root / target-arm / neon_helper.c @ 5d794885
History | View | Annotate | Download (34.2 kB)
1 |
/*
|
---|---|
2 |
* ARM NEON vector operations.
|
3 |
*
|
4 |
* Copyright (c) 2007, 2008 CodeSourcery.
|
5 |
* Written by Paul Brook
|
6 |
*
|
7 |
* This code is licenced under the GNU GPL v2.
|
8 |
*/
|
9 |
#include <stdlib.h> |
10 |
#include <stdio.h> |
11 |
|
12 |
#include "cpu.h" |
13 |
#include "exec-all.h" |
14 |
#include "helpers.h" |
15 |
|
16 |
#define SIGNBIT (uint32_t)0x80000000 |
17 |
#define SIGNBIT64 ((uint64_t)1 << 63) |
18 |
|
19 |
#define SET_QC() env->vfp.xregs[ARM_VFP_FPSCR] = CPSR_Q
|
20 |
|
21 |
static float_status neon_float_status;
|
22 |
#define NFS &neon_float_status
|
23 |
|
24 |
/* Helper routines to perform bitwise copies between float and int. */
|
25 |
static inline float32 vfp_itos(uint32_t i) |
26 |
{ |
27 |
union {
|
28 |
uint32_t i; |
29 |
float32 s; |
30 |
} v; |
31 |
|
32 |
v.i = i; |
33 |
return v.s;
|
34 |
} |
35 |
|
36 |
static inline uint32_t vfp_stoi(float32 s) |
37 |
{ |
38 |
union {
|
39 |
uint32_t i; |
40 |
float32 s; |
41 |
} v; |
42 |
|
43 |
v.s = s; |
44 |
return v.i;
|
45 |
} |
46 |
|
47 |
#define NEON_TYPE1(name, type) \
|
48 |
typedef struct \ |
49 |
{ \ |
50 |
type v1; \ |
51 |
} neon_##name; |
52 |
#ifdef WORDS_BIGENDIAN
|
53 |
#define NEON_TYPE2(name, type) \
|
54 |
typedef struct \ |
55 |
{ \ |
56 |
type v2; \ |
57 |
type v1; \ |
58 |
} neon_##name; |
59 |
#define NEON_TYPE4(name, type) \
|
60 |
typedef struct \ |
61 |
{ \ |
62 |
type v4; \ |
63 |
type v3; \ |
64 |
type v2; \ |
65 |
type v1; \ |
66 |
} neon_##name; |
67 |
#else
|
68 |
#define NEON_TYPE2(name, type) \
|
69 |
typedef struct \ |
70 |
{ \ |
71 |
type v1; \ |
72 |
type v2; \ |
73 |
} neon_##name; |
74 |
#define NEON_TYPE4(name, type) \
|
75 |
typedef struct \ |
76 |
{ \ |
77 |
type v1; \ |
78 |
type v2; \ |
79 |
type v3; \ |
80 |
type v4; \ |
81 |
} neon_##name; |
82 |
#endif
|
83 |
|
84 |
NEON_TYPE4(s8, int8_t) |
85 |
NEON_TYPE4(u8, uint8_t) |
86 |
NEON_TYPE2(s16, int16_t) |
87 |
NEON_TYPE2(u16, uint16_t) |
88 |
NEON_TYPE1(s32, int32_t) |
89 |
NEON_TYPE1(u32, uint32_t) |
90 |
#undef NEON_TYPE4
|
91 |
#undef NEON_TYPE2
|
92 |
#undef NEON_TYPE1
|
93 |
|
94 |
/* Copy from a uint32_t to a vector structure type. */
|
95 |
#define NEON_UNPACK(vtype, dest, val) do { \ |
96 |
union { \
|
97 |
vtype v; \ |
98 |
uint32_t i; \ |
99 |
} conv_u; \ |
100 |
conv_u.i = (val); \ |
101 |
dest = conv_u.v; \ |
102 |
} while(0) |
103 |
|
104 |
/* Copy from a vector structure type to a uint32_t. */
|
105 |
#define NEON_PACK(vtype, dest, val) do { \ |
106 |
union { \
|
107 |
vtype v; \ |
108 |
uint32_t i; \ |
109 |
} conv_u; \ |
110 |
conv_u.v = (val); \ |
111 |
dest = conv_u.i; \ |
112 |
} while(0) |
113 |
|
114 |
#define NEON_DO1 \
|
115 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); |
116 |
#define NEON_DO2 \
|
117 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
118 |
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); |
119 |
#define NEON_DO4 \
|
120 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
121 |
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); \ |
122 |
NEON_FN(vdest.v3, vsrc1.v3, vsrc2.v3); \ |
123 |
NEON_FN(vdest.v4, vsrc1.v4, vsrc2.v4); |
124 |
|
125 |
#define NEON_VOP_BODY(vtype, n) \
|
126 |
{ \ |
127 |
uint32_t res; \ |
128 |
vtype vsrc1; \ |
129 |
vtype vsrc2; \ |
130 |
vtype vdest; \ |
131 |
NEON_UNPACK(vtype, vsrc1, arg1); \ |
132 |
NEON_UNPACK(vtype, vsrc2, arg2); \ |
133 |
NEON_DO##n; \ |
134 |
NEON_PACK(vtype, res, vdest); \ |
135 |
return res; \
|
136 |
} |
137 |
|
138 |
#define NEON_VOP(name, vtype, n) \
|
139 |
uint32_t HELPER(glue(neon_,name))(uint32_t arg1, uint32_t arg2) \ |
140 |
NEON_VOP_BODY(vtype, n) |
141 |
|
142 |
#define NEON_VOP_ENV(name, vtype, n) \
|
143 |
uint32_t HELPER(glue(neon_,name))(CPUState *env, uint32_t arg1, uint32_t arg2) \ |
144 |
NEON_VOP_BODY(vtype, n) |
145 |
|
146 |
/* Pairwise operations. */
|
147 |
/* For 32-bit elements each segment only contains a single element, so
|
148 |
the elementwise and pairwise operations are the same. */
|
149 |
#define NEON_PDO2 \
|
150 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
151 |
NEON_FN(vdest.v2, vsrc2.v1, vsrc2.v2); |
152 |
#define NEON_PDO4 \
|
153 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
154 |
NEON_FN(vdest.v2, vsrc1.v3, vsrc1.v4); \ |
155 |
NEON_FN(vdest.v3, vsrc2.v1, vsrc2.v2); \ |
156 |
NEON_FN(vdest.v4, vsrc2.v3, vsrc2.v4); \ |
157 |
|
158 |
#define NEON_POP(name, vtype, n) \
|
159 |
uint32_t HELPER(glue(neon_,name))(uint32_t arg1, uint32_t arg2) \ |
160 |
{ \ |
161 |
uint32_t res; \ |
162 |
vtype vsrc1; \ |
163 |
vtype vsrc2; \ |
164 |
vtype vdest; \ |
165 |
NEON_UNPACK(vtype, vsrc1, arg1); \ |
166 |
NEON_UNPACK(vtype, vsrc2, arg2); \ |
167 |
NEON_PDO##n; \ |
168 |
NEON_PACK(vtype, res, vdest); \ |
169 |
return res; \
|
170 |
} |
171 |
|
172 |
/* Unary operators. */
|
173 |
#define NEON_VOP1(name, vtype, n) \
|
174 |
uint32_t HELPER(glue(neon_,name))(uint32_t arg) \ |
175 |
{ \ |
176 |
vtype vsrc1; \ |
177 |
vtype vdest; \ |
178 |
NEON_UNPACK(vtype, vsrc1, arg); \ |
179 |
NEON_DO##n; \ |
180 |
NEON_PACK(vtype, arg, vdest); \ |
181 |
return arg; \
|
182 |
} |
183 |
|
184 |
|
185 |
#define NEON_USAT(dest, src1, src2, type) do { \ |
186 |
uint32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
187 |
if (tmp != (type)tmp) { \
|
188 |
SET_QC(); \ |
189 |
dest = ~0; \
|
190 |
} else { \
|
191 |
dest = tmp; \ |
192 |
}} while(0) |
193 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
194 |
NEON_VOP_ENV(qadd_u8, neon_u8, 4)
|
195 |
#undef NEON_FN
|
196 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
197 |
NEON_VOP_ENV(qadd_u16, neon_u16, 2)
|
198 |
#undef NEON_FN
|
199 |
#undef NEON_USAT
|
200 |
|
201 |
#define NEON_SSAT(dest, src1, src2, type) do { \ |
202 |
int32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
203 |
if (tmp != (type)tmp) { \
|
204 |
SET_QC(); \ |
205 |
if (src2 > 0) { \ |
206 |
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
207 |
} else { \
|
208 |
tmp = 1 << (sizeof(type) * 8 - 1); \ |
209 |
} \ |
210 |
} \ |
211 |
dest = tmp; \ |
212 |
} while(0) |
213 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
214 |
NEON_VOP_ENV(qadd_s8, neon_s8, 4)
|
215 |
#undef NEON_FN
|
216 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
217 |
NEON_VOP_ENV(qadd_s16, neon_s16, 2)
|
218 |
#undef NEON_FN
|
219 |
#undef NEON_SSAT
|
220 |
|
221 |
#define NEON_USAT(dest, src1, src2, type) do { \ |
222 |
uint32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
223 |
if (tmp != (type)tmp) { \
|
224 |
SET_QC(); \ |
225 |
dest = 0; \
|
226 |
} else { \
|
227 |
dest = tmp; \ |
228 |
}} while(0) |
229 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
230 |
NEON_VOP_ENV(qsub_u8, neon_u8, 4)
|
231 |
#undef NEON_FN
|
232 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
233 |
NEON_VOP_ENV(qsub_u16, neon_u16, 2)
|
234 |
#undef NEON_FN
|
235 |
#undef NEON_USAT
|
236 |
|
237 |
#define NEON_SSAT(dest, src1, src2, type) do { \ |
238 |
int32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
239 |
if (tmp != (type)tmp) { \
|
240 |
SET_QC(); \ |
241 |
if (src2 < 0) { \ |
242 |
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
243 |
} else { \
|
244 |
tmp = 1 << (sizeof(type) * 8 - 1); \ |
245 |
} \ |
246 |
} \ |
247 |
dest = tmp; \ |
248 |
} while(0) |
249 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
250 |
NEON_VOP_ENV(qsub_s8, neon_s8, 4)
|
251 |
#undef NEON_FN
|
252 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
253 |
NEON_VOP_ENV(qsub_s16, neon_s16, 2)
|
254 |
#undef NEON_FN
|
255 |
#undef NEON_SSAT
|
256 |
|
257 |
#define NEON_FN(dest, src1, src2) dest = (src1 + src2) >> 1 |
258 |
NEON_VOP(hadd_s8, neon_s8, 4)
|
259 |
NEON_VOP(hadd_u8, neon_u8, 4)
|
260 |
NEON_VOP(hadd_s16, neon_s16, 2)
|
261 |
NEON_VOP(hadd_u16, neon_u16, 2)
|
262 |
#undef NEON_FN
|
263 |
|
264 |
int32_t HELPER(neon_hadd_s32)(int32_t src1, int32_t src2) |
265 |
{ |
266 |
int32_t dest; |
267 |
|
268 |
dest = (src1 >> 1) + (src2 >> 1); |
269 |
if (src1 & src2 & 1) |
270 |
dest++; |
271 |
return dest;
|
272 |
} |
273 |
|
274 |
uint32_t HELPER(neon_hadd_u32)(uint32_t src1, uint32_t src2) |
275 |
{ |
276 |
uint32_t dest; |
277 |
|
278 |
dest = (src1 >> 1) + (src2 >> 1); |
279 |
if (src1 & src2 & 1) |
280 |
dest++; |
281 |
return dest;
|
282 |
} |
283 |
|
284 |
#define NEON_FN(dest, src1, src2) dest = (src1 + src2 + 1) >> 1 |
285 |
NEON_VOP(rhadd_s8, neon_s8, 4)
|
286 |
NEON_VOP(rhadd_u8, neon_u8, 4)
|
287 |
NEON_VOP(rhadd_s16, neon_s16, 2)
|
288 |
NEON_VOP(rhadd_u16, neon_u16, 2)
|
289 |
#undef NEON_FN
|
290 |
|
291 |
int32_t HELPER(neon_rhadd_s32)(int32_t src1, int32_t src2) |
292 |
{ |
293 |
int32_t dest; |
294 |
|
295 |
dest = (src1 >> 1) + (src2 >> 1); |
296 |
if ((src1 | src2) & 1) |
297 |
dest++; |
298 |
return dest;
|
299 |
} |
300 |
|
301 |
uint32_t HELPER(neon_rhadd_u32)(uint32_t src1, uint32_t src2) |
302 |
{ |
303 |
uint32_t dest; |
304 |
|
305 |
dest = (src1 >> 1) + (src2 >> 1); |
306 |
if ((src1 | src2) & 1) |
307 |
dest++; |
308 |
return dest;
|
309 |
} |
310 |
|
311 |
#define NEON_FN(dest, src1, src2) dest = (src1 - src2) >> 1 |
312 |
NEON_VOP(hsub_s8, neon_s8, 4)
|
313 |
NEON_VOP(hsub_u8, neon_u8, 4)
|
314 |
NEON_VOP(hsub_s16, neon_s16, 2)
|
315 |
NEON_VOP(hsub_u16, neon_u16, 2)
|
316 |
#undef NEON_FN
|
317 |
|
318 |
int32_t HELPER(neon_hsub_s32)(int32_t src1, int32_t src2) |
319 |
{ |
320 |
int32_t dest; |
321 |
|
322 |
dest = (src1 >> 1) - (src2 >> 1); |
323 |
if ((~src1) & src2 & 1) |
324 |
dest--; |
325 |
return dest;
|
326 |
} |
327 |
|
328 |
uint32_t HELPER(neon_hsub_u32)(uint32_t src1, uint32_t src2) |
329 |
{ |
330 |
uint32_t dest; |
331 |
|
332 |
dest = (src1 >> 1) - (src2 >> 1); |
333 |
if ((~src1) & src2 & 1) |
334 |
dest--; |
335 |
return dest;
|
336 |
} |
337 |
|
338 |
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? ~0 : 0 |
339 |
NEON_VOP(cgt_s8, neon_s8, 4)
|
340 |
NEON_VOP(cgt_u8, neon_u8, 4)
|
341 |
NEON_VOP(cgt_s16, neon_s16, 2)
|
342 |
NEON_VOP(cgt_u16, neon_u16, 2)
|
343 |
NEON_VOP(cgt_s32, neon_s32, 1)
|
344 |
NEON_VOP(cgt_u32, neon_u32, 1)
|
345 |
#undef NEON_FN
|
346 |
|
347 |
#define NEON_FN(dest, src1, src2) dest = (src1 >= src2) ? ~0 : 0 |
348 |
NEON_VOP(cge_s8, neon_s8, 4)
|
349 |
NEON_VOP(cge_u8, neon_u8, 4)
|
350 |
NEON_VOP(cge_s16, neon_s16, 2)
|
351 |
NEON_VOP(cge_u16, neon_u16, 2)
|
352 |
NEON_VOP(cge_s32, neon_s32, 1)
|
353 |
NEON_VOP(cge_u32, neon_u32, 1)
|
354 |
#undef NEON_FN
|
355 |
|
356 |
#define NEON_FN(dest, src1, src2) dest = (src1 < src2) ? src1 : src2
|
357 |
NEON_VOP(min_s8, neon_s8, 4)
|
358 |
NEON_VOP(min_u8, neon_u8, 4)
|
359 |
NEON_VOP(min_s16, neon_s16, 2)
|
360 |
NEON_VOP(min_u16, neon_u16, 2)
|
361 |
NEON_VOP(min_s32, neon_s32, 1)
|
362 |
NEON_VOP(min_u32, neon_u32, 1)
|
363 |
NEON_POP(pmin_s8, neon_s8, 4)
|
364 |
NEON_POP(pmin_u8, neon_u8, 4)
|
365 |
NEON_POP(pmin_s16, neon_s16, 2)
|
366 |
NEON_POP(pmin_u16, neon_u16, 2)
|
367 |
#undef NEON_FN
|
368 |
|
369 |
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? src1 : src2
|
370 |
NEON_VOP(max_s8, neon_s8, 4)
|
371 |
NEON_VOP(max_u8, neon_u8, 4)
|
372 |
NEON_VOP(max_s16, neon_s16, 2)
|
373 |
NEON_VOP(max_u16, neon_u16, 2)
|
374 |
NEON_VOP(max_s32, neon_s32, 1)
|
375 |
NEON_VOP(max_u32, neon_u32, 1)
|
376 |
NEON_POP(pmax_s8, neon_s8, 4)
|
377 |
NEON_POP(pmax_u8, neon_u8, 4)
|
378 |
NEON_POP(pmax_s16, neon_s16, 2)
|
379 |
NEON_POP(pmax_u16, neon_u16, 2)
|
380 |
#undef NEON_FN
|
381 |
|
382 |
#define NEON_FN(dest, src1, src2) \
|
383 |
dest = (src1 > src2) ? (src1 - src2) : (src2 - src1) |
384 |
NEON_VOP(abd_s8, neon_s8, 4)
|
385 |
NEON_VOP(abd_u8, neon_u8, 4)
|
386 |
NEON_VOP(abd_s16, neon_s16, 2)
|
387 |
NEON_VOP(abd_u16, neon_u16, 2)
|
388 |
NEON_VOP(abd_s32, neon_s32, 1)
|
389 |
NEON_VOP(abd_u32, neon_u32, 1)
|
390 |
#undef NEON_FN
|
391 |
|
392 |
#define NEON_FN(dest, src1, src2) do { \ |
393 |
int8_t tmp; \ |
394 |
tmp = (int8_t)src2; \ |
395 |
if (tmp >= sizeof(src1) * 8 || tmp <= -sizeof(src1) * 8) { \ |
396 |
dest = 0; \
|
397 |
} else if (tmp < 0) { \ |
398 |
dest = src1 >> -tmp; \ |
399 |
} else { \
|
400 |
dest = src1 << tmp; \ |
401 |
}} while (0) |
402 |
NEON_VOP(shl_u8, neon_u8, 4)
|
403 |
NEON_VOP(shl_u16, neon_u16, 2)
|
404 |
NEON_VOP(shl_u32, neon_u32, 1)
|
405 |
#undef NEON_FN
|
406 |
|
407 |
uint64_t HELPER(neon_shl_u64)(uint64_t val, uint64_t shiftop) |
408 |
{ |
409 |
int8_t shift = (int8_t)shiftop; |
410 |
if (shift >= 64 || shift <= -64) { |
411 |
val = 0;
|
412 |
} else if (shift < 0) { |
413 |
val >>= -shift; |
414 |
} else {
|
415 |
val <<= shift; |
416 |
} |
417 |
return val;
|
418 |
} |
419 |
|
420 |
#define NEON_FN(dest, src1, src2) do { \ |
421 |
int8_t tmp; \ |
422 |
tmp = (int8_t)src2; \ |
423 |
if (tmp >= sizeof(src1) * 8) { \ |
424 |
dest = 0; \
|
425 |
} else if (tmp <= -sizeof(src1) * 8) { \ |
426 |
dest = src1 >> (sizeof(src1) * 8 - 1); \ |
427 |
} else if (tmp < 0) { \ |
428 |
dest = src1 >> -tmp; \ |
429 |
} else { \
|
430 |
dest = src1 << tmp; \ |
431 |
}} while (0) |
432 |
NEON_VOP(shl_s8, neon_s8, 4)
|
433 |
NEON_VOP(shl_s16, neon_s16, 2)
|
434 |
NEON_VOP(shl_s32, neon_s32, 1)
|
435 |
#undef NEON_FN
|
436 |
|
437 |
uint64_t HELPER(neon_shl_s64)(uint64_t valop, uint64_t shiftop) |
438 |
{ |
439 |
int8_t shift = (int8_t)shiftop; |
440 |
int64_t val = valop; |
441 |
if (shift >= 64) { |
442 |
val = 0;
|
443 |
} else if (shift <= -64) { |
444 |
val >>= 63;
|
445 |
} else if (shift < 0) { |
446 |
val >>= -shift; |
447 |
} else {
|
448 |
val <<= shift; |
449 |
} |
450 |
return val;
|
451 |
} |
452 |
|
453 |
#define NEON_FN(dest, src1, src2) do { \ |
454 |
int8_t tmp; \ |
455 |
tmp = (int8_t)src2; \ |
456 |
if (tmp >= sizeof(src1) * 8) { \ |
457 |
dest = 0; \
|
458 |
} else if (tmp < -sizeof(src1) * 8) { \ |
459 |
dest >>= sizeof(src1) * 8 - 1; \ |
460 |
} else if (tmp == -sizeof(src1) * 8) { \ |
461 |
dest = src1 >> (tmp - 1); \
|
462 |
dest++; \ |
463 |
src2 >>= 1; \
|
464 |
} else if (tmp < 0) { \ |
465 |
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \ |
466 |
} else { \
|
467 |
dest = src1 << tmp; \ |
468 |
}} while (0) |
469 |
NEON_VOP(rshl_s8, neon_s8, 4)
|
470 |
NEON_VOP(rshl_s16, neon_s16, 2)
|
471 |
NEON_VOP(rshl_s32, neon_s32, 1)
|
472 |
#undef NEON_FN
|
473 |
|
474 |
uint64_t HELPER(neon_rshl_s64)(uint64_t valop, uint64_t shiftop) |
475 |
{ |
476 |
int8_t shift = (int8_t)shiftop; |
477 |
int64_t val = valop; |
478 |
if (shift >= 64) { |
479 |
val = 0;
|
480 |
} else if (shift < -64) { |
481 |
val >>= 63;
|
482 |
} else if (shift == -63) { |
483 |
val >>= 63;
|
484 |
val++; |
485 |
val >>= 1;
|
486 |
} else if (shift < 0) { |
487 |
val = (val + ((int64_t)1 << (-1 - shift))) >> -shift; |
488 |
} else {
|
489 |
val <<= shift; |
490 |
} |
491 |
return val;
|
492 |
} |
493 |
|
494 |
#define NEON_FN(dest, src1, src2) do { \ |
495 |
int8_t tmp; \ |
496 |
tmp = (int8_t)src2; \ |
497 |
if (tmp >= sizeof(src1) * 8 || tmp < -sizeof(src1) * 8) { \ |
498 |
dest = 0; \
|
499 |
} else if (tmp == -sizeof(src1) * 8) { \ |
500 |
dest = src1 >> (tmp - 1); \
|
501 |
} else if (tmp < 0) { \ |
502 |
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \ |
503 |
} else { \
|
504 |
dest = src1 << tmp; \ |
505 |
}} while (0) |
506 |
NEON_VOP(rshl_u8, neon_u8, 4)
|
507 |
NEON_VOP(rshl_u16, neon_u16, 2)
|
508 |
NEON_VOP(rshl_u32, neon_u32, 1)
|
509 |
#undef NEON_FN
|
510 |
|
511 |
uint64_t HELPER(neon_rshl_u64)(uint64_t val, uint64_t shiftop) |
512 |
{ |
513 |
int8_t shift = (uint8_t)shiftop; |
514 |
if (shift >= 64 || shift < 64) { |
515 |
val = 0;
|
516 |
} else if (shift == -64) { |
517 |
/* Rounding a 1-bit result just preserves that bit. */
|
518 |
val >>= 63;
|
519 |
} if (shift < 0) { |
520 |
val = (val + ((uint64_t)1 << (-1 - shift))) >> -shift; |
521 |
val >>= -shift; |
522 |
} else {
|
523 |
val <<= shift; |
524 |
} |
525 |
return val;
|
526 |
} |
527 |
|
528 |
#define NEON_FN(dest, src1, src2) do { \ |
529 |
int8_t tmp; \ |
530 |
tmp = (int8_t)src2; \ |
531 |
if (tmp >= sizeof(src1) * 8) { \ |
532 |
if (src1) { \
|
533 |
SET_QC(); \ |
534 |
dest = ~0; \
|
535 |
} else { \
|
536 |
dest = 0; \
|
537 |
} \ |
538 |
} else if (tmp <= -sizeof(src1) * 8) { \ |
539 |
dest = 0; \
|
540 |
} else if (tmp < 0) { \ |
541 |
dest = src1 >> -tmp; \ |
542 |
} else { \
|
543 |
dest = src1 << tmp; \ |
544 |
if ((dest >> tmp) != src1) { \
|
545 |
SET_QC(); \ |
546 |
dest = ~0; \
|
547 |
} \ |
548 |
}} while (0) |
549 |
NEON_VOP_ENV(qshl_u8, neon_u8, 4)
|
550 |
NEON_VOP_ENV(qshl_u16, neon_u16, 2)
|
551 |
NEON_VOP_ENV(qshl_u32, neon_u32, 1)
|
552 |
#undef NEON_FN
|
553 |
|
554 |
uint64_t HELPER(neon_qshl_u64)(CPUState *env, uint64_t val, uint64_t shiftop) |
555 |
{ |
556 |
int8_t shift = (int8_t)shiftop; |
557 |
if (shift >= 64) { |
558 |
if (val) {
|
559 |
val = ~(uint64_t)0;
|
560 |
SET_QC(); |
561 |
} else {
|
562 |
val = 0;
|
563 |
} |
564 |
} else if (shift <= -64) { |
565 |
val = 0;
|
566 |
} else if (shift < 0) { |
567 |
val >>= -shift; |
568 |
} else {
|
569 |
uint64_t tmp = val; |
570 |
val <<= shift; |
571 |
if ((val >> shift) != tmp) {
|
572 |
SET_QC(); |
573 |
val = ~(uint64_t)0;
|
574 |
} |
575 |
} |
576 |
return val;
|
577 |
} |
578 |
|
579 |
#define NEON_FN(dest, src1, src2) do { \ |
580 |
int8_t tmp; \ |
581 |
tmp = (int8_t)src2; \ |
582 |
if (tmp >= sizeof(src1) * 8) { \ |
583 |
if (src1) \
|
584 |
SET_QC(); \ |
585 |
dest = src1 >> 31; \
|
586 |
} else if (tmp <= -sizeof(src1) * 8) { \ |
587 |
dest = src1 >> 31; \
|
588 |
} else if (tmp < 0) { \ |
589 |
dest = src1 >> -tmp; \ |
590 |
} else { \
|
591 |
dest = src1 << tmp; \ |
592 |
if ((dest >> tmp) != src1) { \
|
593 |
SET_QC(); \ |
594 |
dest = src2 >> 31; \
|
595 |
} \ |
596 |
}} while (0) |
597 |
NEON_VOP_ENV(qshl_s8, neon_s8, 4)
|
598 |
NEON_VOP_ENV(qshl_s16, neon_s16, 2)
|
599 |
NEON_VOP_ENV(qshl_s32, neon_s32, 1)
|
600 |
#undef NEON_FN
|
601 |
|
602 |
uint64_t HELPER(neon_qshl_s64)(CPUState *env, uint64_t valop, uint64_t shiftop) |
603 |
{ |
604 |
int8_t shift = (uint8_t)shiftop; |
605 |
int64_t val = valop; |
606 |
if (shift >= 64) { |
607 |
if (val) {
|
608 |
SET_QC(); |
609 |
val = (val >> 63) & ~SIGNBIT64;
|
610 |
} |
611 |
} else if (shift <= 64) { |
612 |
val >>= 63;
|
613 |
} else if (shift < 0) { |
614 |
val >>= -shift; |
615 |
} else {
|
616 |
int64_t tmp = val; |
617 |
val <<= shift; |
618 |
if ((val >> shift) != tmp) {
|
619 |
SET_QC(); |
620 |
val = (tmp >> 63) ^ ~SIGNBIT64;
|
621 |
} |
622 |
} |
623 |
return val;
|
624 |
} |
625 |
|
626 |
|
627 |
/* FIXME: This is wrong. */
|
628 |
#define NEON_FN(dest, src1, src2) do { \ |
629 |
int8_t tmp; \ |
630 |
tmp = (int8_t)src2; \ |
631 |
if (tmp < 0) { \ |
632 |
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \ |
633 |
} else { \
|
634 |
dest = src1 << tmp; \ |
635 |
if ((dest >> tmp) != src1) { \
|
636 |
SET_QC(); \ |
637 |
dest = ~0; \
|
638 |
} \ |
639 |
}} while (0) |
640 |
NEON_VOP_ENV(qrshl_u8, neon_u8, 4)
|
641 |
NEON_VOP_ENV(qrshl_u16, neon_u16, 2)
|
642 |
NEON_VOP_ENV(qrshl_u32, neon_u32, 1)
|
643 |
#undef NEON_FN
|
644 |
|
645 |
uint64_t HELPER(neon_qrshl_u64)(CPUState *env, uint64_t val, uint64_t shiftop) |
646 |
{ |
647 |
int8_t shift = (int8_t)shiftop; |
648 |
if (shift < 0) { |
649 |
val = (val + (1 << (-1 - shift))) >> -shift; |
650 |
} else { \
|
651 |
uint64_t tmp = val; |
652 |
val <<= shift; |
653 |
if ((val >> shift) != tmp) {
|
654 |
SET_QC(); |
655 |
val = ~0;
|
656 |
} |
657 |
} |
658 |
return val;
|
659 |
} |
660 |
|
661 |
#define NEON_FN(dest, src1, src2) do { \ |
662 |
int8_t tmp; \ |
663 |
tmp = (int8_t)src2; \ |
664 |
if (tmp < 0) { \ |
665 |
dest = (src1 + (1 << (-1 - tmp))) >> -tmp; \ |
666 |
} else { \
|
667 |
dest = src1 << tmp; \ |
668 |
if ((dest >> tmp) != src1) { \
|
669 |
SET_QC(); \ |
670 |
dest = src1 >> 31; \
|
671 |
} \ |
672 |
}} while (0) |
673 |
NEON_VOP_ENV(qrshl_s8, neon_s8, 4)
|
674 |
NEON_VOP_ENV(qrshl_s16, neon_s16, 2)
|
675 |
NEON_VOP_ENV(qrshl_s32, neon_s32, 1)
|
676 |
#undef NEON_FN
|
677 |
|
678 |
uint64_t HELPER(neon_qrshl_s64)(CPUState *env, uint64_t valop, uint64_t shiftop) |
679 |
{ |
680 |
int8_t shift = (uint8_t)shiftop; |
681 |
int64_t val = valop; |
682 |
|
683 |
if (shift < 0) { |
684 |
val = (val + (1 << (-1 - shift))) >> -shift; |
685 |
} else {
|
686 |
int64_t tmp = val;; |
687 |
val <<= shift; |
688 |
if ((val >> shift) != tmp) {
|
689 |
SET_QC(); |
690 |
val = tmp >> 31;
|
691 |
} |
692 |
} |
693 |
return val;
|
694 |
} |
695 |
|
696 |
uint32_t HELPER(neon_add_u8)(uint32_t a, uint32_t b) |
697 |
{ |
698 |
uint32_t mask; |
699 |
mask = (a ^ b) & 0x80808080u;
|
700 |
a &= ~0x80808080u;
|
701 |
b &= ~0x80808080u;
|
702 |
return (a + b) ^ mask;
|
703 |
} |
704 |
|
705 |
uint32_t HELPER(neon_add_u16)(uint32_t a, uint32_t b) |
706 |
{ |
707 |
uint32_t mask; |
708 |
mask = (a ^ b) & 0x80008000u;
|
709 |
a &= ~0x80008000u;
|
710 |
b &= ~0x80008000u;
|
711 |
return (a + b) ^ mask;
|
712 |
} |
713 |
|
714 |
#define NEON_FN(dest, src1, src2) dest = src1 + src2
|
715 |
NEON_POP(padd_u8, neon_u8, 4)
|
716 |
NEON_POP(padd_u16, neon_u16, 2)
|
717 |
#undef NEON_FN
|
718 |
|
719 |
#define NEON_FN(dest, src1, src2) dest = src1 - src2
|
720 |
NEON_VOP(sub_u8, neon_u8, 4)
|
721 |
NEON_VOP(sub_u16, neon_u16, 2)
|
722 |
#undef NEON_FN
|
723 |
|
724 |
#define NEON_FN(dest, src1, src2) dest = src1 * src2
|
725 |
NEON_VOP(mul_u8, neon_u8, 4)
|
726 |
NEON_VOP(mul_u16, neon_u16, 2)
|
727 |
#undef NEON_FN
|
728 |
|
729 |
/* Polynomial multiplication is like integer multiplication except the
|
730 |
partial products are XORed, not added. */
|
731 |
uint32_t HELPER(neon_mul_p8)(uint32_t op1, uint32_t op2) |
732 |
{ |
733 |
uint32_t mask; |
734 |
uint32_t result; |
735 |
result = 0;
|
736 |
while (op1) {
|
737 |
mask = 0;
|
738 |
if (op1 & 1) |
739 |
mask |= 0xff;
|
740 |
if (op1 & (1 << 8)) |
741 |
mask |= (0xff << 8); |
742 |
if (op1 & (1 << 16)) |
743 |
mask |= (0xff << 16); |
744 |
if (op1 & (1 << 24)) |
745 |
mask |= (0xff << 24); |
746 |
result ^= op2 & mask; |
747 |
op1 = (op1 >> 1) & 0x7f7f7f7f; |
748 |
op2 = (op2 << 1) & 0xfefefefe; |
749 |
} |
750 |
return result;
|
751 |
} |
752 |
|
753 |
#define NEON_FN(dest, src1, src2) dest = (src1 & src2) ? -1 : 0 |
754 |
NEON_VOP(tst_u8, neon_u8, 4)
|
755 |
NEON_VOP(tst_u16, neon_u16, 2)
|
756 |
NEON_VOP(tst_u32, neon_u32, 1)
|
757 |
#undef NEON_FN
|
758 |
|
759 |
#define NEON_FN(dest, src1, src2) dest = (src1 == src2) ? -1 : 0 |
760 |
NEON_VOP(ceq_u8, neon_u8, 4)
|
761 |
NEON_VOP(ceq_u16, neon_u16, 2)
|
762 |
NEON_VOP(ceq_u32, neon_u32, 1)
|
763 |
#undef NEON_FN
|
764 |
|
765 |
#define NEON_FN(dest, src, dummy) dest = (src < 0) ? -src : src |
766 |
NEON_VOP1(abs_s8, neon_s8, 4)
|
767 |
NEON_VOP1(abs_s16, neon_s16, 2)
|
768 |
#undef NEON_FN
|
769 |
|
770 |
/* Count Leading Sign/Zero Bits. */
|
771 |
static inline int do_clz8(uint8_t x) |
772 |
{ |
773 |
int n;
|
774 |
for (n = 8; x; n--) |
775 |
x >>= 1;
|
776 |
return n;
|
777 |
} |
778 |
|
779 |
static inline int do_clz16(uint16_t x) |
780 |
{ |
781 |
int n;
|
782 |
for (n = 16; x; n--) |
783 |
x >>= 1;
|
784 |
return n;
|
785 |
} |
786 |
|
787 |
#define NEON_FN(dest, src, dummy) dest = do_clz8(src)
|
788 |
NEON_VOP1(clz_u8, neon_u8, 4)
|
789 |
#undef NEON_FN
|
790 |
|
791 |
#define NEON_FN(dest, src, dummy) dest = do_clz16(src)
|
792 |
NEON_VOP1(clz_u16, neon_u16, 2)
|
793 |
#undef NEON_FN
|
794 |
|
795 |
#define NEON_FN(dest, src, dummy) dest = do_clz8((src < 0) ? ~src : src) - 1 |
796 |
NEON_VOP1(cls_s8, neon_s8, 4)
|
797 |
#undef NEON_FN
|
798 |
|
799 |
#define NEON_FN(dest, src, dummy) dest = do_clz16((src < 0) ? ~src : src) - 1 |
800 |
NEON_VOP1(cls_s16, neon_s16, 2)
|
801 |
#undef NEON_FN
|
802 |
|
803 |
uint32_t HELPER(neon_cls_s32)(uint32_t x) |
804 |
{ |
805 |
int count;
|
806 |
if ((int32_t)x < 0) |
807 |
x = ~x; |
808 |
for (count = 32; x; count--) |
809 |
x = x >> 1;
|
810 |
return count - 1; |
811 |
} |
812 |
|
813 |
/* Bit count. */
|
814 |
uint32_t HELPER(neon_cnt_u8)(uint32_t x) |
815 |
{ |
816 |
x = (x & 0x55555555) + ((x >> 1) & 0x55555555); |
817 |
x = (x & 0x33333333) + ((x >> 2) & 0x33333333); |
818 |
x = (x & 0x0f0f0f0f) + ((x >> 4) & 0x0f0f0f0f); |
819 |
return x;
|
820 |
} |
821 |
|
822 |
#define NEON_QDMULH16(dest, src1, src2, round) do { \ |
823 |
uint32_t tmp = (int32_t)(int16_t) src1 * (int16_t) src2; \ |
824 |
if ((tmp ^ (tmp << 1)) & SIGNBIT) { \ |
825 |
SET_QC(); \ |
826 |
tmp = (tmp >> 31) ^ ~SIGNBIT; \
|
827 |
} \ |
828 |
tmp <<= 1; \
|
829 |
if (round) { \
|
830 |
int32_t old = tmp; \ |
831 |
tmp += 1 << 15; \ |
832 |
if ((int32_t)tmp < old) { \
|
833 |
SET_QC(); \ |
834 |
tmp = SIGNBIT - 1; \
|
835 |
} \ |
836 |
} \ |
837 |
dest = tmp >> 16; \
|
838 |
} while(0) |
839 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 0) |
840 |
NEON_VOP_ENV(qdmulh_s16, neon_s16, 2)
|
841 |
#undef NEON_FN
|
842 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 1) |
843 |
NEON_VOP_ENV(qrdmulh_s16, neon_s16, 2)
|
844 |
#undef NEON_FN
|
845 |
#undef NEON_QDMULH16
|
846 |
|
847 |
#define NEON_QDMULH32(dest, src1, src2, round) do { \ |
848 |
uint64_t tmp = (int64_t)(int32_t) src1 * (int32_t) src2; \ |
849 |
if ((tmp ^ (tmp << 1)) & SIGNBIT64) { \ |
850 |
SET_QC(); \ |
851 |
tmp = (tmp >> 63) ^ ~SIGNBIT64; \
|
852 |
} else { \
|
853 |
tmp <<= 1; \
|
854 |
} \ |
855 |
if (round) { \
|
856 |
int64_t old = tmp; \ |
857 |
tmp += (int64_t)1 << 31; \ |
858 |
if ((int64_t)tmp < old) { \
|
859 |
SET_QC(); \ |
860 |
tmp = SIGNBIT64 - 1; \
|
861 |
} \ |
862 |
} \ |
863 |
dest = tmp >> 32; \
|
864 |
} while(0) |
865 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 0) |
866 |
NEON_VOP_ENV(qdmulh_s32, neon_s32, 1)
|
867 |
#undef NEON_FN
|
868 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 1) |
869 |
NEON_VOP_ENV(qrdmulh_s32, neon_s32, 1)
|
870 |
#undef NEON_FN
|
871 |
#undef NEON_QDMULH32
|
872 |
|
873 |
uint32_t HELPER(neon_narrow_u8)(uint64_t x) |
874 |
{ |
875 |
return (x & 0xffu) | ((x >> 8) & 0xff00u) | ((x >> 16) & 0xff0000u) |
876 |
| ((x >> 24) & 0xff000000u); |
877 |
} |
878 |
|
879 |
uint32_t HELPER(neon_narrow_u16)(uint64_t x) |
880 |
{ |
881 |
return (x & 0xffffu) | ((x >> 16) & 0xffff0000u); |
882 |
} |
883 |
|
884 |
uint32_t HELPER(neon_narrow_high_u8)(uint64_t x) |
885 |
{ |
886 |
return ((x >> 8) & 0xff) | ((x >> 16) & 0xff00) |
887 |
| ((x >> 24) & 0xff0000) | ((x >> 32) & 0xff000000); |
888 |
} |
889 |
|
890 |
uint32_t HELPER(neon_narrow_high_u16)(uint64_t x) |
891 |
{ |
892 |
return ((x >> 16) & 0xffff) | ((x >> 32) & 0xffff0000); |
893 |
} |
894 |
|
895 |
uint32_t HELPER(neon_narrow_round_high_u8)(uint64_t x) |
896 |
{ |
897 |
x &= 0xff80ff80ff80ff80ull;
|
898 |
x += 0x0080008000800080ull;
|
899 |
return ((x >> 8) & 0xff) | ((x >> 16) & 0xff00) |
900 |
| ((x >> 24) & 0xff0000) | ((x >> 32) & 0xff000000); |
901 |
} |
902 |
|
903 |
uint32_t HELPER(neon_narrow_round_high_u16)(uint64_t x) |
904 |
{ |
905 |
x &= 0xffff8000ffff8000ull;
|
906 |
x += 0x0000800000008000ull;
|
907 |
return ((x >> 16) & 0xffff) | ((x >> 32) & 0xffff0000); |
908 |
} |
909 |
|
910 |
uint32_t HELPER(neon_narrow_sat_u8)(CPUState *env, uint64_t x) |
911 |
{ |
912 |
uint16_t s; |
913 |
uint8_t d; |
914 |
uint32_t res = 0;
|
915 |
#define SAT8(n) \
|
916 |
s = x >> n; \ |
917 |
if (s > 0xff) { \ |
918 |
d = 0xff; \
|
919 |
SET_QC(); \ |
920 |
} else { \
|
921 |
d = s; \ |
922 |
} \ |
923 |
res |= (uint32_t)d << (n / 2);
|
924 |
|
925 |
SAT8(0);
|
926 |
SAT8(16);
|
927 |
SAT8(32);
|
928 |
SAT8(48);
|
929 |
#undef SAT8
|
930 |
return res;
|
931 |
} |
932 |
|
933 |
uint32_t HELPER(neon_narrow_sat_s8)(CPUState *env, uint64_t x) |
934 |
{ |
935 |
int16_t s; |
936 |
uint8_t d; |
937 |
uint32_t res = 0;
|
938 |
#define SAT8(n) \
|
939 |
s = x >> n; \ |
940 |
if (s != (int8_t)s) { \
|
941 |
d = (s >> 15) ^ 0x7f; \ |
942 |
SET_QC(); \ |
943 |
} else { \
|
944 |
d = s; \ |
945 |
} \ |
946 |
res |= (uint32_t)d << (n / 2);
|
947 |
|
948 |
SAT8(0);
|
949 |
SAT8(16);
|
950 |
SAT8(32);
|
951 |
SAT8(48);
|
952 |
#undef SAT8
|
953 |
return res;
|
954 |
} |
955 |
|
956 |
uint32_t HELPER(neon_narrow_sat_u16)(CPUState *env, uint64_t x) |
957 |
{ |
958 |
uint32_t high; |
959 |
uint32_t low; |
960 |
low = x; |
961 |
if (low > 0xffff) { |
962 |
low = 0xffff;
|
963 |
SET_QC(); |
964 |
} |
965 |
high = x >> 32;
|
966 |
if (high > 0xffff) { |
967 |
high = 0xffff;
|
968 |
SET_QC(); |
969 |
} |
970 |
return low | (high << 16); |
971 |
} |
972 |
|
973 |
uint32_t HELPER(neon_narrow_sat_s16)(CPUState *env, uint64_t x) |
974 |
{ |
975 |
int32_t low; |
976 |
int32_t high; |
977 |
low = x; |
978 |
if (low != (int16_t)low) {
|
979 |
low = (low >> 31) ^ 0x7fff; |
980 |
SET_QC(); |
981 |
} |
982 |
high = x >> 32;
|
983 |
if (high != (int16_t)high) {
|
984 |
high = (high >> 31) ^ 0x7fff; |
985 |
SET_QC(); |
986 |
} |
987 |
return (uint16_t)low | (high << 16); |
988 |
} |
989 |
|
990 |
uint32_t HELPER(neon_narrow_sat_u32)(CPUState *env, uint64_t x) |
991 |
{ |
992 |
if (x > 0xffffffffu) { |
993 |
SET_QC(); |
994 |
return 0xffffffffu; |
995 |
} |
996 |
return x;
|
997 |
} |
998 |
|
999 |
uint32_t HELPER(neon_narrow_sat_s32)(CPUState *env, uint64_t x) |
1000 |
{ |
1001 |
if ((int64_t)x != (int32_t)x) {
|
1002 |
SET_QC(); |
1003 |
return (x >> 63) ^ 0x7fffffff; |
1004 |
} |
1005 |
return x;
|
1006 |
} |
1007 |
|
1008 |
uint64_t HELPER(neon_widen_u8)(uint32_t x) |
1009 |
{ |
1010 |
uint64_t tmp; |
1011 |
uint64_t ret; |
1012 |
ret = (uint8_t)x; |
1013 |
tmp = (uint8_t)(x >> 8);
|
1014 |
ret |= tmp << 16;
|
1015 |
tmp = (uint8_t)(x >> 16);
|
1016 |
ret |= tmp << 32;
|
1017 |
tmp = (uint8_t)(x >> 24);
|
1018 |
ret |= tmp << 48;
|
1019 |
return ret;
|
1020 |
} |
1021 |
|
1022 |
uint64_t HELPER(neon_widen_s8)(uint32_t x) |
1023 |
{ |
1024 |
uint64_t tmp; |
1025 |
uint64_t ret; |
1026 |
ret = (uint16_t)(int8_t)x; |
1027 |
tmp = (uint16_t)(int8_t)(x >> 8);
|
1028 |
ret |= tmp << 16;
|
1029 |
tmp = (uint16_t)(int8_t)(x >> 16);
|
1030 |
ret |= tmp << 32;
|
1031 |
tmp = (uint16_t)(int8_t)(x >> 24);
|
1032 |
ret |= tmp << 48;
|
1033 |
return ret;
|
1034 |
} |
1035 |
|
1036 |
uint64_t HELPER(neon_widen_u16)(uint32_t x) |
1037 |
{ |
1038 |
uint64_t high = (uint16_t)(x >> 16);
|
1039 |
return ((uint16_t)x) | (high << 32); |
1040 |
} |
1041 |
|
1042 |
uint64_t HELPER(neon_widen_s16)(uint32_t x) |
1043 |
{ |
1044 |
uint64_t high = (int16_t)(x >> 16);
|
1045 |
return ((uint32_t)(int16_t)x) | (high << 32); |
1046 |
} |
1047 |
|
1048 |
uint64_t HELPER(neon_addl_u16)(uint64_t a, uint64_t b) |
1049 |
{ |
1050 |
uint64_t mask; |
1051 |
mask = (a ^ b) & 0x8000800080008000ull;
|
1052 |
a &= ~0x8000800080008000ull;
|
1053 |
b &= ~0x8000800080008000ull;
|
1054 |
return (a + b) ^ mask;
|
1055 |
} |
1056 |
|
1057 |
uint64_t HELPER(neon_addl_u32)(uint64_t a, uint64_t b) |
1058 |
{ |
1059 |
uint64_t mask; |
1060 |
mask = (a ^ b) & 0x8000000080000000ull;
|
1061 |
a &= ~0x8000000080000000ull;
|
1062 |
b &= ~0x8000000080000000ull;
|
1063 |
return (a + b) ^ mask;
|
1064 |
} |
1065 |
|
1066 |
uint64_t HELPER(neon_paddl_u16)(uint64_t a, uint64_t b) |
1067 |
{ |
1068 |
uint64_t tmp; |
1069 |
uint64_t tmp2; |
1070 |
|
1071 |
tmp = a & 0x0000ffff0000ffffull;
|
1072 |
tmp += (a >> 16) & 0x0000ffff0000ffffull; |
1073 |
tmp2 = b & 0xffff0000ffff0000ull;
|
1074 |
tmp2 += (b << 16) & 0xffff0000ffff0000ull; |
1075 |
return ( tmp & 0xffff) |
1076 |
| ((tmp >> 16) & 0xffff0000ull) |
1077 |
| ((tmp2 << 16) & 0xffff00000000ull) |
1078 |
| ( tmp2 & 0xffff000000000000ull);
|
1079 |
} |
1080 |
|
1081 |
uint64_t HELPER(neon_paddl_u32)(uint64_t a, uint64_t b) |
1082 |
{ |
1083 |
uint32_t low = a + (a >> 32);
|
1084 |
uint32_t high = b + (b >> 32);
|
1085 |
return low + ((uint64_t)high << 32); |
1086 |
} |
1087 |
|
1088 |
uint64_t HELPER(neon_subl_u16)(uint64_t a, uint64_t b) |
1089 |
{ |
1090 |
uint64_t mask; |
1091 |
mask = (a ^ ~b) & 0x8000800080008000ull;
|
1092 |
a |= 0x8000800080008000ull;
|
1093 |
b &= ~0x8000800080008000ull;
|
1094 |
return (a - b) ^ mask;
|
1095 |
} |
1096 |
|
1097 |
uint64_t HELPER(neon_subl_u32)(uint64_t a, uint64_t b) |
1098 |
{ |
1099 |
uint64_t mask; |
1100 |
mask = (a ^ ~b) & 0x8000000080000000ull;
|
1101 |
a |= 0x8000000080000000ull;
|
1102 |
b &= ~0x8000000080000000ull;
|
1103 |
return (a - b) ^ mask;
|
1104 |
} |
1105 |
|
1106 |
uint64_t HELPER(neon_addl_saturate_s32)(CPUState *env, uint64_t a, uint64_t b) |
1107 |
{ |
1108 |
uint32_t x, y; |
1109 |
uint32_t low, high; |
1110 |
|
1111 |
x = a; |
1112 |
y = b; |
1113 |
low = x + y; |
1114 |
if (((low ^ x) & SIGNBIT) && !((x ^ y) & SIGNBIT)) {
|
1115 |
SET_QC(); |
1116 |
low = ((int32_t)x >> 31) ^ ~SIGNBIT;
|
1117 |
} |
1118 |
x = a >> 32;
|
1119 |
y = b >> 32;
|
1120 |
high = x + y; |
1121 |
if (((high ^ x) & SIGNBIT) && !((x ^ y) & SIGNBIT)) {
|
1122 |
SET_QC(); |
1123 |
high = ((int32_t)x >> 31) ^ ~SIGNBIT;
|
1124 |
} |
1125 |
return low | ((uint64_t)high << 32); |
1126 |
} |
1127 |
|
1128 |
uint64_t HELPER(neon_addl_saturate_s64)(CPUState *env, uint64_t a, uint64_t b) |
1129 |
{ |
1130 |
uint64_t result; |
1131 |
|
1132 |
result = a + b; |
1133 |
if (((result ^ a) & SIGNBIT64) && !((a ^ b) & SIGNBIT64)) {
|
1134 |
SET_QC(); |
1135 |
result = ((int64_t)a >> 63) ^ ~SIGNBIT64;
|
1136 |
} |
1137 |
return result;
|
1138 |
} |
1139 |
|
1140 |
#define DO_ABD(dest, x, y, type) do { \ |
1141 |
type tmp_x = x; \ |
1142 |
type tmp_y = y; \ |
1143 |
dest = ((tmp_x > tmp_y) ? tmp_x - tmp_y : tmp_y - tmp_x); \ |
1144 |
} while(0) |
1145 |
|
1146 |
uint64_t HELPER(neon_abdl_u16)(uint32_t a, uint32_t b) |
1147 |
{ |
1148 |
uint64_t tmp; |
1149 |
uint64_t result; |
1150 |
DO_ABD(result, a, b, uint8_t); |
1151 |
DO_ABD(tmp, a >> 8, b >> 8, uint8_t); |
1152 |
result |= tmp << 16;
|
1153 |
DO_ABD(tmp, a >> 16, b >> 16, uint8_t); |
1154 |
result |= tmp << 32;
|
1155 |
DO_ABD(tmp, a >> 24, b >> 24, uint8_t); |
1156 |
result |= tmp << 48;
|
1157 |
return result;
|
1158 |
} |
1159 |
|
1160 |
uint64_t HELPER(neon_abdl_s16)(uint32_t a, uint32_t b) |
1161 |
{ |
1162 |
uint64_t tmp; |
1163 |
uint64_t result; |
1164 |
DO_ABD(result, a, b, int8_t); |
1165 |
DO_ABD(tmp, a >> 8, b >> 8, int8_t); |
1166 |
result |= tmp << 16;
|
1167 |
DO_ABD(tmp, a >> 16, b >> 16, int8_t); |
1168 |
result |= tmp << 32;
|
1169 |
DO_ABD(tmp, a >> 24, b >> 24, int8_t); |
1170 |
result |= tmp << 48;
|
1171 |
return result;
|
1172 |
} |
1173 |
|
1174 |
uint64_t HELPER(neon_abdl_u32)(uint32_t a, uint32_t b) |
1175 |
{ |
1176 |
uint64_t tmp; |
1177 |
uint64_t result; |
1178 |
DO_ABD(result, a, b, uint16_t); |
1179 |
DO_ABD(tmp, a >> 16, b >> 16, uint16_t); |
1180 |
return result | (tmp << 32); |
1181 |
} |
1182 |
|
1183 |
uint64_t HELPER(neon_abdl_s32)(uint32_t a, uint32_t b) |
1184 |
{ |
1185 |
uint64_t tmp; |
1186 |
uint64_t result; |
1187 |
DO_ABD(result, a, b, int16_t); |
1188 |
DO_ABD(tmp, a >> 16, b >> 16, int16_t); |
1189 |
return result | (tmp << 32); |
1190 |
} |
1191 |
|
1192 |
uint64_t HELPER(neon_abdl_u64)(uint32_t a, uint32_t b) |
1193 |
{ |
1194 |
uint64_t result; |
1195 |
DO_ABD(result, a, b, uint32_t); |
1196 |
return result;
|
1197 |
} |
1198 |
|
1199 |
uint64_t HELPER(neon_abdl_s64)(uint32_t a, uint32_t b) |
1200 |
{ |
1201 |
uint64_t result; |
1202 |
DO_ABD(result, a, b, int32_t); |
1203 |
return result;
|
1204 |
} |
1205 |
#undef DO_ABD
|
1206 |
|
1207 |
/* Widening multiply. Named type is the source type. */
|
1208 |
#define DO_MULL(dest, x, y, type1, type2) do { \ |
1209 |
type1 tmp_x = x; \ |
1210 |
type1 tmp_y = y; \ |
1211 |
dest = (type2)((type2)tmp_x * (type2)tmp_y); \ |
1212 |
} while(0) |
1213 |
|
1214 |
uint64_t HELPER(neon_mull_u8)(uint32_t a, uint32_t b) |
1215 |
{ |
1216 |
uint64_t tmp; |
1217 |
uint64_t result; |
1218 |
|
1219 |
DO_MULL(result, a, b, uint8_t, uint16_t); |
1220 |
DO_MULL(tmp, a >> 8, b >> 8, uint8_t, uint16_t); |
1221 |
result |= tmp << 16;
|
1222 |
DO_MULL(tmp, a >> 16, b >> 16, uint8_t, uint16_t); |
1223 |
result |= tmp << 32;
|
1224 |
DO_MULL(tmp, a >> 24, b >> 24, uint8_t, uint16_t); |
1225 |
result |= tmp << 48;
|
1226 |
return result;
|
1227 |
} |
1228 |
|
1229 |
uint64_t HELPER(neon_mull_s8)(uint32_t a, uint32_t b) |
1230 |
{ |
1231 |
uint64_t tmp; |
1232 |
uint64_t result; |
1233 |
|
1234 |
DO_MULL(result, a, b, int8_t, uint16_t); |
1235 |
DO_MULL(tmp, a >> 8, b >> 8, int8_t, uint16_t); |
1236 |
result |= tmp << 16;
|
1237 |
DO_MULL(tmp, a >> 16, b >> 16, int8_t, uint16_t); |
1238 |
result |= tmp << 32;
|
1239 |
DO_MULL(tmp, a >> 24, b >> 24, int8_t, uint16_t); |
1240 |
result |= tmp << 48;
|
1241 |
return result;
|
1242 |
} |
1243 |
|
1244 |
uint64_t HELPER(neon_mull_u16)(uint32_t a, uint32_t b) |
1245 |
{ |
1246 |
uint64_t tmp; |
1247 |
uint64_t result; |
1248 |
|
1249 |
DO_MULL(result, a, b, uint16_t, uint32_t); |
1250 |
DO_MULL(tmp, a >> 16, b >> 16, uint16_t, uint32_t); |
1251 |
return result | (tmp << 32); |
1252 |
} |
1253 |
|
1254 |
uint64_t HELPER(neon_mull_s16)(uint32_t a, uint32_t b) |
1255 |
{ |
1256 |
uint64_t tmp; |
1257 |
uint64_t result; |
1258 |
|
1259 |
DO_MULL(result, a, b, int16_t, uint32_t); |
1260 |
DO_MULL(tmp, a >> 16, b >> 16, int16_t, uint32_t); |
1261 |
return result | (tmp << 32); |
1262 |
} |
1263 |
|
1264 |
uint64_t HELPER(neon_negl_u16)(uint64_t x) |
1265 |
{ |
1266 |
uint16_t tmp; |
1267 |
uint64_t result; |
1268 |
result = (uint16_t)-x; |
1269 |
tmp = -(x >> 16);
|
1270 |
result |= (uint64_t)tmp << 16;
|
1271 |
tmp = -(x >> 32);
|
1272 |
result |= (uint64_t)tmp << 32;
|
1273 |
tmp = -(x >> 48);
|
1274 |
result |= (uint64_t)tmp << 48;
|
1275 |
return result;
|
1276 |
} |
1277 |
|
1278 |
#include <stdio.h> |
1279 |
uint64_t HELPER(neon_negl_u32)(uint64_t x) |
1280 |
{ |
1281 |
uint32_t low = -x; |
1282 |
uint32_t high = -(x >> 32);
|
1283 |
return low | ((uint64_t)high << 32); |
1284 |
} |
1285 |
|
1286 |
/* FIXME: There should be a native op for this. */
|
1287 |
uint64_t HELPER(neon_negl_u64)(uint64_t x) |
1288 |
{ |
1289 |
return -x;
|
1290 |
} |
1291 |
|
1292 |
/* Saturnating sign manuipulation. */
|
1293 |
/* ??? Make these use NEON_VOP1 */
|
1294 |
#define DO_QABS8(x) do { \ |
1295 |
if (x == (int8_t)0x80) { \ |
1296 |
x = 0x7f; \
|
1297 |
SET_QC(); \ |
1298 |
} else if (x < 0) { \ |
1299 |
x = -x; \ |
1300 |
}} while (0) |
1301 |
uint32_t HELPER(neon_qabs_s8)(CPUState *env, uint32_t x) |
1302 |
{ |
1303 |
neon_s8 vec; |
1304 |
NEON_UNPACK(neon_s8, vec, x); |
1305 |
DO_QABS8(vec.v1); |
1306 |
DO_QABS8(vec.v2); |
1307 |
DO_QABS8(vec.v3); |
1308 |
DO_QABS8(vec.v4); |
1309 |
NEON_PACK(neon_s8, x, vec); |
1310 |
return x;
|
1311 |
} |
1312 |
#undef DO_QABS8
|
1313 |
|
1314 |
#define DO_QNEG8(x) do { \ |
1315 |
if (x == (int8_t)0x80) { \ |
1316 |
x = 0x7f; \
|
1317 |
SET_QC(); \ |
1318 |
} else { \
|
1319 |
x = -x; \ |
1320 |
}} while (0) |
1321 |
uint32_t HELPER(neon_qneg_s8)(CPUState *env, uint32_t x) |
1322 |
{ |
1323 |
neon_s8 vec; |
1324 |
NEON_UNPACK(neon_s8, vec, x); |
1325 |
DO_QNEG8(vec.v1); |
1326 |
DO_QNEG8(vec.v2); |
1327 |
DO_QNEG8(vec.v3); |
1328 |
DO_QNEG8(vec.v4); |
1329 |
NEON_PACK(neon_s8, x, vec); |
1330 |
return x;
|
1331 |
} |
1332 |
#undef DO_QNEG8
|
1333 |
|
1334 |
#define DO_QABS16(x) do { \ |
1335 |
if (x == (int16_t)0x8000) { \ |
1336 |
x = 0x7fff; \
|
1337 |
SET_QC(); \ |
1338 |
} else if (x < 0) { \ |
1339 |
x = -x; \ |
1340 |
}} while (0) |
1341 |
uint32_t HELPER(neon_qabs_s16)(CPUState *env, uint32_t x) |
1342 |
{ |
1343 |
neon_s16 vec; |
1344 |
NEON_UNPACK(neon_s16, vec, x); |
1345 |
DO_QABS16(vec.v1); |
1346 |
DO_QABS16(vec.v2); |
1347 |
NEON_PACK(neon_s16, x, vec); |
1348 |
return x;
|
1349 |
} |
1350 |
#undef DO_QABS16
|
1351 |
|
1352 |
#define DO_QNEG16(x) do { \ |
1353 |
if (x == (int16_t)0x8000) { \ |
1354 |
x = 0x7fff; \
|
1355 |
SET_QC(); \ |
1356 |
} else { \
|
1357 |
x = -x; \ |
1358 |
}} while (0) |
1359 |
uint32_t HELPER(neon_qneg_s16)(CPUState *env, uint32_t x) |
1360 |
{ |
1361 |
neon_s16 vec; |
1362 |
NEON_UNPACK(neon_s16, vec, x); |
1363 |
DO_QNEG16(vec.v1); |
1364 |
DO_QNEG16(vec.v2); |
1365 |
NEON_PACK(neon_s16, x, vec); |
1366 |
return x;
|
1367 |
} |
1368 |
#undef DO_QNEG16
|
1369 |
|
1370 |
uint32_t HELPER(neon_qabs_s32)(CPUState *env, uint32_t x) |
1371 |
{ |
1372 |
if (x == SIGNBIT) {
|
1373 |
SET_QC(); |
1374 |
x = ~SIGNBIT; |
1375 |
} else if ((int32_t)x < 0) { |
1376 |
x = -x; |
1377 |
} |
1378 |
return x;
|
1379 |
} |
1380 |
|
1381 |
uint32_t HELPER(neon_qneg_s32)(CPUState *env, uint32_t x) |
1382 |
{ |
1383 |
if (x == SIGNBIT) {
|
1384 |
SET_QC(); |
1385 |
x = ~SIGNBIT; |
1386 |
} else {
|
1387 |
x = -x; |
1388 |
} |
1389 |
return x;
|
1390 |
} |
1391 |
|
1392 |
/* NEON Float helpers. */
|
1393 |
uint32_t HELPER(neon_min_f32)(uint32_t a, uint32_t b) |
1394 |
{ |
1395 |
float32 f0 = vfp_itos(a); |
1396 |
float32 f1 = vfp_itos(b); |
1397 |
return (float32_compare_quiet(f0, f1, NFS) == -1) ? a : b; |
1398 |
} |
1399 |
|
1400 |
uint32_t HELPER(neon_max_f32)(uint32_t a, uint32_t b) |
1401 |
{ |
1402 |
float32 f0 = vfp_itos(a); |
1403 |
float32 f1 = vfp_itos(b); |
1404 |
return (float32_compare_quiet(f0, f1, NFS) == 1) ? a : b; |
1405 |
} |
1406 |
|
1407 |
uint32_t HELPER(neon_abd_f32)(uint32_t a, uint32_t b) |
1408 |
{ |
1409 |
float32 f0 = vfp_itos(a); |
1410 |
float32 f1 = vfp_itos(b); |
1411 |
return vfp_stoi((float32_compare_quiet(f0, f1, NFS) == 1) |
1412 |
? float32_sub(f0, f1, NFS) |
1413 |
: float32_sub(f1, f0, NFS)); |
1414 |
} |
1415 |
|
1416 |
uint32_t HELPER(neon_add_f32)(uint32_t a, uint32_t b) |
1417 |
{ |
1418 |
return vfp_stoi(float32_add(vfp_itos(a), vfp_itos(b), NFS));
|
1419 |
} |
1420 |
|
1421 |
uint32_t HELPER(neon_sub_f32)(uint32_t a, uint32_t b) |
1422 |
{ |
1423 |
return vfp_stoi(float32_sub(vfp_itos(a), vfp_itos(b), NFS));
|
1424 |
} |
1425 |
|
1426 |
uint32_t HELPER(neon_mul_f32)(uint32_t a, uint32_t b) |
1427 |
{ |
1428 |
return vfp_stoi(float32_mul(vfp_itos(a), vfp_itos(b), NFS));
|
1429 |
} |
1430 |
|
1431 |
/* Floating point comparisons produce an integer result. */
|
1432 |
#define NEON_VOP_FCMP(name, cmp) \
|
1433 |
uint32_t HELPER(neon_##name)(uint32_t a, uint32_t b) \ |
1434 |
{ \ |
1435 |
if (float32_compare_quiet(vfp_itos(a), vfp_itos(b), NFS) cmp 0) \ |
1436 |
return ~0; \ |
1437 |
else \
|
1438 |
return 0; \ |
1439 |
} |
1440 |
|
1441 |
NEON_VOP_FCMP(ceq_f32, ==) |
1442 |
NEON_VOP_FCMP(cge_f32, >=) |
1443 |
NEON_VOP_FCMP(cgt_f32, >) |
1444 |
|
1445 |
uint32_t HELPER(neon_acge_f32)(uint32_t a, uint32_t b) |
1446 |
{ |
1447 |
float32 f0 = float32_abs(vfp_itos(a)); |
1448 |
float32 f1 = float32_abs(vfp_itos(b)); |
1449 |
return (float32_compare_quiet(f0, f1,NFS) >= 0) ? ~0 : 0; |
1450 |
} |
1451 |
|
1452 |
uint32_t HELPER(neon_acgt_f32)(uint32_t a, uint32_t b) |
1453 |
{ |
1454 |
float32 f0 = float32_abs(vfp_itos(a)); |
1455 |
float32 f1 = float32_abs(vfp_itos(b)); |
1456 |
return (float32_compare_quiet(f0, f1, NFS) > 0) ? ~0 : 0; |
1457 |
} |