1 /* Optimized version of the standard memcmp() function.
2 This file is part of the GNU C Library.
3 Copyright (C) 2000-2018 Free Software Foundation, Inc.
4 Contributed by Dan Pop <Dan.Pop@cern.ch>.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <http://www.gnu.org/licenses/>. */
20 /* Return: the result of the comparison
27 In this form, it assumes little endian mode. For big endian mode,
28 the two shifts in .l2 must be inverted:
30 shl tmp1[0] = r[1 + MEMLAT], sh1 // tmp1 = w0 << sh1
31 shr.u tmp2[0] = r[0 + MEMLAT], sh2 // tmp2 = w1 >> sh2
33 and all the mux1 instructions should be replaced by plain mov's. */
58 alloc r2 = ar.pfs, 3, 37, 0, 40
60 .rotr r[MEMLAT + 2], q[MEMLAT + 5], tmp1[4], tmp2[4], val[2]
61 .rotp p[MEMLAT + 4 + 1]
63 mov ret0 = r0 // by default return value = 0
65 mov saved_pr = pr // save the predicate registers
67 mov saved_lc = ar.lc // save the loop counter
69 mov dest = in0 // dest
72 sub tmp = r0, in0 // tmp = -dest
74 and loopcnt = 7, tmp // loopcnt = -dest % 8
75 cmp.ge p6, p0 = OP_T_THRES, len // is len <= OP_T_THRES
76 (p6) br.cond.spnt .cmpfew // compare byte by byte
78 cmp.eq p6, p0 = loopcnt, r0
79 (p6) br.cond.sptk .dest_aligned
80 sub len = len, loopcnt // len -= -dest % 8
81 adds loopcnt = -1, loopcnt // --loopcnt
84 .l1: // copy -dest % 8 bytes
85 ld1 value1 = [src], 1 // value = *src++
86 ld1 value2 = [dest], 1
88 cmp.ne p6, p0 = value1, value2
89 (p6) br.cond.spnt .done
92 and sh1 = 7, src // sh1 = src % 8
93 and tmp = -8, len // tmp = len & -OPSIZ
94 and asrc = -8, src // asrc = src & -OPSIZ -- align src
95 shr.u loopcnt = len, 3 // loopcnt = len / 8
96 and len = 7, len ;; // len = len % 8
97 shl sh1 = sh1, 3 // sh1 = 8 * (src % 8)
98 adds loopcnt = -1, loopcnt // --loopcnt
99 mov pr.rot = 1 << 16 ;; // set rotating predicates
100 sub sh2 = 64, sh1 // sh2 = 64 - sh1
101 mov ar.lc = loopcnt // set LC
102 cmp.eq p6, p0 = sh1, r0 // is the src aligned?
103 (p6) br.cond.sptk .src_aligned
104 add src = src, tmp // src += len & -OPSIZ
105 mov ar.ec = MEMLAT + 4 + 1 // four more passes needed
106 ld8 r[1] = [asrc], 8 ;; // r[1] = w0
109 // We enter this loop with p6 cleared by the above comparison
112 (p[0]) ld8 r[0] = [asrc], 8 // r[0] = w1
113 (p[0]) ld8 q[0] = [dest], 8
114 (p[MEMLAT]) shr.u tmp1[0] = r[1 + MEMLAT], sh1 // tmp1 = w0 >> sh1
115 (p[MEMLAT]) shl tmp2[0] = r[0 + MEMLAT], sh2 // tmp2 = w1 << sh2
116 (p[MEMLAT+4]) cmp.ne p6, p0 = q[MEMLAT + 4], val[1]
117 (p[MEMLAT+3]) or val[0] = tmp1[3], tmp2[3] // val = tmp1 | tmp2
118 (p6) br.cond.spnt .l2exit
122 mux1 value1 = r[MEMLAT], @rev
123 mux1 value2 = q[MEMLAT], @rev
124 cmp.ne p6, p0 = r0, r0 ;; // clear p6
126 (p6) mux1 value1 = val[1], @rev
127 (p6) mux1 value2 = q[MEMLAT + 4], @rev ;;
128 cmp.ltu p6, p7 = value2, value1 ;;
131 mov pr = saved_pr, -1 // restore the predicate registers
132 mov ar.lc = saved_lc // restore the loop counter
135 cmp.ne p6, p0 = r0, r0 // clear p6
136 mov ar.ec = MEMLAT + 1 ;; // set EC
138 (p[0]) ld8 r[0] = [src], 8
139 (p[0]) ld8 q[0] = [dest], 8
140 (p[MEMLAT]) cmp.ne p6, p0 = r[MEMLAT], q[MEMLAT]
141 (p6) br.cond.spnt .l3exit
144 cmp.eq p6, p0 = len, r0 // is len == 0 ?
145 adds len = -1, len // --len;
146 (p6) br.cond.spnt .restore_and_exit ;;
149 ld1 value1 = [src], 1
150 ld1 value2 = [dest], 1
152 cmp.ne p6, p0 = value1, value2
153 (p6) br.cond.spnt .done
156 (p6) sub ret0 = value2, value1 // don't execute it if falling thru
158 mov pr = saved_pr, -1 // restore the predicate registers
159 mov ar.lc = saved_lc // restore the loop counter
163 weak_alias (memcmp, bcmp)
164 libc_hidden_builtin_def (memcmp)