35 typedef float __m128 __attribute__((__vector_size__(16)));  typedef
43 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
44 _mm_add_ss(__m128 a, __m128 b)  in _mm_add_ss()
50 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
51 _mm_add_ps(__m128 a, __m128 b)  in _mm_add_ps()
56 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
57 _mm_sub_ss(__m128 a, __m128 b)  in _mm_sub_ss()
63 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
64 _mm_sub_ps(__m128 a, __m128 b)  in _mm_sub_ps()
69 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
70 _mm_mul_ss(__m128 a, __m128 b)  in _mm_mul_ss()
76 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
77 _mm_mul_ps(__m128 a, __m128 b)  in _mm_mul_ps()
82 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
83 _mm_div_ss(__m128 a, __m128 b)  in _mm_div_ss()
89 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
90 _mm_div_ps(__m128 a, __m128 b)  in _mm_div_ps()
95 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
96 _mm_sqrt_ss(__m128 a)  in _mm_sqrt_ss()
101 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
102 _mm_sqrt_ps(__m128 a)  in _mm_sqrt_ps()
107 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
108 _mm_rcp_ss(__m128 a)  in _mm_rcp_ss()
113 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
114 _mm_rcp_ps(__m128 a)  in _mm_rcp_ps()
119 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
120 _mm_rsqrt_ss(__m128 a)  in _mm_rsqrt_ss()
125 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
126 _mm_rsqrt_ps(__m128 a)  in _mm_rsqrt_ps()
131 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
132 _mm_min_ss(__m128 a, __m128 b)  in _mm_min_ss()
137 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
138 _mm_min_ps(__m128 a, __m128 b)  in _mm_min_ps()
143 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
144 _mm_max_ss(__m128 a, __m128 b)  in _mm_max_ss()
149 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
150 _mm_max_ps(__m128 a, __m128 b)  in _mm_max_ps()
155 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
156 _mm_and_ps(__m128 a, __m128 b)  in _mm_and_ps()
158   return (__m128)((__v4si)a & (__v4si)b);  in _mm_and_ps()
161 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
162 _mm_andnot_ps(__m128 a, __m128 b)  in _mm_andnot_ps()
164   return (__m128)(~(__v4si)a & (__v4si)b);  in _mm_andnot_ps()
167 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
168 _mm_or_ps(__m128 a, __m128 b)  in _mm_or_ps()
170   return (__m128)((__v4si)a | (__v4si)b);  in _mm_or_ps()
173 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
174 _mm_xor_ps(__m128 a, __m128 b)  in _mm_xor_ps()
176   return (__m128)((__v4si)a ^ (__v4si)b);  in _mm_xor_ps()
179 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
180 _mm_cmpeq_ss(__m128 a, __m128 b)  in _mm_cmpeq_ss()
182   return (__m128)__builtin_ia32_cmpss(a, b, 0);  in _mm_cmpeq_ss()
185 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
186 _mm_cmpeq_ps(__m128 a, __m128 b)  in _mm_cmpeq_ps()
188   return (__m128)__builtin_ia32_cmpps(a, b, 0);  in _mm_cmpeq_ps()
191 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
192 _mm_cmplt_ss(__m128 a, __m128 b)  in _mm_cmplt_ss()
194   return (__m128)__builtin_ia32_cmpss(a, b, 1);  in _mm_cmplt_ss()
197 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
198 _mm_cmplt_ps(__m128 a, __m128 b)  in _mm_cmplt_ps()
200   return (__m128)__builtin_ia32_cmpps(a, b, 1);  in _mm_cmplt_ps()
203 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
204 _mm_cmple_ss(__m128 a, __m128 b)  in _mm_cmple_ss()
206   return (__m128)__builtin_ia32_cmpss(a, b, 2);  in _mm_cmple_ss()
209 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
210 _mm_cmple_ps(__m128 a, __m128 b)  in _mm_cmple_ps()
212   return (__m128)__builtin_ia32_cmpps(a, b, 2);  in _mm_cmple_ps()
215 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
216 _mm_cmpgt_ss(__m128 a, __m128 b)  in _mm_cmpgt_ss()
218   return (__m128)__builtin_ia32_cmpss(b, a, 1);  in _mm_cmpgt_ss()
221 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
222 _mm_cmpgt_ps(__m128 a, __m128 b)  in _mm_cmpgt_ps()
224   return (__m128)__builtin_ia32_cmpps(b, a, 1);  in _mm_cmpgt_ps()
227 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
228 _mm_cmpge_ss(__m128 a, __m128 b)  in _mm_cmpge_ss()
230   return (__m128)__builtin_ia32_cmpss(b, a, 2);  in _mm_cmpge_ss()
233 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
234 _mm_cmpge_ps(__m128 a, __m128 b)  in _mm_cmpge_ps()
236   return (__m128)__builtin_ia32_cmpps(b, a, 2);  in _mm_cmpge_ps()
239 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
240 _mm_cmpneq_ss(__m128 a, __m128 b)  in _mm_cmpneq_ss()
242   return (__m128)__builtin_ia32_cmpss(a, b, 4);  in _mm_cmpneq_ss()
245 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
246 _mm_cmpneq_ps(__m128 a, __m128 b)  in _mm_cmpneq_ps()
248   return (__m128)__builtin_ia32_cmpps(a, b, 4);  in _mm_cmpneq_ps()
251 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
252 _mm_cmpnlt_ss(__m128 a, __m128 b)  in _mm_cmpnlt_ss()
254   return (__m128)__builtin_ia32_cmpss(a, b, 5);  in _mm_cmpnlt_ss()
257 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
258 _mm_cmpnlt_ps(__m128 a, __m128 b)  in _mm_cmpnlt_ps()
260   return (__m128)__builtin_ia32_cmpps(a, b, 5);  in _mm_cmpnlt_ps()
263 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
264 _mm_cmpnle_ss(__m128 a, __m128 b)  in _mm_cmpnle_ss()
266   return (__m128)__builtin_ia32_cmpss(a, b, 6);  in _mm_cmpnle_ss()
269 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
270 _mm_cmpnle_ps(__m128 a, __m128 b)  in _mm_cmpnle_ps()
272   return (__m128)__builtin_ia32_cmpps(a, b, 6);  in _mm_cmpnle_ps()
275 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
276 _mm_cmpngt_ss(__m128 a, __m128 b)  in _mm_cmpngt_ss()
278   return (__m128)__builtin_ia32_cmpss(b, a, 5);  in _mm_cmpngt_ss()
281 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
282 _mm_cmpngt_ps(__m128 a, __m128 b)  in _mm_cmpngt_ps()
284   return (__m128)__builtin_ia32_cmpps(b, a, 5);  in _mm_cmpngt_ps()
287 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
288 _mm_cmpnge_ss(__m128 a, __m128 b)  in _mm_cmpnge_ss()
290   return (__m128)__builtin_ia32_cmpss(b, a, 6);  in _mm_cmpnge_ss()
293 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
294 _mm_cmpnge_ps(__m128 a, __m128 b)  in _mm_cmpnge_ps()
296   return (__m128)__builtin_ia32_cmpps(b, a, 6);  in _mm_cmpnge_ps()
299 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
300 _mm_cmpord_ss(__m128 a, __m128 b)  in _mm_cmpord_ss()
302   return (__m128)__builtin_ia32_cmpss(a, b, 7);  in _mm_cmpord_ss()
305 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
306 _mm_cmpord_ps(__m128 a, __m128 b)  in _mm_cmpord_ps()
308   return (__m128)__builtin_ia32_cmpps(a, b, 7);  in _mm_cmpord_ps()
311 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
312 _mm_cmpunord_ss(__m128 a, __m128 b)  in _mm_cmpunord_ss()
314   return (__m128)__builtin_ia32_cmpss(a, b, 3);  in _mm_cmpunord_ss()
317 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
318 _mm_cmpunord_ps(__m128 a, __m128 b)  in _mm_cmpunord_ps()
320   return (__m128)__builtin_ia32_cmpps(a, b, 3);  in _mm_cmpunord_ps()
324 _mm_comieq_ss(__m128 a, __m128 b)  in _mm_comieq_ss()
330 _mm_comilt_ss(__m128 a, __m128 b)  in _mm_comilt_ss()
336 _mm_comile_ss(__m128 a, __m128 b)  in _mm_comile_ss()
342 _mm_comigt_ss(__m128 a, __m128 b)  in _mm_comigt_ss()
348 _mm_comige_ss(__m128 a, __m128 b)  in _mm_comige_ss()
354 _mm_comineq_ss(__m128 a, __m128 b)  in _mm_comineq_ss()
360 _mm_ucomieq_ss(__m128 a, __m128 b)  in _mm_ucomieq_ss()
366 _mm_ucomilt_ss(__m128 a, __m128 b)  in _mm_ucomilt_ss()
372 _mm_ucomile_ss(__m128 a, __m128 b)  in _mm_ucomile_ss()
378 _mm_ucomigt_ss(__m128 a, __m128 b)  in _mm_ucomigt_ss()
384 _mm_ucomige_ss(__m128 a, __m128 b)  in _mm_ucomige_ss()
390 _mm_ucomineq_ss(__m128 a, __m128 b)  in _mm_ucomineq_ss()
396 _mm_cvtss_si32(__m128 a)  in _mm_cvtss_si32()
402 _mm_cvt_ss2si(__m128 a)  in _mm_cvt_ss2si()
410 _mm_cvtss_si64(__m128 a)  in _mm_cvtss_si64()
418 _mm_cvtps_pi32(__m128 a)  in _mm_cvtps_pi32()
424 _mm_cvt_ps2pi(__m128 a)  in _mm_cvt_ps2pi()
430 _mm_cvttss_si32(__m128 a)  in _mm_cvttss_si32()
436 _mm_cvtt_ss2si(__m128 a)  in _mm_cvtt_ss2si()
442 _mm_cvttss_si64(__m128 a)  in _mm_cvttss_si64()
448 _mm_cvttps_pi32(__m128 a)  in _mm_cvttps_pi32()
454 _mm_cvtt_ps2pi(__m128 a)  in _mm_cvtt_ps2pi()
459 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
460 _mm_cvtsi32_ss(__m128 a, int b)  in _mm_cvtsi32_ss()
466 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
467 _mm_cvt_si2ss(__m128 a, int b)  in _mm_cvt_si2ss()
474 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
475 _mm_cvtsi64_ss(__m128 a, long long b)  in _mm_cvtsi64_ss()
483 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
484 _mm_cvtpi32_ps(__m128 a, __m64 b)  in _mm_cvtpi32_ps()
489 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
490 _mm_cvt_pi2ps(__m128 a, __m64 b)  in _mm_cvt_pi2ps()
496 _mm_cvtss_f32(__m128 a)  in _mm_cvtss_f32()
501 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
502 _mm_loadh_pi(__m128 a, const __m64 *p)  in _mm_loadh_pi()
509   __m128 bb = __builtin_shufflevector(b, b, 0, 1, 0, 1);  in _mm_loadh_pi()
513 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
514 _mm_loadl_pi(__m128 a, const __m64 *p)  in _mm_loadl_pi()
521   __m128 bb = __builtin_shufflevector(b, b, 0, 1, 0, 1);  in _mm_loadl_pi()
525 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
532   return (__m128){ u, 0, 0, 0 };  in _mm_load_ss()
535 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
542   return (__m128){ u, u, u, u };  in _mm_load1_ps()
547 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
550   return *(__m128*)p;  in _mm_load_ps()
553 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
557     __m128 v;  in _mm_loadu_ps()
562 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
565   __m128 a = _mm_load_ps(p);  in _mm_loadr_ps()
569 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
572   return (__m128){ w, 0, 0, 0 };  in _mm_set_ss()
575 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
578   return (__m128){ w, w, w, w };  in _mm_set1_ps()
582 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
588 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
591   return (__m128){ w, x, y, z };  in _mm_set_ps()
594 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
597   return (__m128){ z, y, x, w };  in _mm_setr_ps()
600 static __inline__ __m128 __attribute__((__always_inline__))
603   return (__m128){ 0, 0, 0, 0 };  in _mm_setzero_ps()
607 _mm_storeh_pi(__m64 *p, __m128 a)  in _mm_storeh_pi()
613 _mm_storel_pi(__m64 *p, __m128 a)  in _mm_storel_pi()
619 _mm_store_ss(float *p, __m128 a)  in _mm_store_ss()
628 _mm_storeu_ps(float *p, __m128 a)  in _mm_storeu_ps()
634 _mm_store1_ps(float *p, __m128 a)  in _mm_store1_ps()
641 _mm_store_ps1(float *p, __m128 a)  in _mm_store_ps1()
647 _mm_store_ps(float *p, __m128 a)  in _mm_store_ps()
649   *(__m128 *)p = a;  in _mm_store_ps()
653 _mm_storer_ps(float *p, __m128 a)  in _mm_storer_ps()
676 _mm_stream_ps(float *p, __m128 a)  in _mm_stream_ps()
779   __m128 __a = (a); \
780   __m128 __b = (b); \
781   (__m128)__builtin_shufflevector((__v4sf)__a, (__v4sf)__b, \
786 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
787 _mm_unpackhi_ps(__m128 a, __m128 b)  in _mm_unpackhi_ps()
792 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
793 _mm_unpacklo_ps(__m128 a, __m128 b)  in _mm_unpacklo_ps()
798 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
799 _mm_move_ss(__m128 a, __m128 b)  in _mm_move_ss()
804 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
805 _mm_movehl_ps(__m128 a, __m128 b)  in _mm_movehl_ps()
810 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
811 _mm_movelh_ps(__m128 a, __m128 b)  in _mm_movelh_ps()
816 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
820   __m128 r;  in _mm_cvtpi16_ps()
834 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
838   __m128 r;  in _mm_cvtpu16_ps()
851 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
863 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
874 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
877   __m128 c;  in _mm_cvtpi32x2_ps()
887 _mm_cvtps_pi16(__m128 a)  in _mm_cvtps_pi16()
899 _mm_cvtps_pi8(__m128 a)  in _mm_cvtps_pi8()
910 _mm_movemask_ps(__m128 a)  in _mm_movemask_ps()
955   __m128 tmp3, tmp2, tmp1, tmp0; \