root / target-arm / op_neon.h @ 4373f3ce
History | View | Annotate | Download (35.4 kB)
1 | 9ee6e8bb | pbrook | /*
|
---|---|---|---|
2 | 9ee6e8bb | pbrook | * ARM NEON vector operations.
|
3 | 9ee6e8bb | pbrook | *
|
4 | 9ee6e8bb | pbrook | * Copyright (c) 2007 CodeSourcery.
|
5 | 9ee6e8bb | pbrook | * Written by Paul Brook
|
6 | 9ee6e8bb | pbrook | *
|
7 | 9ee6e8bb | pbrook | * This code is licenced under the GPL.
|
8 | 9ee6e8bb | pbrook | */
|
9 | 9ee6e8bb | pbrook | /* Note that for NEON an "l" prefix means it is a wide operation, unlike
|
10 | 9ee6e8bb | pbrook | scalar arm ops where it means a word size operation. */
|
11 | 9ee6e8bb | pbrook | |
12 | 1497c961 | pbrook | #define SIGNBIT (uint32_t)0x80000000 |
13 | 9ee6e8bb | pbrook | /* ??? NEON ops should probably have their own float status. */
|
14 | 9ee6e8bb | pbrook | #define NFS &env->vfp.fp_status
|
15 | 9ee6e8bb | pbrook | #define NEON_OP(name) void OPPROTO op_neon_##name (void) |
16 | 9ee6e8bb | pbrook | |
17 | 4373f3ce | pbrook | /* Helper routines to perform bitwise copies between float and int. */
|
18 | 4373f3ce | pbrook | static inline float32 vfp_itos(uint32_t i) |
19 | 4373f3ce | pbrook | { |
20 | 4373f3ce | pbrook | union {
|
21 | 4373f3ce | pbrook | uint32_t i; |
22 | 4373f3ce | pbrook | float32 s; |
23 | 4373f3ce | pbrook | } v; |
24 | 4373f3ce | pbrook | |
25 | 4373f3ce | pbrook | v.i = i; |
26 | 4373f3ce | pbrook | return v.s;
|
27 | 4373f3ce | pbrook | } |
28 | 4373f3ce | pbrook | |
29 | 4373f3ce | pbrook | static inline uint32_t vfp_stoi(float32 s) |
30 | 4373f3ce | pbrook | { |
31 | 4373f3ce | pbrook | union {
|
32 | 4373f3ce | pbrook | uint32_t i; |
33 | 4373f3ce | pbrook | float32 s; |
34 | 4373f3ce | pbrook | } v; |
35 | 4373f3ce | pbrook | |
36 | 4373f3ce | pbrook | v.s = s; |
37 | 4373f3ce | pbrook | return v.i;
|
38 | 4373f3ce | pbrook | } |
39 | 4373f3ce | pbrook | |
40 | 9ee6e8bb | pbrook | NEON_OP(getreg_T0) |
41 | 9ee6e8bb | pbrook | { |
42 | 9ee6e8bb | pbrook | T0 = *(uint32_t *)((char *) env + PARAM1);
|
43 | 9ee6e8bb | pbrook | } |
44 | 9ee6e8bb | pbrook | |
45 | 9ee6e8bb | pbrook | NEON_OP(getreg_T1) |
46 | 9ee6e8bb | pbrook | { |
47 | 9ee6e8bb | pbrook | T1 = *(uint32_t *)((char *) env + PARAM1);
|
48 | 9ee6e8bb | pbrook | } |
49 | 9ee6e8bb | pbrook | |
50 | 9ee6e8bb | pbrook | NEON_OP(getreg_T2) |
51 | 9ee6e8bb | pbrook | { |
52 | 9ee6e8bb | pbrook | T2 = *(uint32_t *)((char *) env + PARAM1);
|
53 | 9ee6e8bb | pbrook | } |
54 | 9ee6e8bb | pbrook | |
55 | 9ee6e8bb | pbrook | NEON_OP(setreg_T0) |
56 | 9ee6e8bb | pbrook | { |
57 | 9ee6e8bb | pbrook | *(uint32_t *)((char *) env + PARAM1) = T0;
|
58 | 9ee6e8bb | pbrook | } |
59 | 9ee6e8bb | pbrook | |
60 | 9ee6e8bb | pbrook | NEON_OP(setreg_T1) |
61 | 9ee6e8bb | pbrook | { |
62 | 9ee6e8bb | pbrook | *(uint32_t *)((char *) env + PARAM1) = T1;
|
63 | 9ee6e8bb | pbrook | } |
64 | 9ee6e8bb | pbrook | |
65 | 9ee6e8bb | pbrook | NEON_OP(setreg_T2) |
66 | 9ee6e8bb | pbrook | { |
67 | 9ee6e8bb | pbrook | *(uint32_t *)((char *) env + PARAM1) = T2;
|
68 | 9ee6e8bb | pbrook | } |
69 | 9ee6e8bb | pbrook | |
70 | 9ee6e8bb | pbrook | #define NEON_TYPE1(name, type) \
|
71 | 9ee6e8bb | pbrook | typedef struct \ |
72 | 9ee6e8bb | pbrook | { \ |
73 | 9ee6e8bb | pbrook | type v1; \ |
74 | 9ee6e8bb | pbrook | } neon_##name; |
75 | 9ee6e8bb | pbrook | #ifdef WORDS_BIGENDIAN
|
76 | 9ee6e8bb | pbrook | #define NEON_TYPE2(name, type) \
|
77 | 9ee6e8bb | pbrook | typedef struct \ |
78 | 9ee6e8bb | pbrook | { \ |
79 | 9ee6e8bb | pbrook | type v2; \ |
80 | 9ee6e8bb | pbrook | type v1; \ |
81 | 9ee6e8bb | pbrook | } neon_##name; |
82 | 9ee6e8bb | pbrook | #define NEON_TYPE4(name, type) \
|
83 | 9ee6e8bb | pbrook | typedef struct \ |
84 | 9ee6e8bb | pbrook | { \ |
85 | 9ee6e8bb | pbrook | type v4; \ |
86 | 9ee6e8bb | pbrook | type v3; \ |
87 | 9ee6e8bb | pbrook | type v2; \ |
88 | 9ee6e8bb | pbrook | type v1; \ |
89 | 9ee6e8bb | pbrook | } neon_##name; |
90 | 9ee6e8bb | pbrook | #else
|
91 | 9ee6e8bb | pbrook | #define NEON_TYPE2(name, type) \
|
92 | 9ee6e8bb | pbrook | typedef struct \ |
93 | 9ee6e8bb | pbrook | { \ |
94 | 9ee6e8bb | pbrook | type v1; \ |
95 | 9ee6e8bb | pbrook | type v2; \ |
96 | 9ee6e8bb | pbrook | } neon_##name; |
97 | 9ee6e8bb | pbrook | #define NEON_TYPE4(name, type) \
|
98 | 9ee6e8bb | pbrook | typedef struct \ |
99 | 9ee6e8bb | pbrook | { \ |
100 | 9ee6e8bb | pbrook | type v1; \ |
101 | 9ee6e8bb | pbrook | type v2; \ |
102 | 9ee6e8bb | pbrook | type v3; \ |
103 | 9ee6e8bb | pbrook | type v4; \ |
104 | 9ee6e8bb | pbrook | } neon_##name; |
105 | 9ee6e8bb | pbrook | #endif
|
106 | 9ee6e8bb | pbrook | |
107 | 9ee6e8bb | pbrook | NEON_TYPE4(s8, int8_t) |
108 | 9ee6e8bb | pbrook | NEON_TYPE4(u8, uint8_t) |
109 | 9ee6e8bb | pbrook | NEON_TYPE2(s16, int16_t) |
110 | 9ee6e8bb | pbrook | NEON_TYPE2(u16, uint16_t) |
111 | 9ee6e8bb | pbrook | NEON_TYPE1(s32, int32_t) |
112 | 9ee6e8bb | pbrook | NEON_TYPE1(u32, uint32_t) |
113 | 9ee6e8bb | pbrook | #undef NEON_TYPE4
|
114 | 9ee6e8bb | pbrook | #undef NEON_TYPE2
|
115 | 9ee6e8bb | pbrook | #undef NEON_TYPE1
|
116 | 9ee6e8bb | pbrook | |
117 | 9ee6e8bb | pbrook | /* Copy from a uint32_t to a vector structure type. */
|
118 | 9ee6e8bb | pbrook | #define NEON_UNPACK(vtype, dest, val) do { \ |
119 | 9ee6e8bb | pbrook | union { \
|
120 | 9ee6e8bb | pbrook | vtype v; \ |
121 | 9ee6e8bb | pbrook | uint32_t i; \ |
122 | 9ee6e8bb | pbrook | } conv_u; \ |
123 | 9ee6e8bb | pbrook | conv_u.i = (val); \ |
124 | 9ee6e8bb | pbrook | dest = conv_u.v; \ |
125 | 9ee6e8bb | pbrook | } while(0) |
126 | 9ee6e8bb | pbrook | |
127 | 9ee6e8bb | pbrook | /* Copy from a vector structure type to a uint32_t. */
|
128 | 9ee6e8bb | pbrook | #define NEON_PACK(vtype, dest, val) do { \ |
129 | 9ee6e8bb | pbrook | union { \
|
130 | 9ee6e8bb | pbrook | vtype v; \ |
131 | 9ee6e8bb | pbrook | uint32_t i; \ |
132 | 9ee6e8bb | pbrook | } conv_u; \ |
133 | 9ee6e8bb | pbrook | conv_u.v = (val); \ |
134 | 9ee6e8bb | pbrook | dest = conv_u.i; \ |
135 | 9ee6e8bb | pbrook | } while(0) |
136 | 9ee6e8bb | pbrook | |
137 | 9ee6e8bb | pbrook | #define NEON_DO1 \
|
138 | 9ee6e8bb | pbrook | NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); |
139 | 9ee6e8bb | pbrook | #define NEON_DO2 \
|
140 | 9ee6e8bb | pbrook | NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
141 | 9ee6e8bb | pbrook | NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); |
142 | 9ee6e8bb | pbrook | #define NEON_DO4 \
|
143 | 9ee6e8bb | pbrook | NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
144 | 9ee6e8bb | pbrook | NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); \ |
145 | 9ee6e8bb | pbrook | NEON_FN(vdest.v3, vsrc1.v3, vsrc2.v3); \ |
146 | 9ee6e8bb | pbrook | NEON_FN(vdest.v4, vsrc1.v4, vsrc2.v4); |
147 | 9ee6e8bb | pbrook | |
148 | 9ee6e8bb | pbrook | #define NEON_VOP(name, vtype, n) \
|
149 | 9ee6e8bb | pbrook | NEON_OP(name) \ |
150 | 9ee6e8bb | pbrook | { \ |
151 | 9ee6e8bb | pbrook | vtype vsrc1; \ |
152 | 9ee6e8bb | pbrook | vtype vsrc2; \ |
153 | 9ee6e8bb | pbrook | vtype vdest; \ |
154 | 9ee6e8bb | pbrook | NEON_UNPACK(vtype, vsrc1, T0); \ |
155 | 9ee6e8bb | pbrook | NEON_UNPACK(vtype, vsrc2, T1); \ |
156 | 9ee6e8bb | pbrook | NEON_DO##n; \ |
157 | 9ee6e8bb | pbrook | NEON_PACK(vtype, T0, vdest); \ |
158 | 9ee6e8bb | pbrook | FORCE_RET(); \ |
159 | 9ee6e8bb | pbrook | } |
160 | 9ee6e8bb | pbrook | |
161 | 9ee6e8bb | pbrook | #define NEON_VOP1(name, vtype, n) \
|
162 | 9ee6e8bb | pbrook | NEON_OP(name) \ |
163 | 9ee6e8bb | pbrook | { \ |
164 | 9ee6e8bb | pbrook | vtype vsrc1; \ |
165 | 9ee6e8bb | pbrook | vtype vdest; \ |
166 | 9ee6e8bb | pbrook | NEON_UNPACK(vtype, vsrc1, T0); \ |
167 | 9ee6e8bb | pbrook | NEON_DO##n; \ |
168 | 9ee6e8bb | pbrook | NEON_PACK(vtype, T0, vdest); \ |
169 | 9ee6e8bb | pbrook | FORCE_RET(); \ |
170 | 9ee6e8bb | pbrook | } |
171 | 9ee6e8bb | pbrook | |
172 | 9ee6e8bb | pbrook | /* Pairwise operations. */
|
173 | 9ee6e8bb | pbrook | /* For 32-bit elements each segment only contains a single element, so
|
174 | 9ee6e8bb | pbrook | the elementwise and pairwise operations are the same. */
|
175 | 9ee6e8bb | pbrook | #define NEON_PDO2 \
|
176 | 9ee6e8bb | pbrook | NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
177 | 9ee6e8bb | pbrook | NEON_FN(vdest.v2, vsrc2.v1, vsrc2.v2); |
178 | 9ee6e8bb | pbrook | #define NEON_PDO4 \
|
179 | 9ee6e8bb | pbrook | NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
180 | 9ee6e8bb | pbrook | NEON_FN(vdest.v2, vsrc1.v3, vsrc1.v4); \ |
181 | 9ee6e8bb | pbrook | NEON_FN(vdest.v3, vsrc2.v1, vsrc2.v2); \ |
182 | 9ee6e8bb | pbrook | NEON_FN(vdest.v4, vsrc2.v3, vsrc2.v4); \ |
183 | 9ee6e8bb | pbrook | |
184 | 9ee6e8bb | pbrook | #define NEON_POP(name, vtype, n) \
|
185 | 9ee6e8bb | pbrook | NEON_OP(name) \ |
186 | 9ee6e8bb | pbrook | { \ |
187 | 9ee6e8bb | pbrook | vtype vsrc1; \ |
188 | 9ee6e8bb | pbrook | vtype vsrc2; \ |
189 | 9ee6e8bb | pbrook | vtype vdest; \ |
190 | 9ee6e8bb | pbrook | NEON_UNPACK(vtype, vsrc1, T0); \ |
191 | 9ee6e8bb | pbrook | NEON_UNPACK(vtype, vsrc2, T1); \ |
192 | 9ee6e8bb | pbrook | NEON_PDO##n; \ |
193 | 9ee6e8bb | pbrook | NEON_PACK(vtype, T0, vdest); \ |
194 | 9ee6e8bb | pbrook | FORCE_RET(); \ |
195 | 9ee6e8bb | pbrook | } |
196 | 9ee6e8bb | pbrook | |
197 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 + src2) >> 1 |
198 | 9ee6e8bb | pbrook | NEON_VOP(hadd_s8, neon_s8, 4)
|
199 | 9ee6e8bb | pbrook | NEON_VOP(hadd_u8, neon_u8, 4)
|
200 | 9ee6e8bb | pbrook | NEON_VOP(hadd_s16, neon_s16, 2)
|
201 | 9ee6e8bb | pbrook | NEON_VOP(hadd_u16, neon_u16, 2)
|
202 | 9ee6e8bb | pbrook | #undef NEON_FN
|
203 | 9ee6e8bb | pbrook | |
204 | 9ee6e8bb | pbrook | NEON_OP(hadd_s32) |
205 | 9ee6e8bb | pbrook | { |
206 | 9ee6e8bb | pbrook | int32_t src1 = T0; |
207 | 9ee6e8bb | pbrook | int32_t src2 = T1; |
208 | 9ee6e8bb | pbrook | int32_t dest; |
209 | 9ee6e8bb | pbrook | |
210 | 9ee6e8bb | pbrook | dest = (src1 >> 1) + (src2 >> 1); |
211 | 9ee6e8bb | pbrook | if (src1 & src2 & 1) |
212 | 9ee6e8bb | pbrook | dest++; |
213 | 9ee6e8bb | pbrook | T0 = dest; |
214 | 9ee6e8bb | pbrook | FORCE_RET(); |
215 | 9ee6e8bb | pbrook | } |
216 | 9ee6e8bb | pbrook | |
217 | 9ee6e8bb | pbrook | NEON_OP(hadd_u32) |
218 | 9ee6e8bb | pbrook | { |
219 | 9ee6e8bb | pbrook | uint32_t src1 = T0; |
220 | 9ee6e8bb | pbrook | uint32_t src2 = T1; |
221 | 9ee6e8bb | pbrook | uint32_t dest; |
222 | 9ee6e8bb | pbrook | |
223 | 9ee6e8bb | pbrook | dest = (src1 >> 1) + (src2 >> 1); |
224 | 9ee6e8bb | pbrook | if (src1 & src2 & 1) |
225 | 9ee6e8bb | pbrook | dest++; |
226 | 9ee6e8bb | pbrook | T0 = dest; |
227 | 9ee6e8bb | pbrook | FORCE_RET(); |
228 | 9ee6e8bb | pbrook | } |
229 | 9ee6e8bb | pbrook | |
230 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 + src2 + 1) >> 1 |
231 | 9ee6e8bb | pbrook | NEON_VOP(rhadd_s8, neon_s8, 4)
|
232 | 9ee6e8bb | pbrook | NEON_VOP(rhadd_u8, neon_u8, 4)
|
233 | 9ee6e8bb | pbrook | NEON_VOP(rhadd_s16, neon_s16, 2)
|
234 | 9ee6e8bb | pbrook | NEON_VOP(rhadd_u16, neon_u16, 2)
|
235 | 9ee6e8bb | pbrook | #undef NEON_FN
|
236 | 9ee6e8bb | pbrook | |
237 | 9ee6e8bb | pbrook | NEON_OP(rhadd_s32) |
238 | 9ee6e8bb | pbrook | { |
239 | 9ee6e8bb | pbrook | int32_t src1 = T0; |
240 | 9ee6e8bb | pbrook | int32_t src2 = T1; |
241 | 9ee6e8bb | pbrook | int32_t dest; |
242 | 9ee6e8bb | pbrook | |
243 | 9ee6e8bb | pbrook | dest = (src1 >> 1) + (src2 >> 1); |
244 | 9ee6e8bb | pbrook | if ((src1 | src2) & 1) |
245 | 9ee6e8bb | pbrook | dest++; |
246 | 9ee6e8bb | pbrook | T0 = dest; |
247 | 9ee6e8bb | pbrook | FORCE_RET(); |
248 | 9ee6e8bb | pbrook | } |
249 | 9ee6e8bb | pbrook | |
250 | 9ee6e8bb | pbrook | NEON_OP(rhadd_u32) |
251 | 9ee6e8bb | pbrook | { |
252 | 9ee6e8bb | pbrook | uint32_t src1 = T0; |
253 | 9ee6e8bb | pbrook | uint32_t src2 = T1; |
254 | 9ee6e8bb | pbrook | uint32_t dest; |
255 | 9ee6e8bb | pbrook | |
256 | 9ee6e8bb | pbrook | dest = (src1 >> 1) + (src2 >> 1); |
257 | 9ee6e8bb | pbrook | if ((src1 | src2) & 1) |
258 | 9ee6e8bb | pbrook | dest++; |
259 | 9ee6e8bb | pbrook | T0 = dest; |
260 | 9ee6e8bb | pbrook | FORCE_RET(); |
261 | 9ee6e8bb | pbrook | } |
262 | 9ee6e8bb | pbrook | |
263 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 - src2) >> 1 |
264 | 9ee6e8bb | pbrook | NEON_VOP(hsub_s8, neon_s8, 4)
|
265 | 9ee6e8bb | pbrook | NEON_VOP(hsub_u8, neon_u8, 4)
|
266 | 9ee6e8bb | pbrook | NEON_VOP(hsub_s16, neon_s16, 2)
|
267 | 9ee6e8bb | pbrook | NEON_VOP(hsub_u16, neon_u16, 2)
|
268 | 9ee6e8bb | pbrook | #undef NEON_FN
|
269 | 9ee6e8bb | pbrook | |
270 | 9ee6e8bb | pbrook | NEON_OP(hsub_s32) |
271 | 9ee6e8bb | pbrook | { |
272 | 9ee6e8bb | pbrook | int32_t src1 = T0; |
273 | 9ee6e8bb | pbrook | int32_t src2 = T1; |
274 | 9ee6e8bb | pbrook | int32_t dest; |
275 | 9ee6e8bb | pbrook | |
276 | 9ee6e8bb | pbrook | dest = (src1 >> 1) - (src2 >> 1); |
277 | 9ee6e8bb | pbrook | if ((~src1) & src2 & 1) |
278 | 9ee6e8bb | pbrook | dest--; |
279 | 9ee6e8bb | pbrook | T0 = dest; |
280 | 9ee6e8bb | pbrook | FORCE_RET(); |
281 | 9ee6e8bb | pbrook | } |
282 | 9ee6e8bb | pbrook | |
283 | 9ee6e8bb | pbrook | NEON_OP(hsub_u32) |
284 | 9ee6e8bb | pbrook | { |
285 | 9ee6e8bb | pbrook | uint32_t src1 = T0; |
286 | 9ee6e8bb | pbrook | uint32_t src2 = T1; |
287 | 9ee6e8bb | pbrook | uint32_t dest; |
288 | 9ee6e8bb | pbrook | |
289 | 9ee6e8bb | pbrook | dest = (src1 >> 1) - (src2 >> 1); |
290 | 9ee6e8bb | pbrook | if ((~src1) & src2 & 1) |
291 | 9ee6e8bb | pbrook | dest--; |
292 | 9ee6e8bb | pbrook | T0 = dest; |
293 | 9ee6e8bb | pbrook | FORCE_RET(); |
294 | 9ee6e8bb | pbrook | } |
295 | 9ee6e8bb | pbrook | |
296 | 9ee6e8bb | pbrook | /* ??? bsl, bif and bit are all the same op, just with the oparands in a
|
297 | 9ee6e8bb | pbrook | differnet order. It's currently easier to have 3 differnt ops than
|
298 | 9ee6e8bb | pbrook | rearange the operands. */
|
299 | 9ee6e8bb | pbrook | |
300 | 9ee6e8bb | pbrook | /* Bitwise Select. */
|
301 | 9ee6e8bb | pbrook | NEON_OP(bsl) |
302 | 9ee6e8bb | pbrook | { |
303 | 9ee6e8bb | pbrook | T0 = (T0 & T2) | (T1 & ~T2); |
304 | 9ee6e8bb | pbrook | } |
305 | 9ee6e8bb | pbrook | |
306 | 9ee6e8bb | pbrook | /* Bitwise Insert If True. */
|
307 | 9ee6e8bb | pbrook | NEON_OP(bit) |
308 | 9ee6e8bb | pbrook | { |
309 | 9ee6e8bb | pbrook | T0 = (T0 & T1) | (T2 & ~T1); |
310 | 9ee6e8bb | pbrook | } |
311 | 9ee6e8bb | pbrook | |
312 | 9ee6e8bb | pbrook | /* Bitwise Insert If False. */
|
313 | 9ee6e8bb | pbrook | NEON_OP(bif) |
314 | 9ee6e8bb | pbrook | { |
315 | 9ee6e8bb | pbrook | T0 = (T2 & T1) | (T0 & ~T1); |
316 | 9ee6e8bb | pbrook | } |
317 | 9ee6e8bb | pbrook | |
318 | 9ee6e8bb | pbrook | #define NEON_USAT(dest, src1, src2, type) do { \ |
319 | 9ee6e8bb | pbrook | uint32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
320 | 9ee6e8bb | pbrook | if (tmp != (type)tmp) { \
|
321 | 9ee6e8bb | pbrook | env->QF = 1; \
|
322 | 9ee6e8bb | pbrook | dest = ~0; \
|
323 | 9ee6e8bb | pbrook | } else { \
|
324 | 9ee6e8bb | pbrook | dest = tmp; \ |
325 | 9ee6e8bb | pbrook | }} while(0) |
326 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
327 | 9ee6e8bb | pbrook | NEON_VOP(qadd_u8, neon_u8, 4)
|
328 | 9ee6e8bb | pbrook | #undef NEON_FN
|
329 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
330 | 9ee6e8bb | pbrook | NEON_VOP(qadd_u16, neon_u16, 2)
|
331 | 9ee6e8bb | pbrook | #undef NEON_FN
|
332 | 9ee6e8bb | pbrook | #undef NEON_USAT
|
333 | 9ee6e8bb | pbrook | |
334 | 9ee6e8bb | pbrook | #define NEON_SSAT(dest, src1, src2, type) do { \ |
335 | 9ee6e8bb | pbrook | int32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
336 | 9ee6e8bb | pbrook | if (tmp != (type)tmp) { \
|
337 | 9ee6e8bb | pbrook | env->QF = 1; \
|
338 | 9ee6e8bb | pbrook | if (src2 > 0) { \ |
339 | 9ee6e8bb | pbrook | tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
340 | 9ee6e8bb | pbrook | } else { \
|
341 | 9ee6e8bb | pbrook | tmp = 1 << (sizeof(type) * 8 - 1); \ |
342 | 9ee6e8bb | pbrook | } \ |
343 | 9ee6e8bb | pbrook | } \ |
344 | 9ee6e8bb | pbrook | dest = tmp; \ |
345 | 9ee6e8bb | pbrook | } while(0) |
346 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
347 | 9ee6e8bb | pbrook | NEON_VOP(qadd_s8, neon_s8, 4)
|
348 | 9ee6e8bb | pbrook | #undef NEON_FN
|
349 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
350 | 9ee6e8bb | pbrook | NEON_VOP(qadd_s16, neon_s16, 2)
|
351 | 9ee6e8bb | pbrook | #undef NEON_FN
|
352 | 9ee6e8bb | pbrook | #undef NEON_SSAT
|
353 | 9ee6e8bb | pbrook | |
354 | 9ee6e8bb | pbrook | #define NEON_USAT(dest, src1, src2, type) do { \ |
355 | 9ee6e8bb | pbrook | uint32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
356 | 9ee6e8bb | pbrook | if (tmp != (type)tmp) { \
|
357 | 9ee6e8bb | pbrook | env->QF = 1; \
|
358 | 9ee6e8bb | pbrook | dest = 0; \
|
359 | 9ee6e8bb | pbrook | } else { \
|
360 | 9ee6e8bb | pbrook | dest = tmp; \ |
361 | 9ee6e8bb | pbrook | }} while(0) |
362 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
363 | 9ee6e8bb | pbrook | NEON_VOP(qsub_u8, neon_u8, 4)
|
364 | 9ee6e8bb | pbrook | #undef NEON_FN
|
365 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
366 | 9ee6e8bb | pbrook | NEON_VOP(qsub_u16, neon_u16, 2)
|
367 | 9ee6e8bb | pbrook | #undef NEON_FN
|
368 | 9ee6e8bb | pbrook | #undef NEON_USAT
|
369 | 9ee6e8bb | pbrook | |
370 | 9ee6e8bb | pbrook | #define NEON_SSAT(dest, src1, src2, type) do { \ |
371 | 9ee6e8bb | pbrook | int32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
372 | 9ee6e8bb | pbrook | if (tmp != (type)tmp) { \
|
373 | 9ee6e8bb | pbrook | env->QF = 1; \
|
374 | 9ee6e8bb | pbrook | if (src2 < 0) { \ |
375 | 9ee6e8bb | pbrook | tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
376 | 9ee6e8bb | pbrook | } else { \
|
377 | 9ee6e8bb | pbrook | tmp = 1 << (sizeof(type) * 8 - 1); \ |
378 | 9ee6e8bb | pbrook | } \ |
379 | 9ee6e8bb | pbrook | } \ |
380 | 9ee6e8bb | pbrook | dest = tmp; \ |
381 | 9ee6e8bb | pbrook | } while(0) |
382 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
383 | 9ee6e8bb | pbrook | NEON_VOP(qsub_s8, neon_s8, 4)
|
384 | 9ee6e8bb | pbrook | #undef NEON_FN
|
385 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
386 | 9ee6e8bb | pbrook | NEON_VOP(qsub_s16, neon_s16, 2)
|
387 | 9ee6e8bb | pbrook | #undef NEON_FN
|
388 | 9ee6e8bb | pbrook | #undef NEON_SSAT
|
389 | 9ee6e8bb | pbrook | |
390 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? ~0 : 0 |
391 | 9ee6e8bb | pbrook | NEON_VOP(cgt_s8, neon_s8, 4)
|
392 | 9ee6e8bb | pbrook | NEON_VOP(cgt_u8, neon_u8, 4)
|
393 | 9ee6e8bb | pbrook | NEON_VOP(cgt_s16, neon_s16, 2)
|
394 | 9ee6e8bb | pbrook | NEON_VOP(cgt_u16, neon_u16, 2)
|
395 | 9ee6e8bb | pbrook | NEON_VOP(cgt_s32, neon_s32, 1)
|
396 | 9ee6e8bb | pbrook | NEON_VOP(cgt_u32, neon_u32, 1)
|
397 | 9ee6e8bb | pbrook | #undef NEON_FN
|
398 | 9ee6e8bb | pbrook | |
399 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 >= src2) ? ~0 : 0 |
400 | 9ee6e8bb | pbrook | NEON_VOP(cge_s8, neon_s8, 4)
|
401 | 9ee6e8bb | pbrook | NEON_VOP(cge_u8, neon_u8, 4)
|
402 | 9ee6e8bb | pbrook | NEON_VOP(cge_s16, neon_s16, 2)
|
403 | 9ee6e8bb | pbrook | NEON_VOP(cge_u16, neon_u16, 2)
|
404 | 9ee6e8bb | pbrook | NEON_VOP(cge_s32, neon_s32, 1)
|
405 | 9ee6e8bb | pbrook | NEON_VOP(cge_u32, neon_u32, 1)
|
406 | 9ee6e8bb | pbrook | #undef NEON_FN
|
407 | 9ee6e8bb | pbrook | |
408 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
409 | 9ee6e8bb | pbrook | int8_t tmp; \ |
410 | 9ee6e8bb | pbrook | tmp = (int8_t)src2; \ |
411 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
412 | 9ee6e8bb | pbrook | dest = src1 >> -tmp; \ |
413 | 9ee6e8bb | pbrook | } else { \
|
414 | 9ee6e8bb | pbrook | dest = src1 << tmp; \ |
415 | 9ee6e8bb | pbrook | }} while (0) |
416 | 9ee6e8bb | pbrook | NEON_VOP(shl_s8, neon_s8, 4)
|
417 | 9ee6e8bb | pbrook | NEON_VOP(shl_u8, neon_u8, 4)
|
418 | 9ee6e8bb | pbrook | NEON_VOP(shl_s16, neon_s16, 2)
|
419 | 9ee6e8bb | pbrook | NEON_VOP(shl_u16, neon_u16, 2)
|
420 | 9ee6e8bb | pbrook | NEON_VOP(shl_s32, neon_s32, 1)
|
421 | 9ee6e8bb | pbrook | NEON_VOP(shl_u32, neon_u32, 1)
|
422 | 9ee6e8bb | pbrook | #undef NEON_FN
|
423 | 9ee6e8bb | pbrook | |
424 | 9ee6e8bb | pbrook | NEON_OP(shl_u64) |
425 | 9ee6e8bb | pbrook | { |
426 | 9ee6e8bb | pbrook | int8_t shift = T2; |
427 | 9ee6e8bb | pbrook | uint64_t val = T0 | ((uint64_t)T1 << 32);
|
428 | 9ee6e8bb | pbrook | if (shift < 0) { |
429 | 9ee6e8bb | pbrook | val >>= -shift; |
430 | 9ee6e8bb | pbrook | } else {
|
431 | 9ee6e8bb | pbrook | val <<= shift; |
432 | 9ee6e8bb | pbrook | } |
433 | 9ee6e8bb | pbrook | T0 = val; |
434 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
435 | 9ee6e8bb | pbrook | FORCE_RET(); |
436 | 9ee6e8bb | pbrook | } |
437 | 9ee6e8bb | pbrook | |
438 | 9ee6e8bb | pbrook | NEON_OP(shl_s64) |
439 | 9ee6e8bb | pbrook | { |
440 | 9ee6e8bb | pbrook | int8_t shift = T2; |
441 | 9ee6e8bb | pbrook | int64_t val = T0 | ((uint64_t)T1 << 32);
|
442 | 9ee6e8bb | pbrook | if (shift < 0) { |
443 | 9ee6e8bb | pbrook | val >>= -shift; |
444 | 9ee6e8bb | pbrook | } else {
|
445 | 9ee6e8bb | pbrook | val <<= shift; |
446 | 9ee6e8bb | pbrook | } |
447 | 9ee6e8bb | pbrook | T0 = val; |
448 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
449 | 9ee6e8bb | pbrook | FORCE_RET(); |
450 | 9ee6e8bb | pbrook | } |
451 | 9ee6e8bb | pbrook | |
452 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
453 | 9ee6e8bb | pbrook | int8_t tmp; \ |
454 | 9ee6e8bb | pbrook | tmp = (int8_t)src1; \ |
455 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
456 | 9ee6e8bb | pbrook | dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
457 | 9ee6e8bb | pbrook | } else { \
|
458 | 9ee6e8bb | pbrook | dest = src2 << tmp; \ |
459 | 9ee6e8bb | pbrook | }} while (0) |
460 | 9ee6e8bb | pbrook | |
461 | 9ee6e8bb | pbrook | NEON_VOP(rshl_s8, neon_s8, 4)
|
462 | 9ee6e8bb | pbrook | NEON_VOP(rshl_u8, neon_u8, 4)
|
463 | 9ee6e8bb | pbrook | NEON_VOP(rshl_s16, neon_s16, 2)
|
464 | 9ee6e8bb | pbrook | NEON_VOP(rshl_u16, neon_u16, 2)
|
465 | 9ee6e8bb | pbrook | NEON_VOP(rshl_s32, neon_s32, 1)
|
466 | 9ee6e8bb | pbrook | NEON_VOP(rshl_u32, neon_u32, 1)
|
467 | 9ee6e8bb | pbrook | #undef NEON_FN
|
468 | 9ee6e8bb | pbrook | |
469 | 9ee6e8bb | pbrook | NEON_OP(rshl_u64) |
470 | 9ee6e8bb | pbrook | { |
471 | 9ee6e8bb | pbrook | int8_t shift = T2; |
472 | 9ee6e8bb | pbrook | uint64_t val = T0 | ((uint64_t)T1 << 32);
|
473 | 9ee6e8bb | pbrook | if (shift < 0) { |
474 | 9ee6e8bb | pbrook | val = (val + ((uint64_t)1 << (-1 - shift))) >> -shift; |
475 | 9ee6e8bb | pbrook | val >>= -shift; |
476 | 9ee6e8bb | pbrook | } else {
|
477 | 9ee6e8bb | pbrook | val <<= shift; |
478 | 9ee6e8bb | pbrook | } |
479 | 9ee6e8bb | pbrook | T0 = val; |
480 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
481 | 9ee6e8bb | pbrook | FORCE_RET(); |
482 | 9ee6e8bb | pbrook | } |
483 | 9ee6e8bb | pbrook | |
484 | 9ee6e8bb | pbrook | NEON_OP(rshl_s64) |
485 | 9ee6e8bb | pbrook | { |
486 | 9ee6e8bb | pbrook | int8_t shift = T2; |
487 | 9ee6e8bb | pbrook | int64_t val = T0 | ((uint64_t)T1 << 32);
|
488 | 9ee6e8bb | pbrook | if (shift < 0) { |
489 | 9ee6e8bb | pbrook | val = (val + ((int64_t)1 << (-1 - shift))) >> -shift; |
490 | 9ee6e8bb | pbrook | } else {
|
491 | 9ee6e8bb | pbrook | val <<= shift; |
492 | 9ee6e8bb | pbrook | } |
493 | 9ee6e8bb | pbrook | T0 = val; |
494 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
495 | 9ee6e8bb | pbrook | FORCE_RET(); |
496 | 9ee6e8bb | pbrook | } |
497 | 9ee6e8bb | pbrook | |
498 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
499 | 9ee6e8bb | pbrook | int8_t tmp; \ |
500 | 9ee6e8bb | pbrook | tmp = (int8_t)src1; \ |
501 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
502 | 9ee6e8bb | pbrook | dest = src2 >> -tmp; \ |
503 | 9ee6e8bb | pbrook | } else { \
|
504 | 9ee6e8bb | pbrook | dest = src2 << tmp; \ |
505 | 9ee6e8bb | pbrook | if ((dest >> tmp) != src2) { \
|
506 | 9ee6e8bb | pbrook | env->QF = 1; \
|
507 | 9ee6e8bb | pbrook | dest = ~0; \
|
508 | 9ee6e8bb | pbrook | } \ |
509 | 9ee6e8bb | pbrook | }} while (0) |
510 | 9ee6e8bb | pbrook | NEON_VOP(qshl_s8, neon_s8, 4)
|
511 | 9ee6e8bb | pbrook | NEON_VOP(qshl_s16, neon_s16, 2)
|
512 | 9ee6e8bb | pbrook | NEON_VOP(qshl_s32, neon_s32, 1)
|
513 | 9ee6e8bb | pbrook | #undef NEON_FN
|
514 | 9ee6e8bb | pbrook | |
515 | 9ee6e8bb | pbrook | NEON_OP(qshl_s64) |
516 | 9ee6e8bb | pbrook | { |
517 | 9ee6e8bb | pbrook | int8_t shift = T2; |
518 | 9ee6e8bb | pbrook | int64_t val = T0 | ((uint64_t)T1 << 32);
|
519 | 9ee6e8bb | pbrook | if (shift < 0) { |
520 | 9ee6e8bb | pbrook | val >>= -shift; |
521 | 9ee6e8bb | pbrook | } else {
|
522 | 9ee6e8bb | pbrook | int64_t tmp = val; |
523 | 9ee6e8bb | pbrook | val <<= shift; |
524 | 9ee6e8bb | pbrook | if ((val >> shift) != tmp) {
|
525 | 9ee6e8bb | pbrook | env->QF = 1;
|
526 | 9ee6e8bb | pbrook | val = (tmp >> 63) ^ 0x7fffffffffffffffULL; |
527 | 9ee6e8bb | pbrook | } |
528 | 9ee6e8bb | pbrook | } |
529 | 9ee6e8bb | pbrook | T0 = val; |
530 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
531 | 9ee6e8bb | pbrook | FORCE_RET(); |
532 | 9ee6e8bb | pbrook | } |
533 | 9ee6e8bb | pbrook | |
534 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
535 | 9ee6e8bb | pbrook | int8_t tmp; \ |
536 | 9ee6e8bb | pbrook | tmp = (int8_t)src1; \ |
537 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
538 | 9ee6e8bb | pbrook | dest = src2 >> -tmp; \ |
539 | 9ee6e8bb | pbrook | } else { \
|
540 | 9ee6e8bb | pbrook | dest = src2 << tmp; \ |
541 | 9ee6e8bb | pbrook | if ((dest >> tmp) != src2) { \
|
542 | 9ee6e8bb | pbrook | env->QF = 1; \
|
543 | 9ee6e8bb | pbrook | dest = src2 >> 31; \
|
544 | 9ee6e8bb | pbrook | } \ |
545 | 9ee6e8bb | pbrook | }} while (0) |
546 | 9ee6e8bb | pbrook | NEON_VOP(qshl_u8, neon_u8, 4)
|
547 | 9ee6e8bb | pbrook | NEON_VOP(qshl_u16, neon_u16, 2)
|
548 | 9ee6e8bb | pbrook | NEON_VOP(qshl_u32, neon_u32, 1)
|
549 | 9ee6e8bb | pbrook | #undef NEON_FN
|
550 | 9ee6e8bb | pbrook | |
551 | 9ee6e8bb | pbrook | NEON_OP(qshl_u64) |
552 | 9ee6e8bb | pbrook | { |
553 | 9ee6e8bb | pbrook | int8_t shift = T2; |
554 | 9ee6e8bb | pbrook | uint64_t val = T0 | ((uint64_t)T1 << 32);
|
555 | 9ee6e8bb | pbrook | if (shift < 0) { |
556 | 9ee6e8bb | pbrook | val >>= -shift; |
557 | 9ee6e8bb | pbrook | } else {
|
558 | 9ee6e8bb | pbrook | uint64_t tmp = val; |
559 | 9ee6e8bb | pbrook | val <<= shift; |
560 | 9ee6e8bb | pbrook | if ((val >> shift) != tmp) {
|
561 | 9ee6e8bb | pbrook | env->QF = 1;
|
562 | 9ee6e8bb | pbrook | val = ~(uint64_t)0;
|
563 | 9ee6e8bb | pbrook | } |
564 | 9ee6e8bb | pbrook | } |
565 | 9ee6e8bb | pbrook | T0 = val; |
566 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
567 | 9ee6e8bb | pbrook | FORCE_RET(); |
568 | 9ee6e8bb | pbrook | } |
569 | 9ee6e8bb | pbrook | |
570 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
571 | 9ee6e8bb | pbrook | int8_t tmp; \ |
572 | 9ee6e8bb | pbrook | tmp = (int8_t)src1; \ |
573 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
574 | 9ee6e8bb | pbrook | dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
575 | 9ee6e8bb | pbrook | } else { \
|
576 | 9ee6e8bb | pbrook | dest = src2 << tmp; \ |
577 | 9ee6e8bb | pbrook | if ((dest >> tmp) != src2) { \
|
578 | 9ee6e8bb | pbrook | dest = ~0; \
|
579 | 9ee6e8bb | pbrook | } \ |
580 | 9ee6e8bb | pbrook | }} while (0) |
581 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_s8, neon_s8, 4)
|
582 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_s16, neon_s16, 2)
|
583 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_s32, neon_s32, 1)
|
584 | 9ee6e8bb | pbrook | #undef NEON_FN
|
585 | 9ee6e8bb | pbrook | |
586 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) do { \ |
587 | 9ee6e8bb | pbrook | int8_t tmp; \ |
588 | 9ee6e8bb | pbrook | tmp = (int8_t)src1; \ |
589 | 9ee6e8bb | pbrook | if (tmp < 0) { \ |
590 | 9ee6e8bb | pbrook | dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
591 | 9ee6e8bb | pbrook | } else { \
|
592 | 9ee6e8bb | pbrook | dest = src2 << tmp; \ |
593 | 9ee6e8bb | pbrook | if ((dest >> tmp) != src2) { \
|
594 | 9ee6e8bb | pbrook | env->QF = 1; \
|
595 | 9ee6e8bb | pbrook | dest = src2 >> 31; \
|
596 | 9ee6e8bb | pbrook | } \ |
597 | 9ee6e8bb | pbrook | }} while (0) |
598 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_u8, neon_u8, 4)
|
599 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_u16, neon_u16, 2)
|
600 | 9ee6e8bb | pbrook | NEON_VOP(qrshl_u32, neon_u32, 1)
|
601 | 9ee6e8bb | pbrook | #undef NEON_FN
|
602 | 9ee6e8bb | pbrook | |
603 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? src1 : src2
|
604 | 9ee6e8bb | pbrook | NEON_VOP(max_s8, neon_s8, 4)
|
605 | 9ee6e8bb | pbrook | NEON_VOP(max_u8, neon_u8, 4)
|
606 | 9ee6e8bb | pbrook | NEON_VOP(max_s16, neon_s16, 2)
|
607 | 9ee6e8bb | pbrook | NEON_VOP(max_u16, neon_u16, 2)
|
608 | 9ee6e8bb | pbrook | NEON_VOP(max_s32, neon_s32, 1)
|
609 | 9ee6e8bb | pbrook | NEON_VOP(max_u32, neon_u32, 1)
|
610 | 9ee6e8bb | pbrook | NEON_POP(pmax_s8, neon_s8, 4)
|
611 | 9ee6e8bb | pbrook | NEON_POP(pmax_u8, neon_u8, 4)
|
612 | 9ee6e8bb | pbrook | NEON_POP(pmax_s16, neon_s16, 2)
|
613 | 9ee6e8bb | pbrook | NEON_POP(pmax_u16, neon_u16, 2)
|
614 | 9ee6e8bb | pbrook | #undef NEON_FN
|
615 | 9ee6e8bb | pbrook | |
616 | 9ee6e8bb | pbrook | NEON_OP(max_f32) |
617 | 9ee6e8bb | pbrook | { |
618 | 9ee6e8bb | pbrook | float32 f0 = vfp_itos(T0); |
619 | 9ee6e8bb | pbrook | float32 f1 = vfp_itos(T1); |
620 | 9ee6e8bb | pbrook | T0 = (float32_compare_quiet(f0, f1, NFS) == 1) ? T0 : T1;
|
621 | 9ee6e8bb | pbrook | FORCE_RET(); |
622 | 9ee6e8bb | pbrook | } |
623 | 9ee6e8bb | pbrook | |
624 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 < src2) ? src1 : src2
|
625 | 9ee6e8bb | pbrook | NEON_VOP(min_s8, neon_s8, 4)
|
626 | 9ee6e8bb | pbrook | NEON_VOP(min_u8, neon_u8, 4)
|
627 | 9ee6e8bb | pbrook | NEON_VOP(min_s16, neon_s16, 2)
|
628 | 9ee6e8bb | pbrook | NEON_VOP(min_u16, neon_u16, 2)
|
629 | 9ee6e8bb | pbrook | NEON_VOP(min_s32, neon_s32, 1)
|
630 | 9ee6e8bb | pbrook | NEON_VOP(min_u32, neon_u32, 1)
|
631 | 9ee6e8bb | pbrook | NEON_POP(pmin_s8, neon_s8, 4)
|
632 | 9ee6e8bb | pbrook | NEON_POP(pmin_u8, neon_u8, 4)
|
633 | 9ee6e8bb | pbrook | NEON_POP(pmin_s16, neon_s16, 2)
|
634 | 9ee6e8bb | pbrook | NEON_POP(pmin_u16, neon_u16, 2)
|
635 | 9ee6e8bb | pbrook | #undef NEON_FN
|
636 | 9ee6e8bb | pbrook | |
637 | 9ee6e8bb | pbrook | NEON_OP(min_f32) |
638 | 9ee6e8bb | pbrook | { |
639 | 9ee6e8bb | pbrook | float32 f0 = vfp_itos(T0); |
640 | 9ee6e8bb | pbrook | float32 f1 = vfp_itos(T1); |
641 | 9ee6e8bb | pbrook | T0 = (float32_compare_quiet(f0, f1, NFS) == -1) ? T0 : T1;
|
642 | 9ee6e8bb | pbrook | FORCE_RET(); |
643 | 9ee6e8bb | pbrook | } |
644 | 9ee6e8bb | pbrook | |
645 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) \
|
646 | 9ee6e8bb | pbrook | dest = (src1 > src2) ? (src1 - src2) : (src2 - src1) |
647 | 9ee6e8bb | pbrook | NEON_VOP(abd_s8, neon_s8, 4)
|
648 | 9ee6e8bb | pbrook | NEON_VOP(abd_u8, neon_u8, 4)
|
649 | 9ee6e8bb | pbrook | NEON_VOP(abd_s16, neon_s16, 2)
|
650 | 9ee6e8bb | pbrook | NEON_VOP(abd_u16, neon_u16, 2)
|
651 | 9ee6e8bb | pbrook | NEON_VOP(abd_s32, neon_s32, 1)
|
652 | 9ee6e8bb | pbrook | NEON_VOP(abd_u32, neon_u32, 1)
|
653 | 9ee6e8bb | pbrook | #undef NEON_FN
|
654 | 9ee6e8bb | pbrook | |
655 | 9ee6e8bb | pbrook | NEON_OP(abd_f32) |
656 | 9ee6e8bb | pbrook | { |
657 | 9ee6e8bb | pbrook | float32 f0 = vfp_itos(T0); |
658 | 9ee6e8bb | pbrook | float32 f1 = vfp_itos(T1); |
659 | 9ee6e8bb | pbrook | T0 = vfp_stoi((float32_compare_quiet(f0, f1, NFS) == 1)
|
660 | 9ee6e8bb | pbrook | ? float32_sub(f0, f1, NFS) |
661 | 9ee6e8bb | pbrook | : float32_sub(f1, f0, NFS)); |
662 | 9ee6e8bb | pbrook | FORCE_RET(); |
663 | 9ee6e8bb | pbrook | } |
664 | 9ee6e8bb | pbrook | |
665 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = src1 + src2
|
666 | 9ee6e8bb | pbrook | NEON_VOP(add_u8, neon_u8, 4)
|
667 | 9ee6e8bb | pbrook | NEON_VOP(add_u16, neon_u16, 2)
|
668 | 9ee6e8bb | pbrook | NEON_POP(padd_u8, neon_u8, 4)
|
669 | 9ee6e8bb | pbrook | NEON_POP(padd_u16, neon_u16, 2)
|
670 | 9ee6e8bb | pbrook | #undef NEON_FN
|
671 | 9ee6e8bb | pbrook | |
672 | 9ee6e8bb | pbrook | NEON_OP(add_f32) |
673 | 9ee6e8bb | pbrook | { |
674 | 9ee6e8bb | pbrook | T0 = vfp_stoi(float32_add(vfp_itos(T0), vfp_itos(T1), NFS)); |
675 | 9ee6e8bb | pbrook | FORCE_RET(); |
676 | 9ee6e8bb | pbrook | } |
677 | 9ee6e8bb | pbrook | |
678 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = src1 - src2
|
679 | 9ee6e8bb | pbrook | NEON_VOP(sub_u8, neon_u8, 4)
|
680 | 9ee6e8bb | pbrook | NEON_VOP(sub_u16, neon_u16, 2)
|
681 | 9ee6e8bb | pbrook | #undef NEON_FN
|
682 | 9ee6e8bb | pbrook | |
683 | 9ee6e8bb | pbrook | NEON_OP(sub_f32) |
684 | 9ee6e8bb | pbrook | { |
685 | 9ee6e8bb | pbrook | T0 = vfp_stoi(float32_sub(vfp_itos(T0), vfp_itos(T1), NFS)); |
686 | 9ee6e8bb | pbrook | FORCE_RET(); |
687 | 9ee6e8bb | pbrook | } |
688 | 9ee6e8bb | pbrook | |
689 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = src2 - src1
|
690 | 9ee6e8bb | pbrook | NEON_VOP(rsb_u8, neon_u8, 4)
|
691 | 9ee6e8bb | pbrook | NEON_VOP(rsb_u16, neon_u16, 2)
|
692 | 9ee6e8bb | pbrook | #undef NEON_FN
|
693 | 9ee6e8bb | pbrook | |
694 | 9ee6e8bb | pbrook | NEON_OP(rsb_f32) |
695 | 9ee6e8bb | pbrook | { |
696 | 9ee6e8bb | pbrook | T0 = vfp_stoi(float32_sub(vfp_itos(T1), vfp_itos(T0), NFS)); |
697 | 9ee6e8bb | pbrook | FORCE_RET(); |
698 | 9ee6e8bb | pbrook | } |
699 | 9ee6e8bb | pbrook | |
700 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = src1 * src2
|
701 | 9ee6e8bb | pbrook | NEON_VOP(mul_u8, neon_u8, 4)
|
702 | 9ee6e8bb | pbrook | NEON_VOP(mul_u16, neon_u16, 2)
|
703 | 9ee6e8bb | pbrook | #undef NEON_FN
|
704 | 9ee6e8bb | pbrook | |
705 | 9ee6e8bb | pbrook | NEON_OP(mul_f32) |
706 | 9ee6e8bb | pbrook | { |
707 | 9ee6e8bb | pbrook | T0 = vfp_stoi(float32_mul(vfp_itos(T0), vfp_itos(T1), NFS)); |
708 | 9ee6e8bb | pbrook | FORCE_RET(); |
709 | 9ee6e8bb | pbrook | } |
710 | 9ee6e8bb | pbrook | |
711 | 9ee6e8bb | pbrook | NEON_OP(mul_p8) |
712 | 9ee6e8bb | pbrook | { |
713 | 9ee6e8bb | pbrook | T0 = helper_neon_mul_p8(T0, T1); |
714 | 9ee6e8bb | pbrook | } |
715 | 9ee6e8bb | pbrook | |
716 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 & src2) ? -1 : 0 |
717 | 9ee6e8bb | pbrook | NEON_VOP(tst_u8, neon_u8, 4)
|
718 | 9ee6e8bb | pbrook | NEON_VOP(tst_u16, neon_u16, 2)
|
719 | 9ee6e8bb | pbrook | NEON_VOP(tst_u32, neon_u32, 1)
|
720 | 9ee6e8bb | pbrook | #undef NEON_FN
|
721 | 9ee6e8bb | pbrook | |
722 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) dest = (src1 == src2) ? -1 : 0 |
723 | 9ee6e8bb | pbrook | NEON_VOP(ceq_u8, neon_u8, 4)
|
724 | 9ee6e8bb | pbrook | NEON_VOP(ceq_u16, neon_u16, 2)
|
725 | 9ee6e8bb | pbrook | NEON_VOP(ceq_u32, neon_u32, 1)
|
726 | 9ee6e8bb | pbrook | #undef NEON_FN
|
727 | 9ee6e8bb | pbrook | |
728 | 9ee6e8bb | pbrook | #define NEON_QDMULH16(dest, src1, src2, round) do { \ |
729 | 9ee6e8bb | pbrook | uint32_t tmp = (int32_t)(int16_t) src1 * (int16_t) src2; \ |
730 | 9ee6e8bb | pbrook | if ((tmp ^ (tmp << 1)) & SIGNBIT) { \ |
731 | 9ee6e8bb | pbrook | env->QF = 1; \
|
732 | 9ee6e8bb | pbrook | tmp = (tmp >> 31) ^ ~SIGNBIT; \
|
733 | 9ee6e8bb | pbrook | } \ |
734 | 9ee6e8bb | pbrook | tmp <<= 1; \
|
735 | 9ee6e8bb | pbrook | if (round) { \
|
736 | 9ee6e8bb | pbrook | int32_t old = tmp; \ |
737 | 9ee6e8bb | pbrook | tmp += 1 << 15; \ |
738 | 9ee6e8bb | pbrook | if ((int32_t)tmp < old) { \
|
739 | 9ee6e8bb | pbrook | env->QF = 1; \
|
740 | 9ee6e8bb | pbrook | tmp = SIGNBIT - 1; \
|
741 | 9ee6e8bb | pbrook | } \ |
742 | 9ee6e8bb | pbrook | } \ |
743 | 9ee6e8bb | pbrook | dest = tmp >> 16; \
|
744 | 9ee6e8bb | pbrook | } while(0) |
745 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 0) |
746 | 9ee6e8bb | pbrook | NEON_VOP(qdmulh_s16, neon_s16, 2)
|
747 | 9ee6e8bb | pbrook | #undef NEON_FN
|
748 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 1) |
749 | 9ee6e8bb | pbrook | NEON_VOP(qrdmulh_s16, neon_s16, 2)
|
750 | 9ee6e8bb | pbrook | #undef NEON_FN
|
751 | 9ee6e8bb | pbrook | #undef NEON_QDMULH16
|
752 | 9ee6e8bb | pbrook | |
753 | 9ee6e8bb | pbrook | #define SIGNBIT64 ((uint64_t)1 << 63) |
754 | 9ee6e8bb | pbrook | #define NEON_QDMULH32(dest, src1, src2, round) do { \ |
755 | 9ee6e8bb | pbrook | uint64_t tmp = (int64_t)(int32_t) src1 * (int32_t) src2; \ |
756 | 9ee6e8bb | pbrook | if ((tmp ^ (tmp << 1)) & SIGNBIT64) { \ |
757 | 9ee6e8bb | pbrook | env->QF = 1; \
|
758 | 9ee6e8bb | pbrook | tmp = (tmp >> 63) ^ ~SIGNBIT64; \
|
759 | 9ee6e8bb | pbrook | } else { \
|
760 | 9ee6e8bb | pbrook | tmp <<= 1; \
|
761 | 9ee6e8bb | pbrook | } \ |
762 | 9ee6e8bb | pbrook | if (round) { \
|
763 | 9ee6e8bb | pbrook | int64_t old = tmp; \ |
764 | 9ee6e8bb | pbrook | tmp += (int64_t)1 << 31; \ |
765 | 9ee6e8bb | pbrook | if ((int64_t)tmp < old) { \
|
766 | 9ee6e8bb | pbrook | env->QF = 1; \
|
767 | 9ee6e8bb | pbrook | tmp = SIGNBIT64 - 1; \
|
768 | 9ee6e8bb | pbrook | } \ |
769 | 9ee6e8bb | pbrook | } \ |
770 | 9ee6e8bb | pbrook | dest = tmp >> 32; \
|
771 | 9ee6e8bb | pbrook | } while(0) |
772 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 0) |
773 | 9ee6e8bb | pbrook | NEON_VOP(qdmulh_s32, neon_s32, 1)
|
774 | 9ee6e8bb | pbrook | #undef NEON_FN
|
775 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 1) |
776 | 9ee6e8bb | pbrook | NEON_VOP(qrdmulh_s32, neon_s32, 1)
|
777 | 9ee6e8bb | pbrook | #undef NEON_FN
|
778 | 9ee6e8bb | pbrook | #undef NEON_QDMULH32
|
779 | 9ee6e8bb | pbrook | |
780 | 9ee6e8bb | pbrook | /* Floating point comparisons produce an integer result. */
|
781 | 9ee6e8bb | pbrook | #define NEON_VOP_FCMP(name, cmp) \
|
782 | 9ee6e8bb | pbrook | NEON_OP(name) \ |
783 | 9ee6e8bb | pbrook | { \ |
784 | 9ee6e8bb | pbrook | if (float32_compare_quiet(vfp_itos(T0), vfp_itos(T1), NFS) cmp 0) \ |
785 | 9ee6e8bb | pbrook | T0 = -1; \
|
786 | 9ee6e8bb | pbrook | else \
|
787 | 9ee6e8bb | pbrook | T0 = 0; \
|
788 | 9ee6e8bb | pbrook | FORCE_RET(); \ |
789 | 9ee6e8bb | pbrook | } |
790 | 9ee6e8bb | pbrook | |
791 | 9ee6e8bb | pbrook | NEON_VOP_FCMP(ceq_f32, ==) |
792 | 9ee6e8bb | pbrook | NEON_VOP_FCMP(cge_f32, >=) |
793 | 9ee6e8bb | pbrook | NEON_VOP_FCMP(cgt_f32, >) |
794 | 9ee6e8bb | pbrook | |
795 | 9ee6e8bb | pbrook | NEON_OP(acge_f32) |
796 | 9ee6e8bb | pbrook | { |
797 | 9ee6e8bb | pbrook | float32 f0 = float32_abs(vfp_itos(T0)); |
798 | 9ee6e8bb | pbrook | float32 f1 = float32_abs(vfp_itos(T1)); |
799 | 9ee6e8bb | pbrook | T0 = (float32_compare_quiet(f0, f1,NFS) >= 0) ? -1 : 0; |
800 | 9ee6e8bb | pbrook | FORCE_RET(); |
801 | 9ee6e8bb | pbrook | } |
802 | 9ee6e8bb | pbrook | |
803 | 9ee6e8bb | pbrook | NEON_OP(acgt_f32) |
804 | 9ee6e8bb | pbrook | { |
805 | 9ee6e8bb | pbrook | float32 f0 = float32_abs(vfp_itos(T0)); |
806 | 9ee6e8bb | pbrook | float32 f1 = float32_abs(vfp_itos(T1)); |
807 | 9ee6e8bb | pbrook | T0 = (float32_compare_quiet(f0, f1, NFS) > 0) ? -1 : 0; |
808 | 9ee6e8bb | pbrook | FORCE_RET(); |
809 | 9ee6e8bb | pbrook | } |
810 | 9ee6e8bb | pbrook | |
811 | 9ee6e8bb | pbrook | /* Narrowing instructions. The named type is the destination type. */
|
812 | 9ee6e8bb | pbrook | NEON_OP(narrow_u8) |
813 | 9ee6e8bb | pbrook | { |
814 | 9ee6e8bb | pbrook | T0 = (T0 & 0xff) | ((T0 >> 8) & 0xff00) |
815 | 9ee6e8bb | pbrook | | ((T1 << 16) & 0xff0000) | (T1 << 24); |
816 | 9ee6e8bb | pbrook | FORCE_RET(); |
817 | 9ee6e8bb | pbrook | } |
818 | 9ee6e8bb | pbrook | |
819 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_u8) |
820 | 9ee6e8bb | pbrook | { |
821 | 9ee6e8bb | pbrook | neon_u16 src; |
822 | 9ee6e8bb | pbrook | neon_u8 dest; |
823 | 9ee6e8bb | pbrook | #define SAT8(d, s) \
|
824 | 9ee6e8bb | pbrook | if (s > 0xff) { \ |
825 | 9ee6e8bb | pbrook | d = 0xff; \
|
826 | 9ee6e8bb | pbrook | env->QF = 1; \
|
827 | 9ee6e8bb | pbrook | } else { \
|
828 | 9ee6e8bb | pbrook | d = s; \ |
829 | 9ee6e8bb | pbrook | } |
830 | 9ee6e8bb | pbrook | |
831 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_u16, src, T0); |
832 | 9ee6e8bb | pbrook | SAT8(dest.v1, src.v1); |
833 | 9ee6e8bb | pbrook | SAT8(dest.v2, src.v2); |
834 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_u16, src, T1); |
835 | 9ee6e8bb | pbrook | SAT8(dest.v3, src.v1); |
836 | 9ee6e8bb | pbrook | SAT8(dest.v4, src.v2); |
837 | 9ee6e8bb | pbrook | NEON_PACK(neon_u8, T0, dest); |
838 | 9ee6e8bb | pbrook | FORCE_RET(); |
839 | 9ee6e8bb | pbrook | #undef SAT8
|
840 | 9ee6e8bb | pbrook | } |
841 | 9ee6e8bb | pbrook | |
842 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_s8) |
843 | 9ee6e8bb | pbrook | { |
844 | 9ee6e8bb | pbrook | neon_s16 src; |
845 | 9ee6e8bb | pbrook | neon_s8 dest; |
846 | 9ee6e8bb | pbrook | #define SAT8(d, s) \
|
847 | 9ee6e8bb | pbrook | if (s != (uint8_t)s) { \
|
848 | 9ee6e8bb | pbrook | d = (s >> 15) ^ 0x7f; \ |
849 | 9ee6e8bb | pbrook | env->QF = 1; \
|
850 | 9ee6e8bb | pbrook | } else { \
|
851 | 9ee6e8bb | pbrook | d = s; \ |
852 | 9ee6e8bb | pbrook | } |
853 | 9ee6e8bb | pbrook | |
854 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s16, src, T0); |
855 | 9ee6e8bb | pbrook | SAT8(dest.v1, src.v1); |
856 | 9ee6e8bb | pbrook | SAT8(dest.v2, src.v2); |
857 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s16, src, T1); |
858 | 9ee6e8bb | pbrook | SAT8(dest.v3, src.v1); |
859 | 9ee6e8bb | pbrook | SAT8(dest.v4, src.v2); |
860 | 9ee6e8bb | pbrook | NEON_PACK(neon_s8, T0, dest); |
861 | 9ee6e8bb | pbrook | FORCE_RET(); |
862 | 9ee6e8bb | pbrook | #undef SAT8
|
863 | 9ee6e8bb | pbrook | } |
864 | 9ee6e8bb | pbrook | |
865 | 9ee6e8bb | pbrook | NEON_OP(narrow_u16) |
866 | 9ee6e8bb | pbrook | { |
867 | 9ee6e8bb | pbrook | T0 = (T0 & 0xffff) | (T1 << 16); |
868 | 9ee6e8bb | pbrook | } |
869 | 9ee6e8bb | pbrook | |
870 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_u16) |
871 | 9ee6e8bb | pbrook | { |
872 | 9ee6e8bb | pbrook | if (T0 > 0xffff) { |
873 | 9ee6e8bb | pbrook | T0 = 0xffff;
|
874 | 9ee6e8bb | pbrook | env->QF = 1;
|
875 | 9ee6e8bb | pbrook | } |
876 | 9ee6e8bb | pbrook | if (T1 > 0xffff) { |
877 | 9ee6e8bb | pbrook | T1 = 0xffff;
|
878 | 9ee6e8bb | pbrook | env->QF = 1;
|
879 | 9ee6e8bb | pbrook | } |
880 | 9ee6e8bb | pbrook | T0 |= T1 << 16;
|
881 | 9ee6e8bb | pbrook | FORCE_RET(); |
882 | 9ee6e8bb | pbrook | } |
883 | 9ee6e8bb | pbrook | |
884 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_s16) |
885 | 9ee6e8bb | pbrook | { |
886 | 9ee6e8bb | pbrook | if ((int32_t)T0 != (int16_t)T0) {
|
887 | 9ee6e8bb | pbrook | T0 = ((int32_t)T0 >> 31) ^ 0x7fff; |
888 | 9ee6e8bb | pbrook | env->QF = 1;
|
889 | 9ee6e8bb | pbrook | } |
890 | 9ee6e8bb | pbrook | if ((int32_t)T1 != (int16_t) T1) {
|
891 | 9ee6e8bb | pbrook | T1 = ((int32_t)T1 >> 31) ^ 0x7fff; |
892 | 9ee6e8bb | pbrook | env->QF = 1;
|
893 | 9ee6e8bb | pbrook | } |
894 | 9ee6e8bb | pbrook | T0 = (uint16_t)T0 | (T1 << 16);
|
895 | 9ee6e8bb | pbrook | FORCE_RET(); |
896 | 9ee6e8bb | pbrook | } |
897 | 9ee6e8bb | pbrook | |
898 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_u32) |
899 | 9ee6e8bb | pbrook | { |
900 | 9ee6e8bb | pbrook | if (T1) {
|
901 | 9ee6e8bb | pbrook | T0 = 0xffffffffu;
|
902 | 9ee6e8bb | pbrook | env->QF = 1;
|
903 | 9ee6e8bb | pbrook | } |
904 | 9ee6e8bb | pbrook | FORCE_RET(); |
905 | 9ee6e8bb | pbrook | } |
906 | 9ee6e8bb | pbrook | |
907 | 9ee6e8bb | pbrook | NEON_OP(narrow_sat_s32) |
908 | 9ee6e8bb | pbrook | { |
909 | 9ee6e8bb | pbrook | int32_t sign = (int32_t)T1 >> 31;
|
910 | 9ee6e8bb | pbrook | |
911 | 9ee6e8bb | pbrook | if ((int32_t)T1 != sign) {
|
912 | 9ee6e8bb | pbrook | T0 = sign ^ 0x7fffffff;
|
913 | 9ee6e8bb | pbrook | env->QF = 1;
|
914 | 9ee6e8bb | pbrook | } |
915 | 9ee6e8bb | pbrook | FORCE_RET(); |
916 | 9ee6e8bb | pbrook | } |
917 | 9ee6e8bb | pbrook | |
918 | 9ee6e8bb | pbrook | /* Narrowing instructions. Named type is the narrow type. */
|
919 | 9ee6e8bb | pbrook | NEON_OP(narrow_high_u8) |
920 | 9ee6e8bb | pbrook | { |
921 | 9ee6e8bb | pbrook | T0 = ((T0 >> 8) & 0xff) | ((T0 >> 16) & 0xff00) |
922 | 9ee6e8bb | pbrook | | ((T1 << 8) & 0xff0000) | (T1 & 0xff000000); |
923 | 9ee6e8bb | pbrook | FORCE_RET(); |
924 | 9ee6e8bb | pbrook | } |
925 | 9ee6e8bb | pbrook | |
926 | 9ee6e8bb | pbrook | NEON_OP(narrow_high_u16) |
927 | 9ee6e8bb | pbrook | { |
928 | 9ee6e8bb | pbrook | T0 = (T0 >> 16) | (T1 & 0xffff0000); |
929 | 9ee6e8bb | pbrook | FORCE_RET(); |
930 | 9ee6e8bb | pbrook | } |
931 | 9ee6e8bb | pbrook | |
932 | 9ee6e8bb | pbrook | NEON_OP(narrow_high_round_u8) |
933 | 9ee6e8bb | pbrook | { |
934 | 9ee6e8bb | pbrook | T0 = (((T0 + 0x80) >> 8) & 0xff) | (((T0 + 0x800000) >> 16) & 0xff00) |
935 | 9ee6e8bb | pbrook | | (((T1 + 0x80) << 8) & 0xff0000) | ((T1 + 0x800000) & 0xff000000); |
936 | 9ee6e8bb | pbrook | FORCE_RET(); |
937 | 9ee6e8bb | pbrook | } |
938 | 9ee6e8bb | pbrook | |
939 | 9ee6e8bb | pbrook | NEON_OP(narrow_high_round_u16) |
940 | 9ee6e8bb | pbrook | { |
941 | 9ee6e8bb | pbrook | T0 = ((T0 + 0x8000) >> 16) | ((T1 + 0x8000) & 0xffff0000); |
942 | 9ee6e8bb | pbrook | FORCE_RET(); |
943 | 9ee6e8bb | pbrook | } |
944 | 9ee6e8bb | pbrook | |
945 | 9ee6e8bb | pbrook | NEON_OP(narrow_high_round_u32) |
946 | 9ee6e8bb | pbrook | { |
947 | 9ee6e8bb | pbrook | if (T0 >= 0x80000000u) |
948 | 9ee6e8bb | pbrook | T0 = T1 + 1;
|
949 | 9ee6e8bb | pbrook | else
|
950 | 9ee6e8bb | pbrook | T0 = T1; |
951 | 9ee6e8bb | pbrook | FORCE_RET(); |
952 | 9ee6e8bb | pbrook | } |
953 | 9ee6e8bb | pbrook | |
954 | 9ee6e8bb | pbrook | /* Widening instructions. Named type is source type. */
|
955 | 9ee6e8bb | pbrook | NEON_OP(widen_s8) |
956 | 9ee6e8bb | pbrook | { |
957 | 9ee6e8bb | pbrook | uint32_t src; |
958 | 9ee6e8bb | pbrook | |
959 | 9ee6e8bb | pbrook | src = T0; |
960 | 9ee6e8bb | pbrook | T0 = (uint16_t)(int8_t)src | ((int8_t)(src >> 8) << 16); |
961 | 9ee6e8bb | pbrook | T1 = (uint16_t)(int8_t)(src >> 16) | ((int8_t)(src >> 24) << 16); |
962 | 9ee6e8bb | pbrook | } |
963 | 9ee6e8bb | pbrook | |
964 | 9ee6e8bb | pbrook | NEON_OP(widen_u8) |
965 | 9ee6e8bb | pbrook | { |
966 | 9ee6e8bb | pbrook | T1 = ((T0 >> 8) & 0xff0000) | ((T0 >> 16) & 0xff); |
967 | 9ee6e8bb | pbrook | T0 = ((T0 << 8) & 0xff0000) | (T0 & 0xff); |
968 | 9ee6e8bb | pbrook | } |
969 | 9ee6e8bb | pbrook | |
970 | 9ee6e8bb | pbrook | NEON_OP(widen_s16) |
971 | 9ee6e8bb | pbrook | { |
972 | 9ee6e8bb | pbrook | int32_t src; |
973 | 9ee6e8bb | pbrook | |
974 | 9ee6e8bb | pbrook | src = T0; |
975 | 9ee6e8bb | pbrook | T0 = (int16_t)src; |
976 | 9ee6e8bb | pbrook | T1 = src >> 16;
|
977 | 9ee6e8bb | pbrook | } |
978 | 9ee6e8bb | pbrook | |
979 | 9ee6e8bb | pbrook | NEON_OP(widen_u16) |
980 | 9ee6e8bb | pbrook | { |
981 | 9ee6e8bb | pbrook | T1 = T0 >> 16;
|
982 | 9ee6e8bb | pbrook | T0 &= 0xffff;
|
983 | 9ee6e8bb | pbrook | } |
984 | 9ee6e8bb | pbrook | |
985 | 9ee6e8bb | pbrook | NEON_OP(widen_s32) |
986 | 9ee6e8bb | pbrook | { |
987 | 9ee6e8bb | pbrook | T1 = (int32_t)T0 >> 31;
|
988 | 9ee6e8bb | pbrook | FORCE_RET(); |
989 | 9ee6e8bb | pbrook | } |
990 | 9ee6e8bb | pbrook | |
991 | 9ee6e8bb | pbrook | NEON_OP(widen_high_u8) |
992 | 9ee6e8bb | pbrook | { |
993 | 9ee6e8bb | pbrook | T1 = (T0 & 0xff000000) | ((T0 >> 8) & 0xff00); |
994 | 9ee6e8bb | pbrook | T0 = ((T0 << 16) & 0xff000000) | ((T0 << 8) & 0xff00); |
995 | 9ee6e8bb | pbrook | } |
996 | 9ee6e8bb | pbrook | |
997 | 9ee6e8bb | pbrook | NEON_OP(widen_high_u16) |
998 | 9ee6e8bb | pbrook | { |
999 | 9ee6e8bb | pbrook | T1 = T0 & 0xffff0000;
|
1000 | 9ee6e8bb | pbrook | T0 <<= 16;
|
1001 | 9ee6e8bb | pbrook | } |
1002 | 9ee6e8bb | pbrook | |
1003 | 9ee6e8bb | pbrook | /* Long operations. The type is the wide type. */
|
1004 | 9ee6e8bb | pbrook | NEON_OP(shll_u16) |
1005 | 9ee6e8bb | pbrook | { |
1006 | 9ee6e8bb | pbrook | int shift = PARAM1;
|
1007 | 9ee6e8bb | pbrook | uint32_t mask; |
1008 | 9ee6e8bb | pbrook | |
1009 | 9ee6e8bb | pbrook | mask = 0xffff >> (16 - shift); |
1010 | 9ee6e8bb | pbrook | mask |= mask << 16;
|
1011 | 9ee6e8bb | pbrook | mask = ~mask; |
1012 | 9ee6e8bb | pbrook | |
1013 | 9ee6e8bb | pbrook | T0 = (T0 << shift) & mask; |
1014 | 9ee6e8bb | pbrook | T1 = (T1 << shift) & mask; |
1015 | 9ee6e8bb | pbrook | FORCE_RET(); |
1016 | 9ee6e8bb | pbrook | } |
1017 | 9ee6e8bb | pbrook | |
1018 | 9ee6e8bb | pbrook | NEON_OP(shll_u64) |
1019 | 9ee6e8bb | pbrook | { |
1020 | 9ee6e8bb | pbrook | int shift = PARAM1;
|
1021 | 9ee6e8bb | pbrook | |
1022 | 9ee6e8bb | pbrook | T1 <<= shift; |
1023 | 9ee6e8bb | pbrook | T1 |= T0 >> (32 - shift);
|
1024 | 9ee6e8bb | pbrook | T0 <<= shift; |
1025 | 9ee6e8bb | pbrook | FORCE_RET(); |
1026 | 9ee6e8bb | pbrook | } |
1027 | 9ee6e8bb | pbrook | |
1028 | 9ee6e8bb | pbrook | NEON_OP(addl_u16) |
1029 | 9ee6e8bb | pbrook | { |
1030 | 9ee6e8bb | pbrook | uint32_t tmp; |
1031 | 9ee6e8bb | pbrook | uint32_t high; |
1032 | 9ee6e8bb | pbrook | |
1033 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[0];
|
1034 | 9ee6e8bb | pbrook | high = (T0 >> 16) + (tmp >> 16); |
1035 | 9ee6e8bb | pbrook | T0 = (uint16_t)(T0 + tmp); |
1036 | 9ee6e8bb | pbrook | T0 |= (high << 16);
|
1037 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[1];
|
1038 | 9ee6e8bb | pbrook | high = (T1 >> 16) + (tmp >> 16); |
1039 | 9ee6e8bb | pbrook | T1 = (uint16_t)(T1 + tmp); |
1040 | 9ee6e8bb | pbrook | T1 |= (high << 16);
|
1041 | 9ee6e8bb | pbrook | FORCE_RET(); |
1042 | 9ee6e8bb | pbrook | } |
1043 | 9ee6e8bb | pbrook | |
1044 | 9ee6e8bb | pbrook | NEON_OP(addl_u32) |
1045 | 9ee6e8bb | pbrook | { |
1046 | 9ee6e8bb | pbrook | T0 += env->vfp.scratch[0];
|
1047 | 9ee6e8bb | pbrook | T1 += env->vfp.scratch[1];
|
1048 | 9ee6e8bb | pbrook | FORCE_RET(); |
1049 | 9ee6e8bb | pbrook | } |
1050 | 9ee6e8bb | pbrook | |
1051 | 9ee6e8bb | pbrook | NEON_OP(addl_u64) |
1052 | 9ee6e8bb | pbrook | { |
1053 | 9ee6e8bb | pbrook | uint64_t tmp; |
1054 | 9ee6e8bb | pbrook | tmp = T0 | ((uint64_t)T1 << 32);
|
1055 | 9ee6e8bb | pbrook | tmp += env->vfp.scratch[0];
|
1056 | 9ee6e8bb | pbrook | tmp += (uint64_t)env->vfp.scratch[1] << 32; |
1057 | 9ee6e8bb | pbrook | T0 = tmp; |
1058 | 9ee6e8bb | pbrook | T1 = tmp >> 32;
|
1059 | 9ee6e8bb | pbrook | FORCE_RET(); |
1060 | 9ee6e8bb | pbrook | } |
1061 | 9ee6e8bb | pbrook | |
1062 | 9ee6e8bb | pbrook | NEON_OP(subl_u16) |
1063 | 9ee6e8bb | pbrook | { |
1064 | 9ee6e8bb | pbrook | uint32_t tmp; |
1065 | 9ee6e8bb | pbrook | uint32_t high; |
1066 | 9ee6e8bb | pbrook | |
1067 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[0];
|
1068 | 9ee6e8bb | pbrook | high = (T0 >> 16) - (tmp >> 16); |
1069 | 9ee6e8bb | pbrook | T0 = (uint16_t)(T0 - tmp); |
1070 | 9ee6e8bb | pbrook | T0 |= (high << 16);
|
1071 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[1];
|
1072 | 9ee6e8bb | pbrook | high = (T1 >> 16) - (tmp >> 16); |
1073 | 9ee6e8bb | pbrook | T1 = (uint16_t)(T1 - tmp); |
1074 | 9ee6e8bb | pbrook | T1 |= (high << 16);
|
1075 | 9ee6e8bb | pbrook | FORCE_RET(); |
1076 | 9ee6e8bb | pbrook | } |
1077 | 9ee6e8bb | pbrook | |
1078 | 9ee6e8bb | pbrook | NEON_OP(subl_u32) |
1079 | 9ee6e8bb | pbrook | { |
1080 | 9ee6e8bb | pbrook | T0 -= env->vfp.scratch[0];
|
1081 | 9ee6e8bb | pbrook | T1 -= env->vfp.scratch[1];
|
1082 | 9ee6e8bb | pbrook | FORCE_RET(); |
1083 | 9ee6e8bb | pbrook | } |
1084 | 9ee6e8bb | pbrook | |
1085 | 9ee6e8bb | pbrook | NEON_OP(subl_u64) |
1086 | 9ee6e8bb | pbrook | { |
1087 | 9ee6e8bb | pbrook | uint64_t tmp; |
1088 | 9ee6e8bb | pbrook | tmp = T0 | ((uint64_t)T1 << 32);
|
1089 | 9ee6e8bb | pbrook | tmp -= env->vfp.scratch[0];
|
1090 | 9ee6e8bb | pbrook | tmp -= (uint64_t)env->vfp.scratch[1] << 32; |
1091 | 9ee6e8bb | pbrook | T0 = tmp; |
1092 | 9ee6e8bb | pbrook | T1 = tmp >> 32;
|
1093 | 9ee6e8bb | pbrook | FORCE_RET(); |
1094 | 9ee6e8bb | pbrook | } |
1095 | 9ee6e8bb | pbrook | |
1096 | 9ee6e8bb | pbrook | #define DO_ABD(dest, x, y, type) do { \ |
1097 | 9ee6e8bb | pbrook | type tmp_x = x; \ |
1098 | 9ee6e8bb | pbrook | type tmp_y = y; \ |
1099 | 9ee6e8bb | pbrook | dest = ((tmp_x > tmp_y) ? tmp_x - tmp_y : tmp_y - tmp_x); \ |
1100 | 9ee6e8bb | pbrook | } while(0) |
1101 | 9ee6e8bb | pbrook | |
1102 | 9ee6e8bb | pbrook | NEON_OP(abdl_u16) |
1103 | 9ee6e8bb | pbrook | { |
1104 | 9ee6e8bb | pbrook | uint32_t tmp; |
1105 | 9ee6e8bb | pbrook | uint32_t low; |
1106 | 9ee6e8bb | pbrook | uint32_t high; |
1107 | 9ee6e8bb | pbrook | |
1108 | 9ee6e8bb | pbrook | DO_ABD(low, T0, T1, uint8_t); |
1109 | 9ee6e8bb | pbrook | DO_ABD(tmp, T0 >> 8, T1 >> 8, uint8_t); |
1110 | 9ee6e8bb | pbrook | low |= tmp << 16;
|
1111 | 9ee6e8bb | pbrook | DO_ABD(high, T0 >> 16, T1 >> 16, uint8_t); |
1112 | 9ee6e8bb | pbrook | DO_ABD(tmp, T0 >> 24, T1 >> 24, uint8_t); |
1113 | 9ee6e8bb | pbrook | high |= tmp << 16;
|
1114 | 9ee6e8bb | pbrook | T0 = low; |
1115 | 9ee6e8bb | pbrook | T1 = high; |
1116 | 9ee6e8bb | pbrook | FORCE_RET(); |
1117 | 9ee6e8bb | pbrook | } |
1118 | 9ee6e8bb | pbrook | |
1119 | 9ee6e8bb | pbrook | NEON_OP(abdl_s16) |
1120 | 9ee6e8bb | pbrook | { |
1121 | 9ee6e8bb | pbrook | uint32_t tmp; |
1122 | 9ee6e8bb | pbrook | uint32_t low; |
1123 | 9ee6e8bb | pbrook | uint32_t high; |
1124 | 9ee6e8bb | pbrook | |
1125 | 9ee6e8bb | pbrook | DO_ABD(low, T0, T1, int8_t); |
1126 | 9ee6e8bb | pbrook | DO_ABD(tmp, T0 >> 8, T1 >> 8, int8_t); |
1127 | 9ee6e8bb | pbrook | low |= tmp << 16;
|
1128 | 9ee6e8bb | pbrook | DO_ABD(high, T0 >> 16, T1 >> 16, int8_t); |
1129 | 9ee6e8bb | pbrook | DO_ABD(tmp, T0 >> 24, T1 >> 24, int8_t); |
1130 | 9ee6e8bb | pbrook | high |= tmp << 16;
|
1131 | 9ee6e8bb | pbrook | T0 = low; |
1132 | 9ee6e8bb | pbrook | T1 = high; |
1133 | 9ee6e8bb | pbrook | FORCE_RET(); |
1134 | 9ee6e8bb | pbrook | } |
1135 | 9ee6e8bb | pbrook | |
1136 | 9ee6e8bb | pbrook | NEON_OP(abdl_u32) |
1137 | 9ee6e8bb | pbrook | { |
1138 | 9ee6e8bb | pbrook | uint32_t low; |
1139 | 9ee6e8bb | pbrook | uint32_t high; |
1140 | 9ee6e8bb | pbrook | |
1141 | 9ee6e8bb | pbrook | DO_ABD(low, T0, T1, uint16_t); |
1142 | 9ee6e8bb | pbrook | DO_ABD(high, T0 >> 16, T1 >> 16, uint16_t); |
1143 | 9ee6e8bb | pbrook | T0 = low; |
1144 | 9ee6e8bb | pbrook | T1 = high; |
1145 | 9ee6e8bb | pbrook | FORCE_RET(); |
1146 | 9ee6e8bb | pbrook | } |
1147 | 9ee6e8bb | pbrook | |
1148 | 9ee6e8bb | pbrook | NEON_OP(abdl_s32) |
1149 | 9ee6e8bb | pbrook | { |
1150 | 9ee6e8bb | pbrook | uint32_t low; |
1151 | 9ee6e8bb | pbrook | uint32_t high; |
1152 | 9ee6e8bb | pbrook | |
1153 | 9ee6e8bb | pbrook | DO_ABD(low, T0, T1, int16_t); |
1154 | 9ee6e8bb | pbrook | DO_ABD(high, T0 >> 16, T1 >> 16, int16_t); |
1155 | 9ee6e8bb | pbrook | T0 = low; |
1156 | 9ee6e8bb | pbrook | T1 = high; |
1157 | 9ee6e8bb | pbrook | FORCE_RET(); |
1158 | 9ee6e8bb | pbrook | } |
1159 | 9ee6e8bb | pbrook | |
1160 | 9ee6e8bb | pbrook | NEON_OP(abdl_u64) |
1161 | 9ee6e8bb | pbrook | { |
1162 | 9ee6e8bb | pbrook | DO_ABD(T0, T0, T1, uint32_t); |
1163 | 9ee6e8bb | pbrook | T1 = 0;
|
1164 | 9ee6e8bb | pbrook | } |
1165 | 9ee6e8bb | pbrook | |
1166 | 9ee6e8bb | pbrook | NEON_OP(abdl_s64) |
1167 | 9ee6e8bb | pbrook | { |
1168 | 9ee6e8bb | pbrook | DO_ABD(T0, T0, T1, int32_t); |
1169 | 9ee6e8bb | pbrook | T1 = 0;
|
1170 | 9ee6e8bb | pbrook | } |
1171 | 9ee6e8bb | pbrook | #undef DO_ABD
|
1172 | 9ee6e8bb | pbrook | |
1173 | 9ee6e8bb | pbrook | /* Widening multiple. Named type is the source type. */
|
1174 | 9ee6e8bb | pbrook | #define DO_MULL(dest, x, y, type1, type2) do { \ |
1175 | 9ee6e8bb | pbrook | type1 tmp_x = x; \ |
1176 | 9ee6e8bb | pbrook | type1 tmp_y = y; \ |
1177 | 9ee6e8bb | pbrook | dest = (type2)((type2)tmp_x * (type2)tmp_y); \ |
1178 | 9ee6e8bb | pbrook | } while(0) |
1179 | 9ee6e8bb | pbrook | |
1180 | 9ee6e8bb | pbrook | NEON_OP(mull_u8) |
1181 | 9ee6e8bb | pbrook | { |
1182 | 9ee6e8bb | pbrook | uint32_t tmp; |
1183 | 9ee6e8bb | pbrook | uint32_t low; |
1184 | 9ee6e8bb | pbrook | uint32_t high; |
1185 | 9ee6e8bb | pbrook | |
1186 | 9ee6e8bb | pbrook | DO_MULL(low, T0, T1, uint8_t, uint16_t); |
1187 | 9ee6e8bb | pbrook | DO_MULL(tmp, T0 >> 8, T1 >> 8, uint8_t, uint16_t); |
1188 | 9ee6e8bb | pbrook | low |= tmp << 16;
|
1189 | 9ee6e8bb | pbrook | DO_MULL(high, T0 >> 16, T1 >> 16, uint8_t, uint16_t); |
1190 | 9ee6e8bb | pbrook | DO_MULL(tmp, T0 >> 24, T1 >> 24, uint8_t, uint16_t); |
1191 | 9ee6e8bb | pbrook | high |= tmp << 16;
|
1192 | 9ee6e8bb | pbrook | T0 = low; |
1193 | 9ee6e8bb | pbrook | T1 = high; |
1194 | 9ee6e8bb | pbrook | FORCE_RET(); |
1195 | 9ee6e8bb | pbrook | } |
1196 | 9ee6e8bb | pbrook | |
1197 | 9ee6e8bb | pbrook | NEON_OP(mull_s8) |
1198 | 9ee6e8bb | pbrook | { |
1199 | 9ee6e8bb | pbrook | uint32_t tmp; |
1200 | 9ee6e8bb | pbrook | uint32_t low; |
1201 | 9ee6e8bb | pbrook | uint32_t high; |
1202 | 9ee6e8bb | pbrook | |
1203 | 9ee6e8bb | pbrook | DO_MULL(low, T0, T1, int8_t, uint16_t); |
1204 | 9ee6e8bb | pbrook | DO_MULL(tmp, T0 >> 8, T1 >> 8, int8_t, uint16_t); |
1205 | 9ee6e8bb | pbrook | low |= tmp << 16;
|
1206 | 9ee6e8bb | pbrook | DO_MULL(high, T0 >> 16, T1 >> 16, int8_t, uint16_t); |
1207 | 9ee6e8bb | pbrook | DO_MULL(tmp, T0 >> 24, T1 >> 24, int8_t, uint16_t); |
1208 | 9ee6e8bb | pbrook | high |= tmp << 16;
|
1209 | 9ee6e8bb | pbrook | T0 = low; |
1210 | 9ee6e8bb | pbrook | T1 = high; |
1211 | 9ee6e8bb | pbrook | FORCE_RET(); |
1212 | 9ee6e8bb | pbrook | } |
1213 | 9ee6e8bb | pbrook | |
1214 | 9ee6e8bb | pbrook | NEON_OP(mull_u16) |
1215 | 9ee6e8bb | pbrook | { |
1216 | 9ee6e8bb | pbrook | uint32_t low; |
1217 | 9ee6e8bb | pbrook | uint32_t high; |
1218 | 9ee6e8bb | pbrook | |
1219 | 9ee6e8bb | pbrook | DO_MULL(low, T0, T1, uint16_t, uint32_t); |
1220 | 9ee6e8bb | pbrook | DO_MULL(high, T0 >> 16, T1 >> 16, uint16_t, uint32_t); |
1221 | 9ee6e8bb | pbrook | T0 = low; |
1222 | 9ee6e8bb | pbrook | T1 = high; |
1223 | 9ee6e8bb | pbrook | FORCE_RET(); |
1224 | 9ee6e8bb | pbrook | } |
1225 | 9ee6e8bb | pbrook | |
1226 | 9ee6e8bb | pbrook | NEON_OP(mull_s16) |
1227 | 9ee6e8bb | pbrook | { |
1228 | 9ee6e8bb | pbrook | uint32_t low; |
1229 | 9ee6e8bb | pbrook | uint32_t high; |
1230 | 9ee6e8bb | pbrook | |
1231 | 9ee6e8bb | pbrook | DO_MULL(low, T0, T1, int16_t, uint32_t); |
1232 | 9ee6e8bb | pbrook | DO_MULL(high, T0 >> 16, T1 >> 16, int16_t, uint32_t); |
1233 | 9ee6e8bb | pbrook | T0 = low; |
1234 | 9ee6e8bb | pbrook | T1 = high; |
1235 | 9ee6e8bb | pbrook | FORCE_RET(); |
1236 | 9ee6e8bb | pbrook | } |
1237 | 9ee6e8bb | pbrook | |
1238 | 9ee6e8bb | pbrook | NEON_OP(addl_saturate_s32) |
1239 | 9ee6e8bb | pbrook | { |
1240 | 9ee6e8bb | pbrook | uint32_t tmp; |
1241 | 9ee6e8bb | pbrook | uint32_t res; |
1242 | 9ee6e8bb | pbrook | |
1243 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[0];
|
1244 | 9ee6e8bb | pbrook | res = T0 + tmp; |
1245 | 9ee6e8bb | pbrook | if (((res ^ T0) & SIGNBIT) && !((T0 ^ tmp) & SIGNBIT)) {
|
1246 | 9ee6e8bb | pbrook | env->QF = 1;
|
1247 | 9ee6e8bb | pbrook | T0 = (T0 >> 31) ^ 0x7fffffff; |
1248 | 9ee6e8bb | pbrook | } else {
|
1249 | 9ee6e8bb | pbrook | T0 = res; |
1250 | 9ee6e8bb | pbrook | } |
1251 | 9ee6e8bb | pbrook | tmp = env->vfp.scratch[1];
|
1252 | 9ee6e8bb | pbrook | res = T1 + tmp; |
1253 | 9ee6e8bb | pbrook | if (((res ^ T1) & SIGNBIT) && !((T1 ^ tmp) & SIGNBIT)) {
|
1254 | 9ee6e8bb | pbrook | env->QF = 1;
|
1255 | 9ee6e8bb | pbrook | T1 = (T1 >> 31) ^ 0x7fffffff; |
1256 | 9ee6e8bb | pbrook | } else {
|
1257 | 9ee6e8bb | pbrook | T1 = res; |
1258 | 9ee6e8bb | pbrook | } |
1259 | 9ee6e8bb | pbrook | FORCE_RET(); |
1260 | 9ee6e8bb | pbrook | } |
1261 | 9ee6e8bb | pbrook | |
1262 | 9ee6e8bb | pbrook | NEON_OP(addl_saturate_s64) |
1263 | 9ee6e8bb | pbrook | { |
1264 | 9ee6e8bb | pbrook | uint64_t src1; |
1265 | 9ee6e8bb | pbrook | uint64_t src2; |
1266 | 9ee6e8bb | pbrook | uint64_t res; |
1267 | 9ee6e8bb | pbrook | |
1268 | 9ee6e8bb | pbrook | src1 = T0 + ((uint64_t)T1 << 32);
|
1269 | 9ee6e8bb | pbrook | src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
1270 | 9ee6e8bb | pbrook | res = src1 + src2; |
1271 | 9ee6e8bb | pbrook | if (((res ^ src1) & SIGNBIT64) && !((src1 ^ src2) & SIGNBIT64)) {
|
1272 | 9ee6e8bb | pbrook | env->QF = 1;
|
1273 | 9ee6e8bb | pbrook | T0 = ~(int64_t)src1 >> 63;
|
1274 | 9ee6e8bb | pbrook | T1 = T0 ^ 0x80000000;
|
1275 | 9ee6e8bb | pbrook | } else {
|
1276 | 9ee6e8bb | pbrook | T0 = res; |
1277 | 9ee6e8bb | pbrook | T1 = res >> 32;
|
1278 | 9ee6e8bb | pbrook | } |
1279 | 9ee6e8bb | pbrook | FORCE_RET(); |
1280 | 9ee6e8bb | pbrook | } |
1281 | 9ee6e8bb | pbrook | |
1282 | 9ee6e8bb | pbrook | NEON_OP(addl_saturate_u64) |
1283 | 9ee6e8bb | pbrook | { |
1284 | 9ee6e8bb | pbrook | uint64_t src1; |
1285 | 9ee6e8bb | pbrook | uint64_t src2; |
1286 | 9ee6e8bb | pbrook | uint64_t res; |
1287 | 9ee6e8bb | pbrook | |
1288 | 9ee6e8bb | pbrook | src1 = T0 + ((uint64_t)T1 << 32);
|
1289 | 9ee6e8bb | pbrook | src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
1290 | 9ee6e8bb | pbrook | res = src1 + src2; |
1291 | 9ee6e8bb | pbrook | if (res < src1) {
|
1292 | 9ee6e8bb | pbrook | env->QF = 1;
|
1293 | 9ee6e8bb | pbrook | T0 = 0xffffffff;
|
1294 | 9ee6e8bb | pbrook | T1 = 0xffffffff;
|
1295 | 9ee6e8bb | pbrook | } else {
|
1296 | 9ee6e8bb | pbrook | T0 = res; |
1297 | 9ee6e8bb | pbrook | T1 = res >> 32;
|
1298 | 9ee6e8bb | pbrook | } |
1299 | 9ee6e8bb | pbrook | FORCE_RET(); |
1300 | 9ee6e8bb | pbrook | } |
1301 | 9ee6e8bb | pbrook | |
1302 | 9ee6e8bb | pbrook | NEON_OP(subl_saturate_s64) |
1303 | 9ee6e8bb | pbrook | { |
1304 | 9ee6e8bb | pbrook | uint64_t src1; |
1305 | 9ee6e8bb | pbrook | uint64_t src2; |
1306 | 9ee6e8bb | pbrook | uint64_t res; |
1307 | 9ee6e8bb | pbrook | |
1308 | 9ee6e8bb | pbrook | src1 = T0 + ((uint64_t)T1 << 32);
|
1309 | 9ee6e8bb | pbrook | src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
1310 | 9ee6e8bb | pbrook | res = src1 - src2; |
1311 | 9ee6e8bb | pbrook | if (((res ^ src1) & SIGNBIT64) && ((src1 ^ src2) & SIGNBIT64)) {
|
1312 | 9ee6e8bb | pbrook | env->QF = 1;
|
1313 | 9ee6e8bb | pbrook | T0 = ~(int64_t)src1 >> 63;
|
1314 | 9ee6e8bb | pbrook | T1 = T0 ^ 0x80000000;
|
1315 | 9ee6e8bb | pbrook | } else {
|
1316 | 9ee6e8bb | pbrook | T0 = res; |
1317 | 9ee6e8bb | pbrook | T1 = res >> 32;
|
1318 | 9ee6e8bb | pbrook | } |
1319 | 9ee6e8bb | pbrook | FORCE_RET(); |
1320 | 9ee6e8bb | pbrook | } |
1321 | 9ee6e8bb | pbrook | |
1322 | 9ee6e8bb | pbrook | NEON_OP(subl_saturate_u64) |
1323 | 9ee6e8bb | pbrook | { |
1324 | 9ee6e8bb | pbrook | uint64_t src1; |
1325 | 9ee6e8bb | pbrook | uint64_t src2; |
1326 | 9ee6e8bb | pbrook | uint64_t res; |
1327 | 9ee6e8bb | pbrook | |
1328 | 9ee6e8bb | pbrook | src1 = T0 + ((uint64_t)T1 << 32);
|
1329 | 9ee6e8bb | pbrook | src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
1330 | 9ee6e8bb | pbrook | if (src1 < src2) {
|
1331 | 9ee6e8bb | pbrook | env->QF = 1;
|
1332 | 9ee6e8bb | pbrook | T0 = 0;
|
1333 | 9ee6e8bb | pbrook | T1 = 0;
|
1334 | 9ee6e8bb | pbrook | } else {
|
1335 | 9ee6e8bb | pbrook | res = src1 - src2; |
1336 | 9ee6e8bb | pbrook | T0 = res; |
1337 | 9ee6e8bb | pbrook | T1 = res >> 32;
|
1338 | 9ee6e8bb | pbrook | } |
1339 | 9ee6e8bb | pbrook | FORCE_RET(); |
1340 | 9ee6e8bb | pbrook | } |
1341 | 9ee6e8bb | pbrook | |
1342 | 9ee6e8bb | pbrook | NEON_OP(negl_u16) |
1343 | 9ee6e8bb | pbrook | { |
1344 | 9ee6e8bb | pbrook | uint32_t tmp; |
1345 | 9ee6e8bb | pbrook | tmp = T0 >> 16;
|
1346 | 9ee6e8bb | pbrook | tmp = -tmp; |
1347 | 9ee6e8bb | pbrook | T0 = (-T0 & 0xffff) | (tmp << 16); |
1348 | 9ee6e8bb | pbrook | tmp = T1 >> 16;
|
1349 | 9ee6e8bb | pbrook | tmp = -tmp; |
1350 | 9ee6e8bb | pbrook | T1 = (-T1 & 0xffff) | (tmp << 16); |
1351 | 9ee6e8bb | pbrook | FORCE_RET(); |
1352 | 9ee6e8bb | pbrook | } |
1353 | 9ee6e8bb | pbrook | |
1354 | 9ee6e8bb | pbrook | NEON_OP(negl_u32) |
1355 | 9ee6e8bb | pbrook | { |
1356 | 9ee6e8bb | pbrook | T0 = -T0; |
1357 | 9ee6e8bb | pbrook | T1 = -T1; |
1358 | 9ee6e8bb | pbrook | FORCE_RET(); |
1359 | 9ee6e8bb | pbrook | } |
1360 | 9ee6e8bb | pbrook | |
1361 | 9ee6e8bb | pbrook | NEON_OP(negl_u64) |
1362 | 9ee6e8bb | pbrook | { |
1363 | 9ee6e8bb | pbrook | uint64_t val; |
1364 | 9ee6e8bb | pbrook | |
1365 | 9ee6e8bb | pbrook | val = T0 | ((uint64_t)T1 << 32);
|
1366 | 9ee6e8bb | pbrook | val = -val; |
1367 | 9ee6e8bb | pbrook | T0 = val; |
1368 | 9ee6e8bb | pbrook | T1 = val >> 32;
|
1369 | 9ee6e8bb | pbrook | FORCE_RET(); |
1370 | 9ee6e8bb | pbrook | } |
1371 | 9ee6e8bb | pbrook | |
1372 | 9ee6e8bb | pbrook | /* Scalar operations. */
|
1373 | 9ee6e8bb | pbrook | NEON_OP(dup_low16) |
1374 | 9ee6e8bb | pbrook | { |
1375 | 9ee6e8bb | pbrook | T0 = (T0 & 0xffff) | (T0 << 16); |
1376 | 9ee6e8bb | pbrook | FORCE_RET(); |
1377 | 9ee6e8bb | pbrook | } |
1378 | 9ee6e8bb | pbrook | |
1379 | 9ee6e8bb | pbrook | NEON_OP(dup_high16) |
1380 | 9ee6e8bb | pbrook | { |
1381 | 9ee6e8bb | pbrook | T0 = (T0 >> 16) | (T0 & 0xffff0000); |
1382 | 9ee6e8bb | pbrook | FORCE_RET(); |
1383 | 9ee6e8bb | pbrook | } |
1384 | 9ee6e8bb | pbrook | |
1385 | 9ee6e8bb | pbrook | /* Helper for VEXT */
|
1386 | 9ee6e8bb | pbrook | NEON_OP(extract) |
1387 | 9ee6e8bb | pbrook | { |
1388 | 9ee6e8bb | pbrook | int shift = PARAM1;
|
1389 | 9ee6e8bb | pbrook | T0 = (T0 >> shift) | (T1 << (32 - shift));
|
1390 | 9ee6e8bb | pbrook | FORCE_RET(); |
1391 | 9ee6e8bb | pbrook | } |
1392 | 9ee6e8bb | pbrook | |
1393 | 9ee6e8bb | pbrook | /* Pairwise add long. Named type is source type. */
|
1394 | 9ee6e8bb | pbrook | NEON_OP(paddl_s8) |
1395 | 9ee6e8bb | pbrook | { |
1396 | 9ee6e8bb | pbrook | int8_t src1; |
1397 | 9ee6e8bb | pbrook | int8_t src2; |
1398 | 9ee6e8bb | pbrook | uint16_t result; |
1399 | 9ee6e8bb | pbrook | src1 = T0 >> 24;
|
1400 | 9ee6e8bb | pbrook | src2 = T0 >> 16;
|
1401 | 9ee6e8bb | pbrook | result = (uint16_t)src1 + src2; |
1402 | 9ee6e8bb | pbrook | src1 = T0 >> 8;
|
1403 | 9ee6e8bb | pbrook | src2 = T0; |
1404 | 9ee6e8bb | pbrook | T0 = (uint16_t)((uint16_t)src1 + src2) | ((uint32_t)result << 16);
|
1405 | 9ee6e8bb | pbrook | FORCE_RET(); |
1406 | 9ee6e8bb | pbrook | } |
1407 | 9ee6e8bb | pbrook | |
1408 | 9ee6e8bb | pbrook | NEON_OP(paddl_u8) |
1409 | 9ee6e8bb | pbrook | { |
1410 | 9ee6e8bb | pbrook | uint8_t src1; |
1411 | 9ee6e8bb | pbrook | uint8_t src2; |
1412 | 9ee6e8bb | pbrook | uint16_t result; |
1413 | 9ee6e8bb | pbrook | src1 = T0 >> 24;
|
1414 | 9ee6e8bb | pbrook | src2 = T0 >> 16;
|
1415 | 9ee6e8bb | pbrook | result = (uint16_t)src1 + src2; |
1416 | 9ee6e8bb | pbrook | src1 = T0 >> 8;
|
1417 | 9ee6e8bb | pbrook | src2 = T0; |
1418 | 9ee6e8bb | pbrook | T0 = (uint16_t)((uint16_t)src1 + src2) | ((uint32_t)result << 16);
|
1419 | 9ee6e8bb | pbrook | FORCE_RET(); |
1420 | 9ee6e8bb | pbrook | } |
1421 | 9ee6e8bb | pbrook | |
1422 | 9ee6e8bb | pbrook | NEON_OP(paddl_s16) |
1423 | 9ee6e8bb | pbrook | { |
1424 | 9ee6e8bb | pbrook | T0 = (uint32_t)(int16_t)T0 + (uint32_t)(int16_t)(T0 >> 16);
|
1425 | 9ee6e8bb | pbrook | FORCE_RET(); |
1426 | 9ee6e8bb | pbrook | } |
1427 | 9ee6e8bb | pbrook | |
1428 | 9ee6e8bb | pbrook | NEON_OP(paddl_u16) |
1429 | 9ee6e8bb | pbrook | { |
1430 | 9ee6e8bb | pbrook | T0 = (uint32_t)(uint16_t)T0 + (uint32_t)(uint16_t)(T0 >> 16);
|
1431 | 9ee6e8bb | pbrook | FORCE_RET(); |
1432 | 9ee6e8bb | pbrook | } |
1433 | 9ee6e8bb | pbrook | |
1434 | 9ee6e8bb | pbrook | NEON_OP(paddl_s32) |
1435 | 9ee6e8bb | pbrook | { |
1436 | 9ee6e8bb | pbrook | int64_t tmp; |
1437 | 9ee6e8bb | pbrook | tmp = (int64_t)(int32_t)T0 + (int64_t)(int32_t)T1; |
1438 | 9ee6e8bb | pbrook | T0 = tmp; |
1439 | 9ee6e8bb | pbrook | T1 = tmp >> 32;
|
1440 | 9ee6e8bb | pbrook | FORCE_RET(); |
1441 | 9ee6e8bb | pbrook | } |
1442 | 9ee6e8bb | pbrook | |
1443 | 9ee6e8bb | pbrook | NEON_OP(paddl_u32) |
1444 | 9ee6e8bb | pbrook | { |
1445 | 9ee6e8bb | pbrook | uint64_t tmp; |
1446 | 9ee6e8bb | pbrook | tmp = (uint64_t)T0 + (uint64_t)T1; |
1447 | 9ee6e8bb | pbrook | T0 = tmp; |
1448 | 9ee6e8bb | pbrook | T1 = tmp >> 32;
|
1449 | 9ee6e8bb | pbrook | FORCE_RET(); |
1450 | 9ee6e8bb | pbrook | } |
1451 | 9ee6e8bb | pbrook | |
1452 | 9ee6e8bb | pbrook | /* Count Leading Sign/Zero Bits. */
|
1453 | 9ee6e8bb | pbrook | static inline int do_clz8(uint8_t x) |
1454 | 9ee6e8bb | pbrook | { |
1455 | 9ee6e8bb | pbrook | int n;
|
1456 | 9ee6e8bb | pbrook | for (n = 8; x; n--) |
1457 | 9ee6e8bb | pbrook | x >>= 1;
|
1458 | 9ee6e8bb | pbrook | return n;
|
1459 | 9ee6e8bb | pbrook | } |
1460 | 9ee6e8bb | pbrook | |
1461 | 9ee6e8bb | pbrook | static inline int do_clz16(uint16_t x) |
1462 | 9ee6e8bb | pbrook | { |
1463 | 9ee6e8bb | pbrook | int n;
|
1464 | 9ee6e8bb | pbrook | for (n = 16; x; n--) |
1465 | 9ee6e8bb | pbrook | x >>= 1;
|
1466 | 9ee6e8bb | pbrook | return n;
|
1467 | 9ee6e8bb | pbrook | } |
1468 | 9ee6e8bb | pbrook | |
1469 | 9ee6e8bb | pbrook | NEON_OP(clz_u8) |
1470 | 9ee6e8bb | pbrook | { |
1471 | 9ee6e8bb | pbrook | uint32_t result; |
1472 | 9ee6e8bb | pbrook | uint32_t tmp; |
1473 | 9ee6e8bb | pbrook | |
1474 | 9ee6e8bb | pbrook | tmp = T0; |
1475 | 9ee6e8bb | pbrook | result = do_clz8(tmp); |
1476 | 9ee6e8bb | pbrook | result |= do_clz8(tmp >> 8) << 8; |
1477 | 9ee6e8bb | pbrook | result |= do_clz8(tmp >> 16) << 16; |
1478 | 9ee6e8bb | pbrook | result |= do_clz8(tmp >> 24) << 24; |
1479 | 9ee6e8bb | pbrook | T0 = result; |
1480 | 9ee6e8bb | pbrook | FORCE_RET(); |
1481 | 9ee6e8bb | pbrook | } |
1482 | 9ee6e8bb | pbrook | |
1483 | 9ee6e8bb | pbrook | NEON_OP(clz_u16) |
1484 | 9ee6e8bb | pbrook | { |
1485 | 9ee6e8bb | pbrook | uint32_t result; |
1486 | 9ee6e8bb | pbrook | uint32_t tmp; |
1487 | 9ee6e8bb | pbrook | tmp = T0; |
1488 | 9ee6e8bb | pbrook | result = do_clz16(tmp); |
1489 | 9ee6e8bb | pbrook | result |= do_clz16(tmp >> 16) << 16; |
1490 | 9ee6e8bb | pbrook | T0 = result; |
1491 | 9ee6e8bb | pbrook | FORCE_RET(); |
1492 | 9ee6e8bb | pbrook | } |
1493 | 9ee6e8bb | pbrook | |
1494 | 9ee6e8bb | pbrook | NEON_OP(cls_s8) |
1495 | 9ee6e8bb | pbrook | { |
1496 | 9ee6e8bb | pbrook | uint32_t result; |
1497 | 9ee6e8bb | pbrook | int8_t tmp; |
1498 | 9ee6e8bb | pbrook | tmp = T0; |
1499 | 9ee6e8bb | pbrook | result = do_clz8((tmp < 0) ? ~tmp : tmp) - 1; |
1500 | 9ee6e8bb | pbrook | tmp = T0 >> 8;
|
1501 | 9ee6e8bb | pbrook | result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 8; |
1502 | 9ee6e8bb | pbrook | tmp = T0 >> 16;
|
1503 | 9ee6e8bb | pbrook | result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 16; |
1504 | 9ee6e8bb | pbrook | tmp = T0 >> 24;
|
1505 | 9ee6e8bb | pbrook | result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 24; |
1506 | 9ee6e8bb | pbrook | T0 = result; |
1507 | 9ee6e8bb | pbrook | FORCE_RET(); |
1508 | 9ee6e8bb | pbrook | } |
1509 | 9ee6e8bb | pbrook | |
1510 | 9ee6e8bb | pbrook | NEON_OP(cls_s16) |
1511 | 9ee6e8bb | pbrook | { |
1512 | 9ee6e8bb | pbrook | uint32_t result; |
1513 | 9ee6e8bb | pbrook | int16_t tmp; |
1514 | 9ee6e8bb | pbrook | tmp = T0; |
1515 | 9ee6e8bb | pbrook | result = do_clz16((tmp < 0) ? ~tmp : tmp) - 1; |
1516 | 9ee6e8bb | pbrook | tmp = T0 >> 16;
|
1517 | 9ee6e8bb | pbrook | result |= (do_clz16((tmp < 0) ? ~tmp : tmp) - 1) << 16; |
1518 | 9ee6e8bb | pbrook | T0 = result; |
1519 | 9ee6e8bb | pbrook | FORCE_RET(); |
1520 | 9ee6e8bb | pbrook | } |
1521 | 9ee6e8bb | pbrook | |
1522 | 9ee6e8bb | pbrook | NEON_OP(cls_s32) |
1523 | 9ee6e8bb | pbrook | { |
1524 | 9ee6e8bb | pbrook | int count;
|
1525 | 9ee6e8bb | pbrook | if ((int32_t)T0 < 0) |
1526 | 9ee6e8bb | pbrook | T0 = ~T0; |
1527 | 9ee6e8bb | pbrook | for (count = 32; T0 > 0; count--) |
1528 | 9ee6e8bb | pbrook | T0 = T0 >> 1;
|
1529 | 9ee6e8bb | pbrook | T0 = count - 1;
|
1530 | 9ee6e8bb | pbrook | FORCE_RET(); |
1531 | 9ee6e8bb | pbrook | } |
1532 | 9ee6e8bb | pbrook | |
1533 | 9ee6e8bb | pbrook | /* Bit count. */
|
1534 | 9ee6e8bb | pbrook | NEON_OP(cnt_u8) |
1535 | 9ee6e8bb | pbrook | { |
1536 | 9ee6e8bb | pbrook | T0 = (T0 & 0x55555555) + ((T0 >> 1) & 0x55555555); |
1537 | 9ee6e8bb | pbrook | T0 = (T0 & 0x33333333) + ((T0 >> 2) & 0x33333333); |
1538 | 9ee6e8bb | pbrook | T0 = (T0 & 0x0f0f0f0f) + ((T0 >> 4) & 0x0f0f0f0f); |
1539 | 9ee6e8bb | pbrook | FORCE_RET(); |
1540 | 9ee6e8bb | pbrook | } |
1541 | 9ee6e8bb | pbrook | |
1542 | 9ee6e8bb | pbrook | /* Saturnating negation. */
|
1543 | 9ee6e8bb | pbrook | /* ??? Make these use NEON_VOP1 */
|
1544 | 9ee6e8bb | pbrook | #define DO_QABS8(x) do { \ |
1545 | 9ee6e8bb | pbrook | if (x == (int8_t)0x80) { \ |
1546 | 9ee6e8bb | pbrook | x = 0x7f; \
|
1547 | 9ee6e8bb | pbrook | env->QF = 1; \
|
1548 | 9ee6e8bb | pbrook | } else if (x < 0) { \ |
1549 | 9ee6e8bb | pbrook | x = -x; \ |
1550 | 9ee6e8bb | pbrook | }} while (0) |
1551 | 9ee6e8bb | pbrook | NEON_OP(qabs_s8) |
1552 | 9ee6e8bb | pbrook | { |
1553 | 9ee6e8bb | pbrook | neon_s8 vec; |
1554 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s8, vec, T0); |
1555 | 9ee6e8bb | pbrook | DO_QABS8(vec.v1); |
1556 | 9ee6e8bb | pbrook | DO_QABS8(vec.v2); |
1557 | 9ee6e8bb | pbrook | DO_QABS8(vec.v3); |
1558 | 9ee6e8bb | pbrook | DO_QABS8(vec.v4); |
1559 | 9ee6e8bb | pbrook | NEON_PACK(neon_s8, T0, vec); |
1560 | 9ee6e8bb | pbrook | FORCE_RET(); |
1561 | 9ee6e8bb | pbrook | } |
1562 | 9ee6e8bb | pbrook | #undef DO_QABS8
|
1563 | 9ee6e8bb | pbrook | |
1564 | 9ee6e8bb | pbrook | #define DO_QNEG8(x) do { \ |
1565 | 9ee6e8bb | pbrook | if (x == (int8_t)0x80) { \ |
1566 | 9ee6e8bb | pbrook | x = 0x7f; \
|
1567 | 9ee6e8bb | pbrook | env->QF = 1; \
|
1568 | 9ee6e8bb | pbrook | } else { \
|
1569 | 9ee6e8bb | pbrook | x = -x; \ |
1570 | 9ee6e8bb | pbrook | }} while (0) |
1571 | 9ee6e8bb | pbrook | NEON_OP(qneg_s8) |
1572 | 9ee6e8bb | pbrook | { |
1573 | 9ee6e8bb | pbrook | neon_s8 vec; |
1574 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s8, vec, T0); |
1575 | 9ee6e8bb | pbrook | DO_QNEG8(vec.v1); |
1576 | 9ee6e8bb | pbrook | DO_QNEG8(vec.v2); |
1577 | 9ee6e8bb | pbrook | DO_QNEG8(vec.v3); |
1578 | 9ee6e8bb | pbrook | DO_QNEG8(vec.v4); |
1579 | 9ee6e8bb | pbrook | NEON_PACK(neon_s8, T0, vec); |
1580 | 9ee6e8bb | pbrook | FORCE_RET(); |
1581 | 9ee6e8bb | pbrook | } |
1582 | 9ee6e8bb | pbrook | #undef DO_QNEG8
|
1583 | 9ee6e8bb | pbrook | |
1584 | 9ee6e8bb | pbrook | #define DO_QABS16(x) do { \ |
1585 | 9ee6e8bb | pbrook | if (x == (int16_t)0x8000) { \ |
1586 | 9ee6e8bb | pbrook | x = 0x7fff; \
|
1587 | 9ee6e8bb | pbrook | env->QF = 1; \
|
1588 | 9ee6e8bb | pbrook | } else if (x < 0) { \ |
1589 | 9ee6e8bb | pbrook | x = -x; \ |
1590 | 9ee6e8bb | pbrook | }} while (0) |
1591 | 9ee6e8bb | pbrook | NEON_OP(qabs_s16) |
1592 | 9ee6e8bb | pbrook | { |
1593 | 9ee6e8bb | pbrook | neon_s16 vec; |
1594 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s16, vec, T0); |
1595 | 9ee6e8bb | pbrook | DO_QABS16(vec.v1); |
1596 | 9ee6e8bb | pbrook | DO_QABS16(vec.v2); |
1597 | 9ee6e8bb | pbrook | NEON_PACK(neon_s16, T0, vec); |
1598 | 9ee6e8bb | pbrook | FORCE_RET(); |
1599 | 9ee6e8bb | pbrook | } |
1600 | 9ee6e8bb | pbrook | #undef DO_QABS16
|
1601 | 9ee6e8bb | pbrook | |
1602 | 9ee6e8bb | pbrook | #define DO_QNEG16(x) do { \ |
1603 | 9ee6e8bb | pbrook | if (x == (int16_t)0x8000) { \ |
1604 | 9ee6e8bb | pbrook | x = 0x7fff; \
|
1605 | 9ee6e8bb | pbrook | env->QF = 1; \
|
1606 | 9ee6e8bb | pbrook | } else { \
|
1607 | 9ee6e8bb | pbrook | x = -x; \ |
1608 | 9ee6e8bb | pbrook | }} while (0) |
1609 | 9ee6e8bb | pbrook | NEON_OP(qneg_s16) |
1610 | 9ee6e8bb | pbrook | { |
1611 | 9ee6e8bb | pbrook | neon_s16 vec; |
1612 | 9ee6e8bb | pbrook | NEON_UNPACK(neon_s16, vec, T0); |
1613 | 9ee6e8bb | pbrook | DO_QNEG16(vec.v1); |
1614 | 9ee6e8bb | pbrook | DO_QNEG16(vec.v2); |
1615 | 9ee6e8bb | pbrook | NEON_PACK(neon_s16, T0, vec); |
1616 | 9ee6e8bb | pbrook | FORCE_RET(); |
1617 | 9ee6e8bb | pbrook | } |
1618 | 9ee6e8bb | pbrook | #undef DO_QNEG16
|
1619 | 9ee6e8bb | pbrook | |
1620 | 9ee6e8bb | pbrook | NEON_OP(qabs_s32) |
1621 | 9ee6e8bb | pbrook | { |
1622 | 9ee6e8bb | pbrook | if (T0 == 0x80000000) { |
1623 | 9ee6e8bb | pbrook | T0 = 0x7fffffff;
|
1624 | 9ee6e8bb | pbrook | env->QF = 1;
|
1625 | 9ee6e8bb | pbrook | } else if ((int32_t)T0 < 0) { |
1626 | 9ee6e8bb | pbrook | T0 = -T0; |
1627 | 9ee6e8bb | pbrook | } |
1628 | 9ee6e8bb | pbrook | FORCE_RET(); |
1629 | 9ee6e8bb | pbrook | } |
1630 | 9ee6e8bb | pbrook | |
1631 | 9ee6e8bb | pbrook | NEON_OP(qneg_s32) |
1632 | 9ee6e8bb | pbrook | { |
1633 | 9ee6e8bb | pbrook | if (T0 == 0x80000000) { |
1634 | 9ee6e8bb | pbrook | T0 = 0x7fffffff;
|
1635 | 9ee6e8bb | pbrook | env->QF = 1;
|
1636 | 9ee6e8bb | pbrook | } else {
|
1637 | 9ee6e8bb | pbrook | T0 = -T0; |
1638 | 9ee6e8bb | pbrook | } |
1639 | 9ee6e8bb | pbrook | FORCE_RET(); |
1640 | 9ee6e8bb | pbrook | } |
1641 | 9ee6e8bb | pbrook | |
1642 | 9ee6e8bb | pbrook | /* Unary opperations */
|
1643 | 9ee6e8bb | pbrook | #define NEON_FN(dest, src, dummy) dest = (src < 0) ? -src : src |
1644 | 9ee6e8bb | pbrook | NEON_VOP1(abs_s8, neon_s8, 4)
|
1645 | 9ee6e8bb | pbrook | NEON_VOP1(abs_s16, neon_s16, 2)
|
1646 | 9ee6e8bb | pbrook | NEON_OP(abs_s32) |
1647 | 9ee6e8bb | pbrook | { |
1648 | 9ee6e8bb | pbrook | if ((int32_t)T0 < 0) |
1649 | 9ee6e8bb | pbrook | T0 = -T0; |
1650 | 9ee6e8bb | pbrook | FORCE_RET(); |
1651 | 9ee6e8bb | pbrook | } |
1652 | 9ee6e8bb | pbrook | #undef NEON_FN
|
1653 | 9ee6e8bb | pbrook | |
1654 | 9ee6e8bb | pbrook | /* Transpose. Argument order is rather strange to avoid special casing
|
1655 | 9ee6e8bb | pbrook | the tranlation code.
|
1656 | 9ee6e8bb | pbrook | On input T0 = rm, T1 = rd. On output T0 = rd, T1 = rm */
|
1657 | 9ee6e8bb | pbrook | NEON_OP(trn_u8) |
1658 | 9ee6e8bb | pbrook | { |
1659 | 9ee6e8bb | pbrook | uint32_t rd; |
1660 | 9ee6e8bb | pbrook | uint32_t rm; |
1661 | 9ee6e8bb | pbrook | rd = ((T0 & 0x00ff00ff) << 8) | (T1 & 0x00ff00ff); |
1662 | 9ee6e8bb | pbrook | rm = ((T1 & 0xff00ff00) >> 8) | (T0 & 0xff00ff00); |
1663 | 9ee6e8bb | pbrook | T0 = rd; |
1664 | 9ee6e8bb | pbrook | T1 = rm; |
1665 | 9ee6e8bb | pbrook | FORCE_RET(); |
1666 | 9ee6e8bb | pbrook | } |
1667 | 9ee6e8bb | pbrook | |
1668 | 9ee6e8bb | pbrook | NEON_OP(trn_u16) |
1669 | 9ee6e8bb | pbrook | { |
1670 | 9ee6e8bb | pbrook | uint32_t rd; |
1671 | 9ee6e8bb | pbrook | uint32_t rm; |
1672 | 9ee6e8bb | pbrook | rd = (T0 << 16) | (T1 & 0xffff); |
1673 | 9ee6e8bb | pbrook | rm = (T1 >> 16) | (T0 & 0xffff0000); |
1674 | 9ee6e8bb | pbrook | T0 = rd; |
1675 | 9ee6e8bb | pbrook | T1 = rm; |
1676 | 9ee6e8bb | pbrook | FORCE_RET(); |
1677 | 9ee6e8bb | pbrook | } |
1678 | 9ee6e8bb | pbrook | |
1679 | 9ee6e8bb | pbrook | /* Worker routines for zip and unzip. */
|
1680 | 9ee6e8bb | pbrook | NEON_OP(unzip_u8) |
1681 | 9ee6e8bb | pbrook | { |
1682 | 9ee6e8bb | pbrook | uint32_t rd; |
1683 | 9ee6e8bb | pbrook | uint32_t rm; |
1684 | 9ee6e8bb | pbrook | rd = (T0 & 0xff) | ((T0 >> 8) & 0xff00) |
1685 | 9ee6e8bb | pbrook | | ((T1 << 16) & 0xff0000) | ((T1 << 8) & 0xff000000); |
1686 | 9ee6e8bb | pbrook | rm = ((T0 >> 8) & 0xff) | ((T0 >> 16) & 0xff00) |
1687 | 9ee6e8bb | pbrook | | ((T1 << 8) & 0xff0000) | (T1 & 0xff000000); |
1688 | 9ee6e8bb | pbrook | T0 = rd; |
1689 | 9ee6e8bb | pbrook | T1 = rm; |
1690 | 9ee6e8bb | pbrook | FORCE_RET(); |
1691 | 9ee6e8bb | pbrook | } |
1692 | 9ee6e8bb | pbrook | |
1693 | 9ee6e8bb | pbrook | NEON_OP(zip_u8) |
1694 | 9ee6e8bb | pbrook | { |
1695 | 9ee6e8bb | pbrook | uint32_t rd; |
1696 | 9ee6e8bb | pbrook | uint32_t rm; |
1697 | 9ee6e8bb | pbrook | rd = (T0 & 0xff) | ((T1 << 8) & 0xff00) |
1698 | 9ee6e8bb | pbrook | | ((T0 << 16) & 0xff0000) | ((T1 << 24) & 0xff000000); |
1699 | 9ee6e8bb | pbrook | rm = ((T0 >> 16) & 0xff) | ((T1 >> 8) & 0xff00) |
1700 | 9ee6e8bb | pbrook | | ((T0 >> 8) & 0xff0000) | (T1 & 0xff000000); |
1701 | 9ee6e8bb | pbrook | T0 = rd; |
1702 | 9ee6e8bb | pbrook | T1 = rm; |
1703 | 9ee6e8bb | pbrook | FORCE_RET(); |
1704 | 9ee6e8bb | pbrook | } |
1705 | 9ee6e8bb | pbrook | |
1706 | 9ee6e8bb | pbrook | NEON_OP(zip_u16) |
1707 | 9ee6e8bb | pbrook | { |
1708 | 9ee6e8bb | pbrook | uint32_t tmp; |
1709 | 9ee6e8bb | pbrook | |
1710 | 9ee6e8bb | pbrook | tmp = (T0 & 0xffff) | (T1 << 16); |
1711 | 9ee6e8bb | pbrook | T1 = (T1 & 0xffff0000) | (T0 >> 16); |
1712 | 9ee6e8bb | pbrook | T0 = tmp; |
1713 | 9ee6e8bb | pbrook | FORCE_RET(); |
1714 | 9ee6e8bb | pbrook | } |
1715 | 9ee6e8bb | pbrook | |
1716 | 9ee6e8bb | pbrook | /* Table lookup. This accessed the register file directly. */
|
1717 | 9ee6e8bb | pbrook | NEON_OP(tbl) |
1718 | 9ee6e8bb | pbrook | { |
1719 | 9ee6e8bb | pbrook | helper_neon_tbl(PARAM1, PARAM2); |
1720 | 9ee6e8bb | pbrook | } |
1721 | 9ee6e8bb | pbrook | |
1722 | 9ee6e8bb | pbrook | NEON_OP(dup_u8) |
1723 | 9ee6e8bb | pbrook | { |
1724 | 9ee6e8bb | pbrook | T0 = (T0 >> PARAM1) & 0xff;
|
1725 | 9ee6e8bb | pbrook | T0 |= T0 << 8;
|
1726 | 9ee6e8bb | pbrook | T0 |= T0 << 16;
|
1727 | 9ee6e8bb | pbrook | FORCE_RET(); |
1728 | 9ee6e8bb | pbrook | } |
1729 | 9ee6e8bb | pbrook | |
1730 | 9ee6e8bb | pbrook | /* Helpers for element load/store. */
|
1731 | 9ee6e8bb | pbrook | NEON_OP(insert_elt) |
1732 | 9ee6e8bb | pbrook | { |
1733 | 9ee6e8bb | pbrook | int shift = PARAM1;
|
1734 | 9ee6e8bb | pbrook | uint32_t mask = PARAM2; |
1735 | 9ee6e8bb | pbrook | T2 = (T2 & mask) | (T0 << shift); |
1736 | 9ee6e8bb | pbrook | FORCE_RET(); |
1737 | 9ee6e8bb | pbrook | } |
1738 | 9ee6e8bb | pbrook | |
1739 | 9ee6e8bb | pbrook | NEON_OP(extract_elt) |
1740 | 9ee6e8bb | pbrook | { |
1741 | 9ee6e8bb | pbrook | int shift = PARAM1;
|
1742 | 9ee6e8bb | pbrook | uint32_t mask = PARAM2; |
1743 | 9ee6e8bb | pbrook | T0 = (T2 & mask) >> shift; |
1744 | 9ee6e8bb | pbrook | FORCE_RET(); |
1745 | 9ee6e8bb | pbrook | } |