1 /**
2 * SSSE3 intrinsics.
3 * https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#techs=SSSE3
4 *
5 * Copyright: Guillaume Piolat 2021.
6 *            Johan Engelen 2021.
7 * License:   $(LINK2 http://www.boost.org/LICENSE_1_0.txt, Boost License 1.0)
8 */
9 module inteli.tmmintrin;
10 
11 public import inteli.types;
12 import inteli.internals;
13 
14 public import inteli.pmmintrin;
15 import inteli.mmx;
16 
17 nothrow @nogc:
18 
19 
20 // SSSE3 instructions
21 // https://software.intel.com/sites/landingpage/IntrinsicsGuide/#techs=SSSE3
22 // Note: this header will work whether you have SSSE3 enabled or not.
23 // With LDC, use "dflags-ldc": ["-mattr=+ssse3"] or equivalent to actively 
24 // generate SSE3 instructions.
25 // With GDC, use "dflags-gdc": ["-mssse3"] or equivalent to generate SSSE3 instructions.
26 
27 /// Compute the absolute value of packed signed 16-bit integers in `a`.
28 __m128i _mm_abs_epi16 (__m128i a) @trusted
29 {
30     static if (DMD_with_DSIMD)
31     {
32         return cast(__m128i)__simd(XMM.PABSW, a);
33     }
34     else static if (GDC_with_SSSE3)
35     {
36         return cast(__m128i) __builtin_ia32_pabsw128(cast(short8)a);
37     }
38     else static if (LDC_with_ARM64)
39     {
40         return cast(__m128i) vabsq_s16(cast(short8)a);
41     }
42     else
43     {
44         // LDC x86: generate pabsw since LDC 1.1 -O2
45         short8 sa = cast(short8)a;
46         for (int i = 0; i < 8; ++i)
47         {
48             short s = sa.array[i];
49             sa.ptr[i] = s >= 0 ? s : cast(short)(-cast(int)(s));
50         }  
51         return cast(__m128i)sa;
52     }
53 }
54 unittest
55 {
56     __m128i A = _mm_setr_epi16(0, -1, -32768, 32767, 10, -10, 1000, -1000);
57     short8 B = cast(short8) _mm_abs_epi16(A);
58     short[8] correct = [0, 1, -32768, 32767, 10, 10, 1000, 1000];
59     assert(B.array == correct);
60 }
61 
62 /// Compute the absolute value of packed signed 32-bit integers in `a`.
63 __m128i _mm_abs_epi32 (__m128i a) @trusted
64 {
65     static if (DMD_with_DSIMD)
66     {
67         return cast(__m128i)__simd(XMM.PABSD, cast(int4)a);
68     }
69     else static if (GDC_with_SSSE3)
70     {
71         return cast(__m128i) __builtin_ia32_pabsd128(cast(int4)a);
72     }
73     else static if (LDC_with_ARM64)
74     {
75         return cast(__m128i) vabsq_s32(cast(int4)a);
76     }
77     else
78     {
79         // LDC x86: generates pabsd since LDC 1.1 -O2
80         int4 sa = cast(int4)a;
81         for (int i = 0; i < 4; ++i)
82         {
83             int s = sa.array[i];
84             sa.ptr[i] = s >= 0 ? s : -s;
85         }  
86         return cast(__m128i)sa;
87     } 
88 }
89 unittest
90 {
91     __m128i A = _mm_setr_epi32(0, -1, -2_147_483_648, -2_147_483_647);
92     int4 B = cast(int4) _mm_abs_epi32(A);
93     int[4] correct = [0, 1, -2_147_483_648, 2_147_483_647];
94     assert(B.array == correct);
95 }
96 
97 /// Compute the absolute value of packed signed 8-bit integers in `a`.
98 __m128i _mm_abs_epi8 (__m128i a) @trusted
99 {
100     static if (DMD_with_DSIMD)
101     {
102         return cast(__m128i)__simd(XMM.PABSB, cast(byte16)a);
103     }
104     else static if (GDC_with_SSSE3)
105     {
106         alias ubyte16 = __vector(ubyte[16]);
107         return cast(__m128i) __builtin_ia32_pabsb128(cast(ubyte16)a);
108     }
109     else static if (LDC_with_ARM64)
110     {
111         return cast(__m128i) vabsq_s8(cast(byte16)a);
112     }
113     else version(LDC)
114     {
115         // LDC x86: generates pabsb since LDC 1.1 -O1
116         //     arm64: generates abs since LDC 1.8 -O1
117         enum ir = `
118                 %n = sub <16 x i8> <i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0>, %0
119                 %s = icmp slt <16 x i8> <i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0, i8 0>, %0
120                 %r = select <16 x i1> %s, <16 x i8> %0, <16 x i8> %n
121                 ret <16 x i8> %r`;
122         return cast(__m128i) LDCInlineIR!(ir, byte16, byte16)(cast(byte16)a);
123     }
124     else
125     {
126         // A loop version like in _mm_abs_epi16/_mm_abs_epi32 would be very slow 
127         // in LDC x86 and wouldn't vectorize. Doesn't generate pabsb in LDC though.
128         return _mm_min_epu8(a, _mm_sub_epi8(_mm_setzero_si128(), a));
129     }
130 }
131 unittest
132 {
133     __m128i A = _mm_setr_epi8(0, -1, -128, -127, 127, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
134     byte16 B = cast(byte16) _mm_abs_epi8(A);
135     byte[16] correct =       [0,  1, -128,  127, 127, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0];
136     assert(B.array == correct);
137 }
138 
139 /// Compute the absolute value of packed 64-bit floating-point elements in `a`.
140 /// #BONUS.
141 __m128d _mm_abs_pd (__m128d a) @trusted
142 {
143     long2 mask = 0x7fff_ffff_ffff_ffff;
144     return cast(__m128d)((cast(long2)a) & mask);
145 }
146 unittest
147 {
148     __m128d A = _mm_setr_pd(-42.0f, -double.infinity);
149     __m128d R = _mm_abs_pd(A);
150     double[2] correct =    [42.0f, +double.infinity];
151     assert(R.array == correct);
152 }
153 
154 /// Compute the absolute value of packed signed 16-bit integers in `a`.
155 __m64 _mm_abs_pi16 (__m64 a) @trusted
156 {
157     return to_m64(_mm_abs_epi16(to_m128i(a)));
158 }
159 unittest
160 {
161     __m64 A = _mm_setr_pi16(0, -1, -32768, 32767);
162     short4 B = cast(short4) _mm_abs_pi16(A);
163     short[4] correct = [0, 1, -32768, 32767];
164     assert(B.array == correct);
165 }
166 
167 /// Compute the absolute value of packed signed 32-bit integers in `a`.
168 __m64 _mm_abs_pi32 (__m64 a) @trusted
169 {
170      return to_m64(_mm_abs_epi32(to_m128i(a)));
171 }
172 unittest
173 {
174     __m64 A = _mm_setr_pi32(-1, -2_147_483_648);
175     int2 B = cast(int2) _mm_abs_pi32(A);
176     int[2] correct = [1, -2_147_483_648];
177     assert(B.array == correct);
178 }
179 
180 /// Compute the absolute value of packed signed 8-bit integers in `a`.
181 __m64 _mm_abs_pi8 (__m64 a) @trusted
182 {
183     return to_m64(_mm_abs_epi8(to_m128i(a)));
184 }
185 unittest
186 {
187     __m64 A = _mm_setr_pi8(0, -1, -128, -127, 127, 0, 0, 0);
188     byte8 B = cast(byte8) _mm_abs_pi8(A);
189     byte[8] correct =       [0,  1, -128,  127, 127, 0, 0, 0];
190     assert(B.array == correct);
191 }
192 
193 /// Compute the absolute value of packed 32-bit floating-point elements in `a`.
194 /// #BONUS.
195 __m128 _mm_abs_ps (__m128 a) @trusted
196 {
197     __m128i mask = 0x7fffffff;
198     return cast(__m128)((cast(__m128i)a) & mask);
199 }
200 unittest
201 {
202     __m128 A = _mm_setr_ps(-0.0f, 10.0f, -42.0f, -float.infinity);
203     __m128 R = _mm_abs_ps(A);
204     float[4] correct =    [0.0f, 10.0f, 42.0f, +float.infinity];
205     assert(R.array == correct);
206 }
207 
208 /// Concatenate 16-byte blocks in `a` and `b` into a 32-byte temporary result, shift the result right by `count` bytes, and return the low 16 bytes.
209 __m128i _mm_alignr_epi8(ubyte count)(__m128i a, __m128i b) @trusted
210 {
211     // PERF DMD
212     static if (GDC_with_SSSE3)
213     {
214         return cast(__m128i)__builtin_ia32_palignr128(cast(long2)a, cast(long2)b, count * 8);
215     }
216     else version(LDC)
217     {
218         static if (count >= 32)
219         {
220             return _mm_setzero_si128();
221         }
222         else static if (count < 16)
223         {
224             // Generates palignr since LDC 1.1 -O1
225             // Also generates a single ext instruction on arm64.
226             return cast(__m128i) shufflevectorLDC!(byte16, ( 0 + count),
227                                                         ( 1 + count),
228                                                         ( 2 + count),
229                                                         ( 3 + count),
230                                                         ( 4 + count),
231                                                         ( 5 + count),
232                                                         ( 6 + count),
233                                                         ( 7 + count),
234                                                         ( 8 + count),
235                                                         ( 9 + count),
236                                                         (10 + count),
237                                                         (11 + count),
238                                                         (12 + count),
239                                                         (13 + count),
240                                                         (14 + count),
241                                                         (15 + count))(cast(byte16)b, cast(byte16)a);
242         }
243         else
244         {
245             return cast(__m128i) shufflevectorLDC!(byte16, ( 0 + count) % 32,
246                                                         ( 1 + count) % 32,
247                                                         ( 2 + count) % 32,
248                                                         ( 3 + count) % 32,
249                                                         ( 4 + count) % 32,
250                                                         ( 5 + count) % 32,
251                                                         ( 6 + count) % 32,
252                                                         ( 7 + count) % 32,
253                                                         ( 8 + count) % 32,
254                                                         ( 9 + count) % 32,
255                                                         (10 + count) % 32,
256                                                         (11 + count) % 32,
257                                                         (12 + count) % 32,
258                                                         (13 + count) % 32,
259                                                         (14 + count) % 32,
260                                                         (15 + count) % 32)(cast(byte16)_mm_setzero_si128(), cast(byte16)a);
261         }
262     }
263     else
264     {
265         byte16 ab = cast(byte16)a;
266         byte16 bb = cast(byte16)b;
267         byte16 r;
268 
269         for (int i = 0; i < 16; ++i)
270         {
271             const int srcpos = count + cast(int)i;
272             if (srcpos > 31) 
273             {
274                 r.ptr[i] = 0;
275             } 
276             else if (srcpos > 15) 
277             {
278                 r.ptr[i] = ab.array[(srcpos) & 15];
279             } 
280             else 
281             {
282                 r.ptr[i] = bb.array[srcpos];
283             }
284        }
285        return cast(__m128i)r;
286     }
287 }
288 unittest
289 {
290     __m128i A = _mm_setr_epi8(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16);
291     __m128i B = _mm_setr_epi8(17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32);
292 
293     {
294         byte16 C = cast(byte16)_mm_alignr_epi8!0(A ,B);
295         byte[16] correct = [17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32];
296         assert(C.array == correct);
297     }
298     {
299         byte16 C = cast(byte16)_mm_alignr_epi8!20(A ,B);
300         byte[16] correct = [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 0, 0, 0, 0];
301         assert(C.array == correct);
302     }
303     {
304         byte16 C = cast(byte16)_mm_alignr_epi8!34(A ,B);
305         byte[16] correct = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0];
306         assert(C.array == correct);
307     }
308 
309     __m128i D = _mm_setr_epi8(-123, -82, 103, -69, 103, -26, 9, 106, 58, -11, 79, -91, 114, -13, 110, 60);
310     __m128i E = _mm_setr_epi8(25, -51, -32, 91, -85, -39, -125, 31, -116, 104, 5, -101, 127, 82, 14, 81);
311     byte16 F = cast(byte16)_mm_alignr_epi8!8(D, E);
312     byte[16] correct = [-116, 104, 5, -101, 127, 82, 14, 81, -123, -82, 103, -69, 103, -26, 9, 106];
313     assert(F.array == correct);
314 }
315 
316 /// Concatenate 8-byte blocks in `a` and `b` into a 16-byte temporary result, shift the result right by `count` bytes, and return the low 8 bytes.
317 __m64 _mm_alignr_pi8(ubyte count)(__m64 a, __m64 b) @trusted
318 {
319     // PERF DMD
320     static if (GDC_with_SSSE3)
321     {
322         return cast(__m64)__builtin_ia32_palignr(cast(long1)a, cast(long1)b, count * 8);
323     }
324     else version(LDC)
325     {
326         static if (count >= 16)
327         {
328             return _mm_setzero_si64();
329         }
330         else static if (count < 8)
331         {
332             // Note: in LDC x86 this uses a pshufb.
333             // Generates ext in arm64.
334             return cast(__m64) shufflevectorLDC!(byte8, (0 + count),
335                                                      (1 + count),
336                                                      (2 + count),
337                                                      (3 + count),
338                                                      (4 + count),
339                                                      (5 + count),
340                                                      (6 + count),
341                                                      (7 + count))(cast(byte8)b, cast(byte8)a);
342         }
343         else
344         {
345             return cast(__m64) shufflevectorLDC!(byte8, (0 + count)%16,
346                                                      (1 + count)%16,
347                                                      (2 + count)%16,
348                                                      (3 + count)%16,
349                                                      (4 + count)%16,
350                                                      (5 + count)%16,
351                                                      (6 + count)%16,
352                                                      (7 + count)%16)(cast(byte8)_mm_setzero_si64(), cast(byte8)a);
353         }
354     }
355     else
356     {
357         byte8 ab = cast(byte8)a;
358         byte8 bb = cast(byte8)b;
359         byte8 r;
360 
361         for (int i = 0; i < 8; ++i)
362         {
363             const int srcpos = count + cast(int)i;
364             if (srcpos > 15) 
365             {
366                 r.ptr[i] = 0;
367             } 
368             else if (srcpos > 7) 
369             {
370                 r.ptr[i] = ab.array[(srcpos) & 7];
371             } 
372             else 
373             {
374                 r.ptr[i] = bb.array[srcpos];
375             }
376        }
377        return cast(__m64)r;
378     }
379 }
380 unittest
381 {
382     __m64 A = _mm_setr_pi8(1, 2, 3, 4, 5, 6, 7, 8);
383     __m64 B = _mm_setr_pi8(17, 18, 19, 20, 21, 22, 23, 24);
384 
385     {
386         byte8 C = cast(byte8)_mm_alignr_pi8!0(A ,B);
387         byte[8] correct = [17, 18, 19, 20, 21, 22, 23, 24];
388         assert(C.array == correct);
389     }
390 
391     {
392         byte8 C = cast(byte8)_mm_alignr_pi8!3(A ,B);
393         byte[8] correct = [ 20, 21, 22, 23, 24, 1, 2, 3];
394         assert(C.array == correct);
395     }
396     {
397         byte8 C = cast(byte8)_mm_alignr_pi8!11(A ,B);
398         byte[8] correct = [4, 5, 6, 7, 8, 0, 0, 0];
399         assert(C.array == correct);
400     }
401     {
402         byte8 C = cast(byte8)_mm_alignr_pi8!17(A ,B);
403         byte[8] correct = [0, 0, 0, 0, 0, 0, 0, 0];
404         assert(C.array == correct);
405     }
406 }
407 
408 /// Horizontally add adjacent pairs of 16-bit integers in `a` and `b`, and pack the signed 16-bit results.
409 __m128i _mm_hadd_epi16 (__m128i a, __m128i b) @trusted
410 {
411     // PERF DMD
412     static if (GDC_with_SSSE3)
413     {
414         return cast(__m128i)__builtin_ia32_phaddw128(cast(short8)a, cast(short8)b);
415     }
416     else static if (LDC_with_SSSE3)
417     {
418         return cast(__m128i)__builtin_ia32_phaddw128(cast(short8)a, cast(short8)b);
419     }
420     else static if (LDC_with_ARM64)
421     {
422         return cast(__m128i)vpaddq_s16(cast(short8)a, cast(short8)b);
423     }
424     else
425     {
426         short8 sa = cast(short8)a;
427         short8 sb = cast(short8)b;
428         short8 r;
429         r.ptr[0] = cast(short)(sa.array[0] + sa.array[1]);
430         r.ptr[1] = cast(short)(sa.array[2] + sa.array[3]);
431         r.ptr[2] = cast(short)(sa.array[4] + sa.array[5]);
432         r.ptr[3] = cast(short)(sa.array[6] + sa.array[7]);
433         r.ptr[4] = cast(short)(sb.array[0] + sb.array[1]);
434         r.ptr[5] = cast(short)(sb.array[2] + sb.array[3]);
435         r.ptr[6] = cast(short)(sb.array[4] + sb.array[5]);
436         r.ptr[7] = cast(short)(sb.array[6] + sb.array[7]);
437         return cast(__m128i)r;
438     }
439 }
440 unittest
441 {
442     __m128i A = _mm_setr_epi16(1, -2, 4, 8, 16, 32, -1, -32768);
443     short8 C = cast(short8) _mm_hadd_epi16(A, A);
444     short[8] correct = [ -1, 12, 48, 32767, -1, 12, 48, 32767];
445     assert(C.array == correct);
446 }
447 
448 /// Horizontally add adjacent pairs of 32-bit integers in `a` and `b`, and pack the signed 32-bit results.
449 __m128i _mm_hadd_epi32 (__m128i a, __m128i b) @trusted
450 { 
451     // PERF DMD
452     static if (GDC_with_SSSE3)
453     {
454         return cast(__m128i)__builtin_ia32_phaddd128(cast(int4)a, cast(int4)b);
455     }
456     else static if (LDC_with_SSSE3)
457     {
458         return cast(__m128i)__builtin_ia32_phaddd128(cast(int4)a, cast(int4)b);
459     }
460     else static if (LDC_with_ARM64)
461     {
462         return cast(__m128i)vpaddq_s32(cast(int4)a, cast(int4)b);
463     }
464     else
465     {
466         int4 ia = cast(int4)a;
467         int4 ib = cast(int4)b;
468         int4 r;
469         r.ptr[0] = ia.array[0] + ia.array[1];
470         r.ptr[1] = ia.array[2] + ia.array[3];
471         r.ptr[2] = ib.array[0] + ib.array[1];
472         r.ptr[3] = ib.array[2] + ib.array[3];
473         return cast(__m128i)r;
474     }
475 }
476 unittest
477 {
478     __m128i A = _mm_setr_epi32(1, -2, int.min, -1);
479     __m128i B = _mm_setr_epi32(1, int.max, 4, -4);
480     int4 C = cast(int4) _mm_hadd_epi32(A, B);
481     int[4] correct = [ -1, int.max, int.min, 0 ];
482     assert(C.array == correct);
483 }
484 
485 /// Horizontally add adjacent pairs of 16-bit integers in `a` and `b`, and pack the signed 16-bit results.
486 __m64 _mm_hadd_pi16 (__m64 a, __m64 b) @trusted
487 {
488     // PERF DMD
489     static if (GDC_with_SSSE3)
490     {
491         return cast(__m64) __builtin_ia32_phaddw(cast(short4)a, cast(short4)b);
492     }
493     else static if (LDC_with_ARM64)
494     {
495         return cast(__m64) vpadd_s16(cast(short4)a, cast(short4)b);
496     }
497     else
498     {
499         // LDC x86: generates phaddw since LDC 1.24 -O2.
500         short4 r;
501         short4 sa = cast(short4)a;
502         short4 sb = cast(short4)b;
503         r.ptr[0] = cast(short)(sa.array[0] + sa.array[1]); 
504         r.ptr[1] = cast(short)(sa.array[2] + sa.array[3]);
505         r.ptr[2] = cast(short)(sb.array[0] + sb.array[1]);
506         r.ptr[3] = cast(short)(sb.array[2] + sb.array[3]);
507         return cast(__m64)r;
508     }
509 }
510 unittest
511 {
512     __m64 A = _mm_setr_pi16(1, -2, 4, 8);
513     __m64 B = _mm_setr_pi16(16, 32, -1, -32768);
514     short4 C = cast(short4) _mm_hadd_pi16(A, B);
515     short[4] correct = [ -1, 12, 48, 32767 ];
516     assert(C.array == correct);
517 }
518 
519 /// Horizontally add adjacent pairs of 32-bit integers in `a` and `b`, 
520 /// and pack the signed 32-bit results.
521 __m64 _mm_hadd_pi32 (__m64 a, __m64 b) @trusted
522 {
523     // PERF DMD
524     static if (GDC_with_SSSE3)
525     {
526         return cast(__m64) __builtin_ia32_phaddd(cast(int2)a, cast(int2)b);
527     }
528     else static if (LDC_with_ARM64)
529     {
530         return cast(__m64)vpadd_s32(cast(int2)a, cast(int2)b);
531     }
532     else
533     {
534         // LDC x86: generates phaddd since LDC 1.24 -O2
535         int2 ia = cast(int2)a;
536         int2 ib = cast(int2)b;
537         int2 r;
538         r.ptr[0] = ia.array[0] + ia.array[1];
539         r.ptr[1] = ib.array[0] + ib.array[1];
540         return cast(__m64)r;
541     }
542 }
543 unittest
544 {
545     __m64 A = _mm_setr_pi32(int.min, -1);
546     __m64 B = _mm_setr_pi32(1, int.max);
547     int2 C = cast(int2) _mm_hadd_pi32(A, B);
548     int[2] correct = [ int.max, int.min ];
549     assert(C.array == correct);
550 }
551 
552 /// Horizontally add adjacent pairs of signed 16-bit integers in `a` and `b` using saturation, 
553 /// and pack the signed 16-bit results.
554 __m128i _mm_hadds_epi16 (__m128i a, __m128i b) @trusted
555 {
556      // PERF DMD
557     static if (GDC_with_SSSE3)
558     {
559         return cast(__m128i)__builtin_ia32_phaddsw128(cast(short8)a, cast(short8)b);
560     }
561     else static if (LDC_with_SSSE3)
562     {
563         return cast(__m128i)__builtin_ia32_phaddsw128(cast(short8)a, cast(short8)b);
564     }
565     else static if (LDC_with_ARM64)
566     {
567         // uzp1/uzp2/sqadd sequence
568         short8 sa = cast(short8)a;
569         short8 sb = cast(short8)b;
570         short8 c = shufflevectorLDC!(short8, 0, 2, 4, 6, 8, 10, 12, 14)(sa, sb);
571         short8 d = shufflevectorLDC!(short8, 1, 3, 5, 7, 9, 11, 13, 15)(sa, sb);
572         return cast(__m128i)vqaddq_s16(c, d);
573     }
574     else
575     {
576         short8 sa = cast(short8)a;
577         short8 sb = cast(short8)b;
578         short8 r;
579         r.ptr[0] = saturateSignedIntToSignedShort(sa.array[0] + sa.array[1]);
580         r.ptr[1] = saturateSignedIntToSignedShort(sa.array[2] + sa.array[3]);
581         r.ptr[2] = saturateSignedIntToSignedShort(sa.array[4] + sa.array[5]);
582         r.ptr[3] = saturateSignedIntToSignedShort(sa.array[6] + sa.array[7]);
583         r.ptr[4] = saturateSignedIntToSignedShort(sb.array[0] + sb.array[1]);
584         r.ptr[5] = saturateSignedIntToSignedShort(sb.array[2] + sb.array[3]);
585         r.ptr[6] = saturateSignedIntToSignedShort(sb.array[4] + sb.array[5]);
586         r.ptr[7] = saturateSignedIntToSignedShort(sb.array[6] + sb.array[7]);
587         return cast(__m128i)r;
588     }
589 }
590 unittest
591 {
592     __m128i A = _mm_setr_epi16(1, -2, 4, 8, 16, 32, -1, -32768);
593     short8 C = cast(short8) _mm_hadds_epi16(A, A);
594     short[8] correct = [ -1, 12, 48, -32768, -1, 12, 48, -32768];
595     assert(C.array == correct);
596 }
597 
598 /// Horizontally add adjacent pairs of signed 16-bit integers in `a` and `b` using saturation, 
599 /// and pack the signed 16-bit results.
600 __m64 _mm_hadds_pi16 (__m64 a, __m64 b) @trusted
601 {
602     static if (GDC_with_SSSE3)
603     {
604         return cast(__m64)__builtin_ia32_phaddsw(cast(short4)a, cast(short4)b);
605     }
606     else static if (LDC_with_SSSE3)
607     {
608         // Note: LDC doesn't have __builtin_ia32_phaddsw
609         long2 la;
610         la.ptr[0] = a.array[0];
611         long2 lb;
612         lb.ptr[0] = b.array[0];
613         int4 sum = cast(int4)__builtin_ia32_phaddsw128(cast(short8)la, cast(short8)lb);
614         int2 r;
615         r.ptr[0] = sum.array[0];
616         r.ptr[1] = sum.array[2];
617         return cast(__m64)r;
618     }
619     else static if (LDC_with_ARM64)
620     {
621         // uzp1/uzp2/sqadd sequence
622         short4 sa = cast(short4)a;
623         short4 sb = cast(short4)b;
624         short4 c = shufflevectorLDC!(short4, 0, 2, 4, 6)(sa, sb);
625         short4 d = shufflevectorLDC!(short4, 1, 3, 5, 7)(sa, sb);
626         return cast(__m64)vqadd_s16(c, d);
627     }
628     else
629     {
630         short4 sa = cast(short4)a;
631         short4 sb = cast(short4)b;
632         short4 r;
633         r.ptr[0] = saturateSignedIntToSignedShort(sa.array[0] + sa.array[1]);
634         r.ptr[1] = saturateSignedIntToSignedShort(sa.array[2] + sa.array[3]);
635         r.ptr[2] = saturateSignedIntToSignedShort(sb.array[0] + sb.array[1]);
636         r.ptr[3] = saturateSignedIntToSignedShort(sb.array[2] + sb.array[3]);
637         return cast(__m64)r;
638     }
639 }
640 unittest
641 {
642     __m64 A = _mm_setr_pi16(-16, 32, -100, -32768);
643     __m64 B = _mm_setr_pi16( 64, 32,    1,  32767);
644     short4 C = cast(short4) _mm_hadds_pi16(A, B);
645     short[4] correct = [ 16, -32768,  96,  32767];
646     assert(C.array == correct);
647 }
648 
649 
650 /// Horizontally add adjacent pairs of 16-bit integers in `a` and `b`, and pack the signed 16-bit results.
651 __m128i _mm_hsub_epi16 (__m128i a, __m128i b) @trusted
652 {
653     // PERF DMD
654     static if (GDC_with_SSSE3)
655     {
656         return cast(__m128i)__builtin_ia32_phsubw128(cast(short8)a, cast(short8)b);
657     }
658     else static if (LDC_with_SSSE3)
659     {
660         return cast(__m128i)__builtin_ia32_phsubw128(cast(short8)a, cast(short8)b);
661     }
662     else static if (LDC_with_ARM64)
663     {
664         // Produce uzp1 uzp2 sub sequence since LDC 1.8 -O1 
665         short8 sa = cast(short8)a;
666         short8 sb = cast(short8)b;
667         short8 c = shufflevectorLDC!(short8, 0, 2, 4, 6, 8, 10, 12, 14)(sa, sb);
668         short8 d = shufflevectorLDC!(short8, 1, 3, 5, 7, 9, 11, 13, 15)(sa, sb);
669         return cast(__m128i)(c - d);
670     }
671     else 
672     {
673         short8 sa = cast(short8)a;
674         short8 sb = cast(short8)b;
675         short8 r;
676         r.ptr[0] = cast(short)(sa.array[0] - sa.array[1]);
677         r.ptr[1] = cast(short)(sa.array[2] - sa.array[3]);
678         r.ptr[2] = cast(short)(sa.array[4] - sa.array[5]);
679         r.ptr[3] = cast(short)(sa.array[6] - sa.array[7]);
680         r.ptr[4] = cast(short)(sb.array[0] - sb.array[1]);
681         r.ptr[5] = cast(short)(sb.array[2] - sb.array[3]);
682         r.ptr[6] = cast(short)(sb.array[4] - sb.array[5]);
683         r.ptr[7] = cast(short)(sb.array[6] - sb.array[7]);
684         return cast(__m128i)r;
685     }
686 }
687 unittest
688 {
689     __m128i A = _mm_setr_epi16(short.min, 1, 4, 8, 16, 32, 1, -32768);
690     short8 C = cast(short8) _mm_hsub_epi16(A, A);
691     short[8] correct = [ short.max, -4, -16, -32767, short.max, -4, -16, -32767];
692     assert(C.array == correct);
693 }
694 
695 /// Horizontally add adjacent pairs of 32-bit integers in `a` and `b`, and pack the signed 32-bit results.
696 __m128i _mm_hsub_epi32 (__m128i a, __m128i b) @trusted
697 { 
698     // PERF DMD
699     static if (GDC_with_SSSE3)
700     {
701         return cast(__m128i)__builtin_ia32_phsubd128(cast(int4)a, cast(int4)b);
702     }
703     else static if (LDC_with_SSSE3)
704     {
705         return cast(__m128i)__builtin_ia32_phsubd128(cast(int4)a, cast(int4)b);
706     }
707     else static if (LDC_with_ARM64)
708     {
709         // Produce uzp1 uzp2 sub sequence since LDC 1.8 -O1 
710         int4 ia = cast(int4)a;
711         int4 ib = cast(int4)b;
712         int4 c = shufflevectorLDC!(int4, 0, 2, 4, 6)(ia, ib);
713         int4 d = shufflevectorLDC!(int4, 1, 3, 5, 7)(ia, ib);
714         return cast(__m128i)(c - d);
715     }
716     else
717     {
718         int4 ia = cast(int4)a;
719         int4 ib = cast(int4)b;
720         int4 r;
721         r.ptr[0] = ia.array[0] - ia.array[1];
722         r.ptr[1] = ia.array[2] - ia.array[3];
723         r.ptr[2] = ib.array[0] - ib.array[1];
724         r.ptr[3] = ib.array[2] - ib.array[3];
725         return cast(__m128i)r;
726     }
727 }
728 unittest
729 {
730     __m128i A = _mm_setr_epi32(1, 2, int.min, 1);
731     __m128i B = _mm_setr_epi32(int.max, -1, 4, 4);
732     int4 C = cast(int4) _mm_hsub_epi32(A, B);
733     int[4] correct = [ -1, int.max, int.min, 0 ];
734     assert(C.array == correct);
735 }
736 
737 /// Horizontally subtract adjacent pairs of 16-bit integers in `a` and `b`, 
738 /// and pack the signed 16-bit results.
739 __m64 _mm_hsub_pi16 (__m64 a, __m64 b) @trusted
740 {
741     // PERF DMD
742     static if (GDC_with_SSSE3)
743     {
744         return cast(__m64)__builtin_ia32_phsubw(cast(short4)a, cast(short4)b);
745     }
746     else static if (LDC_with_ARM64)
747     {
748         // Produce uzp1 uzp2 sub sequence since LDC 1.3 -O1 
749         short4 sa = cast(short4)a;
750         short4 sb = cast(short4)b;
751         short4 c = shufflevectorLDC!(short4, 0, 2, 4, 6)(sa, sb);
752         short4 d = shufflevectorLDC!(short4, 1, 3, 5, 7)(sa, sb);
753         return cast(__m64)(c - d);
754     }
755     else
756     {
757         // LDC x86: generates phsubw since LDC 1.24 -O2
758         short4 sa = cast(short4)a;
759         short4 sb = cast(short4)b;
760         short4 r;
761         r.ptr[0] = cast(short)(sa.array[0] - sa.array[1]);
762         r.ptr[1] = cast(short)(sa.array[2] - sa.array[3]);
763         r.ptr[2] = cast(short)(sb.array[0] - sb.array[1]);
764         r.ptr[3] = cast(short)(sb.array[2] - sb.array[3]);
765         return cast(__m64)r;
766     }
767 }
768 unittest
769 {
770     __m64 A = _mm_setr_pi16(short.min, 1, 4, 8);
771     __m64 B = _mm_setr_pi16(16, 32, 1, -32768);
772     short4 C = cast(short4) _mm_hsub_pi16(A, B);
773     short[4] correct = [ short.max, -4, -16, -32767];
774     assert(C.array == correct);
775 }
776 
777 /// Horizontally subtract adjacent pairs of 32-bit integers in `a` and `b`, 
778 /// and pack the signed 32-bit results.
779 __m64 _mm_hsub_pi32 (__m64 a, __m64 b) @trusted
780 {
781     // PERF DMD
782     static if (GDC_with_SSSE3)
783     {
784         return cast(__m64)__builtin_ia32_phsubd(cast(int2)a, cast(int2)b);
785     }
786     else static if (LDC_with_ARM64)
787     {
788         // LDC arm64: generates zip1+zip2+sub sequence since LDC 1.8 -O1
789         int2 ia = cast(int2)a;
790         int2 ib = cast(int2)b;
791         int2 c = shufflevectorLDC!(int2, 0, 2)(ia, ib);
792         int2 d = shufflevectorLDC!(int2, 1, 3)(ia, ib);
793         return cast(__m64)(c - d);
794     }
795     else
796     {
797         // LDC x86: generates phsubd since LDC 1.24 -O2
798         int2 ia = cast(int2)a;
799         int2 ib = cast(int2)b;
800         int2 r;
801         r.ptr[0] = ia.array[0] - ia.array[1];
802         r.ptr[1] = ib.array[0] - ib.array[1];
803         return cast(__m64)r;
804     }
805 }
806 unittest
807 {
808     __m64 A = _mm_setr_pi32(int.min, 1);
809     __m64 B = _mm_setr_pi32(int.max, -1);
810     int2 C = cast(int2) _mm_hsub_pi32(A, B);
811     int[2] correct = [ int.max, int.min ];
812     assert(C.array == correct);
813 }
814 
815 /// Horizontally subtract adjacent pairs of signed 16-bit integers in `a` and `b` using saturation, 
816 /// and pack the signed 16-bit results.
817 __m128i _mm_hsubs_epi16 (__m128i a, __m128i b) @trusted
818 {
819      // PERF DMD
820     static if (GDC_with_SSSE3)
821     {
822         return cast(__m128i)__builtin_ia32_phsubsw128(cast(short8)a, cast(short8)b);
823     }
824     else static if (LDC_with_SSSE3)
825     {
826         return cast(__m128i)__builtin_ia32_phsubsw128(cast(short8)a, cast(short8)b);
827     }
828     else static if (LDC_with_ARM64)
829     {
830         // uzp1/uzp2/sqsub sequence
831         short8 sa = cast(short8)a;
832         short8 sb = cast(short8)b;
833         short8 c = shufflevectorLDC!(short8, 0, 2, 4, 6, 8, 10, 12, 14)(sa, sb);
834         short8 d = shufflevectorLDC!(short8, 1, 3, 5, 7, 9, 11, 13, 15)(sa, sb);
835         return cast(__m128i)vqsubq_s16(c, d);
836     }
837     else
838     {
839         short8 sa = cast(short8)a;
840         short8 sb = cast(short8)b;
841         short8 r;
842         r.ptr[0] = saturateSignedIntToSignedShort(sa.array[0] - sa.array[1]);
843         r.ptr[1] = saturateSignedIntToSignedShort(sa.array[2] - sa.array[3]);
844         r.ptr[2] = saturateSignedIntToSignedShort(sa.array[4] - sa.array[5]);
845         r.ptr[3] = saturateSignedIntToSignedShort(sa.array[6] - sa.array[7]);
846         r.ptr[4] = saturateSignedIntToSignedShort(sb.array[0] - sb.array[1]);
847         r.ptr[5] = saturateSignedIntToSignedShort(sb.array[2] - sb.array[3]);
848         r.ptr[6] = saturateSignedIntToSignedShort(sb.array[4] - sb.array[5]);
849         r.ptr[7] = saturateSignedIntToSignedShort(sb.array[6] - sb.array[7]);
850         return cast(__m128i)r;
851     }
852 }
853 unittest
854 {
855     __m128i A = _mm_setr_epi16(1, -2, 4, 8, 32767, -1, -10, 32767);
856     short8 C = cast(short8) _mm_hsubs_epi16(A, A);
857     short[8] correct = [ 3, -4, 32767, -32768, 3, -4, 32767, -32768 ];
858     assert(C.array == correct);
859 }
860 
861 
862 /// Horizontally subtract adjacent pairs of signed 16-bit integers in `a` and `b` using saturation, 
863 /// and pack the signed 16-bit results.
864 __m64 _mm_hsubs_pi16 (__m64 a, __m64 b) @trusted
865 {
866     static if (GDC_with_SSSE3)
867     {
868         return cast(__m64)__builtin_ia32_phsubsw(cast(short4)a, cast(short4)b);
869     }
870     else static if (LDC_with_SSSE3)
871     {
872         // Note: LDC doesn't have __builtin_ia32_phsubsw
873         long2 la;
874         la.ptr[0] = a.array[0];
875         long2 lb;
876         lb.ptr[0] = b.array[0];
877         int4 sum = cast(int4)__builtin_ia32_phsubsw128(cast(short8)la, cast(short8)lb);
878         int2 r;
879         r.ptr[0] = sum.array[0];
880         r.ptr[1] = sum.array[2];
881         return cast(__m64)r;
882     }
883     else static if (LDC_with_ARM64)
884     {
885         // uzp1/uzp2/sqsub sequence in -O1
886         short4 sa = cast(short4)a;
887         short4 sb = cast(short4)b;
888         short4 c = shufflevectorLDC!(short4, 0, 2, 4, 6)(sa, sb);
889         short4 d = shufflevectorLDC!(short4, 1, 3, 5, 7)(sa, sb);
890         return cast(__m64)vqsub_s16(c, d);
891     }
892     else
893     {
894         short4 sa = cast(short4)a;
895         short4 sb = cast(short4)b;
896         short4 r;
897         r.ptr[0] = saturateSignedIntToSignedShort(sa.array[0] - sa.array[1]);
898         r.ptr[1] = saturateSignedIntToSignedShort(sa.array[2] - sa.array[3]);
899         r.ptr[2] = saturateSignedIntToSignedShort(sb.array[0] - sb.array[1]);
900         r.ptr[3] = saturateSignedIntToSignedShort(sb.array[2] - sb.array[3]);
901         return cast(__m64)r;
902     }
903 }
904 unittest
905 {
906     __m64 A = _mm_setr_pi16(-16, 32, 100, -32768);
907     __m64 B = _mm_setr_pi16( 64, 30,   -9,  32767);
908     short4 C = cast(short4) _mm_hsubs_pi16(A, B);
909     short[4] correct = [ -48, 32767,  34,  -32768];
910     assert(C.array == correct);
911 }
912 
913 
914 /// Vertically multiply each unsigned 8-bit integer from `a` with the corresponding 
915 /// signed 8-bit integer from `b`, producing intermediate signed 16-bit integers. 
916 /// Horizontally add adjacent pairs of intermediate signed 16-bit integers, 
917 /// and pack the saturated results.
918 __m128i _mm_maddubs_epi16 (__m128i a, __m128i b) @trusted
919 {
920     static if (GDC_with_SSSE3)
921     {
922         return cast(__m128i)__builtin_ia32_pmaddubsw128(cast(ubyte16)a, cast(ubyte16)b);
923     }
924     else static if (LDC_with_SSSE3)
925     {
926         return cast(__m128i)__builtin_ia32_pmaddubsw128(cast(byte16)a, cast(byte16)b);
927     }
928     else
929     {
930         // zero-extend a to 16-bit
931         __m128i zero = _mm_setzero_si128();
932         __m128i a_lo = _mm_unpacklo_epi8(a, zero);
933         __m128i a_hi = _mm_unpackhi_epi8(a, zero);
934 
935         // sign-extend b to 16-bit
936         __m128i b_lo = _mm_unpacklo_epi8(b, zero);
937         __m128i b_hi = _mm_unpackhi_epi8(b, zero);    
938         b_lo = _mm_srai_epi16( _mm_slli_epi16(b_lo, 8), 8);
939         b_hi = _mm_srai_epi16( _mm_slli_epi16(b_hi, 8), 8); 
940 
941         // Multiply element-wise, no overflow can occur
942         __m128i c_lo = _mm_mullo_epi16(a_lo, b_lo);  
943         __m128i c_hi = _mm_mullo_epi16(a_hi, b_hi);
944 
945         // Add pairwise with saturating horizontal add
946         return _mm_hadds_epi16(c_lo, c_hi);
947     }
948 }
949 unittest
950 {
951     __m128i A = _mm_setr_epi8(  -1,  10, 100, -128, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); // u8
952     __m128i B = _mm_setr_epi8(-128, -30, 100,  127, -1, 2, 4, 6, 0, 0, 0, 0, 0, 0, 0, 0); // i8
953     short8 C = cast(short8) _mm_maddubs_epi16(A, B);
954     short[8] correct =       [   -32768,     26256, 0, 0, 0, 0, 0, 0];
955     assert(C.array == correct);
956 }
957 
958 /// Vertically multiply each unsigned 8-bit integer from `a` with the corresponding 
959 /// signed 8-bit integer from `b`, producing intermediate signed 16-bit integers. 
960 /// Horizontally add adjacent pairs of intermediate signed 16-bit integers, 
961 /// and pack the saturated results.
962 __m64 _mm_maddubs_pi16 (__m64 a, __m64 b) @trusted
963 {
964     static if (GDC_with_SSSE3)
965     {
966         return cast(__m64)__builtin_ia32_pmaddubsw(cast(ubyte8)a, cast(ubyte8)b);
967     }
968     else static if (LDC_with_SSSE3)
969     {
970         __m128i A = to_m128i(a);
971         __m128i B = to_m128i(b);
972         return to_m64( cast(__m128i)__builtin_ia32_pmaddubsw128(cast(byte16) to_m128i(a), cast(byte16) to_m128i(b)));
973     }
974     else
975     {
976         // zero-extend a to 16-bit
977         __m128i zero = _mm_setzero_si128();
978         __m128i A = _mm_unpacklo_epi8(to_m128i(a), zero);
979 
980         // sign-extend b to 16-bit
981         __m128i B = _mm_unpacklo_epi8(to_m128i(b), zero);    
982         B = _mm_srai_epi16( _mm_slli_epi16(B, 8), 8);
983 
984         // Multiply element-wise, no overflow can occur
985         __m128i c = _mm_mullo_epi16(A, B);
986 
987         // Add pairwise with saturating horizontal add
988         return to_m64( _mm_hadds_epi16(c, zero));
989     }
990 }
991 unittest
992 {
993     __m64 A = _mm_setr_pi8(  -1,  10, 100, -128, 0, 0, 0, 0); // u8
994     __m64 B = _mm_setr_pi8(-128, -30, 100,  127, -1, 2, 4, 6); // i8
995     short4 C = cast(short4) _mm_maddubs_pi16(A, B);
996     short[4] correct =       [   -32768,   26256, 0, 0];
997     assert(C.array == correct);
998 }
999 
1000 /// Multiply packed signed 16-bit integers in `a` and `b`, producing intermediate signed 32-bit integers.
1001 /// Truncate each intermediate integer to the 18 most significant bits, round by adding 1, and return bits `[16:1]`.
1002 __m128i _mm_mulhrs_epi16 (__m128i a, __m128i b) @trusted
1003 {
1004     // PERF DMD
1005     static if (GDC_with_SSSE3)
1006     {
1007         return cast(__m128i) __builtin_ia32_pmulhrsw128(cast(short8)a, cast(short8)b);
1008     }
1009     else static if (LDC_with_SSSE3)
1010     {
1011         return cast(__m128i) __builtin_ia32_pmulhrsw128(cast(short8)a, cast(short8)b);
1012     }
1013     else static if (LDC_with_ARM64)
1014     {
1015         int4 mul_lo = vmull_s16(vget_low_s16(cast(short8)a),
1016                                 vget_low_s16(cast(short8)b));
1017         int4 mul_hi = vmull_s16(vget_high_s16(cast(short8)a),
1018                                 vget_high_s16(cast(short8)b));
1019 
1020         // Rounding narrowing shift right
1021         // narrow = (int16_t)((mul + 16384) >> 15);
1022         short4 narrow_lo = vrshrn_n_s32(mul_lo, 15);
1023         short4 narrow_hi = vrshrn_n_s32(mul_hi, 15);
1024 
1025         // Join together.
1026         return cast(__m128i) vcombine_s16(narrow_lo, narrow_hi);
1027     }
1028     else
1029     {
1030         short8 sa = cast(short8)a;
1031         short8 sb = cast(short8)b;
1032         short8 r;
1033 
1034         for (int i = 0; i < 8; ++i)
1035         {
1036             // I doubted it at first, but an exhaustive search show this to be equivalent to Intel pseudocode.
1037             r.ptr[i] = cast(short) ( (sa.array[i] * sb.array[i] + 0x4000) >> 15);
1038         }
1039 
1040         return cast(__m128i)r;
1041     }
1042 }
1043 
1044 unittest
1045 {
1046     __m128i A = _mm_setr_epi16(12345, -32768, 32767, 0, 1, 845, -6999, -1);
1047     __m128i B = _mm_setr_epi16(8877, -24487, 15678, 32760, 1, 0, -149, -1);
1048     short8 C = cast(short8) _mm_mulhrs_epi16(A, B);
1049     short[8] correct = [3344, 24487, 15678, 0, 0, 0, 32, 0];
1050     assert(C.array == correct);
1051 }
1052 
1053 /// Multiply packed signed 16-bit integers in `a` and `b`, producing intermediate signed 32-bit integers.
1054 /// Truncate each intermediate integer to the 18 most significant bits, round by adding 1, and return bits `[16:1]`.
1055 __m64 _mm_mulhrs_pi16 (__m64 a, __m64 b) @trusted
1056 {
1057     // PERF DMD
1058     static if (GDC_with_SSSE3)
1059     {
1060         return cast(__m64) __builtin_ia32_pmulhrsw(cast(short4)a, cast(short4)b);
1061     }
1062     else static if (LDC_with_SSSE3)
1063     {
1064         return cast(__m64) to_m64( cast(__m128i) __builtin_ia32_pmulhrsw128(cast(short8) to_m128i(a), cast(short8) to_m128i(b)));
1065     }
1066     else static if (LDC_with_ARM64)
1067     {
1068         int4 mul = vmull_s16(cast(short4)a, cast(short4)b);
1069 
1070         // Rounding narrowing shift right
1071         // (int16_t)((mul + 16384) >> 15);
1072         return cast(__m64) vrshrn_n_s32(mul, 15);
1073     }
1074     else
1075     {
1076         short4 sa = cast(short4)a;
1077         short4 sb = cast(short4)b;
1078         short4 r;
1079 
1080         for (int i = 0; i < 4; ++i)
1081         {
1082             r.ptr[i] = cast(short) ( (sa.array[i] * sb.array[i] + 0x4000) >> 15);
1083         }
1084         return cast(__m64)r;
1085     }
1086 }
1087 unittest
1088 {
1089     __m64 A = _mm_setr_pi16(12345, -32768, 32767, 0);
1090     __m64 B = _mm_setr_pi16(8877, -24487, 15678, 32760);
1091     short4 C = cast(short4) _mm_mulhrs_pi16(A, B);
1092     short[4] correct = [3344, 24487, 15678, 0];
1093     assert(C.array == correct);
1094 }
1095 
1096 
1097 /// Shuffle packed 8-bit integers in `a` according to shuffle control mask in the corresponding 8-bit element of `b`.
1098 __m128i _mm_shuffle_epi8 (__m128i a, __m128i b) @trusted
1099 {
1100     // This is the lovely pshufb.
1101     // PERF DMD
1102     static if (GDC_with_SSSE3)
1103     {
1104         return cast(__m128i) __builtin_ia32_pshufb128(cast(ubyte16) a, cast(ubyte16) b);
1105     }
1106     else static if (LDC_with_SSSE3)
1107     {
1108         return cast(__m128i) __builtin_ia32_pshufb128(cast(byte16) a, cast(byte16) b);
1109     }
1110     else static if (LDC_with_ARM64)
1111     {
1112         byte16 bb = cast(byte16)b;
1113         byte16 mask;
1114         mask = cast(byte)(0x8F);
1115         bb = bb & mask;
1116         byte16 r = vqtbl1q_s8(cast(byte16)a, bb);
1117         return cast(__m128i)r;
1118     }
1119     else
1120     {
1121         byte16 r;
1122         byte16 ba = cast(byte16)a;
1123         byte16 bb = cast(byte16)b;
1124         for (int i = 0; i < 16; ++i)
1125         {
1126             byte s = bb.array[i];
1127             r.ptr[i] = (s < 0) ? 0 : ba.array[ s & 15 ];
1128         }
1129         return cast(__m128i)r;
1130     }
1131 }
1132 unittest
1133 {
1134     __m128i A = _mm_setr_epi8(15,   14,      13,  12, 11,  10, 9, 8, 7, 6,  5,  4,  3,  2,  1,  0);
1135     __m128i B = _mm_setr_epi8(15, -128, 13 + 16, -12, 11, -10, 9, 8, 7, 6, -5,  4,  3, -2,  1,  0);
1136     byte16 C = cast(byte16) _mm_shuffle_epi8(A, B);
1137     byte[16] correct =         [0,   0,       2,  0,  4,   0, 6, 7, 8, 9,  0, 11, 12,  0, 14, 15];
1138     assert(C.array == correct);
1139 }
1140 
1141 /// Shuffle packed 8-bit integers in `a` according to shuffle control mask in the corresponding 8-bit element of `b`.
1142 __m64 _mm_shuffle_pi8 (__m64 a, __m64 b) @trusted
1143 {
1144     // PERF DMD
1145     static if (GDC_with_SSSE3)
1146     {
1147         alias ubyte8  =__vector(ubyte[8]);
1148         return cast(__m64) __builtin_ia32_pshufb(cast(ubyte8) a, cast(ubyte8) b);
1149     }
1150     else static if (LDC_with_SSSE3)
1151     {
1152         // GDC does proper dance to avoid mmx registers, do it manually in LDC since __builtin_ia32_pshufb doesn't exist there
1153         __m128i A = to_m128i(a);
1154         __m128i index = to_m128i(b);
1155         index = index & _mm_set1_epi32(0xF7F7F7F7);
1156         return to_m64( cast(__m128i) __builtin_ia32_pshufb128(cast(byte16)A, cast(byte16) index) );
1157     }
1158     else static if (LDC_with_ARM64)
1159     {
1160         byte8 bb = cast(byte8)b;
1161         byte8 mask;
1162         mask = cast(byte)(0x87);
1163         bb = bb & mask;
1164         __m128i l = to_m128i(a);
1165         byte8 r = vtbl1_s8(cast(byte16)l, cast(byte8)bb);
1166         return cast(__m64)r;
1167     }
1168     else
1169     {
1170         byte8 r;
1171         byte8 ba = cast(byte8)a;
1172         byte8 bb = cast(byte8)b;
1173         for (int i = 0; i < 8; ++i)
1174         {
1175             byte s = bb.array[i];
1176             r.ptr[i] = (s < 0) ? 0 : ba.array[ s & 7 ];
1177         }
1178         return cast(__m64)r;
1179     }
1180 }
1181 unittest
1182 {
1183     __m64 A = _mm_setr_pi8(7,  6,  5,  4,      3,  2,  1,  0);
1184     __m64 B = _mm_setr_pi8(7,  6, -5,  4,  3 + 8, -2,  1,  0);
1185     byte8 C = cast(byte8) _mm_shuffle_pi8(A, B);
1186     byte[8] correct =    [0,  1,  0,  3,      4,  0,  6,  7];
1187     assert(C.array == correct);
1188 }
1189 
1190 /// Negate packed 16-bit integers in `a` when the corresponding signed 16-bit integer in `b` is negative.
1191 /// Elements in result are zeroed out when the corresponding element in `b` is zero.
1192 __m128i _mm_sign_epi16 (__m128i a, __m128i b) @trusted
1193 {
1194     // PERF DMD
1195     static if (GDC_with_SSSE3)
1196     {
1197         return cast(__m128i) __builtin_ia32_psignw128(cast(short8)a, cast(short8)b);
1198     }
1199     else static if (LDC_with_SSSE3)
1200     {
1201         return cast(__m128i) __builtin_ia32_psignw128(cast(short8)a, cast(short8)b);       
1202     }
1203     else
1204     {
1205         // LDC arm64: 5 instructions
1206         __m128i mask = _mm_srai_epi16(b, 15);
1207         __m128i zeromask = _mm_cmpeq_epi16(b, _mm_setzero_si128());
1208         return _mm_andnot_si128(zeromask, _mm_xor_si128(_mm_add_epi16(a, mask), mask));
1209     }
1210 }
1211 unittest
1212 {
1213     __m128i A = _mm_setr_epi16(-2, -1, 0, 1,  2, short.min, short.min, short.min);
1214     __m128i B = _mm_setr_epi16(-1,  0,-1, 1, -2,       -50,         0,        50);
1215     short8 C = cast(short8) _mm_sign_epi16(A, B);
1216     short[8] correct =        [ 2,  0, 0, 1, -2, short.min,         0, short.min];
1217     assert(C.array == correct);
1218 }
1219 
1220 /// Negate packed 32-bit integers in `a` when the corresponding signed 32-bit integer in `b` is negative. 
1221 /// Elements in result are zeroed out when the corresponding element in `b` is zero.
1222 __m128i _mm_sign_epi32 (__m128i a, __m128i b) @trusted
1223 {
1224     // PERF DMD
1225     static if (GDC_with_SSSE3)
1226     {
1227         return cast(__m128i) __builtin_ia32_psignd128(cast(short8)a, cast(short8)b);
1228     }
1229     else static if (LDC_with_SSSE3)
1230     {
1231         return cast(__m128i) __builtin_ia32_psignd128(cast(short8)a, cast(short8)b);
1232     }
1233     else
1234     {
1235         __m128i mask = _mm_srai_epi32(b, 31);
1236         __m128i zeromask = _mm_cmpeq_epi32(b, _mm_setzero_si128());
1237         return _mm_andnot_si128(zeromask, _mm_xor_si128(_mm_add_epi32(a, mask), mask));
1238     }
1239 }
1240 unittest
1241 {
1242     __m128i A = _mm_setr_epi32(-2, -1,  0, int.max);
1243     __m128i B = _mm_setr_epi32(-1,  0, -1, 1);
1244     int4 C = cast(int4) _mm_sign_epi32(A, B);
1245     int[4] correct =          [ 2,  0, 0, int.max];
1246     assert(C.array == correct);
1247 }
1248 
1249 /// Negate packed 8-bit integers in `a` when the corresponding signed 8-bit integer in `b` is negative. 
1250 /// Elements in result are zeroed out when the corresponding element in `b` is zero.
1251 __m128i _mm_sign_epi8 (__m128i a, __m128i b) @trusted
1252 {
1253     // PERF DMD
1254     static if (GDC_with_SSSE3)
1255     {
1256         return cast(__m128i) __builtin_ia32_psignb128(cast(ubyte16)a, cast(ubyte16)b);
1257     }
1258     else static if (LDC_with_SSSE3)
1259     {
1260         return cast(__m128i) __builtin_ia32_psignb128(cast(byte16)a, cast(byte16)b);
1261     }
1262     else
1263     {
1264         __m128i mask = _mm_cmplt_epi8(b, _mm_setzero_si128()); // extend sign bit
1265         __m128i zeromask = _mm_cmpeq_epi8(b, _mm_setzero_si128());
1266         return _mm_andnot_si128(zeromask, _mm_xor_si128(_mm_add_epi8(a, mask), mask));
1267     }
1268 }
1269 unittest
1270 {
1271     __m128i A = _mm_setr_epi8(-2, -1, 0, 1,  2, byte.min, byte.min, byte.min, -1,  0,-1, 1, -2,      -50,        0,       50);
1272     __m128i B = _mm_setr_epi8(-1,  0,-1, 1, -2,      -50,        0,       50, -2, -1, 0, 1,  2, byte.min, byte.min, byte.min);
1273     byte16  C = cast(byte16) _mm_sign_epi8(A, B);
1274     byte[16] correct =       [ 2,  0, 0, 1, -2, byte.min,        0, byte.min,  1,  0, 0, 1, -2,       50,        0,      -50];
1275     assert(C.array == correct);
1276 }
1277 
1278 /// Negate packed 16-bit integers in `a`  when the corresponding signed 16-bit integer in `b` is negative.
1279 /// Element in result are zeroed out when the corresponding element in `b` is zero.
1280 __m64 _mm_sign_pi16 (__m64 a, __m64 b) @trusted
1281 {
1282     return to_m64( _mm_sign_epi16( to_m128i(a), to_m128i(b)) );
1283 }
1284 unittest
1285 {
1286     __m64 A = _mm_setr_pi16( 2, short.min, short.min, short.min);
1287     __m64 B = _mm_setr_pi16(-2,       -50,         0,        50);
1288     short4 C = cast(short4) _mm_sign_pi16(A, B);
1289     short[4] correct =     [-2, short.min,         0, short.min];
1290     assert(C.array == correct);
1291 }
1292 
1293 /// Negate packed 32-bit integers in `a`  when the corresponding signed 32-bit integer in `b` is negative.
1294 /// Element in result are zeroed out when the corresponding element in `b` is zero.
1295 __m64 _mm_sign_pi32 (__m64 a, __m64 b) @trusted
1296 {
1297     return to_m64( _mm_sign_epi32( to_m128i(a), to_m128i(b)) );
1298 }
1299 unittest
1300 {
1301     __m64 A = _mm_setr_pi32(-2, -100);
1302     __m64 B = _mm_setr_pi32(-1,  0);
1303     int2 C = cast(int2) _mm_sign_pi32(A, B);
1304     int[2] correct =          [ 2,  0];
1305     assert(C.array == correct);
1306 }
1307 
1308 /// Negate packed 8-bit integers in `a` when the corresponding signed 8-bit integer in `b` is negative. 
1309 /// Elements in result are zeroed out when the corresponding element in `b` is zero.
1310 __m64 _mm_sign_pi8 (__m64 a, __m64 b) @trusted
1311 {
1312     return to_m64( _mm_sign_epi8( to_m128i(a), to_m128i(b)) );
1313 }
1314 unittest
1315 {
1316     __m64 A = _mm_setr_pi8(-2, -1, 0, 1,  2, byte.min, byte.min, byte.min);
1317     __m64 B = _mm_setr_pi8(-1,  0,-1, 1, -2,      -50,        0,       50);
1318     byte8  C = cast(byte8) _mm_sign_pi8(A, B);
1319     byte[8] correct =     [ 2,  0, 0, 1, -2, byte.min,        0, byte.min];
1320     assert(C.array == correct);
1321 }