51 typedef float __v4sf __attribute__ ((__vector_size__ (16)));
107   return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B);
113   return (__m128) __builtin_ia32_subss ((__v4sf)__A, (__v4sf)__B);
119   return (__m128) __builtin_ia32_mulss ((__v4sf)__A, (__v4sf)__B);
125   return (__m128) __builtin_ia32_divss ((__v4sf)__A, (__v4sf)__B);
131   return (__m128) __builtin_ia32_sqrtss ((__v4sf)__A);
137   return (__m128) __builtin_ia32_rcpss ((__v4sf)__A);
143   return (__m128) __builtin_ia32_rsqrtss ((__v4sf)__A);
149   return (__m128) __builtin_ia32_minss ((__v4sf)__A, (__v4sf)__B);
155   return (__m128) __builtin_ia32_maxss ((__v4sf)__A, (__v4sf)__B);
163   return (__m128) __builtin_ia32_addps ((__v4sf)__A, (__v4sf)__B);
169   return (__m128) __builtin_ia32_subps ((__v4sf)__A, (__v4sf)__B);
175   return (__m128) __builtin_ia32_mulps ((__v4sf)__A, (__v4sf)__B);
181   return (__m128) __builtin_ia32_divps ((__v4sf)__A, (__v4sf)__B);
187   return (__m128) __builtin_ia32_sqrtps ((__v4sf)__A);
193   return (__m128) __builtin_ia32_rcpps ((__v4sf)__A);
199   return (__m128) __builtin_ia32_rsqrtps ((__v4sf)__A);
205   return (__m128) __builtin_ia32_minps ((__v4sf)__A, (__v4sf)__B);
211   return (__m128) __builtin_ia32_maxps ((__v4sf)__A, (__v4sf)__B);
247   return (__m128) __builtin_ia32_cmpeqss ((__v4sf)__A, (__v4sf)__B);
253   return (__m128) __builtin_ia32_cmpltss ((__v4sf)__A, (__v4sf)__B);
259   return (__m128) __builtin_ia32_cmpless ((__v4sf)__A, (__v4sf)__B);
265   return (__m128) __builtin_ia32_movss ((__v4sf) __A,
266 					(__v4sf)
267 					__builtin_ia32_cmpltss ((__v4sf) __B,
268 								(__v4sf)
275   return (__m128) __builtin_ia32_movss ((__v4sf) __A,
276 					(__v4sf)
277 					__builtin_ia32_cmpless ((__v4sf) __B,
278 								(__v4sf)
285   return (__m128) __builtin_ia32_cmpneqss ((__v4sf)__A, (__v4sf)__B);
291   return (__m128) __builtin_ia32_cmpnltss ((__v4sf)__A, (__v4sf)__B);
297   return (__m128) __builtin_ia32_cmpnless ((__v4sf)__A, (__v4sf)__B);
303   return (__m128) __builtin_ia32_movss ((__v4sf) __A,
304 					(__v4sf)
305 					__builtin_ia32_cmpnltss ((__v4sf) __B,
306 								 (__v4sf)
313   return (__m128) __builtin_ia32_movss ((__v4sf) __A,
314 					(__v4sf)
315 					__builtin_ia32_cmpnless ((__v4sf) __B,
316 								 (__v4sf)
323   return (__m128) __builtin_ia32_cmpordss ((__v4sf)__A, (__v4sf)__B);
329   return (__m128) __builtin_ia32_cmpunordss ((__v4sf)__A, (__v4sf)__B);
339   return (__m128) __builtin_ia32_cmpeqps ((__v4sf)__A, (__v4sf)__B);
345   return (__m128) __builtin_ia32_cmpltps ((__v4sf)__A, (__v4sf)__B);
351   return (__m128) __builtin_ia32_cmpleps ((__v4sf)__A, (__v4sf)__B);
357   return (__m128) __builtin_ia32_cmpgtps ((__v4sf)__A, (__v4sf)__B);
363   return (__m128) __builtin_ia32_cmpgeps ((__v4sf)__A, (__v4sf)__B);
369   return (__m128) __builtin_ia32_cmpneqps ((__v4sf)__A, (__v4sf)__B);
375   return (__m128) __builtin_ia32_cmpnltps ((__v4sf)__A, (__v4sf)__B);
381   return (__m128) __builtin_ia32_cmpnleps ((__v4sf)__A, (__v4sf)__B);
387   return (__m128) __builtin_ia32_cmpngtps ((__v4sf)__A, (__v4sf)__B);
393   return (__m128) __builtin_ia32_cmpngeps ((__v4sf)__A, (__v4sf)__B);
399   return (__m128) __builtin_ia32_cmpordps ((__v4sf)__A, (__v4sf)__B);
405   return (__m128) __builtin_ia32_cmpunordps ((__v4sf)__A, (__v4sf)__B);
414   return __builtin_ia32_comieq ((__v4sf)__A, (__v4sf)__B);
420   return __builtin_ia32_comilt ((__v4sf)__A, (__v4sf)__B);
426   return __builtin_ia32_comile ((__v4sf)__A, (__v4sf)__B);
432   return __builtin_ia32_comigt ((__v4sf)__A, (__v4sf)__B);
438   return __builtin_ia32_comige ((__v4sf)__A, (__v4sf)__B);
444   return __builtin_ia32_comineq ((__v4sf)__A, (__v4sf)__B);
450   return __builtin_ia32_ucomieq ((__v4sf)__A, (__v4sf)__B);
456   return __builtin_ia32_ucomilt ((__v4sf)__A, (__v4sf)__B);
462   return __builtin_ia32_ucomile ((__v4sf)__A, (__v4sf)__B);
468   return __builtin_ia32_ucomigt ((__v4sf)__A, (__v4sf)__B);
474   return __builtin_ia32_ucomige ((__v4sf)__A, (__v4sf)__B);
480   return __builtin_ia32_ucomineq ((__v4sf)__A, (__v4sf)__B);
488   return __builtin_ia32_cvtss2si ((__v4sf) __A);
505   return __builtin_ia32_cvtss2si64 ((__v4sf) __A);
512   return __builtin_ia32_cvtss2si64 ((__v4sf) __A);
521   return (__m64) __builtin_ia32_cvtps2pi ((__v4sf) __A);
534   return __builtin_ia32_cvttss2si ((__v4sf) __A);
550   return __builtin_ia32_cvttss2si64 ((__v4sf) __A);
557   return __builtin_ia32_cvttss2si64 ((__v4sf) __A);
566   return (__m64) __builtin_ia32_cvttps2pi ((__v4sf) __A);
579   return (__m128) __builtin_ia32_cvtsi2ss ((__v4sf) __A, __B);
595   return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B);
602   return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B);
611   return (__m128) __builtin_ia32_cvtpi2ps ((__v4sf) __A, (__v2si)__B);
626   __v4sf __r;
638   __r = (__v4sf) _mm_setzero_ps ();
651   __v4sf __r;
658   __r = (__v4sf) _mm_setzero_ps ();
695   __v4sf __zero = (__v4sf) _mm_setzero_ps ();
696   __v4sf __sfa = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__A);
697   __v4sf __sfb = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__B);
705   __v4sf __hisf = (__v4sf)__A;
706   __v4sf __losf = __builtin_ia32_movhlps (__hisf, __hisf);
725   return (__m128) __builtin_ia32_shufps ((__v4sf)__A, (__v4sf)__B, __mask);
729  ((__m128) __builtin_ia32_shufps ((__v4sf)(A), (__v4sf)(B), (MASK)))
737   return (__m128) __builtin_ia32_unpckhps ((__v4sf)__A, (__v4sf)__B);
744   return (__m128) __builtin_ia32_unpcklps ((__v4sf)__A, (__v4sf)__B);
752   return (__m128) __builtin_ia32_loadhps ((__v4sf)__A, (__v2si *)__P);
759   __builtin_ia32_storehps ((__v2si *)__P, (__v4sf)__A);
766   return (__m128) __builtin_ia32_movhlps ((__v4sf)__A, (__v4sf)__B);
773   return (__m128) __builtin_ia32_movlhps ((__v4sf)__A, (__v4sf)__B);
781   return (__m128) __builtin_ia32_loadlps ((__v4sf)__A, (__v2si *)__P);
788   __builtin_ia32_storelps ((__v2si *)__P, (__v4sf)__A);
795   return __builtin_ia32_movmskps ((__v4sf)__A);
866   return __extension__ (__m128)(__v4sf){ __F, 0, 0, 0 };
873   return __extension__ (__m128)(__v4sf){ __F, __F, __F, __F };
906   return (__m128) *(__v4sf *)__P;
920   __v4sf __tmp = *(__v4sf *)__P;
928   return __extension__ (__m128)(__v4sf){ __W, __X, __Y, __Z };
935   return __extension__ (__m128)(__v4sf){ __Z, __Y, __X, __W };
942   *__P = __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0);
948   return __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0);
955   *(__v4sf *)__P = (__v4sf)__A;
962   __builtin_ia32_storeups (__P, (__v4sf)__A);
969   __v4sf __va = (__v4sf)__A;
970   __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,0,0,0));
984   __v4sf __va = (__v4sf)__A;
985   __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,1,2,3));
993   return (__m128) __builtin_ia32_movss ((__v4sf)__A, (__v4sf)__B);
1213   __builtin_ia32_movntps (__P, (__v4sf)__A);
1236   __v4sf __r0 = (row0), __r1 = (row1), __r2 = (row2), __r3 = (row3);	\
1237   __v4sf __t0 = __builtin_ia32_unpcklps (__r0, __r1);			\
1238   __v4sf __t1 = __builtin_ia32_unpcklps (__r2, __r3);			\
1239   __v4sf __t2 = __builtin_ia32_unpckhps (__r0, __r1);			\
1240   __v4sf __t3 = __builtin_ia32_unpckhps (__r2, __r3);			\

Indexes created Wed Jul 03 21:38:27 MDT 2024