Lines Matching full:xmm8
103 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
104 "vfmadd132ps %%xmm7, %%xmm8, %%xmm9;"
106 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
109 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
110 "vfmadd132ps (%2), %%xmm8, %%xmm9;"
112 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
115 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
116 "vfmadd213ps %%xmm7, %%xmm8, %%xmm9;"
118 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
121 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
122 "vfmadd213ps (%3), %%xmm8, %%xmm9;"
124 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
127 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
128 "vfmadd231ps %%xmm7, %%xmm8, %%xmm9;"
130 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
133 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
134 "vfmadd231ps (%2), %%xmm8, %%xmm9;"
136 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
147 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
148 "vfnmsub132ps %%xmm7, %%xmm8, %%xmm9;"
150 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
153 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
154 "vfnmsub132ps (%2), %%xmm8, %%xmm9;"
156 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
159 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
160 "vfnmsub213ps %%xmm7, %%xmm8, %%xmm9;"
162 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
165 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
166 "vfnmsub213ps (%3), %%xmm8, %%xmm9;"
168 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
171 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
172 "vfnmsub231ps %%xmm7, %%xmm8, %%xmm9;"
174 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
177 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
178 "vfnmsub231ps (%2), %%xmm8, %%xmm9;"
180 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
195 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
196 "vfmsub132ps %%xmm7, %%xmm8, %%xmm9;"
198 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
201 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
202 "vfmsub132ps (%2), %%xmm8, %%xmm9;"
204 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
207 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
208 "vfmsub213ps %%xmm7, %%xmm8, %%xmm9;"
210 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
213 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
214 "vfmsub213ps (%3), %%xmm8, %%xmm9;"
216 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
219 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
220 "vfmsub231ps %%xmm7, %%xmm8, %%xmm9;"
222 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
225 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
226 "vfmsub231ps (%2), %%xmm8, %%xmm9;"
228 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
239 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
240 "vfnmadd132ps %%xmm7, %%xmm8, %%xmm9;"
242 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
245 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
246 "vfnmadd132ps (%2), %%xmm8, %%xmm9;"
248 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
251 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
252 "vfnmadd213ps %%xmm7, %%xmm8, %%xmm9;"
254 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
257 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
258 "vfnmadd213ps (%3), %%xmm8, %%xmm9;"
260 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
263 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
264 "vfnmadd231ps %%xmm7, %%xmm8, %%xmm9;"
266 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
269 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
270 "vfnmadd231ps (%2), %%xmm8, %%xmm9;"
272 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
287 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
288 "vfmaddsub132ps %%xmm7, %%xmm8, %%xmm9;"
290 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
293 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
294 "vfmaddsub132ps (%2), %%xmm8, %%xmm9;"
296 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
299 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
300 "vfmaddsub213ps %%xmm7, %%xmm8, %%xmm9;"
302 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
305 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
306 "vfmaddsub213ps (%3), %%xmm8, %%xmm9;"
308 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
311 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
312 "vfmaddsub231ps %%xmm7, %%xmm8, %%xmm9;"
314 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
317 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
318 "vfmaddsub231ps (%2), %%xmm8, %%xmm9;"
320 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
335 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
336 "vfmsubadd132ps %%xmm7, %%xmm8, %%xmm9;"
338 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
341 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
342 "vfmsubadd132ps (%2), %%xmm8, %%xmm9;"
344 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
347 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
348 "vfmsubadd213ps %%xmm7, %%xmm8, %%xmm9;"
350 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
353 __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
354 "vfmsubadd213ps (%3), %%xmm8, %%xmm9;"
356 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
359 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
360 "vfmsubadd231ps %%xmm7, %%xmm8, %%xmm9;"
362 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
365 __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
366 "vfmsubadd231ps (%2), %%xmm8, %%xmm9;"
368 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
386 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
392 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
398 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
404 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
410 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
416 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
430 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
436 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
442 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
448 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
454 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
460 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
478 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
484 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
490 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
496 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
502 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
508 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
522 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
528 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
534 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
540 xmm8", "xmm9");
546 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
552 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
570 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
576 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
582 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
588 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
594 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
600 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
618 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
624 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
630 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
636 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
642 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
648 "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
752 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
753 "vfmadd132pd %%xmm7, %%xmm8, %%xmm9;"
755 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
758 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
759 "vfmadd132pd (%2), %%xmm8, %%xmm9;"
761 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
764 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
765 "vfmadd213pd %%xmm7, %%xmm8, %%xmm9;"
767 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
770 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
771 "vfmadd213pd (%3), %%xmm8, %%xmm9;"
773 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
776 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
777 "vfmadd231pd %%xmm7, %%xmm8, %%xmm9;"
779 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
782 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
783 "vfmadd231pd (%2), %%xmm8, %%xmm9;"
785 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
796 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
797 "vfnmsub132pd %%xmm7, %%xmm8, %%xmm9;"
799 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
802 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
803 "vfnmsub132pd (%2), %%xmm8, %%xmm9;"
805 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
808 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
809 "vfnmsub213pd %%xmm7, %%xmm8, %%xmm9;"
811 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
814 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
815 "vfnmsub213pd (%3), %%xmm8, %%xmm9;"
817 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
820 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
821 "vfnmsub231pd %%xmm7, %%xmm8, %%xmm9;"
823 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
826 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
827 "vfnmsub231pd (%2), %%xmm8, %%xmm9;"
829 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
844 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
845 "vfmsub132pd %%xmm7, %%xmm8, %%xmm9;"
847 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
850 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
851 "vfmsub132pd (%2), %%xmm8, %%xmm9;"
853 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
856 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
857 "vfmsub213pd %%xmm7, %%xmm8, %%xmm9;"
859 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
862 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
863 "vfmsub213pd (%3), %%xmm8, %%xmm9;"
865 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
868 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
869 "vfmsub231pd %%xmm7, %%xmm8, %%xmm9;"
871 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
874 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
875 "vfmsub231pd (%2), %%xmm8, %%xmm9;"
877 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
888 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
889 "vfnmadd132pd %%xmm7, %%xmm8, %%xmm9;"
891 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
894 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
895 "vfnmadd132pd (%2), %%xmm8, %%xmm9;"
897 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
900 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
901 "vfnmadd213pd %%xmm7, %%xmm8, %%xmm9;"
903 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
906 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
907 "vfnmadd213pd (%3), %%xmm8, %%xmm9;"
909 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
912 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
913 "vfnmadd231pd %%xmm7, %%xmm8, %%xmm9;"
915 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
918 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
919 "vfnmadd231pd (%2), %%xmm8, %%xmm9;"
921 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
936 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
937 "vfmaddsub132pd %%xmm7, %%xmm8, %%xmm9;"
939 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
942 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
943 "vfmaddsub132pd (%2), %%xmm8, %%xmm9;"
945 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
948 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
949 "vfmaddsub213pd %%xmm7, %%xmm8, %%xmm9;"
951 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
954 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
955 "vfmaddsub213pd (%3), %%xmm8, %%xmm9;"
957 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
960 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
961 "vfmaddsub231pd %%xmm7, %%xmm8, %%xmm9;"
963 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
966 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
967 "vfmaddsub231pd (%2), %%xmm8, %%xmm9;"
969 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
984 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
985 "vfmsubadd132pd %%xmm7, %%xmm8, %%xmm9;"
987 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
990 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
991 "vfmsubadd132pd (%2), %%xmm8, %%xmm9;"
993 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
996 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
997 "vfmsubadd213pd %%xmm7, %%xmm8, %%xmm9;"
999 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1002 __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
1003 "vfmsubadd213pd (%3), %%xmm8, %%xmm9;"
1005 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1008 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
1009 "vfmsubadd231pd %%xmm7, %%xmm8, %%xmm9;"
1011 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1014 __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
1015 "vfmsubadd231pd (%2), %%xmm8, %%xmm9;"
1017 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1035 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1041 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1047 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1053 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1059 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1065 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1079 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1085 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1091 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1097 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1103 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1109 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1127 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1133 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1139 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1145 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1151 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1157 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1171 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1177 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1183 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1189 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1195 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1201 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1219 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1225 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1231 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1237 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1243 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1249 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1267 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1273 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1279 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1285 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1291 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1297 "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
