xref: /aosp_15_r20/external/libaom/test/simd_impl.h (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <tuple>
13 
14 #define SIMD_CHECK 1
15 #include "aom_dsp/aom_simd_inline.h"
16 #include "aom_dsp/simd/v256_intrinsics_c.h"
17 #include "gtest/gtest.h"
18 #include "test/register_state_check.h"
19 
20 namespace SIMD_NAMESPACE {
21 
22 template <typename param_signature>
23 class TestIntrinsic : public ::testing::TestWithParam<param_signature> {
24  public:
25   ~TestIntrinsic() override = default;
SetUp()26   void SetUp() override {
27     mask = std::get<0>(this->GetParam());
28     maskwidth = std::get<1>(this->GetParam());
29     name = std::get<2>(this->GetParam());
30   }
31 
32  protected:
33   uint32_t mask, maskwidth;
34   const char *name;
35 };
36 
37 // Create one typedef for each function signature
38 #define TYPEDEF_SIMD(name)                                             \
39   typedef TestIntrinsic<std::tuple<uint32_t, uint32_t, const char *> > \
40   ARCH_POSTFIX(name)
41 
42 TYPEDEF_SIMD(V64_U8);
43 TYPEDEF_SIMD(V64_U16);
44 TYPEDEF_SIMD(V64_U32);
45 TYPEDEF_SIMD(V64_V64);
46 TYPEDEF_SIMD(U32_V64);
47 TYPEDEF_SIMD(S32_V64);
48 TYPEDEF_SIMD(U64_V64);
49 TYPEDEF_SIMD(S64_V64);
50 TYPEDEF_SIMD(V64_U32U32);
51 TYPEDEF_SIMD(V64_V64V64);
52 TYPEDEF_SIMD(S64_V64V64);
53 TYPEDEF_SIMD(V64_V64U32);
54 TYPEDEF_SIMD(U32_V64V64);
55 TYPEDEF_SIMD(V128_V64);
56 TYPEDEF_SIMD(V128_V128);
57 TYPEDEF_SIMD(U32_V128);
58 TYPEDEF_SIMD(U64_V128);
59 TYPEDEF_SIMD(V64_V128);
60 TYPEDEF_SIMD(V128_U8);
61 TYPEDEF_SIMD(V128_U16);
62 TYPEDEF_SIMD(V128_U32);
63 TYPEDEF_SIMD(V128_U64);
64 TYPEDEF_SIMD(V128_U64U64);
65 TYPEDEF_SIMD(V128_V64V64);
66 TYPEDEF_SIMD(V128_V128V128);
67 TYPEDEF_SIMD(V128_V128V128V128);
68 TYPEDEF_SIMD(S64_V128V128);
69 TYPEDEF_SIMD(V128_V128U32);
70 TYPEDEF_SIMD(U32_V128V128);
71 TYPEDEF_SIMD(U64_V128V128);
72 TYPEDEF_SIMD(V256_V128);
73 TYPEDEF_SIMD(V256_V256);
74 TYPEDEF_SIMD(U64_V256);
75 TYPEDEF_SIMD(V256_V128V128);
76 TYPEDEF_SIMD(V256_V256V256);
77 TYPEDEF_SIMD(V256_V256V256V256);
78 TYPEDEF_SIMD(U64_V256V256);
79 TYPEDEF_SIMD(S64_V256V256);
80 TYPEDEF_SIMD(V256_V256U32);
81 TYPEDEF_SIMD(U32_V256V256);
82 TYPEDEF_SIMD(V256_U8);
83 TYPEDEF_SIMD(V256_U16);
84 TYPEDEF_SIMD(V256_U32);
85 TYPEDEF_SIMD(V256_U64);
86 TYPEDEF_SIMD(U32_V256);
87 TYPEDEF_SIMD(V64_V256);
88 
89 // Google Test allows up to 50 tests per case, so split the largest
90 typedef ARCH_POSTFIX(V64_V64) ARCH_POSTFIX(V64_V64_Part2);
91 typedef ARCH_POSTFIX(V64_V64V64) ARCH_POSTFIX(V64_V64V64_Part2);
92 typedef ARCH_POSTFIX(V128_V128) ARCH_POSTFIX(V128_V128_Part2);
93 typedef ARCH_POSTFIX(V128_V128) ARCH_POSTFIX(V128_V128_Part3);
94 typedef ARCH_POSTFIX(V128_V128) ARCH_POSTFIX(V128_V128_Part4);
95 typedef ARCH_POSTFIX(V128_V128V128) ARCH_POSTFIX(V128_V128V128_Part2);
96 typedef ARCH_POSTFIX(V256_V256) ARCH_POSTFIX(V256_V256_Part2);
97 typedef ARCH_POSTFIX(V256_V256) ARCH_POSTFIX(V256_V256_Part3);
98 typedef ARCH_POSTFIX(V256_V256) ARCH_POSTFIX(V256_V256_Part4);
99 typedef ARCH_POSTFIX(V256_V256) ARCH_POSTFIX(V256_V256_Part5);
100 typedef ARCH_POSTFIX(V256_V256V256) ARCH_POSTFIX(V256_V256V256_Part2);
101 
102 // These functions are machine tuned located elsewhere
103 template <typename c_ret, typename c_arg>
104 void TestSimd1Arg(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
105                   const char *name);
106 
107 template <typename c_ret, typename c_arg1, typename c_arg2>
108 void TestSimd2Args(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
109                    const char *name);
110 
111 template <typename c_ret, typename c_arg1, typename c_arg2, typename c_arg3>
112 void TestSimd3Args(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
113                    const char *name);
114 
115 const int kIterations = 65536;
116 
117 // Add a macro layer since TEST_P will quote the name so we need to
118 // expand it first with the prefix.
119 #define MY_TEST_P(name, test) TEST_P(name, test)
120 
MY_TEST_P(ARCH_POSTFIX (V64_U8),TestIntrinsics)121 MY_TEST_P(ARCH_POSTFIX(V64_U8), TestIntrinsics) {
122   TestSimd1Arg<c_v64, uint8_t>(kIterations, mask, maskwidth, name);
123 }
124 
MY_TEST_P(ARCH_POSTFIX (V64_U16),TestIntrinsics)125 MY_TEST_P(ARCH_POSTFIX(V64_U16), TestIntrinsics) {
126   TestSimd1Arg<c_v64, uint16_t>(kIterations, mask, maskwidth, name);
127 }
128 
MY_TEST_P(ARCH_POSTFIX (V64_U32),TestIntrinsics)129 MY_TEST_P(ARCH_POSTFIX(V64_U32), TestIntrinsics) {
130   TestSimd1Arg<c_v64, uint32_t>(kIterations, mask, maskwidth, name);
131 }
132 
MY_TEST_P(ARCH_POSTFIX (V64_V64),TestIntrinsics)133 MY_TEST_P(ARCH_POSTFIX(V64_V64), TestIntrinsics) {
134   TestSimd1Arg<c_v64, c_v64>(kIterations, mask, maskwidth, name);
135 }
136 
MY_TEST_P(ARCH_POSTFIX (U64_V64),TestIntrinsics)137 MY_TEST_P(ARCH_POSTFIX(U64_V64), TestIntrinsics) {
138   TestSimd1Arg<uint64_t, c_v64>(kIterations, mask, maskwidth, name);
139 }
140 
MY_TEST_P(ARCH_POSTFIX (S64_V64),TestIntrinsics)141 MY_TEST_P(ARCH_POSTFIX(S64_V64), TestIntrinsics) {
142   TestSimd1Arg<int64_t, c_v64>(kIterations, mask, maskwidth, name);
143 }
144 
MY_TEST_P(ARCH_POSTFIX (U32_V64),TestIntrinsics)145 MY_TEST_P(ARCH_POSTFIX(U32_V64), TestIntrinsics) {
146   TestSimd1Arg<uint32_t, c_v64>(kIterations, mask, maskwidth, name);
147 }
148 
MY_TEST_P(ARCH_POSTFIX (S32_V64),TestIntrinsics)149 MY_TEST_P(ARCH_POSTFIX(S32_V64), TestIntrinsics) {
150   TestSimd1Arg<int32_t, c_v64>(kIterations, mask, maskwidth, name);
151 }
152 
MY_TEST_P(ARCH_POSTFIX (V64_U32U32),TestIntrinsics)153 MY_TEST_P(ARCH_POSTFIX(V64_U32U32), TestIntrinsics) {
154   TestSimd2Args<c_v64, uint32_t, uint32_t>(kIterations, mask, maskwidth, name);
155 }
156 
MY_TEST_P(ARCH_POSTFIX (V64_V64V64),TestIntrinsics)157 MY_TEST_P(ARCH_POSTFIX(V64_V64V64), TestIntrinsics) {
158   TestSimd2Args<c_v64, c_v64, c_v64>(kIterations, mask, maskwidth, name);
159 }
160 
MY_TEST_P(ARCH_POSTFIX (S64_V64V64),TestIntrinsics)161 MY_TEST_P(ARCH_POSTFIX(S64_V64V64), TestIntrinsics) {
162   TestSimd2Args<int64_t, c_v64, c_v64>(kIterations, mask, maskwidth, name);
163 }
164 
MY_TEST_P(ARCH_POSTFIX (U32_V64V64),TestIntrinsics)165 MY_TEST_P(ARCH_POSTFIX(U32_V64V64), TestIntrinsics) {
166   TestSimd2Args<uint32_t, c_v64, c_v64>(kIterations, mask, maskwidth, name);
167 }
168 
MY_TEST_P(ARCH_POSTFIX (V64_V64U32),TestIntrinsics)169 MY_TEST_P(ARCH_POSTFIX(V64_V64U32), TestIntrinsics) {
170   TestSimd2Args<c_v64, c_v64, uint32_t>(kIterations, mask, maskwidth, name);
171 }
172 
173 // Google Test allows up to 50 tests per case, so split the largest
MY_TEST_P(ARCH_POSTFIX (V64_V64_Part2),TestIntrinsics)174 MY_TEST_P(ARCH_POSTFIX(V64_V64_Part2), TestIntrinsics) {
175   TestSimd1Arg<c_v64, c_v64>(kIterations, mask, maskwidth, name);
176 }
177 
MY_TEST_P(ARCH_POSTFIX (V64_V64V64_Part2),TestIntrinsics)178 MY_TEST_P(ARCH_POSTFIX(V64_V64V64_Part2), TestIntrinsics) {
179   TestSimd2Args<c_v64, c_v64, c_v64>(kIterations, mask, maskwidth, name);
180 }
181 
MY_TEST_P(ARCH_POSTFIX (U32_V128),TestIntrinsics)182 MY_TEST_P(ARCH_POSTFIX(U32_V128), TestIntrinsics) {
183   TestSimd1Arg<uint32_t, c_v128>(kIterations, mask, maskwidth, name);
184 }
185 
MY_TEST_P(ARCH_POSTFIX (U64_V128),TestIntrinsics)186 MY_TEST_P(ARCH_POSTFIX(U64_V128), TestIntrinsics) {
187   TestSimd1Arg<uint64_t, c_v128>(kIterations, mask, maskwidth, name);
188 }
189 
MY_TEST_P(ARCH_POSTFIX (V64_V128),TestIntrinsics)190 MY_TEST_P(ARCH_POSTFIX(V64_V128), TestIntrinsics) {
191   TestSimd1Arg<c_v64, c_v128>(kIterations, mask, maskwidth, name);
192 }
193 
MY_TEST_P(ARCH_POSTFIX (V128_V128),TestIntrinsics)194 MY_TEST_P(ARCH_POSTFIX(V128_V128), TestIntrinsics) {
195   TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
196 }
197 
MY_TEST_P(ARCH_POSTFIX (V128_U8),TestIntrinsics)198 MY_TEST_P(ARCH_POSTFIX(V128_U8), TestIntrinsics) {
199   TestSimd1Arg<c_v128, uint8_t>(kIterations, mask, maskwidth, name);
200 }
201 
MY_TEST_P(ARCH_POSTFIX (V128_U16),TestIntrinsics)202 MY_TEST_P(ARCH_POSTFIX(V128_U16), TestIntrinsics) {
203   TestSimd1Arg<c_v128, uint16_t>(kIterations, mask, maskwidth, name);
204 }
205 
MY_TEST_P(ARCH_POSTFIX (V128_U32),TestIntrinsics)206 MY_TEST_P(ARCH_POSTFIX(V128_U32), TestIntrinsics) {
207   TestSimd1Arg<c_v128, uint32_t>(kIterations, mask, maskwidth, name);
208 }
209 
MY_TEST_P(ARCH_POSTFIX (V128_U64),TestIntrinsics)210 MY_TEST_P(ARCH_POSTFIX(V128_U64), TestIntrinsics) {
211   TestSimd1Arg<c_v128, uint64_t>(kIterations, mask, maskwidth, name);
212 }
213 
MY_TEST_P(ARCH_POSTFIX (V128_V64),TestIntrinsics)214 MY_TEST_P(ARCH_POSTFIX(V128_V64), TestIntrinsics) {
215   TestSimd1Arg<c_v128, c_v64>(kIterations, mask, maskwidth, name);
216 }
217 
MY_TEST_P(ARCH_POSTFIX (V128_V128V128),TestIntrinsics)218 MY_TEST_P(ARCH_POSTFIX(V128_V128V128), TestIntrinsics) {
219   TestSimd2Args<c_v128, c_v128, c_v128>(kIterations, mask, maskwidth, name);
220 }
221 
MY_TEST_P(ARCH_POSTFIX (V128_V128V128V128),TestIntrinsics)222 MY_TEST_P(ARCH_POSTFIX(V128_V128V128V128), TestIntrinsics) {
223   TestSimd3Args<c_v128, c_v128, c_v128, c_v128>(kIterations, mask, maskwidth,
224                                                 name);
225 }
226 
MY_TEST_P(ARCH_POSTFIX (U32_V128V128),TestIntrinsics)227 MY_TEST_P(ARCH_POSTFIX(U32_V128V128), TestIntrinsics) {
228   TestSimd2Args<uint32_t, c_v128, c_v128>(kIterations, mask, maskwidth, name);
229 }
230 
MY_TEST_P(ARCH_POSTFIX (U64_V128V128),TestIntrinsics)231 MY_TEST_P(ARCH_POSTFIX(U64_V128V128), TestIntrinsics) {
232   TestSimd2Args<uint64_t, c_v128, c_v128>(kIterations, mask, maskwidth, name);
233 }
234 
MY_TEST_P(ARCH_POSTFIX (S64_V128V128),TestIntrinsics)235 MY_TEST_P(ARCH_POSTFIX(S64_V128V128), TestIntrinsics) {
236   TestSimd2Args<int64_t, c_v128, c_v128>(kIterations, mask, maskwidth, name);
237 }
238 
MY_TEST_P(ARCH_POSTFIX (V128_U64U64),TestIntrinsics)239 MY_TEST_P(ARCH_POSTFIX(V128_U64U64), TestIntrinsics) {
240   TestSimd2Args<c_v128, uint64_t, uint64_t>(kIterations, mask, maskwidth, name);
241 }
242 
MY_TEST_P(ARCH_POSTFIX (V128_V64V64),TestIntrinsics)243 MY_TEST_P(ARCH_POSTFIX(V128_V64V64), TestIntrinsics) {
244   TestSimd2Args<c_v128, c_v64, c_v64>(kIterations, mask, maskwidth, name);
245 }
246 
MY_TEST_P(ARCH_POSTFIX (V128_V128U32),TestIntrinsics)247 MY_TEST_P(ARCH_POSTFIX(V128_V128U32), TestIntrinsics) {
248   TestSimd2Args<c_v128, c_v128, uint32_t>(kIterations, mask, maskwidth, name);
249 }
250 
MY_TEST_P(ARCH_POSTFIX (V128_V128V128_Part2),TestIntrinsics)251 MY_TEST_P(ARCH_POSTFIX(V128_V128V128_Part2), TestIntrinsics) {
252   TestSimd2Args<c_v128, c_v128, c_v128>(kIterations, mask, maskwidth, name);
253 }
254 
MY_TEST_P(ARCH_POSTFIX (V128_V128_Part2),TestIntrinsics)255 MY_TEST_P(ARCH_POSTFIX(V128_V128_Part2), TestIntrinsics) {
256   TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
257 }
258 
MY_TEST_P(ARCH_POSTFIX (V128_V128_Part3),TestIntrinsics)259 MY_TEST_P(ARCH_POSTFIX(V128_V128_Part3), TestIntrinsics) {
260   TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
261 }
262 
MY_TEST_P(ARCH_POSTFIX (V128_V128_Part4),TestIntrinsics)263 MY_TEST_P(ARCH_POSTFIX(V128_V128_Part4), TestIntrinsics) {
264   TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
265 }
266 
MY_TEST_P(ARCH_POSTFIX (U64_V256),TestIntrinsics)267 MY_TEST_P(ARCH_POSTFIX(U64_V256), TestIntrinsics) {
268   TestSimd1Arg<uint64_t, c_v256>(kIterations, mask, maskwidth, name);
269 }
270 
MY_TEST_P(ARCH_POSTFIX (V256_V256),TestIntrinsics)271 MY_TEST_P(ARCH_POSTFIX(V256_V256), TestIntrinsics) {
272   TestSimd1Arg<c_v256, c_v256>(kIterations, mask, maskwidth, name);
273 }
274 
MY_TEST_P(ARCH_POSTFIX (V256_V128),TestIntrinsics)275 MY_TEST_P(ARCH_POSTFIX(V256_V128), TestIntrinsics) {
276   TestSimd1Arg<c_v256, c_v128>(kIterations, mask, maskwidth, name);
277 }
278 
MY_TEST_P(ARCH_POSTFIX (V256_V256V256),TestIntrinsics)279 MY_TEST_P(ARCH_POSTFIX(V256_V256V256), TestIntrinsics) {
280   TestSimd2Args<c_v256, c_v256, c_v256>(kIterations, mask, maskwidth, name);
281 }
282 
MY_TEST_P(ARCH_POSTFIX (V256_V256V256V256),TestIntrinsics)283 MY_TEST_P(ARCH_POSTFIX(V256_V256V256V256), TestIntrinsics) {
284   TestSimd3Args<c_v256, c_v256, c_v256, c_v256>(kIterations, mask, maskwidth,
285                                                 name);
286 }
287 
MY_TEST_P(ARCH_POSTFIX (V256_V128V128),TestIntrinsics)288 MY_TEST_P(ARCH_POSTFIX(V256_V128V128), TestIntrinsics) {
289   TestSimd2Args<c_v256, c_v128, c_v128>(kIterations, mask, maskwidth, name);
290 }
291 
MY_TEST_P(ARCH_POSTFIX (U32_V256V256),TestIntrinsics)292 MY_TEST_P(ARCH_POSTFIX(U32_V256V256), TestIntrinsics) {
293   TestSimd2Args<uint32_t, c_v256, c_v256>(kIterations, mask, maskwidth, name);
294 }
295 
MY_TEST_P(ARCH_POSTFIX (U64_V256V256),TestIntrinsics)296 MY_TEST_P(ARCH_POSTFIX(U64_V256V256), TestIntrinsics) {
297   TestSimd2Args<uint64_t, c_v256, c_v256>(kIterations, mask, maskwidth, name);
298 }
299 
MY_TEST_P(ARCH_POSTFIX (S64_V256V256),TestIntrinsics)300 MY_TEST_P(ARCH_POSTFIX(S64_V256V256), TestIntrinsics) {
301   TestSimd2Args<int64_t, c_v256, c_v256>(kIterations, mask, maskwidth, name);
302 }
303 
MY_TEST_P(ARCH_POSTFIX (V256_V256V256_Part2),TestIntrinsics)304 MY_TEST_P(ARCH_POSTFIX(V256_V256V256_Part2), TestIntrinsics) {
305   TestSimd2Args<c_v256, c_v256, c_v256>(kIterations, mask, maskwidth, name);
306 }
307 
MY_TEST_P(ARCH_POSTFIX (V256_V256U32),TestIntrinsics)308 MY_TEST_P(ARCH_POSTFIX(V256_V256U32), TestIntrinsics) {
309   TestSimd2Args<c_v256, c_v256, uint32_t>(kIterations, mask, maskwidth, name);
310 }
311 
MY_TEST_P(ARCH_POSTFIX (V256_V256_Part2),TestIntrinsics)312 MY_TEST_P(ARCH_POSTFIX(V256_V256_Part2), TestIntrinsics) {
313   TestSimd1Arg<c_v256, c_v256>(kIterations, mask, maskwidth, name);
314 }
315 
MY_TEST_P(ARCH_POSTFIX (V256_V256_Part3),TestIntrinsics)316 MY_TEST_P(ARCH_POSTFIX(V256_V256_Part3), TestIntrinsics) {
317   TestSimd1Arg<c_v256, c_v256>(kIterations, mask, maskwidth, name);
318 }
319 
MY_TEST_P(ARCH_POSTFIX (V256_V256_Part4),TestIntrinsics)320 MY_TEST_P(ARCH_POSTFIX(V256_V256_Part4), TestIntrinsics) {
321   TestSimd1Arg<c_v256, c_v256>(kIterations, mask, maskwidth, name);
322 }
323 
MY_TEST_P(ARCH_POSTFIX (V256_V256_Part5),TestIntrinsics)324 MY_TEST_P(ARCH_POSTFIX(V256_V256_Part5), TestIntrinsics) {
325   TestSimd1Arg<c_v256, c_v256>(kIterations, mask, maskwidth, name);
326 }
327 
MY_TEST_P(ARCH_POSTFIX (V256_U8),TestIntrinsics)328 MY_TEST_P(ARCH_POSTFIX(V256_U8), TestIntrinsics) {
329   TestSimd1Arg<c_v256, uint8_t>(kIterations, mask, maskwidth, name);
330 }
331 
MY_TEST_P(ARCH_POSTFIX (V256_U16),TestIntrinsics)332 MY_TEST_P(ARCH_POSTFIX(V256_U16), TestIntrinsics) {
333   TestSimd1Arg<c_v256, uint16_t>(kIterations, mask, maskwidth, name);
334 }
335 
MY_TEST_P(ARCH_POSTFIX (V256_U32),TestIntrinsics)336 MY_TEST_P(ARCH_POSTFIX(V256_U32), TestIntrinsics) {
337   TestSimd1Arg<c_v256, uint32_t>(kIterations, mask, maskwidth, name);
338 }
339 
MY_TEST_P(ARCH_POSTFIX (V256_U64),TestIntrinsics)340 MY_TEST_P(ARCH_POSTFIX(V256_U64), TestIntrinsics) {
341   TestSimd1Arg<c_v256, uint64_t>(kIterations, mask, maskwidth, name);
342 }
343 
MY_TEST_P(ARCH_POSTFIX (U32_V256),TestIntrinsics)344 MY_TEST_P(ARCH_POSTFIX(U32_V256), TestIntrinsics) {
345   TestSimd1Arg<uint32_t, c_v256>(kIterations, mask, maskwidth, name);
346 }
347 
MY_TEST_P(ARCH_POSTFIX (V64_V256),TestIntrinsics)348 MY_TEST_P(ARCH_POSTFIX(V64_V256), TestIntrinsics) {
349   TestSimd1Arg<c_v64, c_v256>(kIterations, mask, maskwidth, name);
350 }
351 
352 // Add a macro layer since INSTANTIATE_TEST_SUITE_P will quote the name
353 // so we need to expand it first with the prefix
354 #define INSTANTIATE(name, type, ...) \
355   INSTANTIATE_TEST_SUITE_P(name, type, ::testing::Values(__VA_ARGS__))
356 
357 #define SIMD_TUPLE(name, mask, maskwidth) \
358   std::make_tuple(mask, maskwidth, static_cast<const char *>(#name))
359 
360 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V64V64), SIMD_TUPLE(v64_sad_u8, 0U, 0U),
361             SIMD_TUPLE(v64_ssd_u8, 0U, 0U));
362 
363 INSTANTIATE(
364     ARCH, ARCH_POSTFIX(V64_V64V64), SIMD_TUPLE(v64_add_8, 0U, 0U),
365     SIMD_TUPLE(v64_add_16, 0U, 0U), SIMD_TUPLE(v64_sadd_s16, 0U, 0U),
366     SIMD_TUPLE(v64_add_32, 0U, 0U), SIMD_TUPLE(v64_sub_8, 0U, 0U),
367     SIMD_TUPLE(v64_ssub_u8, 0U, 0U), SIMD_TUPLE(v64_ssub_s8, 0U, 0U),
368     SIMD_TUPLE(v64_sub_16, 0U, 0U), SIMD_TUPLE(v64_ssub_s16, 0U, 0U),
369     SIMD_TUPLE(v64_ssub_u16, 0U, 0U), SIMD_TUPLE(v64_sub_32, 0U, 0U),
370     SIMD_TUPLE(v64_ziplo_8, 0U, 0U), SIMD_TUPLE(v64_ziphi_8, 0U, 0U),
371     SIMD_TUPLE(v64_ziplo_16, 0U, 0U), SIMD_TUPLE(v64_ziphi_16, 0U, 0U),
372     SIMD_TUPLE(v64_ziplo_32, 0U, 0U), SIMD_TUPLE(v64_ziphi_32, 0U, 0U),
373     SIMD_TUPLE(v64_pack_s32_s16, 0U, 0U), SIMD_TUPLE(v64_pack_s16_u8, 0U, 0U),
374     SIMD_TUPLE(v64_pack_s16_s8, 0U, 0U), SIMD_TUPLE(v64_unziphi_8, 0U, 0U),
375     SIMD_TUPLE(v64_unziplo_8, 0U, 0U), SIMD_TUPLE(v64_unziphi_16, 0U, 0U),
376     SIMD_TUPLE(v64_unziplo_16, 0U, 0U), SIMD_TUPLE(v64_or, 0U, 0U),
377     SIMD_TUPLE(v64_xor, 0U, 0U), SIMD_TUPLE(v64_and, 0U, 0U),
378     SIMD_TUPLE(v64_andn, 0U, 0U), SIMD_TUPLE(v64_mullo_s16, 0U, 0U),
379     SIMD_TUPLE(v64_mulhi_s16, 0U, 0U), SIMD_TUPLE(v64_mullo_s32, 0U, 0U),
380     SIMD_TUPLE(v64_madd_s16, 0U, 0U), SIMD_TUPLE(v64_madd_us8, 0U, 0U),
381     SIMD_TUPLE(v64_avg_u8, 0U, 0U), SIMD_TUPLE(v64_rdavg_u8, 0U, 0U),
382     SIMD_TUPLE(v64_avg_u16, 0U, 0U), SIMD_TUPLE(v64_min_u8, 0U, 0U),
383     SIMD_TUPLE(v64_max_u8, 0U, 0U), SIMD_TUPLE(v64_min_s8, 0U, 0U),
384     SIMD_TUPLE(v64_max_s8, 0U, 0U), SIMD_TUPLE(v64_min_s16, 0U, 0U),
385     SIMD_TUPLE(v64_max_s16, 0U, 0U), SIMD_TUPLE(v64_cmpgt_s8, 0U, 0U),
386     SIMD_TUPLE(v64_cmplt_s8, 0U, 0U), SIMD_TUPLE(v64_cmpeq_8, 0U, 0U),
387     SIMD_TUPLE(v64_cmpgt_s16, 0U, 0U), SIMD_TUPLE(v64_cmplt_s16, 0U, 0U),
388     SIMD_TUPLE(v64_cmpeq_16, 0U, 0U));
389 
390 INSTANTIATE(
391     ARCH, ARCH_POSTFIX(V64_V64V64_Part2), SIMD_TUPLE(v64_shuffle_8, 7U, 8U),
392     SIMD_TUPLE(v64_pack_s32_u16, 0U, 0U), SIMD_TUPLE(v64_rdavg_u16, 0U, 0U),
393     SIMD_TUPLE(v64_sadd_s8, 0U, 0U), SIMD_TUPLE(v64_sadd_u8, 0U, 0U),
394     SIMD_TUPLE(imm_v64_align<1>, 0U, 0U), SIMD_TUPLE(imm_v64_align<2>, 0U, 0U),
395     SIMD_TUPLE(imm_v64_align<3>, 0U, 0U), SIMD_TUPLE(imm_v64_align<4>, 0U, 0U),
396     SIMD_TUPLE(imm_v64_align<5>, 0U, 0U), SIMD_TUPLE(imm_v64_align<6>, 0U, 0U),
397     SIMD_TUPLE(imm_v64_align<7>, 0U, 0U));
398 
399 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64), SIMD_TUPLE(v64_abs_s8, 0U, 0U),
400             SIMD_TUPLE(v64_abs_s16, 0U, 0U),
401             SIMD_TUPLE(v64_unpacklo_u8_s16, 0U, 0U),
402             SIMD_TUPLE(v64_unpackhi_u8_s16, 0U, 0U),
403             SIMD_TUPLE(v64_unpacklo_s8_s16, 0U, 0U),
404             SIMD_TUPLE(v64_unpackhi_s8_s16, 0U, 0U),
405             SIMD_TUPLE(v64_unpacklo_u16_s32, 0U, 0U),
406             SIMD_TUPLE(v64_unpacklo_s16_s32, 0U, 0U),
407             SIMD_TUPLE(v64_unpackhi_u16_s32, 0U, 0U),
408             SIMD_TUPLE(v64_unpackhi_s16_s32, 0U, 0U),
409             SIMD_TUPLE(imm_v64_shr_n_byte<1>, 0U, 0U),
410             SIMD_TUPLE(imm_v64_shr_n_byte<2>, 0U, 0U),
411             SIMD_TUPLE(imm_v64_shr_n_byte<3>, 0U, 0U),
412             SIMD_TUPLE(imm_v64_shr_n_byte<4>, 0U, 0U),
413             SIMD_TUPLE(imm_v64_shr_n_byte<5>, 0U, 0U),
414             SIMD_TUPLE(imm_v64_shr_n_byte<6>, 0U, 0U),
415             SIMD_TUPLE(imm_v64_shr_n_byte<7>, 0U, 0U),
416             SIMD_TUPLE(imm_v64_shl_n_byte<1>, 0U, 0U),
417             SIMD_TUPLE(imm_v64_shl_n_byte<2>, 0U, 0U),
418             SIMD_TUPLE(imm_v64_shl_n_byte<3>, 0U, 0U),
419             SIMD_TUPLE(imm_v64_shl_n_byte<4>, 0U, 0U),
420             SIMD_TUPLE(imm_v64_shl_n_byte<5>, 0U, 0U),
421             SIMD_TUPLE(imm_v64_shl_n_byte<6>, 0U, 0U),
422             SIMD_TUPLE(imm_v64_shl_n_byte<7>, 0U, 0U),
423             SIMD_TUPLE(imm_v64_shl_n_8<1>, 0U, 0U),
424             SIMD_TUPLE(imm_v64_shl_n_8<2>, 0U, 0U),
425             SIMD_TUPLE(imm_v64_shl_n_8<3>, 0U, 0U),
426             SIMD_TUPLE(imm_v64_shl_n_8<4>, 0U, 0U),
427             SIMD_TUPLE(imm_v64_shl_n_8<5>, 0U, 0U),
428             SIMD_TUPLE(imm_v64_shl_n_8<6>, 0U, 0U),
429             SIMD_TUPLE(imm_v64_shl_n_8<7>, 0U, 0U),
430             SIMD_TUPLE(imm_v64_shr_n_u8<1>, 0U, 0U),
431             SIMD_TUPLE(imm_v64_shr_n_u8<2>, 0U, 0U),
432             SIMD_TUPLE(imm_v64_shr_n_u8<3>, 0U, 0U),
433             SIMD_TUPLE(imm_v64_shr_n_u8<4>, 0U, 0U),
434             SIMD_TUPLE(imm_v64_shr_n_u8<5>, 0U, 0U),
435             SIMD_TUPLE(imm_v64_shr_n_u8<6>, 0U, 0U),
436             SIMD_TUPLE(imm_v64_shr_n_u8<7>, 0U, 0U),
437             SIMD_TUPLE(imm_v64_shr_n_s8<1>, 0U, 0U),
438             SIMD_TUPLE(imm_v64_shr_n_s8<2>, 0U, 0U),
439             SIMD_TUPLE(imm_v64_shr_n_s8<3>, 0U, 0U),
440             SIMD_TUPLE(imm_v64_shr_n_s8<4>, 0U, 0U),
441             SIMD_TUPLE(imm_v64_shr_n_s8<5>, 0U, 0U),
442             SIMD_TUPLE(imm_v64_shr_n_s8<6>, 0U, 0U),
443             SIMD_TUPLE(imm_v64_shr_n_s8<7>, 0U, 0U),
444             SIMD_TUPLE(imm_v64_shl_n_16<1>, 0U, 0U),
445             SIMD_TUPLE(imm_v64_shl_n_16<2>, 0U, 0U),
446             SIMD_TUPLE(imm_v64_shl_n_16<4>, 0U, 0U),
447             SIMD_TUPLE(imm_v64_shl_n_16<6>, 0U, 0U),
448             SIMD_TUPLE(imm_v64_shl_n_16<8>, 0U, 0U));
449 
450 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64_Part2),
451             SIMD_TUPLE(imm_v64_shl_n_16<10>, 0U, 0U),
452             SIMD_TUPLE(imm_v64_shl_n_16<12>, 0U, 0U),
453             SIMD_TUPLE(imm_v64_shl_n_16<14>, 0U, 0U),
454             SIMD_TUPLE(imm_v64_shr_n_u16<1>, 0U, 0U),
455             SIMD_TUPLE(imm_v64_shr_n_u16<2>, 0U, 0U),
456             SIMD_TUPLE(imm_v64_shr_n_u16<4>, 0U, 0U),
457             SIMD_TUPLE(imm_v64_shr_n_u16<6>, 0U, 0U),
458             SIMD_TUPLE(imm_v64_shr_n_u16<8>, 0U, 0U),
459             SIMD_TUPLE(imm_v64_shr_n_u16<10>, 0U, 0U),
460             SIMD_TUPLE(imm_v64_shr_n_u16<12>, 0U, 0U),
461             SIMD_TUPLE(imm_v64_shr_n_u16<14>, 0U, 0U),
462             SIMD_TUPLE(imm_v64_shr_n_s16<1>, 0U, 0U),
463             SIMD_TUPLE(imm_v64_shr_n_s16<2>, 0U, 0U),
464             SIMD_TUPLE(imm_v64_shr_n_s16<4>, 0U, 0U),
465             SIMD_TUPLE(imm_v64_shr_n_s16<6>, 0U, 0U),
466             SIMD_TUPLE(imm_v64_shr_n_s16<8>, 0U, 0U),
467             SIMD_TUPLE(imm_v64_shr_n_s16<10>, 0U, 0U),
468             SIMD_TUPLE(imm_v64_shr_n_s16<12>, 0U, 0U),
469             SIMD_TUPLE(imm_v64_shr_n_s16<14>, 0U, 0U),
470             SIMD_TUPLE(imm_v64_shl_n_32<1>, 0U, 0U),
471             SIMD_TUPLE(imm_v64_shl_n_32<4>, 0U, 0U),
472             SIMD_TUPLE(imm_v64_shl_n_32<8>, 0U, 0U),
473             SIMD_TUPLE(imm_v64_shl_n_32<12>, 0U, 0U),
474             SIMD_TUPLE(imm_v64_shl_n_32<16>, 0U, 0U),
475             SIMD_TUPLE(imm_v64_shl_n_32<20>, 0U, 0U),
476             SIMD_TUPLE(imm_v64_shl_n_32<24>, 0U, 0U),
477             SIMD_TUPLE(imm_v64_shl_n_32<28>, 0U, 0U),
478             SIMD_TUPLE(imm_v64_shr_n_u32<1>, 0U, 0U),
479             SIMD_TUPLE(imm_v64_shr_n_u32<4>, 0U, 0U),
480             SIMD_TUPLE(imm_v64_shr_n_u32<8>, 0U, 0U),
481             SIMD_TUPLE(imm_v64_shr_n_u32<12>, 0U, 0U),
482             SIMD_TUPLE(imm_v64_shr_n_u32<16>, 0U, 0U),
483             SIMD_TUPLE(imm_v64_shr_n_u32<20>, 0U, 0U),
484             SIMD_TUPLE(imm_v64_shr_n_u32<24>, 0U, 0U),
485             SIMD_TUPLE(imm_v64_shr_n_u32<28>, 0U, 0U),
486             SIMD_TUPLE(imm_v64_shr_n_s32<1>, 0U, 0U),
487             SIMD_TUPLE(imm_v64_shr_n_s32<4>, 0U, 0U),
488             SIMD_TUPLE(imm_v64_shr_n_s32<8>, 0U, 0U),
489             SIMD_TUPLE(imm_v64_shr_n_s32<12>, 0U, 0U),
490             SIMD_TUPLE(imm_v64_shr_n_s32<16>, 0U, 0U),
491             SIMD_TUPLE(imm_v64_shr_n_s32<20>, 0U, 0U),
492             SIMD_TUPLE(imm_v64_shr_n_s32<24>, 0U, 0U),
493             SIMD_TUPLE(imm_v64_shr_n_s32<28>, 0U, 0U));
494 
495 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64U32), SIMD_TUPLE(v64_shl_8, 7U, 32U),
496             SIMD_TUPLE(v64_shr_u8, 7U, 32U), SIMD_TUPLE(v64_shr_s8, 7U, 32U),
497             SIMD_TUPLE(v64_shl_16, 15U, 32U), SIMD_TUPLE(v64_shr_u16, 15U, 32U),
498             SIMD_TUPLE(v64_shr_s16, 15U, 32U), SIMD_TUPLE(v64_shl_32, 31U, 32U),
499             SIMD_TUPLE(v64_shr_u32, 31U, 32U),
500             SIMD_TUPLE(v64_shr_s32, 31U, 32U));
501 
502 INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V64), SIMD_TUPLE(v64_hadd_u8, 0U, 0U),
503             SIMD_TUPLE(v64_u64, 0U, 0U));
504 
505 INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V64), SIMD_TUPLE(v64_hadd_s16, 0U, 0U));
506 
507 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V64), SIMD_TUPLE(v64_low_u32, 0U, 0U),
508             SIMD_TUPLE(v64_high_u32, 0U, 0U));
509 
510 INSTANTIATE(ARCH, ARCH_POSTFIX(S32_V64), SIMD_TUPLE(v64_low_s32, 0U, 0U),
511             SIMD_TUPLE(v64_high_s32, 0U, 0U));
512 
513 INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V64V64), SIMD_TUPLE(v64_dotp_s16, 0U, 0U),
514             SIMD_TUPLE(v64_dotp_su8, 0U, 0U));
515 
516 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U8), SIMD_TUPLE(v64_dup_8, 0U, 0U));
517 
518 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U16), SIMD_TUPLE(v64_dup_16, 0U, 0U));
519 
520 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U32), SIMD_TUPLE(v64_dup_32, 0U, 0U));
521 
522 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U32U32), SIMD_TUPLE(v64_from_32, 0U, 0U));
523 
524 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V128V128), SIMD_TUPLE(v128_sad_u8, 0U, 0U),
525             SIMD_TUPLE(v128_ssd_u8, 0U, 0U), SIMD_TUPLE(v128_sad_u16, 0U, 0U));
526 INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V128V128), SIMD_TUPLE(v128_ssd_s16, 0U, 0U));
527 
528 INSTANTIATE(
529     ARCH, ARCH_POSTFIX(V128_V128V128), SIMD_TUPLE(v128_add_8, 0U, 0U),
530     SIMD_TUPLE(v128_add_16, 0U, 0U), SIMD_TUPLE(v128_sadd_s16, 0U, 0U),
531     SIMD_TUPLE(v128_add_32, 0U, 0U), SIMD_TUPLE(v128_sub_8, 0U, 0U),
532     SIMD_TUPLE(v128_ssub_u8, 0U, 0U), SIMD_TUPLE(v128_ssub_s8, 0U, 0U),
533     SIMD_TUPLE(v128_sub_16, 0U, 0U), SIMD_TUPLE(v128_ssub_s16, 0U, 0U),
534     SIMD_TUPLE(v128_ssub_u16, 0U, 0U), SIMD_TUPLE(v128_sub_32, 0U, 0U),
535     SIMD_TUPLE(v128_ziplo_8, 0U, 0U), SIMD_TUPLE(v128_ziphi_8, 0U, 0U),
536     SIMD_TUPLE(v128_ziplo_16, 0U, 0U), SIMD_TUPLE(v128_ziphi_16, 0U, 0U),
537     SIMD_TUPLE(v128_ziplo_32, 0U, 0U), SIMD_TUPLE(v128_ziphi_32, 0U, 0U),
538     SIMD_TUPLE(v128_ziplo_64, 0U, 0U), SIMD_TUPLE(v128_ziphi_64, 0U, 0U),
539     SIMD_TUPLE(v128_unziphi_8, 0U, 0U), SIMD_TUPLE(v128_unziplo_8, 0U, 0U),
540     SIMD_TUPLE(v128_unziphi_16, 0U, 0U), SIMD_TUPLE(v128_unziplo_16, 0U, 0U),
541     SIMD_TUPLE(v128_unziphi_32, 0U, 0U), SIMD_TUPLE(v128_unziplo_32, 0U, 0U),
542     SIMD_TUPLE(v128_pack_s32_s16, 0U, 0U), SIMD_TUPLE(v128_pack_s16_u8, 0U, 0U),
543     SIMD_TUPLE(v128_pack_s16_s8, 0U, 0U), SIMD_TUPLE(v128_or, 0U, 0U),
544     SIMD_TUPLE(v128_xor, 0U, 0U), SIMD_TUPLE(v128_and, 0U, 0U),
545     SIMD_TUPLE(v128_andn, 0U, 0U), SIMD_TUPLE(v128_mullo_s16, 0U, 0U),
546     SIMD_TUPLE(v128_mulhi_s16, 0U, 0U), SIMD_TUPLE(v128_mullo_s32, 0U, 0U),
547     SIMD_TUPLE(v128_madd_s16, 0U, 0U), SIMD_TUPLE(v128_madd_us8, 0U, 0U),
548     SIMD_TUPLE(v128_avg_u8, 0U, 0U), SIMD_TUPLE(v128_rdavg_u8, 0U, 0U),
549     SIMD_TUPLE(v128_avg_u16, 0U, 0U), SIMD_TUPLE(v128_min_u8, 0U, 0U),
550     SIMD_TUPLE(v128_max_u8, 0U, 0U), SIMD_TUPLE(v128_min_s8, 0U, 0U),
551     SIMD_TUPLE(v128_max_s8, 0U, 0U), SIMD_TUPLE(v128_min_s16, 0U, 0U),
552     SIMD_TUPLE(v128_max_s16, 0U, 0U), SIMD_TUPLE(v128_cmpgt_s8, 0U, 0U),
553     SIMD_TUPLE(v128_cmplt_s8, 0U, 0U), SIMD_TUPLE(v128_cmpeq_8, 0U, 0U),
554     SIMD_TUPLE(v128_cmpgt_s16, 0U, 0U));
555 
556 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128V128_Part2),
557             SIMD_TUPLE(v128_pack_s32_u16, 0U, 0U),
558             SIMD_TUPLE(v128_rdavg_u16, 0U, 0U), SIMD_TUPLE(v128_add_64, 0U, 0U),
559             SIMD_TUPLE(v128_sub_64, 0U, 0U), SIMD_TUPLE(v128_sadd_s8, 0U, 0U),
560             SIMD_TUPLE(v128_sadd_u8, 0U, 0U), SIMD_TUPLE(v128_cmpeq_16, 0U, 0U),
561             SIMD_TUPLE(v128_cmplt_s16, 0U, 0U),
562             SIMD_TUPLE(v128_cmplt_s32, 0U, 0U),
563             SIMD_TUPLE(v128_cmpeq_32, 0U, 0U),
564             SIMD_TUPLE(v128_cmpgt_s32, 0U, 0U),
565             SIMD_TUPLE(v128_shuffle_8, 15U, 8U),
566             SIMD_TUPLE(v128_min_s32, 0U, 0U), SIMD_TUPLE(v128_max_s32, 0U, 0U),
567             SIMD_TUPLE(imm_v128_align<1>, 0U, 0U),
568             SIMD_TUPLE(imm_v128_align<2>, 0U, 0U),
569             SIMD_TUPLE(imm_v128_align<3>, 0U, 0U),
570             SIMD_TUPLE(imm_v128_align<4>, 0U, 0U),
571             SIMD_TUPLE(imm_v128_align<5>, 0U, 0U),
572             SIMD_TUPLE(imm_v128_align<6>, 0U, 0U),
573             SIMD_TUPLE(imm_v128_align<7>, 0U, 0U),
574             SIMD_TUPLE(imm_v128_align<8>, 0U, 0U),
575             SIMD_TUPLE(imm_v128_align<9>, 0U, 0U),
576             SIMD_TUPLE(imm_v128_align<10>, 0U, 0U),
577             SIMD_TUPLE(imm_v128_align<11>, 0U, 0U),
578             SIMD_TUPLE(imm_v128_align<12>, 0U, 0U),
579             SIMD_TUPLE(imm_v128_align<13>, 0U, 0U),
580             SIMD_TUPLE(imm_v128_align<14>, 0U, 0U),
581             SIMD_TUPLE(imm_v128_align<15>, 0U, 0U));
582 
583 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128V128V128),
584             SIMD_TUPLE(v128_blend_8, 0U, 0U));
585 
586 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128), SIMD_TUPLE(v128_abs_s8, 0U, 0U),
587             SIMD_TUPLE(v128_abs_s16, 0U, 0U), SIMD_TUPLE(v128_padd_s16, 0U, 0U),
588             SIMD_TUPLE(v128_unpacklo_u8_s16, 0U, 0U),
589             SIMD_TUPLE(v128_unpacklo_s8_s16, 0U, 0U),
590             SIMD_TUPLE(v128_unpacklo_u16_s32, 0U, 0U),
591             SIMD_TUPLE(v128_unpacklo_s16_s32, 0U, 0U),
592             SIMD_TUPLE(v128_unpackhi_u8_s16, 0U, 0U),
593             SIMD_TUPLE(v128_unpackhi_s8_s16, 0U, 0U),
594             SIMD_TUPLE(v128_unpackhi_u16_s32, 0U, 0U),
595             SIMD_TUPLE(v128_unpackhi_s16_s32, 0U, 0U),
596             SIMD_TUPLE(imm_v128_shr_n_byte<1>, 0U, 0U),
597             SIMD_TUPLE(imm_v128_shr_n_byte<2>, 0U, 0U),
598             SIMD_TUPLE(imm_v128_shr_n_byte<3>, 0U, 0U),
599             SIMD_TUPLE(imm_v128_shr_n_byte<4>, 0U, 0U),
600             SIMD_TUPLE(imm_v128_shr_n_byte<5>, 0U, 0U),
601             SIMD_TUPLE(imm_v128_shr_n_byte<6>, 0U, 0U),
602             SIMD_TUPLE(imm_v128_shr_n_byte<7>, 0U, 0U),
603             SIMD_TUPLE(imm_v128_shr_n_byte<8>, 0U, 0U),
604             SIMD_TUPLE(imm_v128_shr_n_byte<9>, 0U, 0U),
605             SIMD_TUPLE(imm_v128_shr_n_byte<10>, 0U, 0U),
606             SIMD_TUPLE(imm_v128_shr_n_byte<11>, 0U, 0U),
607             SIMD_TUPLE(imm_v128_shr_n_byte<12>, 0U, 0U),
608             SIMD_TUPLE(imm_v128_shr_n_byte<13>, 0U, 0U),
609             SIMD_TUPLE(imm_v128_shr_n_byte<14>, 0U, 0U),
610             SIMD_TUPLE(imm_v128_shr_n_byte<15>, 0U, 0U),
611             SIMD_TUPLE(imm_v128_shl_n_byte<1>, 0U, 0U),
612             SIMD_TUPLE(imm_v128_shl_n_byte<2>, 0U, 0U),
613             SIMD_TUPLE(imm_v128_shl_n_byte<3>, 0U, 0U),
614             SIMD_TUPLE(imm_v128_shl_n_byte<4>, 0U, 0U),
615             SIMD_TUPLE(imm_v128_shl_n_byte<5>, 0U, 0U),
616             SIMD_TUPLE(imm_v128_shl_n_byte<6>, 0U, 0U),
617             SIMD_TUPLE(imm_v128_shl_n_byte<7>, 0U, 0U),
618             SIMD_TUPLE(imm_v128_shl_n_byte<8>, 0U, 0U),
619             SIMD_TUPLE(imm_v128_shl_n_byte<9>, 0U, 0U),
620             SIMD_TUPLE(imm_v128_shl_n_byte<10>, 0U, 0U),
621             SIMD_TUPLE(imm_v128_shl_n_byte<11>, 0U, 0U),
622             SIMD_TUPLE(imm_v128_shl_n_byte<12>, 0U, 0U),
623             SIMD_TUPLE(imm_v128_shl_n_byte<13>, 0U, 0U),
624             SIMD_TUPLE(imm_v128_shl_n_byte<14>, 0U, 0U),
625             SIMD_TUPLE(imm_v128_shl_n_byte<15>, 0U, 0U),
626             SIMD_TUPLE(imm_v128_shl_n_8<1>, 0U, 0U),
627             SIMD_TUPLE(imm_v128_shl_n_8<2>, 0U, 0U),
628             SIMD_TUPLE(imm_v128_shl_n_8<3>, 0U, 0U),
629             SIMD_TUPLE(imm_v128_shl_n_8<4>, 0U, 0U),
630             SIMD_TUPLE(imm_v128_shl_n_8<5>, 0U, 0U),
631             SIMD_TUPLE(imm_v128_shl_n_8<6>, 0U, 0U),
632             SIMD_TUPLE(imm_v128_shl_n_8<7>, 0U, 0U),
633             SIMD_TUPLE(imm_v128_shr_n_u8<1>, 0U, 0U));
634 
635 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128_Part2),
636             SIMD_TUPLE(imm_v128_shr_n_u8<2>, 0U, 0U),
637             SIMD_TUPLE(imm_v128_shr_n_u8<3>, 0U, 0U),
638             SIMD_TUPLE(imm_v128_shr_n_u8<4>, 0U, 0U),
639             SIMD_TUPLE(imm_v128_shr_n_u8<5>, 0U, 0U),
640             SIMD_TUPLE(imm_v128_shr_n_u8<6>, 0U, 0U),
641             SIMD_TUPLE(imm_v128_shr_n_u8<7>, 0U, 0U),
642             SIMD_TUPLE(imm_v128_shr_n_s8<1>, 0U, 0U),
643             SIMD_TUPLE(imm_v128_shr_n_s8<2>, 0U, 0U),
644             SIMD_TUPLE(imm_v128_shr_n_s8<3>, 0U, 0U),
645             SIMD_TUPLE(imm_v128_shr_n_s8<4>, 0U, 0U),
646             SIMD_TUPLE(imm_v128_shr_n_s8<5>, 0U, 0U),
647             SIMD_TUPLE(imm_v128_shr_n_s8<6>, 0U, 0U),
648             SIMD_TUPLE(imm_v128_shr_n_s8<7>, 0U, 0U),
649             SIMD_TUPLE(imm_v128_shl_n_16<1>, 0U, 0U),
650             SIMD_TUPLE(imm_v128_shl_n_16<2>, 0U, 0U),
651             SIMD_TUPLE(imm_v128_shl_n_16<4>, 0U, 0U),
652             SIMD_TUPLE(imm_v128_shl_n_16<6>, 0U, 0U),
653             SIMD_TUPLE(imm_v128_shl_n_16<8>, 0U, 0U),
654             SIMD_TUPLE(imm_v128_shl_n_16<10>, 0U, 0U),
655             SIMD_TUPLE(imm_v128_shl_n_16<12>, 0U, 0U),
656             SIMD_TUPLE(imm_v128_shl_n_16<14>, 0U, 0U),
657             SIMD_TUPLE(imm_v128_shr_n_u16<1>, 0U, 0U),
658             SIMD_TUPLE(imm_v128_shr_n_u16<2>, 0U, 0U),
659             SIMD_TUPLE(imm_v128_shr_n_u16<4>, 0U, 0U),
660             SIMD_TUPLE(imm_v128_shr_n_u16<6>, 0U, 0U),
661             SIMD_TUPLE(imm_v128_shr_n_u16<8>, 0U, 0U),
662             SIMD_TUPLE(imm_v128_shr_n_u16<10>, 0U, 0U),
663             SIMD_TUPLE(imm_v128_shr_n_u16<12>, 0U, 0U),
664             SIMD_TUPLE(imm_v128_shr_n_u16<14>, 0U, 0U),
665             SIMD_TUPLE(imm_v128_shr_n_s16<1>, 0U, 0U),
666             SIMD_TUPLE(imm_v128_shr_n_s16<2>, 0U, 0U),
667             SIMD_TUPLE(imm_v128_shr_n_s16<4>, 0U, 0U),
668             SIMD_TUPLE(imm_v128_shr_n_s16<6>, 0U, 0U),
669             SIMD_TUPLE(imm_v128_shr_n_s16<8>, 0U, 0U),
670             SIMD_TUPLE(imm_v128_shr_n_s16<10>, 0U, 0U),
671             SIMD_TUPLE(imm_v128_shr_n_s16<12>, 0U, 0U),
672             SIMD_TUPLE(imm_v128_shr_n_s16<14>, 0U, 0U),
673             SIMD_TUPLE(imm_v128_shl_n_32<1>, 0U, 0U),
674             SIMD_TUPLE(imm_v128_shl_n_32<4>, 0U, 0U),
675             SIMD_TUPLE(imm_v128_shl_n_32<8>, 0U, 0U),
676             SIMD_TUPLE(imm_v128_shl_n_32<12>, 0U, 0U),
677             SIMD_TUPLE(imm_v128_shl_n_32<16>, 0U, 0U),
678             SIMD_TUPLE(imm_v128_shl_n_32<20>, 0U, 0U),
679             SIMD_TUPLE(imm_v128_shl_n_32<24>, 0U, 0U),
680             SIMD_TUPLE(imm_v128_shl_n_32<28>, 0U, 0U),
681             SIMD_TUPLE(imm_v128_shr_n_u32<1>, 0U, 0U),
682             SIMD_TUPLE(imm_v128_shr_n_u32<4>, 0U, 0U));
683 
684 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128_Part3),
685             SIMD_TUPLE(imm_v128_shr_n_u32<8>, 0U, 0U),
686             SIMD_TUPLE(imm_v128_shr_n_u32<12>, 0U, 0U),
687             SIMD_TUPLE(imm_v128_shr_n_u32<16>, 0U, 0U),
688             SIMD_TUPLE(imm_v128_shr_n_u32<20>, 0U, 0U),
689             SIMD_TUPLE(imm_v128_shr_n_u32<24>, 0U, 0U),
690             SIMD_TUPLE(imm_v128_shr_n_u32<28>, 0U, 0U),
691             SIMD_TUPLE(imm_v128_shr_n_s32<1>, 0U, 0U),
692             SIMD_TUPLE(imm_v128_shr_n_s32<4>, 0U, 0U),
693             SIMD_TUPLE(imm_v128_shr_n_s32<8>, 0U, 0U),
694             SIMD_TUPLE(imm_v128_shr_n_s32<12>, 0U, 0U),
695             SIMD_TUPLE(imm_v128_shr_n_s32<16>, 0U, 0U),
696             SIMD_TUPLE(imm_v128_shr_n_s32<20>, 0U, 0U),
697             SIMD_TUPLE(imm_v128_shr_n_s32<24>, 0U, 0U),
698             SIMD_TUPLE(imm_v128_shr_n_s32<28>, 0U, 0U));
699 
700 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128_Part4),
701             SIMD_TUPLE(imm_v128_shl_n_64<1>, 0U, 0U),
702             SIMD_TUPLE(imm_v128_shl_n_64<4>, 0U, 0U),
703             SIMD_TUPLE(imm_v128_shl_n_64<8>, 0U, 0U),
704             SIMD_TUPLE(imm_v128_shl_n_64<12>, 0U, 0U),
705             SIMD_TUPLE(imm_v128_shl_n_64<16>, 0U, 0U),
706             SIMD_TUPLE(imm_v128_shl_n_64<20>, 0U, 0U),
707             SIMD_TUPLE(imm_v128_shl_n_64<24>, 0U, 0U),
708             SIMD_TUPLE(imm_v128_shl_n_64<28>, 0U, 0U),
709             SIMD_TUPLE(imm_v128_shl_n_64<32>, 0U, 0U),
710             SIMD_TUPLE(imm_v128_shl_n_64<36>, 0U, 0U),
711             SIMD_TUPLE(imm_v128_shl_n_64<40>, 0U, 0U),
712             SIMD_TUPLE(imm_v128_shl_n_64<44>, 0U, 0U),
713             SIMD_TUPLE(imm_v128_shl_n_64<48>, 0U, 0U),
714             SIMD_TUPLE(imm_v128_shl_n_64<52>, 0U, 0U),
715             SIMD_TUPLE(imm_v128_shl_n_64<56>, 0U, 0U),
716             SIMD_TUPLE(imm_v128_shl_n_64<60>, 0U, 0U),
717             SIMD_TUPLE(imm_v128_shr_n_u64<1>, 0U, 0U),
718             SIMD_TUPLE(imm_v128_shr_n_u64<4>, 0U, 0U),
719             SIMD_TUPLE(imm_v128_shr_n_u64<8>, 0U, 0U),
720             SIMD_TUPLE(imm_v128_shr_n_u64<12>, 0U, 0U),
721             SIMD_TUPLE(imm_v128_shr_n_u64<16>, 0U, 0U),
722             SIMD_TUPLE(imm_v128_shr_n_u64<20>, 0U, 0U),
723             SIMD_TUPLE(imm_v128_shr_n_u64<24>, 0U, 0U),
724             SIMD_TUPLE(imm_v128_shr_n_u64<28>, 0U, 0U),
725             SIMD_TUPLE(imm_v128_shr_n_u64<32>, 0U, 0U),
726             SIMD_TUPLE(imm_v128_shr_n_u64<36>, 0U, 0U),
727             SIMD_TUPLE(imm_v128_shr_n_u64<40>, 0U, 0U),
728             SIMD_TUPLE(imm_v128_shr_n_u64<44>, 0U, 0U),
729             SIMD_TUPLE(imm_v128_shr_n_u64<48>, 0U, 0U),
730             SIMD_TUPLE(imm_v128_shr_n_u64<52>, 0U, 0U),
731             SIMD_TUPLE(imm_v128_shr_n_u64<56>, 0U, 0U),
732             SIMD_TUPLE(imm_v128_shr_n_u64<60>, 0U, 0U),
733             SIMD_TUPLE(imm_v128_shr_n_s64<1>, 0U, 0U),
734             SIMD_TUPLE(imm_v128_shr_n_s64<4>, 0U, 0U),
735             SIMD_TUPLE(imm_v128_shr_n_s64<8>, 0U, 0U),
736             SIMD_TUPLE(imm_v128_shr_n_s64<12>, 0U, 0U),
737             SIMD_TUPLE(imm_v128_shr_n_s64<16>, 0U, 0U),
738             SIMD_TUPLE(imm_v128_shr_n_s64<20>, 0U, 0U),
739             SIMD_TUPLE(imm_v128_shr_n_s64<24>, 0U, 0U),
740             SIMD_TUPLE(imm_v128_shr_n_s64<28>, 0U, 0U),
741             SIMD_TUPLE(imm_v128_shr_n_s64<32>, 0U, 0U),
742             SIMD_TUPLE(imm_v128_shr_n_s64<36>, 0U, 0U),
743             SIMD_TUPLE(imm_v128_shr_n_s64<40>, 0U, 0U),
744             SIMD_TUPLE(imm_v128_shr_n_s64<44>, 0U, 0U),
745             SIMD_TUPLE(imm_v128_shr_n_s64<48>, 0U, 0U),
746             SIMD_TUPLE(imm_v128_shr_n_s64<52>, 0U, 0U),
747             SIMD_TUPLE(imm_v128_shr_n_s64<56>, 0U, 0U),
748             SIMD_TUPLE(imm_v128_shr_n_s64<60>, 0U, 0U),
749             SIMD_TUPLE(v128_padd_u8, 0U, 0U));
750 
751 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V64V64), SIMD_TUPLE(v128_from_v64, 0U, 0U),
752             SIMD_TUPLE(v128_zip_8, 0U, 0U), SIMD_TUPLE(v128_zip_16, 0U, 0U),
753             SIMD_TUPLE(v128_zip_32, 0U, 0U), SIMD_TUPLE(v128_mul_s16, 0U, 0U));
754 
755 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U64U64), SIMD_TUPLE(v128_from_64, 0U, 0U));
756 
757 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V64),
758             SIMD_TUPLE(v128_unpack_u8_s16, 0U, 0U),
759             SIMD_TUPLE(v128_unpack_s8_s16, 0U, 0U),
760             SIMD_TUPLE(v128_unpack_u16_s32, 0U, 0U),
761             SIMD_TUPLE(v128_unpack_s16_s32, 0U, 0U));
762 
763 INSTANTIATE(
764     ARCH, ARCH_POSTFIX(V128_V128U32), SIMD_TUPLE(v128_shl_8, 7U, 32U),
765     SIMD_TUPLE(v128_shr_u8, 7U, 32U), SIMD_TUPLE(v128_shr_s8, 7U, 32U),
766     SIMD_TUPLE(v128_shl_16, 15U, 32U), SIMD_TUPLE(v128_shr_u16, 15U, 32U),
767     SIMD_TUPLE(v128_shr_s16, 15U, 32U), SIMD_TUPLE(v128_shl_32, 31U, 32U),
768     SIMD_TUPLE(v128_shr_u32, 31U, 32U), SIMD_TUPLE(v128_shr_s32, 31U, 32U),
769     SIMD_TUPLE(v128_shl_64, 63U, 32U), SIMD_TUPLE(v128_shr_u64, 63U, 32U),
770     SIMD_TUPLE(v128_shr_s64, 63U, 32U));
771 
772 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V128), SIMD_TUPLE(v128_low_u32, 0U, 0U),
773             SIMD_TUPLE(v128_movemask_8, 0U, 0U));
774 
775 INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V128), SIMD_TUPLE(v128_hadd_u8, 0U, 0U));
776 
777 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V128), SIMD_TUPLE(v128_low_v64, 0U, 0U),
778             SIMD_TUPLE(v128_high_v64, 0U, 0U));
779 
780 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U8), SIMD_TUPLE(v128_dup_8, 0U, 0U));
781 
782 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U16), SIMD_TUPLE(v128_dup_16, 0U, 0U));
783 
784 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U32), SIMD_TUPLE(v128_dup_32, 0U, 0U));
785 
786 INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U64), SIMD_TUPLE(v128_dup_64, 0U, 0U));
787 
788 INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V128V128), SIMD_TUPLE(v128_dotp_s16, 0U, 0U),
789             SIMD_TUPLE(v128_dotp_s32, 0U, 0U),
790             SIMD_TUPLE(v128_dotp_su8, 0U, 0U));
791 
792 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V256V256), SIMD_TUPLE(v256_sad_u8, 0U, 0U),
793             SIMD_TUPLE(v256_ssd_u8, 0U, 0U), SIMD_TUPLE(v256_sad_u16, 0U, 0U));
794 
795 INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V256), SIMD_TUPLE(v256_hadd_u8, 0U, 0U),
796             SIMD_TUPLE(v256_low_u64, 0U, 0U));
797 
798 INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V256V256), SIMD_TUPLE(v256_dotp_s16, 0U, 0U),
799             SIMD_TUPLE(v256_dotp_s32, 0U, 0U),
800             SIMD_TUPLE(v256_dotp_su8, 0U, 0U));
801 
802 INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V256V256), SIMD_TUPLE(v256_ssd_s16, 0U, 0U));
803 
804 INSTANTIATE(
805     ARCH, ARCH_POSTFIX(V256_V256V256), SIMD_TUPLE(v256_add_8, 0U, 0U),
806     SIMD_TUPLE(v256_add_16, 0U, 0U), SIMD_TUPLE(v256_sadd_s16, 0U, 0U),
807     SIMD_TUPLE(v256_add_32, 0U, 0U), SIMD_TUPLE(v256_sub_8, 0U, 0U),
808     SIMD_TUPLE(v256_ssub_u8, 0U, 0U), SIMD_TUPLE(v256_ssub_s8, 0U, 0U),
809     SIMD_TUPLE(v256_sub_16, 0U, 0U), SIMD_TUPLE(v256_ssub_s16, 0U, 0U),
810     SIMD_TUPLE(v256_ssub_u16, 0U, 0U), SIMD_TUPLE(v256_sub_32, 0U, 0U),
811     SIMD_TUPLE(v256_ziplo_8, 0U, 0U), SIMD_TUPLE(v256_ziphi_8, 0U, 0U),
812     SIMD_TUPLE(v256_ziplo_16, 0U, 0U), SIMD_TUPLE(v256_ziphi_16, 0U, 0U),
813     SIMD_TUPLE(v256_ziplo_32, 0U, 0U), SIMD_TUPLE(v256_ziphi_32, 0U, 0U),
814     SIMD_TUPLE(v256_ziplo_64, 0U, 0U), SIMD_TUPLE(v256_ziphi_64, 0U, 0U),
815     SIMD_TUPLE(v256_ziplo_128, 0U, 0U), SIMD_TUPLE(v256_ziphi_128, 0U, 0U),
816     SIMD_TUPLE(v256_unziphi_8, 0U, 0U), SIMD_TUPLE(v256_unziplo_8, 0U, 0U),
817     SIMD_TUPLE(v256_unziphi_16, 0U, 0U), SIMD_TUPLE(v256_unziplo_16, 0U, 0U),
818     SIMD_TUPLE(v256_unziphi_32, 0U, 0U), SIMD_TUPLE(v256_unziplo_32, 0U, 0U),
819     SIMD_TUPLE(v256_pack_s32_s16, 0U, 0U), SIMD_TUPLE(v256_pack_s16_u8, 0U, 0U),
820     SIMD_TUPLE(v256_pack_s16_s8, 0U, 0U), SIMD_TUPLE(v256_or, 0U, 0U),
821     SIMD_TUPLE(v256_xor, 0U, 0U), SIMD_TUPLE(v256_and, 0U, 0U),
822     SIMD_TUPLE(v256_andn, 0U, 0U), SIMD_TUPLE(v256_mullo_s16, 0U, 0U),
823     SIMD_TUPLE(v256_mulhi_s16, 0U, 0U), SIMD_TUPLE(v256_mullo_s32, 0U, 0U),
824     SIMD_TUPLE(v256_madd_s16, 0U, 0U), SIMD_TUPLE(v256_madd_us8, 0U, 0U),
825     SIMD_TUPLE(v256_avg_u8, 0U, 0U), SIMD_TUPLE(v256_rdavg_u8, 0U, 0U),
826     SIMD_TUPLE(v256_avg_u16, 0U, 0U), SIMD_TUPLE(v256_min_u8, 0U, 0U),
827     SIMD_TUPLE(v256_max_u8, 0U, 0U), SIMD_TUPLE(v256_min_s8, 0U, 0U),
828     SIMD_TUPLE(v256_max_s8, 0U, 0U), SIMD_TUPLE(v256_min_s16, 0U, 0U),
829     SIMD_TUPLE(v256_max_s16, 0U, 0U), SIMD_TUPLE(v256_cmpgt_s8, 0U, 0U),
830     SIMD_TUPLE(v256_cmplt_s8, 0U, 0U));
831 
832 INSTANTIATE(
833     ARCH, ARCH_POSTFIX(V256_V256V256_Part2), SIMD_TUPLE(v256_cmpeq_8, 0U, 0U),
834     SIMD_TUPLE(v256_min_s32, 0U, 0U), SIMD_TUPLE(v256_max_s32, 0U, 0U),
835     SIMD_TUPLE(v256_add_64, 0U, 0U), SIMD_TUPLE(v256_sub_64, 0U, 0U),
836     SIMD_TUPLE(v256_cmpgt_s16, 0U, 0U), SIMD_TUPLE(v256_cmplt_s16, 0U, 0U),
837     SIMD_TUPLE(v256_cmpeq_16, 0U, 0U), SIMD_TUPLE(v256_cmpgt_s32, 0U, 0U),
838     SIMD_TUPLE(v256_cmplt_s32, 0U, 0U), SIMD_TUPLE(v256_cmpeq_32, 0U, 0U),
839     SIMD_TUPLE(v256_shuffle_8, 31U, 8U), SIMD_TUPLE(v256_pshuffle_8, 15U, 8U),
840     SIMD_TUPLE(imm_v256_align<1>, 0U, 0U), SIMD_TUPLE(v256_sadd_s8, 0U, 0U),
841     SIMD_TUPLE(v256_sadd_u8, 0U, 0U), SIMD_TUPLE(v256_pack_s32_u16, 0U, 0U),
842     SIMD_TUPLE(v256_rdavg_u16, 0U, 0U), SIMD_TUPLE(imm_v256_align<2>, 0U, 0U),
843     SIMD_TUPLE(v256_unziphi_64, 0U, 0U), SIMD_TUPLE(v256_unziplo_64, 0U, 0U),
844     SIMD_TUPLE(imm_v256_align<3>, 0U, 0U),
845     SIMD_TUPLE(imm_v256_align<4>, 0U, 0U),
846     SIMD_TUPLE(imm_v256_align<5>, 0U, 0U),
847     SIMD_TUPLE(imm_v256_align<6>, 0U, 0U),
848     SIMD_TUPLE(imm_v256_align<7>, 0U, 0U),
849     SIMD_TUPLE(imm_v256_align<8>, 0U, 0U),
850     SIMD_TUPLE(imm_v256_align<9>, 0U, 0U),
851     SIMD_TUPLE(imm_v256_align<10>, 0U, 0U),
852     SIMD_TUPLE(imm_v256_align<11>, 0U, 0U),
853     SIMD_TUPLE(imm_v256_align<12>, 0U, 0U),
854     SIMD_TUPLE(imm_v256_align<13>, 0U, 0U),
855     SIMD_TUPLE(imm_v256_align<14>, 0U, 0U),
856     SIMD_TUPLE(imm_v256_align<15>, 0U, 0U),
857     SIMD_TUPLE(imm_v256_align<16>, 0U, 0U),
858     SIMD_TUPLE(imm_v256_align<17>, 0U, 0U),
859     SIMD_TUPLE(imm_v256_align<18>, 0U, 0U),
860     SIMD_TUPLE(imm_v256_align<19>, 0U, 0U),
861     SIMD_TUPLE(imm_v256_align<20>, 0U, 0U),
862     SIMD_TUPLE(imm_v256_align<21>, 0U, 0U),
863     SIMD_TUPLE(imm_v256_align<22>, 0U, 0U),
864     SIMD_TUPLE(imm_v256_align<23>, 0U, 0U),
865     SIMD_TUPLE(imm_v256_align<24>, 0U, 0U),
866     SIMD_TUPLE(imm_v256_align<25>, 0U, 0U),
867     SIMD_TUPLE(imm_v256_align<26>, 0U, 0U),
868     SIMD_TUPLE(imm_v256_align<27>, 0U, 0U),
869     SIMD_TUPLE(imm_v256_align<28>, 0U, 0U),
870     SIMD_TUPLE(imm_v256_align<29>, 0U, 0U),
871     SIMD_TUPLE(imm_v256_align<30>, 0U, 0U),
872     SIMD_TUPLE(imm_v256_align<31>, 0U, 0U));
873 
874 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V128V128),
875             SIMD_TUPLE(v256_from_v128, 0U, 0U), SIMD_TUPLE(v256_zip_8, 0U, 0U),
876             SIMD_TUPLE(v256_zip_16, 0U, 0U), SIMD_TUPLE(v256_zip_32, 0U, 0U),
877             SIMD_TUPLE(v256_mul_s16, 0U, 0U));
878 
879 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V128),
880             SIMD_TUPLE(v256_unpack_u8_s16, 0U, 0U),
881             SIMD_TUPLE(v256_unpack_s8_s16, 0U, 0U),
882             SIMD_TUPLE(v256_unpack_u16_s32, 0U, 0U),
883             SIMD_TUPLE(v256_unpack_s16_s32, 0U, 0U));
884 
885 INSTANTIATE(
886     ARCH, ARCH_POSTFIX(V256_V256U32), SIMD_TUPLE(v256_shl_8, 7U, 32U),
887     SIMD_TUPLE(v256_shr_u8, 7U, 32U), SIMD_TUPLE(v256_shr_s8, 7U, 32U),
888     SIMD_TUPLE(v256_shl_16, 15U, 32U), SIMD_TUPLE(v256_shr_u16, 15U, 32U),
889     SIMD_TUPLE(v256_shr_s16, 15U, 32U), SIMD_TUPLE(v256_shl_32, 31U, 32U),
890     SIMD_TUPLE(v256_shr_u32, 31U, 32U), SIMD_TUPLE(v256_shr_s32, 31U, 32U),
891     SIMD_TUPLE(v256_shl_64, 63U, 32U), SIMD_TUPLE(v256_shr_u64, 63U, 32U),
892     SIMD_TUPLE(v256_shr_s64, 63U, 32U));
893 
894 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256), SIMD_TUPLE(v256_abs_s8, 0U, 0U),
895             SIMD_TUPLE(v256_abs_s16, 0U, 0U), SIMD_TUPLE(v256_padd_s16, 0U, 0U),
896             SIMD_TUPLE(v256_unpacklo_u8_s16, 0U, 0U),
897             SIMD_TUPLE(v256_unpacklo_s8_s16, 0U, 0U),
898             SIMD_TUPLE(v256_unpacklo_u16_s32, 0U, 0U),
899             SIMD_TUPLE(v256_unpacklo_s16_s32, 0U, 0U),
900             SIMD_TUPLE(v256_unpackhi_u8_s16, 0U, 0U),
901             SIMD_TUPLE(v256_unpackhi_s8_s16, 0U, 0U),
902             SIMD_TUPLE(v256_unpackhi_u16_s32, 0U, 0U),
903             SIMD_TUPLE(v256_unpackhi_s16_s32, 0U, 0U),
904             SIMD_TUPLE(imm_v256_shr_n_byte<1>, 0U, 0U),
905             SIMD_TUPLE(imm_v256_shr_n_byte<2>, 0U, 0U),
906             SIMD_TUPLE(imm_v256_shr_n_byte<3>, 0U, 0U),
907             SIMD_TUPLE(imm_v256_shr_n_byte<4>, 0U, 0U),
908             SIMD_TUPLE(imm_v256_shr_n_byte<5>, 0U, 0U),
909             SIMD_TUPLE(imm_v256_shr_n_byte<6>, 0U, 0U),
910             SIMD_TUPLE(imm_v256_shr_n_byte<7>, 0U, 0U),
911             SIMD_TUPLE(imm_v256_shr_n_byte<8>, 0U, 0U),
912             SIMD_TUPLE(imm_v256_shr_n_byte<9>, 0U, 0U),
913             SIMD_TUPLE(imm_v256_shr_n_byte<10>, 0U, 0U),
914             SIMD_TUPLE(imm_v256_shr_n_byte<11>, 0U, 0U),
915             SIMD_TUPLE(imm_v256_shr_n_byte<12>, 0U, 0U),
916             SIMD_TUPLE(imm_v256_shr_n_byte<13>, 0U, 0U),
917             SIMD_TUPLE(imm_v256_shr_n_byte<14>, 0U, 0U),
918             SIMD_TUPLE(imm_v256_shr_n_byte<15>, 0U, 0U),
919             SIMD_TUPLE(imm_v256_shr_n_byte<16>, 0U, 0U),
920             SIMD_TUPLE(imm_v256_shr_n_byte<17>, 0U, 0U),
921             SIMD_TUPLE(imm_v256_shr_n_byte<18>, 0U, 0U),
922             SIMD_TUPLE(imm_v256_shr_n_byte<19>, 0U, 0U),
923             SIMD_TUPLE(imm_v256_shr_n_byte<20>, 0U, 0U),
924             SIMD_TUPLE(imm_v256_shr_n_byte<21>, 0U, 0U),
925             SIMD_TUPLE(imm_v256_shr_n_byte<22>, 0U, 0U),
926             SIMD_TUPLE(imm_v256_shr_n_byte<23>, 0U, 0U),
927             SIMD_TUPLE(imm_v256_shr_n_byte<24>, 0U, 0U),
928             SIMD_TUPLE(imm_v256_shr_n_byte<25>, 0U, 0U),
929             SIMD_TUPLE(imm_v256_shr_n_byte<26>, 0U, 0U),
930             SIMD_TUPLE(imm_v256_shr_n_byte<27>, 0U, 0U),
931             SIMD_TUPLE(imm_v256_shr_n_byte<28>, 0U, 0U),
932             SIMD_TUPLE(imm_v256_shr_n_byte<29>, 0U, 0U),
933             SIMD_TUPLE(imm_v256_shr_n_byte<30>, 0U, 0U),
934             SIMD_TUPLE(imm_v256_shr_n_byte<31>, 0U, 0U),
935             SIMD_TUPLE(imm_v256_shl_n_byte<1>, 0U, 0U),
936             SIMD_TUPLE(imm_v256_shl_n_byte<2>, 0U, 0U),
937             SIMD_TUPLE(imm_v256_shl_n_byte<3>, 0U, 0U),
938             SIMD_TUPLE(imm_v256_shl_n_byte<4>, 0U, 0U),
939             SIMD_TUPLE(imm_v256_shl_n_byte<5>, 0U, 0U),
940             SIMD_TUPLE(imm_v256_shl_n_byte<6>, 0U, 0U),
941             SIMD_TUPLE(imm_v256_shl_n_byte<7>, 0U, 0U),
942             SIMD_TUPLE(imm_v256_shl_n_byte<8>, 0U, 0U));
943 
944 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256_Part2),
945             SIMD_TUPLE(imm_v256_shl_n_byte<9>, 0U, 0U),
946             SIMD_TUPLE(imm_v256_shl_n_byte<10>, 0U, 0U),
947             SIMD_TUPLE(imm_v256_shl_n_byte<11>, 0U, 0U),
948             SIMD_TUPLE(imm_v256_shl_n_byte<12>, 0U, 0U),
949             SIMD_TUPLE(imm_v256_shl_n_byte<13>, 0U, 0U),
950             SIMD_TUPLE(imm_v256_shl_n_byte<14>, 0U, 0U),
951             SIMD_TUPLE(imm_v256_shl_n_byte<15>, 0U, 0U),
952             SIMD_TUPLE(imm_v256_shl_n_byte<16>, 0U, 0U),
953             SIMD_TUPLE(imm_v256_shl_n_byte<17>, 0U, 0U),
954             SIMD_TUPLE(imm_v256_shl_n_byte<18>, 0U, 0U),
955             SIMD_TUPLE(imm_v256_shl_n_byte<19>, 0U, 0U),
956             SIMD_TUPLE(imm_v256_shl_n_byte<20>, 0U, 0U),
957             SIMD_TUPLE(imm_v256_shl_n_byte<21>, 0U, 0U),
958             SIMD_TUPLE(imm_v256_shl_n_byte<22>, 0U, 0U),
959             SIMD_TUPLE(imm_v256_shl_n_byte<23>, 0U, 0U),
960             SIMD_TUPLE(imm_v256_shl_n_byte<24>, 0U, 0U),
961             SIMD_TUPLE(imm_v256_shl_n_byte<25>, 0U, 0U),
962             SIMD_TUPLE(imm_v256_shl_n_byte<26>, 0U, 0U),
963             SIMD_TUPLE(imm_v256_shl_n_byte<27>, 0U, 0U),
964             SIMD_TUPLE(imm_v256_shl_n_byte<28>, 0U, 0U),
965             SIMD_TUPLE(imm_v256_shl_n_byte<29>, 0U, 0U),
966             SIMD_TUPLE(imm_v256_shl_n_byte<30>, 0U, 0U),
967             SIMD_TUPLE(imm_v256_shl_n_byte<31>, 0U, 0U),
968             SIMD_TUPLE(imm_v256_shl_n_8<1>, 0U, 0U),
969             SIMD_TUPLE(imm_v256_shl_n_8<2>, 0U, 0U),
970             SIMD_TUPLE(imm_v256_shl_n_8<3>, 0U, 0U),
971             SIMD_TUPLE(imm_v256_shl_n_8<4>, 0U, 0U),
972             SIMD_TUPLE(imm_v256_shl_n_8<5>, 0U, 0U),
973             SIMD_TUPLE(imm_v256_shl_n_8<6>, 0U, 0U),
974             SIMD_TUPLE(imm_v256_shl_n_8<7>, 0U, 0U),
975             SIMD_TUPLE(imm_v256_shr_n_u8<1>, 0U, 0U),
976             SIMD_TUPLE(imm_v256_shr_n_u8<2>, 0U, 0U),
977             SIMD_TUPLE(imm_v256_shr_n_u8<3>, 0U, 0U),
978             SIMD_TUPLE(imm_v256_shr_n_u8<4>, 0U, 0U),
979             SIMD_TUPLE(imm_v256_shr_n_u8<5>, 0U, 0U),
980             SIMD_TUPLE(imm_v256_shr_n_u8<6>, 0U, 0U),
981             SIMD_TUPLE(imm_v256_shr_n_u8<7>, 0U, 0U),
982             SIMD_TUPLE(imm_v256_shr_n_s8<1>, 0U, 0U),
983             SIMD_TUPLE(imm_v256_shr_n_s8<2>, 0U, 0U),
984             SIMD_TUPLE(imm_v256_shr_n_s8<3>, 0U, 0U),
985             SIMD_TUPLE(imm_v256_shr_n_s8<4>, 0U, 0U),
986             SIMD_TUPLE(imm_v256_shr_n_s8<5>, 0U, 0U),
987             SIMD_TUPLE(imm_v256_shr_n_s8<6>, 0U, 0U),
988             SIMD_TUPLE(imm_v256_shr_n_s8<7>, 0U, 0U),
989             SIMD_TUPLE(imm_v256_shl_n_16<1>, 0U, 0U),
990             SIMD_TUPLE(imm_v256_shl_n_16<2>, 0U, 0U),
991             SIMD_TUPLE(imm_v256_shl_n_16<4>, 0U, 0U),
992             SIMD_TUPLE(imm_v256_shl_n_16<6>, 0U, 0U),
993             SIMD_TUPLE(imm_v256_shl_n_16<8>, 0U, 0U),
994             SIMD_TUPLE(imm_v256_shl_n_16<10>, 0U, 0U));
995 
996 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256_Part3),
997             SIMD_TUPLE(imm_v256_shl_n_16<12>, 0U, 0U),
998             SIMD_TUPLE(imm_v256_shl_n_16<14>, 0U, 0U),
999             SIMD_TUPLE(imm_v256_shr_n_u16<1>, 0U, 0U),
1000             SIMD_TUPLE(imm_v256_shr_n_u16<2>, 0U, 0U),
1001             SIMD_TUPLE(imm_v256_shr_n_u16<4>, 0U, 0U),
1002             SIMD_TUPLE(imm_v256_shr_n_u16<6>, 0U, 0U),
1003             SIMD_TUPLE(imm_v256_shr_n_u16<8>, 0U, 0U),
1004             SIMD_TUPLE(imm_v256_shr_n_u16<10>, 0U, 0U),
1005             SIMD_TUPLE(imm_v256_shr_n_u16<12>, 0U, 0U),
1006             SIMD_TUPLE(imm_v256_shr_n_u16<14>, 0U, 0U),
1007             SIMD_TUPLE(imm_v256_shr_n_s16<1>, 0U, 0U),
1008             SIMD_TUPLE(imm_v256_shr_n_s16<2>, 0U, 0U),
1009             SIMD_TUPLE(imm_v256_shr_n_s16<4>, 0U, 0U),
1010             SIMD_TUPLE(imm_v256_shr_n_s16<6>, 0U, 0U),
1011             SIMD_TUPLE(imm_v256_shr_n_s16<8>, 0U, 0U),
1012             SIMD_TUPLE(imm_v256_shr_n_s16<10>, 0U, 0U),
1013             SIMD_TUPLE(imm_v256_shr_n_s16<12>, 0U, 0U),
1014             SIMD_TUPLE(imm_v256_shr_n_s16<14>, 0U, 0U),
1015             SIMD_TUPLE(imm_v256_shl_n_32<1>, 0U, 0U),
1016             SIMD_TUPLE(imm_v256_shl_n_32<4>, 0U, 0U),
1017             SIMD_TUPLE(imm_v256_shl_n_32<8>, 0U, 0U),
1018             SIMD_TUPLE(imm_v256_shl_n_32<12>, 0U, 0U),
1019             SIMD_TUPLE(imm_v256_shl_n_32<16>, 0U, 0U),
1020             SIMD_TUPLE(imm_v256_shl_n_32<20>, 0U, 0U),
1021             SIMD_TUPLE(imm_v256_shl_n_32<24>, 0U, 0U),
1022             SIMD_TUPLE(imm_v256_shl_n_32<28>, 0U, 0U),
1023             SIMD_TUPLE(imm_v256_shr_n_u32<1>, 0U, 0U),
1024             SIMD_TUPLE(imm_v256_shr_n_u32<4>, 0U, 0U),
1025             SIMD_TUPLE(imm_v256_shr_n_u32<8>, 0U, 0U),
1026             SIMD_TUPLE(imm_v256_shr_n_u32<12>, 0U, 0U),
1027             SIMD_TUPLE(imm_v256_shr_n_u32<16>, 0U, 0U),
1028             SIMD_TUPLE(imm_v256_shr_n_u32<20>, 0U, 0U),
1029             SIMD_TUPLE(imm_v256_shr_n_u32<24>, 0U, 0U),
1030             SIMD_TUPLE(imm_v256_shr_n_u32<28>, 0U, 0U),
1031             SIMD_TUPLE(imm_v256_shr_n_s32<1>, 0U, 0U),
1032             SIMD_TUPLE(imm_v256_shr_n_s32<4>, 0U, 0U),
1033             SIMD_TUPLE(imm_v256_shr_n_s32<8>, 0U, 0U),
1034             SIMD_TUPLE(imm_v256_shr_n_s32<12>, 0U, 0U),
1035             SIMD_TUPLE(imm_v256_shr_n_s32<16>, 0U, 0U),
1036             SIMD_TUPLE(imm_v256_shr_n_s32<20>, 0U, 0U),
1037             SIMD_TUPLE(imm_v256_shr_n_s32<24>, 0U, 0U),
1038             SIMD_TUPLE(imm_v256_shr_n_s32<28>, 0U, 0U));
1039 
1040 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256_Part4),
1041             SIMD_TUPLE(imm_v256_shl_n_64<1>, 0U, 0U),
1042             SIMD_TUPLE(imm_v256_shl_n_64<4>, 0U, 0U),
1043             SIMD_TUPLE(imm_v256_shl_n_64<8>, 0U, 0U),
1044             SIMD_TUPLE(imm_v256_shl_n_64<12>, 0U, 0U),
1045             SIMD_TUPLE(imm_v256_shl_n_64<16>, 0U, 0U),
1046             SIMD_TUPLE(imm_v256_shl_n_64<20>, 0U, 0U),
1047             SIMD_TUPLE(imm_v256_shl_n_64<24>, 0U, 0U),
1048             SIMD_TUPLE(imm_v256_shl_n_64<28>, 0U, 0U),
1049             SIMD_TUPLE(imm_v256_shl_n_64<32>, 0U, 0U),
1050             SIMD_TUPLE(imm_v256_shl_n_64<36>, 0U, 0U),
1051             SIMD_TUPLE(imm_v256_shl_n_64<40>, 0U, 0U),
1052             SIMD_TUPLE(imm_v256_shl_n_64<44>, 0U, 0U),
1053             SIMD_TUPLE(imm_v256_shl_n_64<48>, 0U, 0U),
1054             SIMD_TUPLE(imm_v256_shl_n_64<52>, 0U, 0U),
1055             SIMD_TUPLE(imm_v256_shl_n_64<56>, 0U, 0U),
1056             SIMD_TUPLE(imm_v256_shl_n_64<60>, 0U, 0U),
1057             SIMD_TUPLE(imm_v256_shr_n_u64<1>, 0U, 0U),
1058             SIMD_TUPLE(imm_v256_shr_n_u64<4>, 0U, 0U),
1059             SIMD_TUPLE(imm_v256_shr_n_u64<8>, 0U, 0U),
1060             SIMD_TUPLE(imm_v256_shr_n_u64<12>, 0U, 0U),
1061             SIMD_TUPLE(imm_v256_shr_n_u64<16>, 0U, 0U),
1062             SIMD_TUPLE(imm_v256_shr_n_u64<20>, 0U, 0U),
1063             SIMD_TUPLE(imm_v256_shr_n_u64<24>, 0U, 0U),
1064             SIMD_TUPLE(imm_v256_shr_n_u64<28>, 0U, 0U),
1065             SIMD_TUPLE(imm_v256_shr_n_u64<32>, 0U, 0U),
1066             SIMD_TUPLE(imm_v256_shr_n_u64<36>, 0U, 0U),
1067             SIMD_TUPLE(imm_v256_shr_n_u64<40>, 0U, 0U),
1068             SIMD_TUPLE(imm_v256_shr_n_u64<44>, 0U, 0U),
1069             SIMD_TUPLE(imm_v256_shr_n_u64<48>, 0U, 0U),
1070             SIMD_TUPLE(imm_v256_shr_n_u64<52>, 0U, 0U),
1071             SIMD_TUPLE(imm_v256_shr_n_u64<56>, 0U, 0U),
1072             SIMD_TUPLE(imm_v256_shr_n_u64<60>, 0U, 0U),
1073             SIMD_TUPLE(imm_v256_shr_n_s64<1>, 0U, 0U),
1074             SIMD_TUPLE(imm_v256_shr_n_s64<4>, 0U, 0U),
1075             SIMD_TUPLE(imm_v256_shr_n_s64<8>, 0U, 0U),
1076             SIMD_TUPLE(imm_v256_shr_n_s64<12>, 0U, 0U),
1077             SIMD_TUPLE(imm_v256_shr_n_s64<16>, 0U, 0U),
1078             SIMD_TUPLE(imm_v256_shr_n_s64<20>, 0U, 0U),
1079             SIMD_TUPLE(imm_v256_shr_n_s64<24>, 0U, 0U),
1080             SIMD_TUPLE(imm_v256_shr_n_s64<28>, 0U, 0U),
1081             SIMD_TUPLE(imm_v256_shr_n_s64<32>, 0U, 0U),
1082             SIMD_TUPLE(imm_v256_shr_n_s64<36>, 0U, 0U),
1083             SIMD_TUPLE(imm_v256_shr_n_s64<40>, 0U, 0U),
1084             SIMD_TUPLE(imm_v256_shr_n_s64<44>, 0U, 0U),
1085             SIMD_TUPLE(imm_v256_shr_n_s64<48>, 0U, 0U),
1086             SIMD_TUPLE(imm_v256_shr_n_s64<52>, 0U, 0U),
1087             SIMD_TUPLE(imm_v256_shr_n_s64<56>, 0U, 0U),
1088             SIMD_TUPLE(imm_v256_shr_n_s64<60>, 0U, 0U),
1089             SIMD_TUPLE(v256_padd_u8, 0U, 0U));
1090 
1091 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256_Part5),
1092             SIMD_TUPLE(imm_v256_shr_n_word<1>, 0U, 0U),
1093             SIMD_TUPLE(imm_v256_shr_n_word<2>, 0U, 0U),
1094             SIMD_TUPLE(imm_v256_shr_n_word<3>, 0U, 0U),
1095             SIMD_TUPLE(imm_v256_shr_n_word<4>, 0U, 0U),
1096             SIMD_TUPLE(imm_v256_shr_n_word<5>, 0U, 0U),
1097             SIMD_TUPLE(imm_v256_shr_n_word<6>, 0U, 0U),
1098             SIMD_TUPLE(imm_v256_shr_n_word<7>, 0U, 0U),
1099             SIMD_TUPLE(imm_v256_shr_n_word<8>, 0U, 0U),
1100             SIMD_TUPLE(imm_v256_shr_n_word<9>, 0U, 0U),
1101             SIMD_TUPLE(imm_v256_shr_n_word<10>, 0U, 0U),
1102             SIMD_TUPLE(imm_v256_shr_n_word<11>, 0U, 0U),
1103             SIMD_TUPLE(imm_v256_shr_n_word<12>, 0U, 0U),
1104             SIMD_TUPLE(imm_v256_shr_n_word<13>, 0U, 0U),
1105             SIMD_TUPLE(imm_v256_shr_n_word<14>, 0U, 0U),
1106             SIMD_TUPLE(imm_v256_shr_n_word<15>, 0U, 0U),
1107             SIMD_TUPLE(imm_v256_shl_n_word<1>, 0U, 0U),
1108             SIMD_TUPLE(imm_v256_shl_n_word<2>, 0U, 0U),
1109             SIMD_TUPLE(imm_v256_shl_n_word<3>, 0U, 0U),
1110             SIMD_TUPLE(imm_v256_shl_n_word<4>, 0U, 0U),
1111             SIMD_TUPLE(imm_v256_shl_n_word<5>, 0U, 0U),
1112             SIMD_TUPLE(imm_v256_shl_n_word<6>, 0U, 0U),
1113             SIMD_TUPLE(imm_v256_shl_n_word<7>, 0U, 0U),
1114             SIMD_TUPLE(imm_v256_shl_n_word<8>, 0U, 0U),
1115             SIMD_TUPLE(imm_v256_shl_n_word<9>, 0U, 0U),
1116             SIMD_TUPLE(imm_v256_shl_n_word<10>, 0U, 0U),
1117             SIMD_TUPLE(imm_v256_shl_n_word<11>, 0U, 0U),
1118             SIMD_TUPLE(imm_v256_shl_n_word<12>, 0U, 0U),
1119             SIMD_TUPLE(imm_v256_shl_n_word<13>, 0U, 0U),
1120             SIMD_TUPLE(imm_v256_shl_n_word<14>, 0U, 0U),
1121             SIMD_TUPLE(imm_v256_shl_n_word<15>, 0U, 0U));
1122 
1123 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_V256V256V256),
1124             SIMD_TUPLE(v256_blend_8, 0U, 0U),
1125             SIMD_TUPLE(v256_wideshuffle_8, 63U, 8U));
1126 
1127 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_U8), SIMD_TUPLE(v256_dup_8, 0U, 0U));
1128 
1129 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_U16), SIMD_TUPLE(v256_dup_16, 0U, 0U));
1130 
1131 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_U32), SIMD_TUPLE(v256_dup_32, 0U, 0U));
1132 
1133 INSTANTIATE(ARCH, ARCH_POSTFIX(V256_U64), SIMD_TUPLE(v256_dup_64, 0U, 0U));
1134 
1135 INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V256), SIMD_TUPLE(v256_low_u32, 0U, 0U),
1136             SIMD_TUPLE(v256_movemask_8, 0U, 0U));
1137 
1138 INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V256), SIMD_TUPLE(v256_low_v64, 0U, 0U));
1139 
1140 }  // namespace SIMD_NAMESPACE
1141