root / host-utils.h @ ed23fbd9
History | View | Annotate | Download (5.1 kB)
1 | 05f778c8 | ths | /*
|
---|---|---|---|
2 | 05f778c8 | ths | * Utility compute operations used by translated code.
|
3 | 05f778c8 | ths | *
|
4 | 05f778c8 | ths | * Copyright (c) 2007 Thiemo Seufer
|
5 | 05f778c8 | ths | * Copyright (c) 2007 Jocelyn Mayer
|
6 | 05f778c8 | ths | *
|
7 | 05f778c8 | ths | * Permission is hereby granted, free of charge, to any person obtaining a copy
|
8 | 05f778c8 | ths | * of this software and associated documentation files (the "Software"), to deal
|
9 | 05f778c8 | ths | * in the Software without restriction, including without limitation the rights
|
10 | 05f778c8 | ths | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
11 | 05f778c8 | ths | * copies of the Software, and to permit persons to whom the Software is
|
12 | 05f778c8 | ths | * furnished to do so, subject to the following conditions:
|
13 | 05f778c8 | ths | *
|
14 | 05f778c8 | ths | * The above copyright notice and this permission notice shall be included in
|
15 | 05f778c8 | ths | * all copies or substantial portions of the Software.
|
16 | 05f778c8 | ths | *
|
17 | 05f778c8 | ths | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
18 | 05f778c8 | ths | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
19 | 05f778c8 | ths | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
|
20 | 05f778c8 | ths | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
21 | 05f778c8 | ths | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
22 | 05f778c8 | ths | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
23 | 05f778c8 | ths | * THE SOFTWARE.
|
24 | 05f778c8 | ths | */
|
25 | 05f778c8 | ths | |
26 | cebdff77 | ths | #include "osdep.h" |
27 | cebdff77 | ths | |
28 | 7a51ad82 | j_mayer | #if defined(__x86_64__)
|
29 | 7a51ad82 | j_mayer | #define __HAVE_FAST_MULU64__
|
30 | 7a51ad82 | j_mayer | static always_inline void mulu64 (uint64_t *plow, uint64_t *phigh, |
31 | 7a51ad82 | j_mayer | uint64_t a, uint64_t b) |
32 | 7a51ad82 | j_mayer | { |
33 | 7a51ad82 | j_mayer | __asm__ ("mul %0\n\t"
|
34 | 7a51ad82 | j_mayer | : "=d" (*phigh), "=a" (*plow) |
35 | 7a51ad82 | j_mayer | : "a" (a), "0" (b)); |
36 | 7a51ad82 | j_mayer | } |
37 | 7a51ad82 | j_mayer | #define __HAVE_FAST_MULS64__
|
38 | 7a51ad82 | j_mayer | static always_inline void muls64 (uint64_t *plow, uint64_t *phigh, |
39 | 7a51ad82 | j_mayer | int64_t a, int64_t b) |
40 | 7a51ad82 | j_mayer | { |
41 | 7a51ad82 | j_mayer | __asm__ ("imul %0\n\t"
|
42 | 7a51ad82 | j_mayer | : "=d" (*phigh), "=a" (*plow) |
43 | 7a51ad82 | j_mayer | : "a" (a), "0" (b)); |
44 | 7a51ad82 | j_mayer | } |
45 | 7a51ad82 | j_mayer | #else
|
46 | 05e1d830 | j_mayer | void muls64(uint64_t *phigh, uint64_t *plow, int64_t a, int64_t b);
|
47 | 7a51ad82 | j_mayer | void mulu64(uint64_t *phigh, uint64_t *plow, uint64_t a, uint64_t b);
|
48 | 7a51ad82 | j_mayer | #endif
|
49 | 7a51ad82 | j_mayer | |
50 | 05f778c8 | ths | /* Note that some of those functions may end up calling libgcc functions,
|
51 | 05f778c8 | ths | depending on the host machine. It is up to the target emulation to
|
52 | 05f778c8 | ths | cope with that. */
|
53 | 05f778c8 | ths | |
54 | 05f778c8 | ths | /* Binary search for leading zeros. */
|
55 | 05f778c8 | ths | |
56 | 05f778c8 | ths | static always_inline int clz32(uint32_t val) |
57 | 05f778c8 | ths | { |
58 | 05f778c8 | ths | int cnt = 0; |
59 | 05f778c8 | ths | |
60 | 05f778c8 | ths | if (!(val & 0xFFFF0000U)) { |
61 | 05f778c8 | ths | cnt += 16;
|
62 | 05f778c8 | ths | val <<= 16;
|
63 | 05f778c8 | ths | } |
64 | 05f778c8 | ths | if (!(val & 0xFF000000U)) { |
65 | 05f778c8 | ths | cnt += 8;
|
66 | 05f778c8 | ths | val <<= 8;
|
67 | 05f778c8 | ths | } |
68 | 05f778c8 | ths | if (!(val & 0xF0000000U)) { |
69 | 05f778c8 | ths | cnt += 4;
|
70 | 05f778c8 | ths | val <<= 4;
|
71 | 05f778c8 | ths | } |
72 | 05f778c8 | ths | if (!(val & 0xC0000000U)) { |
73 | 05f778c8 | ths | cnt += 2;
|
74 | 05f778c8 | ths | val <<= 2;
|
75 | 05f778c8 | ths | } |
76 | 05f778c8 | ths | if (!(val & 0x80000000U)) { |
77 | 05f778c8 | ths | cnt++; |
78 | 05f778c8 | ths | val <<= 1;
|
79 | 05f778c8 | ths | } |
80 | 05f778c8 | ths | if (!(val & 0x80000000U)) { |
81 | 05f778c8 | ths | cnt++; |
82 | 05f778c8 | ths | } |
83 | 05f778c8 | ths | return cnt;
|
84 | 05f778c8 | ths | } |
85 | 05f778c8 | ths | |
86 | 05f778c8 | ths | static always_inline int clo32(uint32_t val) |
87 | 05f778c8 | ths | { |
88 | 05f778c8 | ths | return clz32(~val);
|
89 | 05f778c8 | ths | } |
90 | 05f778c8 | ths | |
91 | 05f778c8 | ths | static always_inline int clz64(uint64_t val) |
92 | 05f778c8 | ths | { |
93 | 05f778c8 | ths | int cnt = 0; |
94 | 05f778c8 | ths | |
95 | 7a51ad82 | j_mayer | if (!(val >> 32)) { |
96 | 05f778c8 | ths | cnt += 32;
|
97 | 7a51ad82 | j_mayer | } else {
|
98 | 7a51ad82 | j_mayer | val >>= 32;
|
99 | 05f778c8 | ths | } |
100 | 7a51ad82 | j_mayer | |
101 | 7a51ad82 | j_mayer | return cnt + clz32(val);
|
102 | 05f778c8 | ths | } |
103 | 05f778c8 | ths | |
104 | 05f778c8 | ths | static always_inline int clo64(uint64_t val) |
105 | 05f778c8 | ths | { |
106 | 05f778c8 | ths | return clz64(~val);
|
107 | 05f778c8 | ths | } |
108 | b9ef45ff | j_mayer | |
109 | b9ef45ff | j_mayer | static always_inline int ctz32 (uint32_t val) |
110 | b9ef45ff | j_mayer | { |
111 | b9ef45ff | j_mayer | int cnt;
|
112 | b9ef45ff | j_mayer | |
113 | b9ef45ff | j_mayer | cnt = 0;
|
114 | b9ef45ff | j_mayer | if (!(val & 0x0000FFFFUL)) { |
115 | b9ef45ff | j_mayer | cnt += 16;
|
116 | b9ef45ff | j_mayer | val >>= 16;
|
117 | b9ef45ff | j_mayer | } |
118 | b9ef45ff | j_mayer | if (!(val & 0x000000FFUL)) { |
119 | b9ef45ff | j_mayer | cnt += 8;
|
120 | b9ef45ff | j_mayer | val >>= 8;
|
121 | b9ef45ff | j_mayer | } |
122 | b9ef45ff | j_mayer | if (!(val & 0x0000000FUL)) { |
123 | b9ef45ff | j_mayer | cnt += 4;
|
124 | b9ef45ff | j_mayer | val >>= 4;
|
125 | b9ef45ff | j_mayer | } |
126 | b9ef45ff | j_mayer | if (!(val & 0x00000003UL)) { |
127 | b9ef45ff | j_mayer | cnt += 2;
|
128 | b9ef45ff | j_mayer | val >>= 2;
|
129 | b9ef45ff | j_mayer | } |
130 | b9ef45ff | j_mayer | if (!(val & 0x00000001UL)) { |
131 | b9ef45ff | j_mayer | cnt++; |
132 | b9ef45ff | j_mayer | val >>= 1;
|
133 | b9ef45ff | j_mayer | } |
134 | b9ef45ff | j_mayer | if (!(val & 0x00000001UL)) { |
135 | b9ef45ff | j_mayer | cnt++; |
136 | b9ef45ff | j_mayer | } |
137 | b9ef45ff | j_mayer | |
138 | b9ef45ff | j_mayer | return cnt;
|
139 | b9ef45ff | j_mayer | } |
140 | b9ef45ff | j_mayer | |
141 | b9ef45ff | j_mayer | static always_inline int cto32 (uint32_t val) |
142 | b9ef45ff | j_mayer | { |
143 | b9ef45ff | j_mayer | return ctz32(~val);
|
144 | b9ef45ff | j_mayer | } |
145 | b9ef45ff | j_mayer | |
146 | b9ef45ff | j_mayer | static always_inline int ctz64 (uint64_t val) |
147 | b9ef45ff | j_mayer | { |
148 | b9ef45ff | j_mayer | int cnt;
|
149 | b9ef45ff | j_mayer | |
150 | b9ef45ff | j_mayer | cnt = 0;
|
151 | b9ef45ff | j_mayer | if (!((uint32_t)val)) {
|
152 | b9ef45ff | j_mayer | cnt += 32;
|
153 | b9ef45ff | j_mayer | val >>= 32;
|
154 | b9ef45ff | j_mayer | } |
155 | b9ef45ff | j_mayer | |
156 | b9ef45ff | j_mayer | return cnt + ctz32(val);
|
157 | b9ef45ff | j_mayer | } |
158 | b9ef45ff | j_mayer | |
159 | b9ef45ff | j_mayer | static always_inline int cto64 (uint64_t val) |
160 | b9ef45ff | j_mayer | { |
161 | b9ef45ff | j_mayer | return ctz64(~val);
|
162 | b9ef45ff | j_mayer | } |
163 | b9ef45ff | j_mayer | |
164 | b9ef45ff | j_mayer | static always_inline int ctpop8 (uint8_t val) |
165 | b9ef45ff | j_mayer | { |
166 | b9ef45ff | j_mayer | val = (val & 0x55) + ((val >> 1) & 0x55); |
167 | b9ef45ff | j_mayer | val = (val & 0x33) + ((val >> 2) & 0x33); |
168 | b9ef45ff | j_mayer | val = (val & 0x0f) + ((val >> 4) & 0x0f); |
169 | b9ef45ff | j_mayer | |
170 | b9ef45ff | j_mayer | return val;
|
171 | b9ef45ff | j_mayer | } |
172 | b9ef45ff | j_mayer | |
173 | b9ef45ff | j_mayer | static always_inline int ctpop16 (uint16_t val) |
174 | b9ef45ff | j_mayer | { |
175 | b9ef45ff | j_mayer | val = (val & 0x5555) + ((val >> 1) & 0x5555); |
176 | b9ef45ff | j_mayer | val = (val & 0x3333) + ((val >> 2) & 0x3333); |
177 | b9ef45ff | j_mayer | val = (val & 0x0f0f) + ((val >> 4) & 0x0f0f); |
178 | b9ef45ff | j_mayer | val = (val & 0x00ff) + ((val >> 8) & 0x00ff); |
179 | b9ef45ff | j_mayer | |
180 | b9ef45ff | j_mayer | return val;
|
181 | b9ef45ff | j_mayer | } |
182 | b9ef45ff | j_mayer | |
183 | b9ef45ff | j_mayer | static always_inline int ctpop32 (uint32_t val) |
184 | b9ef45ff | j_mayer | { |
185 | b9ef45ff | j_mayer | val = (val & 0x55555555) + ((val >> 1) & 0x55555555); |
186 | b9ef45ff | j_mayer | val = (val & 0x33333333) + ((val >> 2) & 0x33333333); |
187 | b9ef45ff | j_mayer | val = (val & 0x0f0f0f0f) + ((val >> 4) & 0x0f0f0f0f); |
188 | b9ef45ff | j_mayer | val = (val & 0x00ff00ff) + ((val >> 8) & 0x00ff00ff); |
189 | b9ef45ff | j_mayer | val = (val & 0x0000ffff) + ((val >> 16) & 0x0000ffff); |
190 | b9ef45ff | j_mayer | |
191 | b9ef45ff | j_mayer | return val;
|
192 | b9ef45ff | j_mayer | } |
193 | b9ef45ff | j_mayer | |
194 | b9ef45ff | j_mayer | static always_inline int ctpop64 (uint64_t val) |
195 | b9ef45ff | j_mayer | { |
196 | b9ef45ff | j_mayer | val = (val & 0x5555555555555555ULL) + ((val >> 1) & 0x5555555555555555ULL); |
197 | b9ef45ff | j_mayer | val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL); |
198 | b9ef45ff | j_mayer | val = (val & 0x0f0f0f0f0f0f0f0fULL) + ((val >> 4) & 0x0f0f0f0f0f0f0f0fULL); |
199 | b9ef45ff | j_mayer | val = (val & 0x00ff00ff00ff00ffULL) + ((val >> 8) & 0x00ff00ff00ff00ffULL); |
200 | b9ef45ff | j_mayer | val = (val & 0x0000ffff0000ffffULL) + ((val >> 16) & 0x0000ffff0000ffffULL); |
201 | b9ef45ff | j_mayer | val = (val & 0x00000000ffffffffULL) + ((val >> 32) & 0x00000000ffffffffULL); |
202 | b9ef45ff | j_mayer | |
203 | b9ef45ff | j_mayer | return val;
|
204 | 3800af9e | ths | } |