1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
|
/* The gcc-provided loongson intrinsic functions are way too fucking broken
* to be of any use, otherwise I'd use them.
*
* - The hardware instructions are very similar to MMX or iwMMXt. Certainly
* close enough that they could have implemented the _mm_*-style intrinsic
* interface and had a ton of optimized code available to them. Instead they
* implemented something much, much worse.
*
* - pshuf takes a dead first argument, causing extra instructions to be
* generated.
*
* - There are no 64-bit shift or logical intrinsics, which means you have
* to implement them with inline assembly, but this is a nightmare because
* gcc doesn't understand that the integer vector datatypes are actually in
* floating-point registers, so you end up with braindead code like
*
* punpcklwd $f9,$f9,$f5
* dmtc1 v0,$f8
* punpcklwd $f19,$f19,$f5
* dmfc1 t9,$f9
* dmtc1 v0,$f9
* dmtc1 t9,$f20
* dmfc1 s0,$f19
* punpcklbh $f20,$f20,$f2
*
* where crap just gets copied back and forth between integer and floating-
* point registers ad nauseum.
*
* Instead of trying to workaround the problems from these crap intrinsics, I
* just implement the _mm_* intrinsics needed for pixman-mmx.c using inline
* assembly.
*/
#include <stdint.h>
/* vectors are stored in 64-bit floating-point registers */
typedef double __m64;
/* having a 32-bit datatype allows us to use 32-bit loads in places like load8888 */
typedef float __m32;
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_setzero_si64 (void)
{
return 0.0;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_adds_pu16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("paddush %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_adds_pu8 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("paddusb %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_and_si64 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("and %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_empty (void)
{
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_madd_pi16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("pmaddhw %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_mulhi_pu16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("pmulhuh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_mullo_pi16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("pmullh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_or_si64 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("or %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_packs_pu16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("packushb %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_shuffle_pi16 (__m64 __m, int64_t __n)
{
__m64 ret;
asm("pshufh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m), "f" (*(__m64 *)&__n)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_slli_si64 (__m64 __m, int64_t __count)
{
__m64 ret;
asm("dsll %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m), "f" (*(__m64 *)&__count)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_srli_pi16 (__m64 __m, int64_t __count)
{
__m64 ret;
asm("psrlh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m), "f" (*(__m64 *)&__count)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_srli_si64 (__m64 __m, int64_t __count)
{
__m64 ret;
asm("dsrl %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m), "f" (*(__m64 *)&__count)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_unpackhi_pi8 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("punpckhbh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_unpackhi_pi16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("punpckhhw %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_unpacklo_pi8 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("punpcklbh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
/* Since punpcklbh doesn't care about the high 32-bits, we use the __m32 datatype which
* allows load8888 to use 32-bit loads */
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_unpacklo_pi8_f (__m32 __m1, __m64 __m2)
{
__m64 ret;
asm("punpcklbh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_unpacklo_pi16 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("punpcklhw %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_xor_si64 (__m64 __m1, __m64 __m2)
{
__m64 ret;
asm("xor %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
loongson_extract_pi16 (__m64 __m, int64_t __pos)
{
__m64 ret;
asm("pextrh %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m), "f" (*(__m64 *)&__pos)
);
return ret;
}
extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
loongson_insert_pi16 (__m64 __m1, __m64 __m2, int64_t __pos)
{
__m64 ret;
asm("pinsrh_%3 %0, %1, %2\n\t"
: "=f" (ret)
: "f" (__m1), "f" (__m2), "i" (__pos)
);
return ret;
}
|
