]>
Commit | Line | Data |
---|---|---|
d5efd131 MF |
1 | /* Optimized version of the standard memcmp() function. |
2 | This file is part of the GNU C Library. | |
2b778ceb | 3 | Copyright (C) 2000-2021 Free Software Foundation, Inc. |
d5efd131 MF |
4 | Contributed by Dan Pop <Dan.Pop@cern.ch>. |
5 | ||
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. | |
10 | ||
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. | |
15 | ||
16 | You should have received a copy of the GNU Lesser General Public | |
75efb018 | 17 | License along with the GNU C Library; if not, see |
5a82c748 | 18 | <https://www.gnu.org/licenses/>. */ |
d5efd131 MF |
19 | |
20 | /* Return: the result of the comparison | |
21 | ||
22 | Inputs: | |
23 | in0: dest (aka s1) | |
24 | in1: src (aka s2) | |
25 | in2: byte count | |
26 | ||
27 | In this form, it assumes little endian mode. For big endian mode, | |
28 | the two shifts in .l2 must be inverted: | |
29 | ||
30 | shl tmp1[0] = r[1 + MEMLAT], sh1 // tmp1 = w0 << sh1 | |
31 | shr.u tmp2[0] = r[0 + MEMLAT], sh2 // tmp2 = w1 >> sh2 | |
32 | ||
33 | and all the mux1 instructions should be replaced by plain mov's. */ | |
34 | ||
35 | #include <sysdep.h> | |
36 | #undef ret | |
37 | ||
38 | #define OP_T_THRES 16 | |
39 | #define OPSIZ 8 | |
40 | #define MEMLAT 2 | |
41 | ||
42 | #define start r15 | |
43 | #define saved_pr r17 | |
44 | #define saved_lc r18 | |
45 | #define dest r19 | |
46 | #define src r20 | |
47 | #define len r21 | |
48 | #define asrc r22 | |
49 | #define tmp r23 | |
50 | #define value1 r24 | |
51 | #define value2 r25 | |
52 | #define sh2 r28 | |
53 | #define sh1 r29 | |
54 | #define loopcnt r30 | |
55 | ||
56 | ENTRY(memcmp) | |
57 | .prologue | |
58 | alloc r2 = ar.pfs, 3, 37, 0, 40 | |
59 | ||
60 | .rotr r[MEMLAT + 2], q[MEMLAT + 5], tmp1[4], tmp2[4], val[2] | |
61 | .rotp p[MEMLAT + 4 + 1] | |
62 | ||
63 | mov ret0 = r0 // by default return value = 0 | |
64 | .save pr, saved_pr | |
65 | mov saved_pr = pr // save the predicate registers | |
66 | .save ar.lc, saved_lc | |
67 | mov saved_lc = ar.lc // save the loop counter | |
68 | .body | |
69 | mov dest = in0 // dest | |
70 | mov src = in1 // src | |
71 | mov len = in2 // len | |
72 | sub tmp = r0, in0 // tmp = -dest | |
73 | ;; | |
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 | |
77 | ;; | |
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 | |
82 | ;; | |
83 | mov ar.lc = loopcnt | |
84 | .l1: // copy -dest % 8 bytes | |
85 | ld1 value1 = [src], 1 // value = *src++ | |
86 | ld1 value2 = [dest], 1 | |
87 | ;; | |
88 | cmp.ne p6, p0 = value1, value2 | |
89 | (p6) br.cond.spnt .done | |
90 | br.cloop.dptk .l1 | |
91 | .dest_aligned: | |
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 | |
107 | .align 32 | |
108 | ||
109 | // We enter this loop with p6 cleared by the above comparison | |
110 | ||
111 | .l2: | |
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 | |
119 | br.ctop.sptk .l2 | |
120 | br.cond.sptk .cmpfew | |
121 | .l3exit: | |
122 | mux1 value1 = r[MEMLAT], @rev | |
123 | mux1 value2 = q[MEMLAT], @rev | |
124 | cmp.ne p6, p0 = r0, r0 ;; // clear p6 | |
125 | .l2exit: | |
126 | (p6) mux1 value1 = val[1], @rev | |
127 | (p6) mux1 value2 = q[MEMLAT + 4], @rev ;; | |
128 | cmp.ltu p6, p7 = value2, value1 ;; | |
129 | (p6) mov ret0 = -1 | |
130 | (p7) mov ret0 = 1 | |
131 | mov pr = saved_pr, -1 // restore the predicate registers | |
132 | mov ar.lc = saved_lc // restore the loop counter | |
133 | br.ret.sptk.many b0 | |
134 | .src_aligned: | |
135 | cmp.ne p6, p0 = r0, r0 // clear p6 | |
136 | mov ar.ec = MEMLAT + 1 ;; // set EC | |
137 | .l3: | |
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 | |
142 | br.ctop.dptk .l3 ;; | |
143 | .cmpfew: | |
144 | cmp.eq p6, p0 = len, r0 // is len == 0 ? | |
145 | adds len = -1, len // --len; | |
146 | (p6) br.cond.spnt .restore_and_exit ;; | |
147 | mov ar.lc = len | |
148 | .l4: | |
149 | ld1 value1 = [src], 1 | |
150 | ld1 value2 = [dest], 1 | |
151 | ;; | |
152 | cmp.ne p6, p0 = value1, value2 | |
153 | (p6) br.cond.spnt .done | |
154 | br.cloop.dptk .l4 ;; | |
155 | .done: | |
156 | (p6) sub ret0 = value2, value1 // don't execute it if falling thru | |
157 | .restore_and_exit: | |
158 | mov pr = saved_pr, -1 // restore the predicate registers | |
159 | mov ar.lc = saved_lc // restore the loop counter | |
160 | br.ret.sptk.many b0 | |
161 | END(memcmp) | |
162 | ||
163 | weak_alias (memcmp, bcmp) | |
164 | libc_hidden_builtin_def (memcmp) |