  Optimized for the
0006  * Alpha architecture:
0007  *
0008  *  - memory accessed as aligned quadwords only
0009  *  - uses bcmpge to compare 8 bytes in parallel
0010  *  - does binary search to find 0 byte in last
0011  *    quadword (HAKMEM needed 12 instructions to
0012  *    do this instead of the 9 instructions that
0013  *    binary search needs).
0014  */
0015 #include <asm/export.h>
0016     .set noreorder
0017     .set noat
0018
0019     .align 3
0020
0021     .globl  strlen
0022     .ent    strlen
0023
0024 strlen:
0025     ldq_u   $1, 0($16)  # load first quadword ($16  may be misaligned)
0026     lda $2, -1($31)
0027     insqh   $2, $16, $2
0028     andnot  $16, 7, $0
0029     or  $2, $1, $1
0030     cmpbge  $31, $1, $2 # $2  <- bitmask: bit i == 1 <==> i-th byte == 0
0031     bne $2, found
0032
0033 loop:   ldq $1, 8($0)
0034     addq    $0, 8, $0   # addr += 8
0035     nop         # helps dual issue last two insns
0036     cmpbge  $31, $1, $2
0037     beq $2, loop
0038
0039 found:  blbs    $2, done    # make aligned case fast
0040     negq    $2, $3
0041     and $2, $3, $2
0042
0043     and $2, 0x0f, $1
0044     addq    $0, 4, $3
0045     cmoveq  $1, $3, $0
0046
0047     and $2, 0x33, $1
0048     addq    $0, 2, $3
0049     cmoveq  $1, $3, $0
0050
0051     and $2, 0x55, $1
0052     addq    $0, 1, $3
0053     cmoveq  $1, $3, $0
0054
0055 done:   subq    $0, $16, $0
0056     ret $31, ($26)
0057
0058     .end    strlen
0059     EXPORT_SYMBOL(strlen)