|
Packit |
6c4009 |
/* Optimized version of the standard memchr() function.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
Copyright (C) 2000-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
Contributed by Dan Pop <Dan.Pop@cern.ch>.
|
|
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 |
/* Return: the address of the first occurence of chr in str or NULL
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Inputs:
|
|
Packit |
6c4009 |
in0: str
|
|
Packit |
6c4009 |
in1: chr
|
|
Packit |
6c4009 |
in2: byte count
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
This implementation assumes little endian mode. For big endian mode,
|
|
Packit |
6c4009 |
the instruction czx1.r should be replaced by czx1.l.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The algorithm is fairly straightforward: search byte by byte until we
|
|
Packit |
6c4009 |
we get to a word aligned address, then search word by word as much as
|
|
Packit |
6c4009 |
possible; the remaining few bytes are searched one at a time.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The word by word search is performed by xor-ing the word with a word
|
|
Packit |
6c4009 |
containing chr in every byte. If there is a hit, the result will
|
|
Packit |
6c4009 |
contain a zero byte in the corresponding position. The presence and
|
|
Packit |
6c4009 |
position of that zero byte is detected with a czx instruction.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
All the loops in this function could have had the internal branch removed
|
|
Packit |
6c4009 |
if br.ctop and br.cloop could be predicated :-(. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
#undef ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define saved_pr r15
|
|
Packit |
6c4009 |
#define saved_lc r16
|
|
Packit |
6c4009 |
#define chr r17
|
|
Packit |
6c4009 |
#define len r18
|
|
Packit |
6c4009 |
#define last r20
|
|
Packit |
6c4009 |
#define val r21
|
|
Packit |
6c4009 |
#define tmp r24
|
|
Packit |
6c4009 |
#define chrx8 r25
|
|
Packit |
6c4009 |
#define loopcnt r30
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#define str in0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ENTRY(__memchr)
|
|
Packit |
6c4009 |
.prologue
|
|
Packit |
6c4009 |
alloc r2 = ar.pfs, 3, 0, 29, 32
|
|
Packit |
6c4009 |
#include "softpipe.h"
|
|
Packit |
6c4009 |
.rotr value[MEMLAT+1], addr[MEMLAT+3], aux[2], poschr[2]
|
|
Packit |
6c4009 |
.rotp p[MEMLAT+3]
|
|
Packit |
6c4009 |
.save ar.lc, saved_lc
|
|
Packit |
6c4009 |
mov saved_lc = ar.lc // save the loop counter
|
|
Packit |
6c4009 |
.save pr, saved_pr
|
|
Packit |
6c4009 |
mov saved_pr = pr // save the predicates
|
|
Packit |
6c4009 |
.body
|
|
Packit |
6c4009 |
mov ret0 = str
|
|
Packit |
6c4009 |
add last = str, in2 // last byte
|
|
Packit |
6c4009 |
;;
|
|
Packit |
6c4009 |
cmp.ltu p6, p0 = last, str
|
|
Packit |
6c4009 |
;;
|
|
Packit |
6c4009 |
(p6) mov last = -1
|
|
Packit |
6c4009 |
and tmp = 7, str // tmp = str % 8
|
|
Packit |
6c4009 |
cmp.ne p7, p0 = r0, r0 // clear p7
|
|
Packit |
6c4009 |
extr.u chr = in1, 0, 8 // chr = (unsigned char) in1
|
|
Packit |
6c4009 |
mov len = in2
|
|
Packit |
6c4009 |
cmp.gtu p6, p0 = 16, in2 // use a simple loop for short
|
|
Packit |
6c4009 |
(p6) br.cond.spnt .srchfew ;; // searches
|
|
Packit |
6c4009 |
sub loopcnt = 8, tmp // loopcnt = 8 - tmp
|
|
Packit |
6c4009 |
cmp.eq p6, p0 = tmp, r0
|
|
Packit |
6c4009 |
(p6) br.cond.sptk .str_aligned;;
|
|
Packit |
6c4009 |
sub len = len, loopcnt
|
|
Packit |
6c4009 |
adds loopcnt = -1, loopcnt;;
|
|
Packit |
6c4009 |
mov ar.lc = loopcnt
|
|
Packit |
6c4009 |
.l1:
|
|
Packit |
6c4009 |
ld1 val = [ret0], 1
|
|
Packit |
6c4009 |
;;
|
|
Packit |
6c4009 |
cmp.eq p6, p0 = val, chr
|
|
Packit |
6c4009 |
(p6) br.cond.spnt .foundit
|
|
Packit |
6c4009 |
br.cloop.sptk .l1 ;;
|
|
Packit |
6c4009 |
.str_aligned:
|
|
Packit |
6c4009 |
cmp.ne p6, p0 = r0, r0 // clear p6
|
|
Packit |
6c4009 |
shr.u loopcnt = len, 3 // loopcnt = len / 8
|
|
Packit |
6c4009 |
and len = 7, len ;; // remaining len = len & 7
|
|
Packit |
6c4009 |
adds loopcnt = -1, loopcnt
|
|
Packit |
6c4009 |
mov ar.ec = MEMLAT + 3
|
|
Packit |
6c4009 |
mux1 chrx8 = chr, @brcst ;; // get a word full of chr
|
|
Packit |
6c4009 |
mov ar.lc = loopcnt
|
|
Packit |
6c4009 |
mov pr.rot = 1 << 16 ;;
|
|
Packit |
6c4009 |
.l2:
|
|
Packit |
6c4009 |
(p[0]) mov addr[0] = ret0
|
|
Packit |
6c4009 |
(p[0]) ld8.s value[0] = [ret0], 8 // speculative load
|
|
Packit |
6c4009 |
(p[MEMLAT]) chk.s value[MEMLAT], .recovery // check and recovery
|
|
Packit |
6c4009 |
(p[MEMLAT]) xor aux[0] = value[MEMLAT], chrx8
|
|
Packit |
6c4009 |
(p[MEMLAT+1]) czx1.r poschr[0] = aux[1]
|
|
Packit |
6c4009 |
(p[MEMLAT+2]) cmp.ne p7, p0 = 8, poschr[1]
|
|
Packit |
6c4009 |
(p7) br.cond.dpnt .foundit
|
|
Packit |
6c4009 |
br.ctop.dptk .l2
|
|
Packit |
6c4009 |
.srchfew:
|
|
Packit |
6c4009 |
adds loopcnt = -1, len
|
|
Packit |
6c4009 |
cmp.eq p6, p0 = len, r0
|
|
Packit |
6c4009 |
(p6) br.cond.spnt .notfound ;;
|
|
Packit |
6c4009 |
mov ar.lc = loopcnt
|
|
Packit |
6c4009 |
.l3:
|
|
Packit |
6c4009 |
ld1 val = [ret0], 1
|
|
Packit |
6c4009 |
;;
|
|
Packit |
6c4009 |
cmp.eq p6, p0 = val, chr
|
|
Packit |
6c4009 |
(p6) br.cond.dpnt .foundit
|
|
Packit |
6c4009 |
br.cloop.sptk .l3 ;;
|
|
Packit |
6c4009 |
.notfound:
|
|
Packit |
6c4009 |
cmp.ne p6, p0 = r0, r0 // clear p6 (p7 was already 0 when we got here)
|
|
Packit |
6c4009 |
mov ret0 = r0 ;; // return NULL
|
|
Packit |
6c4009 |
.foundit:
|
|
Packit |
6c4009 |
.pred.rel "mutex" p6, p7
|
|
Packit |
6c4009 |
(p6) adds ret0 = -1, ret0 // if we got here from l1 or l3
|
|
Packit |
6c4009 |
(p7) add ret0 = addr[MEMLAT+2], poschr[1] // if we got here from l2
|
|
Packit |
6c4009 |
mov pr = saved_pr, -1
|
|
Packit |
6c4009 |
mov ar.lc = saved_lc
|
|
Packit |
6c4009 |
br.ret.sptk.many b0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.recovery:
|
|
Packit |
6c4009 |
#if MEMLAT != 6
|
|
Packit |
6c4009 |
# error "MEMLAT must be 6!"
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
(p[MEMLAT-6]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT-5]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT-4]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT-3]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT-2]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT-1]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT+1]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
(p[MEMLAT+2]) add ret0 = -8, ret0;;
|
|
Packit |
6c4009 |
.l4:
|
|
Packit |
6c4009 |
mov addr[MEMLAT+2] = ret0
|
|
Packit |
6c4009 |
ld8 tmp = [ret0];; // load the first unchecked 8byte
|
|
Packit |
6c4009 |
xor aux[1] = tmp, chrx8;;
|
|
Packit |
6c4009 |
czx1.r poschr[1] = aux[1];;
|
|
Packit |
6c4009 |
cmp.ne p7, p0 = 8, poschr[1];;
|
|
Packit |
6c4009 |
(p7) add ret0 = addr[MEMLAT+2], poschr[1];;
|
|
Packit |
6c4009 |
(p7) cmp.geu p6, p7 = ret0, last // don't go over the last byte
|
|
Packit |
6c4009 |
(p6) br.cond.spnt .notfound;;
|
|
Packit |
6c4009 |
(p7) br.cond.spnt .foundit;;
|
|
Packit |
6c4009 |
adds ret0 = 8, ret0 // load the next unchecked 8byte
|
|
Packit |
6c4009 |
br.sptk .l4;;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END(__memchr)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
weak_alias (__memchr, memchr)
|
|
Packit |
6c4009 |
libc_hidden_builtin_def (memchr)
|