41       "dup v4.4s, %w[input_range_min]\n"  in Transform()
42       "dup v5.4s, %w[output_range_min]\n"  in Transform()
43       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
44       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
45       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
46       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
52       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
54       "scvtf v0.4s, v0.4s\n"  in Transform()
55       "scvtf v1.4s, v1.4s\n"  in Transform()
56       "scvtf v2.4s, v2.4s\n"  in Transform()
57       "scvtf v3.4s, v3.4s\n"  in Transform()
58       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
59       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
60       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
61       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
62       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
63       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
64       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
65       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
66       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
67       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
68       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
69       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
70       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
71       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
72       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
73       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
74       "fcvtzs v0.4s, v0.4s\n"  in Transform()
75       "fcvtzs v1.4s, v1.4s\n"  in Transform()
76       "fcvtzs v2.4s, v2.4s\n"  in Transform()
77       "fcvtzs v3.4s, v3.4s\n"  in Transform()
78       "sqxtn v0.4h, v0.4s\n"  in Transform()
79       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
80       "sqxtn v2.4h, v2.4s\n"  in Transform()
81       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
85       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
114       "dup v4.4s, %w[input_range_min]\n"  in Transform()
115       "dup v5.4s, %w[output_range_min]\n"  in Transform()
116       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
117       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
118       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
119       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
129       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
131       "scvtf v0.4s, v0.4s\n"  in Transform()
132       "scvtf v1.4s, v1.4s\n"  in Transform()
133       "scvtf v2.4s, v2.4s\n"  in Transform()
134       "scvtf v3.4s, v3.4s\n"  in Transform()
135       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
136       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
137       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
138       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
139       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
140       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
141       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
142       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
143       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
144       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
145       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
146       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
147       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
148       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
149       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
150       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
151       "fcvtzs v0.4s, v0.4s\n"  in Transform()
152       "fcvtzs v1.4s, v1.4s\n"  in Transform()
153       "fcvtzs v2.4s, v2.4s\n"  in Transform()
154       "fcvtzs v3.4s, v3.4s\n"  in Transform()
155       "sqxtn v0.4h, v0.4s\n"  in Transform()
156       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
157       "sqxtn v2.4h, v2.4s\n"  in Transform()
158       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
162       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
171       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
173       "scvtf v0.4s, v0.4s\n"  in Transform()
174       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
175       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
176       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
177       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
178       "fcvtzs v0.4s, v0.4s\n"  in Transform()
179       "sqxtn v0.4h, v0.4s\n"  in Transform()
209       "dup v4.4s, %w[input_range_min]\n"  in Transform()
210       "dup v5.4s, %w[output_range_min]\n"  in Transform()
211       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
212       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
213       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
214       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
224       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
226       "scvtf v0.4s, v0.4s\n"  in Transform()
227       "scvtf v1.4s, v1.4s\n"  in Transform()
228       "scvtf v2.4s, v2.4s\n"  in Transform()
229       "scvtf v3.4s, v3.4s\n"  in Transform()
230       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
231       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
232       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
233       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
234       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
235       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
236       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
237       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
238       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
239       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
240       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
241       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
242       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
243       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
244       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
245       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
246       "fcvtzs v0.4s, v0.4s\n"  in Transform()
247       "fcvtzs v1.4s, v1.4s\n"  in Transform()
248       "fcvtzs v2.4s, v2.4s\n"  in Transform()
249       "fcvtzs v3.4s, v3.4s\n"  in Transform()
250       "sqxtn v0.4h, v0.4s\n"  in Transform()
251       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
252       "sqxtn v2.4h, v2.4s\n"  in Transform()
253       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
257       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
266       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
268       "scvtf v0.4s, v0.4s\n"  in Transform()
269       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
270       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
271       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
272       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
273       "fcvtzs v0.4s, v0.4s\n"  in Transform()
274       "sqxtn v0.4h, v0.4s\n"  in Transform()
304       "dup v4.4s, %w[input_range_min]\n"  in Transform()
305       "dup v5.4s, %w[output_range_min]\n"  in Transform()
306       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
307       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
308       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
309       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
319       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
321       "scvtf v0.4s, v0.4s\n"  in Transform()
322       "scvtf v1.4s, v1.4s\n"  in Transform()
323       "scvtf v2.4s, v2.4s\n"  in Transform()
324       "scvtf v3.4s, v3.4s\n"  in Transform()
325       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
326       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
327       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
328       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
329       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
330       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
331       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
332       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
333       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
334       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
335       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
336       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
337       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
338       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
339       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
340       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
341       "fcvtzs v0.4s, v0.4s\n"  in Transform()
342       "fcvtzs v1.4s, v1.4s\n"  in Transform()
343       "fcvtzs v2.4s, v2.4s\n"  in Transform()
344       "fcvtzs v3.4s, v3.4s\n"  in Transform()
345       "sqxtn v0.4h, v0.4s\n"  in Transform()
346       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
347       "sqxtn v2.4h, v2.4s\n"  in Transform()
348       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
352       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
361       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
362       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
364       "scvtf v0.4s, v0.4s\n"  in Transform()
365       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
366       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
367       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
368       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
369       "fcvtzs v0.4s, v0.4s\n"  in Transform()
370       "sqxtn v0.4h, v0.4s\n"  in Transform()
401       "dup v4.4s, %w[input_range_min]\n"  in Transform()
402       "dup v5.4s, %w[output_range_min]\n"  in Transform()
403       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
404       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
405       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
406       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
416       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
418       "scvtf v0.4s, v0.4s\n"  in Transform()
419       "scvtf v1.4s, v1.4s\n"  in Transform()
420       "scvtf v2.4s, v2.4s\n"  in Transform()
421       "scvtf v3.4s, v3.4s\n"  in Transform()
422       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
423       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
424       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
425       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
426       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
427       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
428       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
429       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
430       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
431       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
432       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
433       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
434       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
435       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
436       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
437       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
438       "fcvtzs v0.4s, v0.4s\n"  in Transform()
439       "fcvtzs v1.4s, v1.4s\n"  in Transform()
440       "fcvtzs v2.4s, v2.4s\n"  in Transform()
441       "fcvtzs v3.4s, v3.4s\n"  in Transform()
442       "sqxtn v0.4h, v0.4s\n"  in Transform()
443       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
444       "sqxtn v2.4h, v2.4s\n"  in Transform()
445       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
449       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
458       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
460       "scvtf v0.4s, v0.4s\n"  in Transform()
461       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
462       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
463       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
464       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
465       "fcvtzs v0.4s, v0.4s\n"  in Transform()
466       "sqxtn v0.4h, v0.4s\n"  in Transform()
469       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
496       "dup v4.4s, %w[input_range_min]\n"  in Transform()
497       "dup v5.4s, %w[output_range_min]\n"  in Transform()
498       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
499       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
500       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
501       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
511       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
513       "scvtf v0.4s, v0.4s\n"  in Transform()
514       "scvtf v1.4s, v1.4s\n"  in Transform()
515       "scvtf v2.4s, v2.4s\n"  in Transform()
516       "scvtf v3.4s, v3.4s\n"  in Transform()
517       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
518       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
519       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
520       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
521       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
522       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
523       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
524       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
525       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
526       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
527       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
528       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
529       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
530       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
531       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
532       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
533       "fcvtzs v0.4s, v0.4s\n"  in Transform()
534       "fcvtzs v1.4s, v1.4s\n"  in Transform()
535       "fcvtzs v2.4s, v2.4s\n"  in Transform()
536       "fcvtzs v3.4s, v3.4s\n"  in Transform()
537       "sqxtn v0.4h, v0.4s\n"  in Transform()
538       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
539       "sqxtn v2.4h, v2.4s\n"  in Transform()
540       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
544       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
553       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
554       "ld1 {v1.s}[0], [%x[input]], #4\n"  in Transform()
556       "scvtf v0.4s, v0.4s\n"  in Transform()
557       "scvtf v1.4s, v1.4s\n"  in Transform()
558       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
559       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
560       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
561       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
562       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
563       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
564       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
565       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
566       "fcvtzs v0.4s, v0.4s\n"  in Transform()
567       "fcvtzs v1.4s, v1.4s\n"  in Transform()
568       "sqxtn v0.4h, v0.4s\n"  in Transform()
569       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
572       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
600       "dup v4.4s, %w[input_range_min]\n"  in Transform()
601       "dup v5.4s, %w[output_range_min]\n"  in Transform()
602       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
603       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
604       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
605       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
615       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
617       "scvtf v0.4s, v0.4s\n"  in Transform()
618       "scvtf v1.4s, v1.4s\n"  in Transform()
619       "scvtf v2.4s, v2.4s\n"  in Transform()
620       "scvtf v3.4s, v3.4s\n"  in Transform()
621       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
622       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
623       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
624       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
625       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
626       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
627       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
628       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
629       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
630       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
631       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
632       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
633       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
634       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
635       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
636       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
637       "fcvtzs v0.4s, v0.4s\n"  in Transform()
638       "fcvtzs v1.4s, v1.4s\n"  in Transform()
639       "fcvtzs v2.4s, v2.4s\n"  in Transform()
640       "fcvtzs v3.4s, v3.4s\n"  in Transform()
641       "sqxtn v0.4h, v0.4s\n"  in Transform()
642       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
643       "sqxtn v2.4h, v2.4s\n"  in Transform()
644       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
648       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
657       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
658       "ld1 {v1.2s}, [%x[input]], #8\n"  in Transform()
660       "scvtf v0.4s, v0.4s\n"  in Transform()
661       "scvtf v1.4s, v1.4s\n"  in Transform()
662       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
663       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
664       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
665       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
666       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
667       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
668       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
669       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
670       "fcvtzs v0.4s, v0.4s\n"  in Transform()
671       "fcvtzs v1.4s, v1.4s\n"  in Transform()
672       "sqxtn v0.4h, v0.4s\n"  in Transform()
673       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
676       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
704       "dup v4.4s, %w[input_range_min]\n"  in Transform()
705       "dup v5.4s, %w[output_range_min]\n"  in Transform()
706       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
707       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
708       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
709       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
719       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
721       "scvtf v0.4s, v0.4s\n"  in Transform()
722       "scvtf v1.4s, v1.4s\n"  in Transform()
723       "scvtf v2.4s, v2.4s\n"  in Transform()
724       "scvtf v3.4s, v3.4s\n"  in Transform()
725       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
726       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
727       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
728       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
729       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
730       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
731       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
732       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
733       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
734       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
735       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
736       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
737       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
738       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
739       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
740       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
741       "fcvtzs v0.4s, v0.4s\n"  in Transform()
742       "fcvtzs v1.4s, v1.4s\n"  in Transform()
743       "fcvtzs v2.4s, v2.4s\n"  in Transform()
744       "fcvtzs v3.4s, v3.4s\n"  in Transform()
745       "sqxtn v0.4h, v0.4s\n"  in Transform()
746       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
747       "sqxtn v2.4h, v2.4s\n"  in Transform()
748       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
752       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
761       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
762       "ld1 {v1.2s}, [%x[input]], #8\n"  in Transform()
763       "ld1 {v1.s}[2], [%x[input]], #4\n"  in Transform()
765       "scvtf v0.4s, v0.4s\n"  in Transform()
766       "scvtf v1.4s, v1.4s\n"  in Transform()
767       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
768       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
769       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
770       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
771       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
772       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
773       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
774       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
775       "fcvtzs v0.4s, v0.4s\n"  in Transform()
776       "fcvtzs v1.4s, v1.4s\n"  in Transform()
777       "sqxtn v0.4h, v0.4s\n"  in Transform()
778       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
781       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
810       "dup v4.4s, %w[input_range_min]\n"  in Transform()
811       "dup v5.4s, %w[output_range_min]\n"  in Transform()
812       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
813       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
814       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
815       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
825       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
827       "scvtf v0.4s, v0.4s\n"  in Transform()
828       "scvtf v1.4s, v1.4s\n"  in Transform()
829       "scvtf v2.4s, v2.4s\n"  in Transform()
830       "scvtf v3.4s, v3.4s\n"  in Transform()
831       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
832       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
833       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
834       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
835       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
836       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
837       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
838       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
839       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
840       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
841       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
842       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
843       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
844       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
845       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
846       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
847       "fcvtzs v0.4s, v0.4s\n"  in Transform()
848       "fcvtzs v1.4s, v1.4s\n"  in Transform()
849       "fcvtzs v2.4s, v2.4s\n"  in Transform()
850       "fcvtzs v3.4s, v3.4s\n"  in Transform()
851       "sqxtn v0.4h, v0.4s\n"  in Transform()
852       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
853       "sqxtn v2.4h, v2.4s\n"  in Transform()
854       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
858       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
867       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
869       "scvtf v0.4s, v0.4s\n"  in Transform()
870       "scvtf v1.4s, v1.4s\n"  in Transform()
871       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
872       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
873       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
874       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
875       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
876       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
877       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
878       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
879       "fcvtzs v0.4s, v0.4s\n"  in Transform()
880       "fcvtzs v1.4s, v1.4s\n"  in Transform()
881       "sqxtn v0.4h, v0.4s\n"  in Transform()
882       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
885       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
912       "dup v4.4s, %w[input_range_min]\n"  in Transform()
913       "dup v5.4s, %w[output_range_min]\n"  in Transform()
914       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
915       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
916       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
917       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
927       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
929       "scvtf v0.4s, v0.4s\n"  in Transform()
930       "scvtf v1.4s, v1.4s\n"  in Transform()
931       "scvtf v2.4s, v2.4s\n"  in Transform()
932       "scvtf v3.4s, v3.4s\n"  in Transform()
933       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
934       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
935       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
936       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
937       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
938       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
939       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
940       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
941       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
942       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
943       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
944       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
945       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
946       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
947       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
948       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
949       "fcvtzs v0.4s, v0.4s\n"  in Transform()
950       "fcvtzs v1.4s, v1.4s\n"  in Transform()
951       "fcvtzs v2.4s, v2.4s\n"  in Transform()
952       "fcvtzs v3.4s, v3.4s\n"  in Transform()
953       "sqxtn v0.4h, v0.4s\n"  in Transform()
954       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
955       "sqxtn v2.4h, v2.4s\n"  in Transform()
956       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
960       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
969       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
970       "ld1 {v2.s}[0], [%x[input]], #4\n"  in Transform()
972       "scvtf v0.4s, v0.4s\n"  in Transform()
973       "scvtf v1.4s, v1.4s\n"  in Transform()
974       "scvtf v2.4s, v2.4s\n"  in Transform()
975       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
976       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
977       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
978       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
979       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
980       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
981       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
982       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
983       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
984       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
985       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
986       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
987       "fcvtzs v0.4s, v0.4s\n"  in Transform()
988       "fcvtzs v1.4s, v1.4s\n"  in Transform()
989       "fcvtzs v2.4s, v2.4s\n"  in Transform()
990       "sqxtn v0.4h, v0.4s\n"  in Transform()
991       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
992       "sqxtn v2.4h, v2.4s\n"  in Transform()
996       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1024       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1025       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1026       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1027       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1028       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1029       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1039       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1041       "scvtf v0.4s, v0.4s\n"  in Transform()
1042       "scvtf v1.4s, v1.4s\n"  in Transform()
1043       "scvtf v2.4s, v2.4s\n"  in Transform()
1044       "scvtf v3.4s, v3.4s\n"  in Transform()
1045       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1046       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1047       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1048       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1049       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1050       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1051       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1052       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1053       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1054       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1055       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1056       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1057       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1058       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1059       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1060       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1061       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1062       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1063       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1064       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1065       "sqxtn v0.4h, v0.4s\n"  in Transform()
1066       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1067       "sqxtn v2.4h, v2.4s\n"  in Transform()
1068       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1072       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1081       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
1082       "ld1 {v2.2s}, [%x[input]], #8\n"  in Transform()
1084       "scvtf v0.4s, v0.4s\n"  in Transform()
1085       "scvtf v1.4s, v1.4s\n"  in Transform()
1086       "scvtf v2.4s, v2.4s\n"  in Transform()
1087       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1088       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1089       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1090       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1091       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1092       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1093       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1094       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1095       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1096       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1097       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1098       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1099       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1100       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1101       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1102       "sqxtn v0.4h, v0.4s\n"  in Transform()
1103       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1104       "sqxtn v2.4h, v2.4s\n"  in Transform()
1108       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1136       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1137       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1138       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1139       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1140       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1141       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1151       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1153       "scvtf v0.4s, v0.4s\n"  in Transform()
1154       "scvtf v1.4s, v1.4s\n"  in Transform()
1155       "scvtf v2.4s, v2.4s\n"  in Transform()
1156       "scvtf v3.4s, v3.4s\n"  in Transform()
1157       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1158       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1159       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1160       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1161       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1162       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1163       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1164       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1165       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1166       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1167       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1168       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1169       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1170       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1171       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1172       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1173       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1174       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1175       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1176       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1177       "sqxtn v0.4h, v0.4s\n"  in Transform()
1178       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1179       "sqxtn v2.4h, v2.4s\n"  in Transform()
1180       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1184       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1193       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
1194       "ld1 {v2.2s}, [%x[input]], #8\n"  in Transform()
1195       "ld1 {v2.s}[2], [%x[input]], #4\n"  in Transform()
1197       "scvtf v0.4s, v0.4s\n"  in Transform()
1198       "scvtf v1.4s, v1.4s\n"  in Transform()
1199       "scvtf v2.4s, v2.4s\n"  in Transform()
1200       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1201       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1202       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1203       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1204       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1205       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1206       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1207       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1208       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1209       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1210       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1211       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1212       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1213       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1214       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1215       "sqxtn v0.4h, v0.4s\n"  in Transform()
1216       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1217       "sqxtn v2.4h, v2.4s\n"  in Transform()
1221       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1250       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1251       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1252       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1253       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1254       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1255       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1265       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1267       "scvtf v0.4s, v0.4s\n"  in Transform()
1268       "scvtf v1.4s, v1.4s\n"  in Transform()
1269       "scvtf v2.4s, v2.4s\n"  in Transform()
1270       "scvtf v3.4s, v3.4s\n"  in Transform()
1271       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1272       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1273       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1274       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1275       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1276       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1277       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1278       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1279       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1280       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1281       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1282       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1283       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1284       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1285       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1286       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1287       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1288       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1289       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1290       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1291       "sqxtn v0.4h, v0.4s\n"  in Transform()
1292       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1293       "sqxtn v2.4h, v2.4s\n"  in Transform()
1294       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1298       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1307       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
1309       "scvtf v0.4s, v0.4s\n"  in Transform()
1310       "scvtf v1.4s, v1.4s\n"  in Transform()
1311       "scvtf v2.4s, v2.4s\n"  in Transform()
1312       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1313       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1314       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1315       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1316       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1317       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1318       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1319       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1320       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1321       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1322       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1323       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1324       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1325       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1326       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1327       "sqxtn v0.4h, v0.4s\n"  in Transform()
1328       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1329       "sqxtn v2.4h, v2.4s\n"  in Transform()
1333       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1334       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
1361       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1362       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1363       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1364       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1365       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1366       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1376       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1378       "scvtf v0.4s, v0.4s\n"  in Transform()
1379       "scvtf v1.4s, v1.4s\n"  in Transform()
1380       "scvtf v2.4s, v2.4s\n"  in Transform()
1381       "scvtf v3.4s, v3.4s\n"  in Transform()
1382       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1383       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1384       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1385       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1386       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1387       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1388       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1389       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1390       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1391       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1392       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1393       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1394       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1395       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1396       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1397       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1398       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1399       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1400       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1401       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1402       "sqxtn v0.4h, v0.4s\n"  in Transform()
1403       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1404       "sqxtn v2.4h, v2.4s\n"  in Transform()
1405       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1409       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1418       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
1419       "ld1 {v3.s}[0], [%x[input]], #4\n"  in Transform()
1421       "scvtf v0.4s, v0.4s\n"  in Transform()
1422       "scvtf v1.4s, v1.4s\n"  in Transform()
1423       "scvtf v2.4s, v2.4s\n"  in Transform()
1424       "scvtf v3.4s, v3.4s\n"  in Transform()
1425       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1426       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1427       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1428       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1429       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1430       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1431       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1432       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1433       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1434       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1435       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1436       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1437       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1438       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1439       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1440       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1441       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1442       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1443       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1444       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1445       "sqxtn v0.4h, v0.4s\n"  in Transform()
1446       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1447       "sqxtn v2.4h, v2.4s\n"  in Transform()
1448       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1452       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1453       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
1481       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1482       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1483       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1484       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1485       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1486       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1496       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1498       "scvtf v0.4s, v0.4s\n"  in Transform()
1499       "scvtf v1.4s, v1.4s\n"  in Transform()
1500       "scvtf v2.4s, v2.4s\n"  in Transform()
1501       "scvtf v3.4s, v3.4s\n"  in Transform()
1502       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1503       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1504       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1505       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1506       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1507       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1508       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1509       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1510       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1511       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1512       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1513       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1514       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1515       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1516       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1517       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1518       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1519       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1520       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1521       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1522       "sqxtn v0.4h, v0.4s\n"  in Transform()
1523       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1524       "sqxtn v2.4h, v2.4s\n"  in Transform()
1525       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1529       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1538       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
1539       "ld1 {v3.2s}, [%x[input]], #8\n"  in Transform()
1541       "scvtf v0.4s, v0.4s\n"  in Transform()
1542       "scvtf v1.4s, v1.4s\n"  in Transform()
1543       "scvtf v2.4s, v2.4s\n"  in Transform()
1544       "scvtf v3.4s, v3.4s\n"  in Transform()
1545       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1546       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1547       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1548       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1549       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1550       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1551       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1552       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1553       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1554       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1555       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1556       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1557       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1558       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1559       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1560       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1561       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1562       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1563       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1564       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1565       "sqxtn v0.4h, v0.4s\n"  in Transform()
1566       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1567       "sqxtn v2.4h, v2.4s\n"  in Transform()
1568       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1572       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1573       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
1601       "dup v4.4s, %w[input_range_min]\n"  in Transform()
1602       "dup v5.4s, %w[output_range_min]\n"  in Transform()
1603       "dup v6.4s, %w[input_range_offset]\n"  in Transform()
1604       "dup v7.4s, %w[input_range_scale]\n"  in Transform()
1605       "dup v8.4s, %w[one_over_output_range_scale]\n"  in Transform()
1606       "fsub v4.4s, v4.4s, v5.4s\n"  in Transform()
1616       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1618       "scvtf v0.4s, v0.4s\n"  in Transform()
1619       "scvtf v1.4s, v1.4s\n"  in Transform()
1620       "scvtf v2.4s, v2.4s\n"  in Transform()
1621       "scvtf v3.4s, v3.4s\n"  in Transform()
1622       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1623       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1624       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1625       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1626       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1627       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1628       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1629       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1630       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1631       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1632       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1633       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1634       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1635       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1636       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1637       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1638       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1639       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1640       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1641       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1642       "sqxtn v0.4h, v0.4s\n"  in Transform()
1643       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1644       "sqxtn v2.4h, v2.4s\n"  in Transform()
1645       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1649       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1658       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
1659       "ld1 {v3.2s}, [%x[input]], #8\n"  in Transform()
1660       "ld1 {v3.s}[2], [%x[input]], #4\n"  in Transform()
1662       "scvtf v0.4s, v0.4s\n"  in Transform()
1663       "scvtf v1.4s, v1.4s\n"  in Transform()
1664       "scvtf v2.4s, v2.4s\n"  in Transform()
1665       "scvtf v3.4s, v3.4s\n"  in Transform()
1666       "fsub v0.4s, v0.4s, v6.4s\n"  in Transform()
1667       "fsub v1.4s, v1.4s, v6.4s\n"  in Transform()
1668       "fsub v2.4s, v2.4s, v6.4s\n"  in Transform()
1669       "fsub v3.4s, v3.4s, v6.4s\n"  in Transform()
1670       "fmul v0.4s, v0.4s, v7.4s\n"  in Transform()
1671       "fmul v1.4s, v1.4s, v7.4s\n"  in Transform()
1672       "fmul v2.4s, v2.4s, v7.4s\n"  in Transform()
1673       "fmul v3.4s, v3.4s, v7.4s\n"  in Transform()
1674       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
1675       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
1676       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
1677       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
1678       "fmul v0.4s, v0.4s, v8.4s\n"  in Transform()
1679       "fmul v1.4s, v1.4s, v8.4s\n"  in Transform()
1680       "fmul v2.4s, v2.4s, v8.4s\n"  in Transform()
1681       "fmul v3.4s, v3.4s, v8.4s\n"  in Transform()
1682       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1683       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1684       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1685       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1686       "sqxtn v0.4h, v0.4s\n"  in Transform()
1687       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1688       "sqxtn v2.4h, v2.4s\n"  in Transform()
1689       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1693       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
1694       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
1723       "dup v4.4s, %w[range_min]\n"  in Transform()
1724       "dup v5.4s, %w[range_offset]\n"  in Transform()
1725       "dup v6.4s, %w[range_scale]\n"  in Transform()
1731       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1733       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1734       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
1735       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
1736       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
1737       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1738       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
1739       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
1740       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
1741       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1742       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
1743       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
1744       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
1745       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1746       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1747       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1748       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1749       "sqxtn v0.4h, v0.4s\n"  in Transform()
1750       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1751       "sqxtn v2.4h, v2.4s\n"  in Transform()
1752       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1756       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1782       "dup v4.4s, %w[range_min]\n"  in Transform()
1783       "dup v5.4s, %w[range_offset]\n"  in Transform()
1784       "dup v6.4s, %w[range_scale]\n"  in Transform()
1794       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1796       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1797       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
1798       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
1799       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
1800       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1801       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
1802       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
1803       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
1804       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1805       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
1806       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
1807       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
1808       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1809       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1810       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1811       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1812       "sqxtn v0.4h, v0.4s\n"  in Transform()
1813       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1814       "sqxtn v2.4h, v2.4s\n"  in Transform()
1815       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1819       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1828       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
1830       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1831       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1832       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1833       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1834       "sqxtn v0.4h, v0.4s\n"  in Transform()
1861       "dup v4.4s, %w[range_min]\n"  in Transform()
1862       "dup v5.4s, %w[range_offset]\n"  in Transform()
1863       "dup v6.4s, %w[range_scale]\n"  in Transform()
1873       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1875       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1876       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
1877       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
1878       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
1879       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1880       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
1881       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
1882       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
1883       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1884       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
1885       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
1886       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
1887       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1888       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1889       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1890       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1891       "sqxtn v0.4h, v0.4s\n"  in Transform()
1892       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1893       "sqxtn v2.4h, v2.4s\n"  in Transform()
1894       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1898       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1907       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
1909       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1910       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1911       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1912       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1913       "sqxtn v0.4h, v0.4s\n"  in Transform()
1940       "dup v4.4s, %w[range_min]\n"  in Transform()
1941       "dup v5.4s, %w[range_offset]\n"  in Transform()
1942       "dup v6.4s, %w[range_scale]\n"  in Transform()
1952       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
1954       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1955       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
1956       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
1957       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
1958       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1959       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
1960       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
1961       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
1962       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1963       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
1964       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
1965       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
1966       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1967       "fcvtzs v1.4s, v1.4s\n"  in Transform()
1968       "fcvtzs v2.4s, v2.4s\n"  in Transform()
1969       "fcvtzs v3.4s, v3.4s\n"  in Transform()
1970       "sqxtn v0.4h, v0.4s\n"  in Transform()
1971       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
1972       "sqxtn v2.4h, v2.4s\n"  in Transform()
1973       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
1977       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
1986       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
1987       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
1989       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
1990       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
1991       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
1992       "fcvtzs v0.4s, v0.4s\n"  in Transform()
1993       "sqxtn v0.4h, v0.4s\n"  in Transform()
2021       "dup v4.4s, %w[range_min]\n"  in Transform()
2022       "dup v5.4s, %w[range_offset]\n"  in Transform()
2023       "dup v6.4s, %w[range_scale]\n"  in Transform()
2033       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2035       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2036       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2037       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2038       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2039       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2040       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2041       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2042       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2043       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2044       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2045       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2046       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2047       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2048       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2049       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2050       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2051       "sqxtn v0.4h, v0.4s\n"  in Transform()
2052       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2053       "sqxtn v2.4h, v2.4s\n"  in Transform()
2054       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2058       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2067       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
2069       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2070       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2071       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2072       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2073       "sqxtn v0.4h, v0.4s\n"  in Transform()
2076       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
2100       "dup v4.4s, %w[range_min]\n"  in Transform()
2101       "dup v5.4s, %w[range_offset]\n"  in Transform()
2102       "dup v6.4s, %w[range_scale]\n"  in Transform()
2112       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2114       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2115       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2116       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2117       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2118       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2119       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2120       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2121       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2122       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2123       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2124       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2125       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2126       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2127       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2128       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2129       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2130       "sqxtn v0.4h, v0.4s\n"  in Transform()
2131       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2132       "sqxtn v2.4h, v2.4s\n"  in Transform()
2133       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2137       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2146       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
2147       "ld1 {v1.s}[0], [%x[input]], #4\n"  in Transform()
2149       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2150       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2151       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2152       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2153       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2154       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2155       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2156       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2157       "sqxtn v0.4h, v0.4s\n"  in Transform()
2158       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2161       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
2186       "dup v4.4s, %w[range_min]\n"  in Transform()
2187       "dup v5.4s, %w[range_offset]\n"  in Transform()
2188       "dup v6.4s, %w[range_scale]\n"  in Transform()
2198       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2200       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2201       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2202       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2203       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2204       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2205       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2206       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2207       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2208       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2209       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2210       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2211       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2212       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2213       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2214       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2215       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2216       "sqxtn v0.4h, v0.4s\n"  in Transform()
2217       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2218       "sqxtn v2.4h, v2.4s\n"  in Transform()
2219       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2223       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2232       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
2233       "ld1 {v1.2s}, [%x[input]], #8\n"  in Transform()
2235       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2236       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2237       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2238       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2239       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2240       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2241       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2242       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2243       "sqxtn v0.4h, v0.4s\n"  in Transform()
2244       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2247       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
2272       "dup v4.4s, %w[range_min]\n"  in Transform()
2273       "dup v5.4s, %w[range_offset]\n"  in Transform()
2274       "dup v6.4s, %w[range_scale]\n"  in Transform()
2284       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2286       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2287       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2288       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2289       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2290       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2291       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2292       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2293       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2294       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2295       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2296       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2297       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2298       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2299       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2300       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2301       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2302       "sqxtn v0.4h, v0.4s\n"  in Transform()
2303       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2304       "sqxtn v2.4h, v2.4s\n"  in Transform()
2305       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2309       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2318       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
2319       "ld1 {v1.2s}, [%x[input]], #8\n"  in Transform()
2320       "ld1 {v1.s}[2], [%x[input]], #4\n"  in Transform()
2322       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2323       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2324       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2325       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2326       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2327       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2328       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2329       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2330       "sqxtn v0.4h, v0.4s\n"  in Transform()
2331       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2334       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
2360       "dup v4.4s, %w[range_min]\n"  in Transform()
2361       "dup v5.4s, %w[range_offset]\n"  in Transform()
2362       "dup v6.4s, %w[range_scale]\n"  in Transform()
2372       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2374       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2375       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2376       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2377       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2378       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2379       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2380       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2381       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2382       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2383       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2384       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2385       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2386       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2387       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2388       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2389       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2390       "sqxtn v0.4h, v0.4s\n"  in Transform()
2391       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2392       "sqxtn v2.4h, v2.4s\n"  in Transform()
2393       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2397       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2406       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
2408       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2409       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2410       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2411       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2412       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2413       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2414       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2415       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2416       "sqxtn v0.4h, v0.4s\n"  in Transform()
2417       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2420       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2444       "dup v4.4s, %w[range_min]\n"  in Transform()
2445       "dup v5.4s, %w[range_offset]\n"  in Transform()
2446       "dup v6.4s, %w[range_scale]\n"  in Transform()
2456       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2458       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2459       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2460       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2461       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2462       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2463       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2464       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2465       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2466       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2467       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2468       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2469       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2470       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2471       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2472       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2473       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2474       "sqxtn v0.4h, v0.4s\n"  in Transform()
2475       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2476       "sqxtn v2.4h, v2.4s\n"  in Transform()
2477       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2481       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2490       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
2491       "ld1 {v2.s}[0], [%x[input]], #4\n"  in Transform()
2493       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2494       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2495       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2496       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2497       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2498       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2499       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2500       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2501       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2502       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2503       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2504       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2505       "sqxtn v0.4h, v0.4s\n"  in Transform()
2506       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2507       "sqxtn v2.4h, v2.4s\n"  in Transform()
2511       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2536       "dup v4.4s, %w[range_min]\n"  in Transform()
2537       "dup v5.4s, %w[range_offset]\n"  in Transform()
2538       "dup v6.4s, %w[range_scale]\n"  in Transform()
2548       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2550       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2551       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2552       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2553       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2554       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2555       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2556       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2557       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2558       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2559       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2560       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2561       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2562       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2563       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2564       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2565       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2566       "sqxtn v0.4h, v0.4s\n"  in Transform()
2567       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2568       "sqxtn v2.4h, v2.4s\n"  in Transform()
2569       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2573       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2582       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
2583       "ld1 {v2.2s}, [%x[input]], #8\n"  in Transform()
2585       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2586       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2587       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2588       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2589       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2590       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2591       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2592       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2593       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2594       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2595       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2596       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2597       "sqxtn v0.4h, v0.4s\n"  in Transform()
2598       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2599       "sqxtn v2.4h, v2.4s\n"  in Transform()
2603       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2628       "dup v4.4s, %w[range_min]\n"  in Transform()
2629       "dup v5.4s, %w[range_offset]\n"  in Transform()
2630       "dup v6.4s, %w[range_scale]\n"  in Transform()
2640       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2642       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2643       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2644       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2645       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2646       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2647       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2648       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2649       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2650       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2651       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2652       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2653       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2654       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2655       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2656       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2657       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2658       "sqxtn v0.4h, v0.4s\n"  in Transform()
2659       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2660       "sqxtn v2.4h, v2.4s\n"  in Transform()
2661       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2665       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2674       "ld1 {v0.4s, v1.4s}, [%x[input]], #32\n"  in Transform()
2675       "ld1 {v2.2s}, [%x[input]], #8\n"  in Transform()
2676       "ld1 {v2.s}[2], [%x[input]], #4\n"  in Transform()
2678       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2679       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2680       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2681       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2682       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2683       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2684       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2685       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2686       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2687       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2688       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2689       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2690       "sqxtn v0.4h, v0.4s\n"  in Transform()
2691       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2692       "sqxtn v2.4h, v2.4s\n"  in Transform()
2696       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2722       "dup v4.4s, %w[range_min]\n"  in Transform()
2723       "dup v5.4s, %w[range_offset]\n"  in Transform()
2724       "dup v6.4s, %w[range_scale]\n"  in Transform()
2734       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2736       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2737       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2738       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2739       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2740       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2741       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2742       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2743       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2744       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2745       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2746       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2747       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2748       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2749       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2750       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2751       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2752       "sqxtn v0.4h, v0.4s\n"  in Transform()
2753       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2754       "sqxtn v2.4h, v2.4s\n"  in Transform()
2755       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2759       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2768       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
2770       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2771       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2772       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2773       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2774       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2775       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2776       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2777       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2778       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2779       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2780       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2781       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2782       "sqxtn v0.4h, v0.4s\n"  in Transform()
2783       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2784       "sqxtn v2.4h, v2.4s\n"  in Transform()
2788       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2789       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
2813       "dup v4.4s, %w[range_min]\n"  in Transform()
2814       "dup v5.4s, %w[range_offset]\n"  in Transform()
2815       "dup v6.4s, %w[range_scale]\n"  in Transform()
2825       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2827       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2828       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2829       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2830       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2831       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2832       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2833       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2834       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2835       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2836       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2837       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2838       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2839       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2840       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2841       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2842       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2843       "sqxtn v0.4h, v0.4s\n"  in Transform()
2844       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2845       "sqxtn v2.4h, v2.4s\n"  in Transform()
2846       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2850       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2859       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
2860       "ld1 {v3.s}[0], [%x[input]], #4\n"  in Transform()
2862       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2863       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2864       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2865       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2866       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2867       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2868       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2869       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2870       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2871       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2872       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2873       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2874       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2875       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2876       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2877       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2878       "sqxtn v0.4h, v0.4s\n"  in Transform()
2879       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2880       "sqxtn v2.4h, v2.4s\n"  in Transform()
2881       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2885       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2886       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
2911       "dup v4.4s, %w[range_min]\n"  in Transform()
2912       "dup v5.4s, %w[range_offset]\n"  in Transform()
2913       "dup v6.4s, %w[range_scale]\n"  in Transform()
2923       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
2925       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2926       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2927       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2928       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2929       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2930       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2931       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2932       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2933       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2934       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2935       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2936       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2937       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2938       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2939       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2940       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2941       "sqxtn v0.4h, v0.4s\n"  in Transform()
2942       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2943       "sqxtn v2.4h, v2.4s\n"  in Transform()
2944       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2948       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
2957       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
2958       "ld1 {v3.2s}, [%x[input]], #8\n"  in Transform()
2960       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
2961       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
2962       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
2963       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
2964       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
2965       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
2966       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
2967       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
2968       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
2969       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
2970       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
2971       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
2972       "fcvtzs v0.4s, v0.4s\n"  in Transform()
2973       "fcvtzs v1.4s, v1.4s\n"  in Transform()
2974       "fcvtzs v2.4s, v2.4s\n"  in Transform()
2975       "fcvtzs v3.4s, v3.4s\n"  in Transform()
2976       "sqxtn v0.4h, v0.4s\n"  in Transform()
2977       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
2978       "sqxtn v2.4h, v2.4s\n"  in Transform()
2979       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
2983       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
2984       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
3009       "dup v4.4s, %w[range_min]\n"  in Transform()
3010       "dup v5.4s, %w[range_offset]\n"  in Transform()
3011       "dup v6.4s, %w[range_scale]\n"  in Transform()
3021       "ld1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[input]], #64\n"  in Transform()
3023       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
3024       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
3025       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
3026       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
3027       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3028       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3029       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3030       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3031       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
3032       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
3033       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
3034       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
3035       "fcvtzs v0.4s, v0.4s\n"  in Transform()
3036       "fcvtzs v1.4s, v1.4s\n"  in Transform()
3037       "fcvtzs v2.4s, v2.4s\n"  in Transform()
3038       "fcvtzs v3.4s, v3.4s\n"  in Transform()
3039       "sqxtn v0.4h, v0.4s\n"  in Transform()
3040       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
3041       "sqxtn v2.4h, v2.4s\n"  in Transform()
3042       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
3046       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
3055       "ld1 {v0.4s, v1.4s, v2.4s}, [%x[input]], #48\n"  in Transform()
3056       "ld1 {v3.2s}, [%x[input]], #8\n"  in Transform()
3057       "ld1 {v3.s}[2], [%x[input]], #4\n"  in Transform()
3059       "fsub v0.4s, v0.4s, v4.4s\n"  in Transform()
3060       "fsub v1.4s, v1.4s, v4.4s\n"  in Transform()
3061       "fsub v2.4s, v2.4s, v4.4s\n"  in Transform()
3062       "fsub v3.4s, v3.4s, v4.4s\n"  in Transform()
3063       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3064       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3065       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3066       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3067       "fadd v0.4s, v0.4s, v5.4s\n"  in Transform()
3068       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
3069       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
3070       "fadd v3.4s, v3.4s, v5.4s\n"  in Transform()
3071       "fcvtzs v0.4s, v0.4s\n"  in Transform()
3072       "fcvtzs v1.4s, v1.4s\n"  in Transform()
3073       "fcvtzs v2.4s, v2.4s\n"  in Transform()
3074       "fcvtzs v3.4s, v3.4s\n"  in Transform()
3075       "sqxtn v0.4h, v0.4s\n"  in Transform()
3076       "sqxtn2 v0.8h, v1.4s\n"  in Transform()
3077       "sqxtn v2.4h, v2.4s\n"  in Transform()
3078       "sqxtn2 v2.8h, v3.4s\n"  in Transform()
3082       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
3083       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
3109       "dup v4.4s, %w[range_min]\n"  in Transform()
3110       "dup v5.4s, %w[range_offset]\n"  in Transform()
3111       "dup v6.4s, %w[range_scale]\n"  in Transform()
3117       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3121       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3122       "sxtl v2.4s, v1.4h\n"  in Transform()
3123       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3124       "sxtl v0.4s, v0.4h\n"  in Transform()
3125       "scvtf v0.4s, v0.4s\n"  in Transform()
3126       "scvtf v1.4s, v1.4s\n"  in Transform()
3127       "scvtf v2.4s, v2.4s\n"  in Transform()
3128       "scvtf v3.4s, v3.4s\n"  in Transform()
3129       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3130       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3131       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3132       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3133       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3134       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3135       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3136       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3137       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3138       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3139       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3140       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3142       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3168       "dup v4.4s, %w[range_min]\n"  in Transform()
3169       "dup v5.4s, %w[range_offset]\n"  in Transform()
3170       "dup v6.4s, %w[range_scale]\n"  in Transform()
3180       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3184       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3185       "sxtl v2.4s, v1.4h\n"  in Transform()
3186       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3187       "sxtl v0.4s, v0.4h\n"  in Transform()
3188       "scvtf v0.4s, v0.4s\n"  in Transform()
3189       "scvtf v1.4s, v1.4s\n"  in Transform()
3190       "scvtf v2.4s, v2.4s\n"  in Transform()
3191       "scvtf v3.4s, v3.4s\n"  in Transform()
3192       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3193       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3194       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3195       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3196       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3197       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3198       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3199       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3200       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3201       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3202       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3203       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3205       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3217       "sxtl v0.4s, v0.4h\n"  in Transform()
3218       "scvtf v0.4s, v0.4s\n"  in Transform()
3219       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3220       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3221       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3223       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
3247       "dup v4.4s, %w[range_min]\n"  in Transform()
3248       "dup v5.4s, %w[range_offset]\n"  in Transform()
3249       "dup v6.4s, %w[range_scale]\n"  in Transform()
3259       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3263       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3264       "sxtl v2.4s, v1.4h\n"  in Transform()
3265       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3266       "sxtl v0.4s, v0.4h\n"  in Transform()
3267       "scvtf v0.4s, v0.4s\n"  in Transform()
3268       "scvtf v1.4s, v1.4s\n"  in Transform()
3269       "scvtf v2.4s, v2.4s\n"  in Transform()
3270       "scvtf v3.4s, v3.4s\n"  in Transform()
3271       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3272       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3273       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3274       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3275       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3276       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3277       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3278       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3279       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3280       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3281       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3282       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3284       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3296       "sxtl v0.4s, v0.4h\n"  in Transform()
3297       "scvtf v0.4s, v0.4s\n"  in Transform()
3298       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3299       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3300       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3302       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
3326       "dup v4.4s, %w[range_min]\n"  in Transform()
3327       "dup v5.4s, %w[range_offset]\n"  in Transform()
3328       "dup v6.4s, %w[range_scale]\n"  in Transform()
3338       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3342       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3343       "sxtl v2.4s, v1.4h\n"  in Transform()
3344       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3345       "sxtl v0.4s, v0.4h\n"  in Transform()
3346       "scvtf v0.4s, v0.4s\n"  in Transform()
3347       "scvtf v1.4s, v1.4s\n"  in Transform()
3348       "scvtf v2.4s, v2.4s\n"  in Transform()
3349       "scvtf v3.4s, v3.4s\n"  in Transform()
3350       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3351       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3352       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3353       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3354       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3355       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3356       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3357       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3358       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3359       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3360       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3361       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3363       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3376       "sxtl v0.4s, v0.4h\n"  in Transform()
3377       "scvtf v0.4s, v0.4s\n"  in Transform()
3378       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3379       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3380       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3382       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
3383       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
3407       "dup v4.4s, %w[range_min]\n"  in Transform()
3408       "dup v5.4s, %w[range_offset]\n"  in Transform()
3409       "dup v6.4s, %w[range_scale]\n"  in Transform()
3419       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3423       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3424       "sxtl v2.4s, v1.4h\n"  in Transform()
3425       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3426       "sxtl v0.4s, v0.4h\n"  in Transform()
3427       "scvtf v0.4s, v0.4s\n"  in Transform()
3428       "scvtf v1.4s, v1.4s\n"  in Transform()
3429       "scvtf v2.4s, v2.4s\n"  in Transform()
3430       "scvtf v3.4s, v3.4s\n"  in Transform()
3431       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3432       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3433       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3434       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3435       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3436       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3437       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3438       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3439       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3440       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3441       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3442       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3444       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3453       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
3456       "sxtl v0.4s, v0.4h\n"  in Transform()
3457       "scvtf v0.4s, v0.4s\n"  in Transform()
3458       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3459       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3460       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3462       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
3486       "dup v4.4s, %w[range_min]\n"  in Transform()
3487       "dup v5.4s, %w[range_offset]\n"  in Transform()
3488       "dup v6.4s, %w[range_scale]\n"  in Transform()
3498       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3502       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3503       "sxtl v2.4s, v1.4h\n"  in Transform()
3504       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3505       "sxtl v0.4s, v0.4h\n"  in Transform()
3506       "scvtf v0.4s, v0.4s\n"  in Transform()
3507       "scvtf v1.4s, v1.4s\n"  in Transform()
3508       "scvtf v2.4s, v2.4s\n"  in Transform()
3509       "scvtf v3.4s, v3.4s\n"  in Transform()
3510       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3511       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3512       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3513       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3514       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3515       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3516       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3517       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3518       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3519       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3520       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3521       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3523       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3532       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
3536       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3537       "sxtl v0.4s, v0.4h\n"  in Transform()
3538       "scvtf v0.4s, v0.4s\n"  in Transform()
3539       "scvtf v1.4s, v1.4s\n"  in Transform()
3540       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3541       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3542       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3543       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3544       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3545       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3547       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
3548       "st1 {v1.s}[0], [%x[output]], #4\n"  in Transform()
3572       "dup v4.4s, %w[range_min]\n"  in Transform()
3573       "dup v5.4s, %w[range_offset]\n"  in Transform()
3574       "dup v6.4s, %w[range_scale]\n"  in Transform()
3584       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3588       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3589       "sxtl v2.4s, v1.4h\n"  in Transform()
3590       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3591       "sxtl v0.4s, v0.4h\n"  in Transform()
3592       "scvtf v0.4s, v0.4s\n"  in Transform()
3593       "scvtf v1.4s, v1.4s\n"  in Transform()
3594       "scvtf v2.4s, v2.4s\n"  in Transform()
3595       "scvtf v3.4s, v3.4s\n"  in Transform()
3596       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3597       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3598       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3599       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3600       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3601       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3602       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3603       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3604       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3605       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3606       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3607       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3609       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3618       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
3622       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3623       "sxtl v0.4s, v0.4h\n"  in Transform()
3624       "scvtf v0.4s, v0.4s\n"  in Transform()
3625       "scvtf v1.4s, v1.4s\n"  in Transform()
3626       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3627       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3628       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3629       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3630       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3631       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3633       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
3634       "st1 {v1.2s}, [%x[output]], #8\n"  in Transform()
3658       "dup v4.4s, %w[range_min]\n"  in Transform()
3659       "dup v5.4s, %w[range_offset]\n"  in Transform()
3660       "dup v6.4s, %w[range_scale]\n"  in Transform()
3670       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3674       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3675       "sxtl v2.4s, v1.4h\n"  in Transform()
3676       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3677       "sxtl v0.4s, v0.4h\n"  in Transform()
3678       "scvtf v0.4s, v0.4s\n"  in Transform()
3679       "scvtf v1.4s, v1.4s\n"  in Transform()
3680       "scvtf v2.4s, v2.4s\n"  in Transform()
3681       "scvtf v3.4s, v3.4s\n"  in Transform()
3682       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3683       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3684       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3685       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3686       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3687       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3688       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3689       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3690       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3691       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3692       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3693       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3695       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3704       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
3709       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3710       "sxtl v0.4s, v0.4h\n"  in Transform()
3711       "scvtf v0.4s, v0.4s\n"  in Transform()
3712       "scvtf v1.4s, v1.4s\n"  in Transform()
3713       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3714       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3715       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3716       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3717       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3718       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3720       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
3721       "st1 {v1.2s}, [%x[output]], #8\n"  in Transform()
3722       "st1 {v1.s}[2], [%x[output]], #4\n"  in Transform()
3746       "dup v4.4s, %w[range_min]\n"  in Transform()
3747       "dup v5.4s, %w[range_offset]\n"  in Transform()
3748       "dup v6.4s, %w[range_scale]\n"  in Transform()
3758       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3762       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3763       "sxtl v2.4s, v1.4h\n"  in Transform()
3764       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3765       "sxtl v0.4s, v0.4h\n"  in Transform()
3766       "scvtf v0.4s, v0.4s\n"  in Transform()
3767       "scvtf v1.4s, v1.4s\n"  in Transform()
3768       "scvtf v2.4s, v2.4s\n"  in Transform()
3769       "scvtf v3.4s, v3.4s\n"  in Transform()
3770       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3771       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3772       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3773       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3774       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3775       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3776       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3777       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3778       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3779       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3780       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3781       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3783       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3792       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
3795       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3796       "sxtl v0.4s, v0.4h\n"  in Transform()
3797       "scvtf v0.4s, v0.4s\n"  in Transform()
3798       "scvtf v1.4s, v1.4s\n"  in Transform()
3799       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3800       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3801       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3802       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3803       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3804       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3806       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
3830       "dup v4.4s, %w[range_min]\n"  in Transform()
3831       "dup v5.4s, %w[range_offset]\n"  in Transform()
3832       "dup v6.4s, %w[range_scale]\n"  in Transform()
3842       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3846       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3847       "sxtl v2.4s, v1.4h\n"  in Transform()
3848       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3849       "sxtl v0.4s, v0.4h\n"  in Transform()
3850       "scvtf v0.4s, v0.4s\n"  in Transform()
3851       "scvtf v1.4s, v1.4s\n"  in Transform()
3852       "scvtf v2.4s, v2.4s\n"  in Transform()
3853       "scvtf v3.4s, v3.4s\n"  in Transform()
3854       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3855       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3856       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3857       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3858       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3859       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3860       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3861       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3862       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3863       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3864       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3865       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3867       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3876       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
3881       "sxtl v2.4s, v1.4h\n"  in Transform()
3882       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3883       "sxtl v0.4s, v0.4h\n"  in Transform()
3884       "scvtf v0.4s, v0.4s\n"  in Transform()
3885       "scvtf v1.4s, v1.4s\n"  in Transform()
3886       "scvtf v2.4s, v2.4s\n"  in Transform()
3887       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3888       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3889       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3890       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3891       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3892       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3893       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3894       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3895       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3897       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
3898       "st1 {v2.s}[0], [%x[output]], #4\n"  in Transform()
3922       "dup v4.4s, %w[range_min]\n"  in Transform()
3923       "dup v5.4s, %w[range_offset]\n"  in Transform()
3924       "dup v6.4s, %w[range_scale]\n"  in Transform()
3934       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
3938       "sxtl2 v3.4s, v1.8h\n"  in Transform()
3939       "sxtl v2.4s, v1.4h\n"  in Transform()
3940       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3941       "sxtl v0.4s, v0.4h\n"  in Transform()
3942       "scvtf v0.4s, v0.4s\n"  in Transform()
3943       "scvtf v1.4s, v1.4s\n"  in Transform()
3944       "scvtf v2.4s, v2.4s\n"  in Transform()
3945       "scvtf v3.4s, v3.4s\n"  in Transform()
3946       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3947       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3948       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3949       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
3950       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3951       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3952       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3953       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
3954       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3955       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3956       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3957       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
3959       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
3968       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
3973       "sxtl v2.4s, v1.4h\n"  in Transform()
3974       "sxtl2 v1.4s, v0.8h\n"  in Transform()
3975       "sxtl v0.4s, v0.4h\n"  in Transform()
3976       "scvtf v0.4s, v0.4s\n"  in Transform()
3977       "scvtf v1.4s, v1.4s\n"  in Transform()
3978       "scvtf v2.4s, v2.4s\n"  in Transform()
3979       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
3980       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
3981       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
3982       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
3983       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
3984       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
3985       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
3986       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
3987       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
3989       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
3990       "st1 {v2.2s}, [%x[output]], #8\n"  in Transform()
4014       "dup v4.4s, %w[range_min]\n"  in Transform()
4015       "dup v5.4s, %w[range_offset]\n"  in Transform()
4016       "dup v6.4s, %w[range_scale]\n"  in Transform()
4026       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4030       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4031       "sxtl v2.4s, v1.4h\n"  in Transform()
4032       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4033       "sxtl v0.4s, v0.4h\n"  in Transform()
4034       "scvtf v0.4s, v0.4s\n"  in Transform()
4035       "scvtf v1.4s, v1.4s\n"  in Transform()
4036       "scvtf v2.4s, v2.4s\n"  in Transform()
4037       "scvtf v3.4s, v3.4s\n"  in Transform()
4038       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4039       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4040       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4041       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4042       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4043       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4044       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4045       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4046       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4047       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4048       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4049       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4051       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
4060       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4066       "sxtl v2.4s, v1.4h\n"  in Transform()
4067       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4068       "sxtl v0.4s, v0.4h\n"  in Transform()
4069       "scvtf v0.4s, v0.4s\n"  in Transform()
4070       "scvtf v1.4s, v1.4s\n"  in Transform()
4071       "scvtf v2.4s, v2.4s\n"  in Transform()
4072       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4073       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4074       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4075       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4076       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4077       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4078       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4079       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4080       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4082       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
4083       "st1 {v2.2s}, [%x[output]], #8\n"  in Transform()
4084       "st1 {v2.s}[2], [%x[output]], #4\n"  in Transform()
4108       "dup v4.4s, %w[range_min]\n"  in Transform()
4109       "dup v5.4s, %w[range_offset]\n"  in Transform()
4110       "dup v6.4s, %w[range_scale]\n"  in Transform()
4120       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4124       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4125       "sxtl v2.4s, v1.4h\n"  in Transform()
4126       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4127       "sxtl v0.4s, v0.4h\n"  in Transform()
4128       "scvtf v0.4s, v0.4s\n"  in Transform()
4129       "scvtf v1.4s, v1.4s\n"  in Transform()
4130       "scvtf v2.4s, v2.4s\n"  in Transform()
4131       "scvtf v3.4s, v3.4s\n"  in Transform()
4132       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4133       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4134       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4135       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4136       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4137       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4138       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4139       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4140       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4141       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4142       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4143       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4145       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
4154       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4155       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
4159       "sxtl v2.4s, v1.4h\n"  in Transform()
4160       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4161       "sxtl v0.4s, v0.4h\n"  in Transform()
4162       "scvtf v0.4s, v0.4s\n"  in Transform()
4163       "scvtf v1.4s, v1.4s\n"  in Transform()
4164       "scvtf v2.4s, v2.4s\n"  in Transform()
4165       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4166       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4167       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4168       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4169       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4170       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4171       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4172       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4173       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4175       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
4199       "dup v4.4s, %w[range_min]\n"  in Transform()
4200       "dup v5.4s, %w[range_offset]\n"  in Transform()
4201       "dup v6.4s, %w[range_scale]\n"  in Transform()
4211       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4215       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4216       "sxtl v2.4s, v1.4h\n"  in Transform()
4217       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4218       "sxtl v0.4s, v0.4h\n"  in Transform()
4219       "scvtf v0.4s, v0.4s\n"  in Transform()
4220       "scvtf v1.4s, v1.4s\n"  in Transform()
4221       "scvtf v2.4s, v2.4s\n"  in Transform()
4222       "scvtf v3.4s, v3.4s\n"  in Transform()
4223       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4224       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4225       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4226       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4227       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4228       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4229       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4230       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4231       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4232       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4233       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4234       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4236       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
4245       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4246       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
4251       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4252       "sxtl v2.4s, v1.4h\n"  in Transform()
4253       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4254       "sxtl v0.4s, v0.4h\n"  in Transform()
4255       "scvtf v0.4s, v0.4s\n"  in Transform()
4256       "scvtf v1.4s, v1.4s\n"  in Transform()
4257       "scvtf v2.4s, v2.4s\n"  in Transform()
4258       "scvtf v3.4s, v3.4s\n"  in Transform()
4259       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4260       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4261       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4262       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4263       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4264       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4265       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4266       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4267       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4268       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4269       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4270       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4272       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
4273       "st1 {v3.s}[0], [%x[output]], #4\n"  in Transform()
4297       "dup v4.4s, %w[range_min]\n"  in Transform()
4298       "dup v5.4s, %w[range_offset]\n"  in Transform()
4299       "dup v6.4s, %w[range_scale]\n"  in Transform()
4309       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4313       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4314       "sxtl v2.4s, v1.4h\n"  in Transform()
4315       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4316       "sxtl v0.4s, v0.4h\n"  in Transform()
4317       "scvtf v0.4s, v0.4s\n"  in Transform()
4318       "scvtf v1.4s, v1.4s\n"  in Transform()
4319       "scvtf v2.4s, v2.4s\n"  in Transform()
4320       "scvtf v3.4s, v3.4s\n"  in Transform()
4321       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4322       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4323       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4324       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4325       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4326       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4327       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4328       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4329       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4330       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4331       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4332       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4334       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
4343       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4344       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
4349       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4350       "sxtl v2.4s, v1.4h\n"  in Transform()
4351       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4352       "sxtl v0.4s, v0.4h\n"  in Transform()
4353       "scvtf v0.4s, v0.4s\n"  in Transform()
4354       "scvtf v1.4s, v1.4s\n"  in Transform()
4355       "scvtf v2.4s, v2.4s\n"  in Transform()
4356       "scvtf v3.4s, v3.4s\n"  in Transform()
4357       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4358       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4359       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4360       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4361       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4362       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4363       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4364       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4365       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4366       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4367       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4368       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4370       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
4371       "st1 {v3.2s}, [%x[output]], #8\n"  in Transform()
4395       "dup v4.4s, %w[range_min]\n"  in Transform()
4396       "dup v5.4s, %w[range_offset]\n"  in Transform()
4397       "dup v6.4s, %w[range_scale]\n"  in Transform()
4407       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4411       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4412       "sxtl v2.4s, v1.4h\n"  in Transform()
4413       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4414       "sxtl v0.4s, v0.4h\n"  in Transform()
4415       "scvtf v0.4s, v0.4s\n"  in Transform()
4416       "scvtf v1.4s, v1.4s\n"  in Transform()
4417       "scvtf v2.4s, v2.4s\n"  in Transform()
4418       "scvtf v3.4s, v3.4s\n"  in Transform()
4419       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4420       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4421       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4422       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4423       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4424       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4425       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4426       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4427       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4428       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4429       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4430       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4432       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
4441       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4442       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
4448       "sxtl2 v3.4s, v1.8h\n"  in Transform()
4449       "sxtl v2.4s, v1.4h\n"  in Transform()
4450       "sxtl2 v1.4s, v0.8h\n"  in Transform()
4451       "sxtl v0.4s, v0.4h\n"  in Transform()
4452       "scvtf v0.4s, v0.4s\n"  in Transform()
4453       "scvtf v1.4s, v1.4s\n"  in Transform()
4454       "scvtf v2.4s, v2.4s\n"  in Transform()
4455       "scvtf v3.4s, v3.4s\n"  in Transform()
4456       "fsub v0.4s, v0.4s, v5.4s\n"  in Transform()
4457       "fsub v1.4s, v1.4s, v5.4s\n"  in Transform()
4458       "fsub v2.4s, v2.4s, v5.4s\n"  in Transform()
4459       "fsub v3.4s, v3.4s, v5.4s\n"  in Transform()
4460       "fmul v0.4s, v0.4s, v6.4s\n"  in Transform()
4461       "fmul v1.4s, v1.4s, v6.4s\n"  in Transform()
4462       "fmul v2.4s, v2.4s, v6.4s\n"  in Transform()
4463       "fmul v3.4s, v3.4s, v6.4s\n"  in Transform()
4464       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
4465       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
4466       "fadd v2.4s, v2.4s, v4.4s\n"  in Transform()
4467       "fadd v3.4s, v3.4s, v4.4s\n"  in Transform()
4469       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
4470       "st1 {v3.2s}, [%x[output]], #8\n"  in Transform()
4471       "st1 {v3.s}[2], [%x[output]], #4\n"  in Transform()
4505       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4510       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4549       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4554       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4605       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4610       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4661       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4666       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4719       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4724       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4733       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
4738       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
4775       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4780       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4789       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
4795       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
4833       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4838       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4847       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
4853       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
4891       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4896       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4905       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
4912       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
4951       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
4956       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
4965       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
4970       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5007       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5012       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5021       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5027       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5065       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5070       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5079       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5085       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5123       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5128       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5137       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5144       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5183       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5188       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5197       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5198       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
5203       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5204       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
5241       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5246       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5255       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5256       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
5262       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5263       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
5301       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5306       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5315       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5316       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
5322       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5323       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
5361       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5366       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
5375       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
5376       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
5383       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5384       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
5411       "dup v8.4s, w0\n"  in Transform()
5413       "dup v9.4s, w0\n"  in Transform()
5415       "dup v10.4s, w0\n"  in Transform()
5417       "dup v11.4s, w0\n"  in Transform()
5419       "dup v12.4s, w0\n"  in Transform()
5421       "dup v13.4s, w0\n"  in Transform()
5423       "dup v14.4s, w0\n"  in Transform()
5431       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5432       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
5438       "sxtl2 v3.4s, v1.8h\n"  in Transform()
5439       "sxtl v2.4s, v1.4h\n"  in Transform()
5440       "sxtl2 v7.4s, v5.8h\n"  in Transform()
5441       "sxtl v6.4s, v5.4h\n"  in Transform()
5442       "sxtl2 v1.4s, v0.8h\n"  in Transform()
5443       "sxtl v0.4s, v0.4h\n"  in Transform()
5444       "sxtl2 v5.4s, v4.8h\n"  in Transform()
5445       "sxtl v4.4s, v4.4h\n"  in Transform()
5446       "scvtf v0.4s, v0.4s\n"  in Transform()
5447       "scvtf v1.4s, v1.4s\n"  in Transform()
5448       "scvtf v2.4s, v2.4s\n"  in Transform()
5449       "scvtf v3.4s, v3.4s\n"  in Transform()
5450       "scvtf v4.4s, v4.4s\n"  in Transform()
5451       "scvtf v5.4s, v5.4s\n"  in Transform()
5452       "scvtf v6.4s, v6.4s\n"  in Transform()
5453       "scvtf v7.4s, v7.4s\n"  in Transform()
5454       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5455       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
5456       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
5457       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
5458       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
5459       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
5460       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
5461       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
5462       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5463       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
5464       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
5465       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
5466       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
5467       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
5468       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
5469       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
5470       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
5471       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
5472       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
5473       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
5474       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5475       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
5476       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
5477       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
5478       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5479       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
5480       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
5481       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
5482       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5483       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
5484       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
5485       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
5486       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5487       "fcvtzs v1.4s, v1.4s\n"  in Transform()
5488       "fcvtzs v2.4s, v2.4s\n"  in Transform()
5489       "fcvtzs v3.4s, v3.4s\n"  in Transform()
5491       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
5526       "dup v8.4s, w0\n"  in Transform()
5528       "dup v9.4s, w0\n"  in Transform()
5530       "dup v10.4s, w0\n"  in Transform()
5532       "dup v11.4s, w0\n"  in Transform()
5534       "dup v12.4s, w0\n"  in Transform()
5536       "dup v13.4s, w0\n"  in Transform()
5538       "dup v14.4s, w0\n"  in Transform()
5548       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5549       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
5555       "sxtl2 v3.4s, v1.8h\n"  in Transform()
5556       "sxtl v2.4s, v1.4h\n"  in Transform()
5557       "sxtl2 v7.4s, v5.8h\n"  in Transform()
5558       "sxtl v6.4s, v5.4h\n"  in Transform()
5559       "sxtl2 v1.4s, v0.8h\n"  in Transform()
5560       "sxtl v0.4s, v0.4h\n"  in Transform()
5561       "sxtl2 v5.4s, v4.8h\n"  in Transform()
5562       "sxtl v4.4s, v4.4h\n"  in Transform()
5563       "scvtf v0.4s, v0.4s\n"  in Transform()
5564       "scvtf v1.4s, v1.4s\n"  in Transform()
5565       "scvtf v2.4s, v2.4s\n"  in Transform()
5566       "scvtf v3.4s, v3.4s\n"  in Transform()
5567       "scvtf v4.4s, v4.4s\n"  in Transform()
5568       "scvtf v5.4s, v5.4s\n"  in Transform()
5569       "scvtf v6.4s, v6.4s\n"  in Transform()
5570       "scvtf v7.4s, v7.4s\n"  in Transform()
5571       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5572       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
5573       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
5574       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
5575       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
5576       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
5577       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
5578       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
5579       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5580       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
5581       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
5582       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
5583       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
5584       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
5585       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
5586       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
5587       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
5588       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
5589       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
5590       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
5591       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5592       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
5593       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
5594       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
5595       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5596       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
5597       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
5598       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
5599       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5600       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
5601       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
5602       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
5603       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5604       "fcvtzs v1.4s, v1.4s\n"  in Transform()
5605       "fcvtzs v2.4s, v2.4s\n"  in Transform()
5606       "fcvtzs v3.4s, v3.4s\n"  in Transform()
5608       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
5619       "sxtl v0.4s, v0.4h\n"  in Transform()
5620       "sxtl v1.4s, v1.4h\n"  in Transform()
5621       "scvtf v0.4s, v0.4s\n"  in Transform()
5622       "scvtf v1.4s, v1.4s\n"  in Transform()
5623       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5624       "fmul v1.4s, v1.4s, v11.4s\n"  in Transform()
5625       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5626       "fadd v1.4s, v1.4s, v10.4s\n"  in Transform()
5627       "fadd v0.4s, v0.4s, v1.4s\n"  in Transform()
5628       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5629       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5630       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5631       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5633       "st1 {v0.s}[0], [%x[output]], #4\n"  in Transform()
5667       "dup v8.4s, w0\n"  in Transform()
5669       "dup v9.4s, w0\n"  in Transform()
5671       "dup v10.4s, w0\n"  in Transform()
5673       "dup v11.4s, w0\n"  in Transform()
5675       "dup v12.4s, w0\n"  in Transform()
5677       "dup v13.4s, w0\n"  in Transform()
5679       "dup v14.4s, w0\n"  in Transform()
5689       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5690       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
5696       "sxtl2 v3.4s, v1.8h\n"  in Transform()
5697       "sxtl v2.4s, v1.4h\n"  in Transform()
5698       "sxtl2 v7.4s, v5.8h\n"  in Transform()
5699       "sxtl v6.4s, v5.4h\n"  in Transform()
5700       "sxtl2 v1.4s, v0.8h\n"  in Transform()
5701       "sxtl v0.4s, v0.4h\n"  in Transform()
5702       "sxtl2 v5.4s, v4.8h\n"  in Transform()
5703       "sxtl v4.4s, v4.4h\n"  in Transform()
5704       "scvtf v0.4s, v0.4s\n"  in Transform()
5705       "scvtf v1.4s, v1.4s\n"  in Transform()
5706       "scvtf v2.4s, v2.4s\n"  in Transform()
5707       "scvtf v3.4s, v3.4s\n"  in Transform()
5708       "scvtf v4.4s, v4.4s\n"  in Transform()
5709       "scvtf v5.4s, v5.4s\n"  in Transform()
5710       "scvtf v6.4s, v6.4s\n"  in Transform()
5711       "scvtf v7.4s, v7.4s\n"  in Transform()
5712       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5713       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
5714       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
5715       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
5716       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
5717       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
5718       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
5719       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
5720       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5721       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
5722       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
5723       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
5724       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
5725       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
5726       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
5727       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
5728       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
5729       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
5730       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
5731       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
5732       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5733       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
5734       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
5735       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
5736       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5737       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
5738       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
5739       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
5740       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5741       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
5742       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
5743       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
5744       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5745       "fcvtzs v1.4s, v1.4s\n"  in Transform()
5746       "fcvtzs v2.4s, v2.4s\n"  in Transform()
5747       "fcvtzs v3.4s, v3.4s\n"  in Transform()
5749       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
5760       "sxtl v0.4s, v0.4h\n"  in Transform()
5761       "sxtl v1.4s, v1.4h\n"  in Transform()
5762       "scvtf v0.4s, v0.4s\n"  in Transform()
5763       "scvtf v1.4s, v1.4s\n"  in Transform()
5764       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5765       "fmul v1.4s, v1.4s, v11.4s\n"  in Transform()
5766       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5767       "fadd v1.4s, v1.4s, v10.4s\n"  in Transform()
5768       "fadd v0.4s, v0.4s, v1.4s\n"  in Transform()
5769       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5770       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5771       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5772       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5774       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5808       "dup v8.4s, w0\n"  in Transform()
5810       "dup v9.4s, w0\n"  in Transform()
5812       "dup v10.4s, w0\n"  in Transform()
5814       "dup v11.4s, w0\n"  in Transform()
5816       "dup v12.4s, w0\n"  in Transform()
5818       "dup v13.4s, w0\n"  in Transform()
5820       "dup v14.4s, w0\n"  in Transform()
5830       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5831       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
5837       "sxtl2 v3.4s, v1.8h\n"  in Transform()
5838       "sxtl v2.4s, v1.4h\n"  in Transform()
5839       "sxtl2 v7.4s, v5.8h\n"  in Transform()
5840       "sxtl v6.4s, v5.4h\n"  in Transform()
5841       "sxtl2 v1.4s, v0.8h\n"  in Transform()
5842       "sxtl v0.4s, v0.4h\n"  in Transform()
5843       "sxtl2 v5.4s, v4.8h\n"  in Transform()
5844       "sxtl v4.4s, v4.4h\n"  in Transform()
5845       "scvtf v0.4s, v0.4s\n"  in Transform()
5846       "scvtf v1.4s, v1.4s\n"  in Transform()
5847       "scvtf v2.4s, v2.4s\n"  in Transform()
5848       "scvtf v3.4s, v3.4s\n"  in Transform()
5849       "scvtf v4.4s, v4.4s\n"  in Transform()
5850       "scvtf v5.4s, v5.4s\n"  in Transform()
5851       "scvtf v6.4s, v6.4s\n"  in Transform()
5852       "scvtf v7.4s, v7.4s\n"  in Transform()
5853       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5854       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
5855       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
5856       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
5857       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
5858       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
5859       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
5860       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
5861       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5862       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
5863       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
5864       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
5865       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
5866       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
5867       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
5868       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
5869       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
5870       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
5871       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
5872       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
5873       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5874       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
5875       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
5876       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
5877       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5878       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
5879       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
5880       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
5881       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5882       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
5883       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
5884       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
5885       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5886       "fcvtzs v1.4s, v1.4s\n"  in Transform()
5887       "fcvtzs v2.4s, v2.4s\n"  in Transform()
5888       "fcvtzs v3.4s, v3.4s\n"  in Transform()
5890       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
5903       "sxtl v0.4s, v0.4h\n"  in Transform()
5904       "sxtl v1.4s, v1.4h\n"  in Transform()
5905       "scvtf v0.4s, v0.4s\n"  in Transform()
5906       "scvtf v1.4s, v1.4s\n"  in Transform()
5907       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5908       "fmul v1.4s, v1.4s, v11.4s\n"  in Transform()
5909       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
5910       "fadd v1.4s, v1.4s, v10.4s\n"  in Transform()
5911       "fadd v0.4s, v0.4s, v1.4s\n"  in Transform()
5912       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
5913       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
5914       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
5915       "fcvtzs v0.4s, v0.4s\n"  in Transform()
5917       "st1 {v0.2s}, [%x[output]], #8\n"  in Transform()
5918       "st1 {v0.s}[2], [%x[output]], #4\n"  in Transform()
5952       "dup v8.4s, w0\n"  in Transform()
5954       "dup v9.4s, w0\n"  in Transform()
5956       "dup v10.4s, w0\n"  in Transform()
5958       "dup v11.4s, w0\n"  in Transform()
5960       "dup v12.4s, w0\n"  in Transform()
5962       "dup v13.4s, w0\n"  in Transform()
5964       "dup v14.4s, w0\n"  in Transform()
5974       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
5975       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
5981       "sxtl2 v3.4s, v1.8h\n"  in Transform()
5982       "sxtl v2.4s, v1.4h\n"  in Transform()
5983       "sxtl2 v7.4s, v5.8h\n"  in Transform()
5984       "sxtl v6.4s, v5.4h\n"  in Transform()
5985       "sxtl2 v1.4s, v0.8h\n"  in Transform()
5986       "sxtl v0.4s, v0.4h\n"  in Transform()
5987       "sxtl2 v5.4s, v4.8h\n"  in Transform()
5988       "sxtl v4.4s, v4.4h\n"  in Transform()
5989       "scvtf v0.4s, v0.4s\n"  in Transform()
5990       "scvtf v1.4s, v1.4s\n"  in Transform()
5991       "scvtf v2.4s, v2.4s\n"  in Transform()
5992       "scvtf v3.4s, v3.4s\n"  in Transform()
5993       "scvtf v4.4s, v4.4s\n"  in Transform()
5994       "scvtf v5.4s, v5.4s\n"  in Transform()
5995       "scvtf v6.4s, v6.4s\n"  in Transform()
5996       "scvtf v7.4s, v7.4s\n"  in Transform()
5997       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
5998       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
5999       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6000       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6001       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6002       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6003       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6004       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6005       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6006       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6007       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6008       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6009       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6010       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6011       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6012       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6013       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6014       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6015       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6016       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6017       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6018       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6019       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6020       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6021       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6022       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6023       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6024       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6025       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6026       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6027       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6028       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6029       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6030       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6031       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6032       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6034       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6040       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
6041       "ld1 {v1.s}[0], [x1], #4\n"  in Transform()
6045       "sxtl v0.4s, v0.4h\n"  in Transform()
6046       "sxtl v1.4s, v1.4h\n"  in Transform()
6047       "scvtf v0.4s, v0.4s\n"  in Transform()
6048       "scvtf v1.4s, v1.4s\n"  in Transform()
6049       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6050       "fmul v1.4s, v1.4s, v11.4s\n"  in Transform()
6051       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6052       "fadd v1.4s, v1.4s, v10.4s\n"  in Transform()
6053       "fadd v0.4s, v0.4s, v1.4s\n"  in Transform()
6054       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6055       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6056       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6057       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6059       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
6093       "dup v8.4s, w0\n"  in Transform()
6095       "dup v9.4s, w0\n"  in Transform()
6097       "dup v10.4s, w0\n"  in Transform()
6099       "dup v11.4s, w0\n"  in Transform()
6101       "dup v12.4s, w0\n"  in Transform()
6103       "dup v13.4s, w0\n"  in Transform()
6105       "dup v14.4s, w0\n"  in Transform()
6115       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6116       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6122       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6123       "sxtl v2.4s, v1.4h\n"  in Transform()
6124       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6125       "sxtl v6.4s, v5.4h\n"  in Transform()
6126       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6127       "sxtl v0.4s, v0.4h\n"  in Transform()
6128       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6129       "sxtl v4.4s, v4.4h\n"  in Transform()
6130       "scvtf v0.4s, v0.4s\n"  in Transform()
6131       "scvtf v1.4s, v1.4s\n"  in Transform()
6132       "scvtf v2.4s, v2.4s\n"  in Transform()
6133       "scvtf v3.4s, v3.4s\n"  in Transform()
6134       "scvtf v4.4s, v4.4s\n"  in Transform()
6135       "scvtf v5.4s, v5.4s\n"  in Transform()
6136       "scvtf v6.4s, v6.4s\n"  in Transform()
6137       "scvtf v7.4s, v7.4s\n"  in Transform()
6138       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6139       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6140       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6141       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6142       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6143       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6144       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6145       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6146       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6147       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6148       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6149       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6150       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6151       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6152       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6153       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6154       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6155       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6156       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6157       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6158       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6159       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6160       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6161       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6162       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6163       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6164       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6165       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6166       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6167       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6168       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6169       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6170       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6171       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6172       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6173       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6175       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6181       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
6183       "ld1 {v2.s}[0], [x1], #4\n"  in Transform()
6188       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6189       "sxtl v0.4s, v0.4h\n"  in Transform()
6190       "sxtl2 v3.4s, v2.8h\n"  in Transform()
6191       "sxtl v2.4s, v2.4h\n"  in Transform()
6192       "scvtf v0.4s, v0.4s\n"  in Transform()
6193       "scvtf v1.4s, v1.4s\n"  in Transform()
6194       "scvtf v2.4s, v2.4s\n"  in Transform()
6195       "scvtf v3.4s, v3.4s\n"  in Transform()
6196       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6197       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6198       "fmul v2.4s, v2.4s, v11.4s\n"  in Transform()
6199       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
6200       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6201       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6202       "fadd v2.4s, v2.4s, v10.4s\n"  in Transform()
6203       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
6204       "fadd v0.4s, v0.4s, v2.4s\n"  in Transform()
6205       "fadd v1.4s, v1.4s, v3.4s\n"  in Transform()
6206       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6207       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6208       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6209       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6210       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6211       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6212       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6213       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6215       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
6216       "st1 {v1.s}[0], [%x[output]], #4\n"  in Transform()
6250       "dup v8.4s, w0\n"  in Transform()
6252       "dup v9.4s, w0\n"  in Transform()
6254       "dup v10.4s, w0\n"  in Transform()
6256       "dup v11.4s, w0\n"  in Transform()
6258       "dup v12.4s, w0\n"  in Transform()
6260       "dup v13.4s, w0\n"  in Transform()
6262       "dup v14.4s, w0\n"  in Transform()
6272       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6273       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6279       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6280       "sxtl v2.4s, v1.4h\n"  in Transform()
6281       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6282       "sxtl v6.4s, v5.4h\n"  in Transform()
6283       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6284       "sxtl v0.4s, v0.4h\n"  in Transform()
6285       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6286       "sxtl v4.4s, v4.4h\n"  in Transform()
6287       "scvtf v0.4s, v0.4s\n"  in Transform()
6288       "scvtf v1.4s, v1.4s\n"  in Transform()
6289       "scvtf v2.4s, v2.4s\n"  in Transform()
6290       "scvtf v3.4s, v3.4s\n"  in Transform()
6291       "scvtf v4.4s, v4.4s\n"  in Transform()
6292       "scvtf v5.4s, v5.4s\n"  in Transform()
6293       "scvtf v6.4s, v6.4s\n"  in Transform()
6294       "scvtf v7.4s, v7.4s\n"  in Transform()
6295       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6296       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6297       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6298       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6299       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6300       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6301       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6302       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6303       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6304       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6305       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6306       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6307       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6308       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6309       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6310       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6311       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6312       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6313       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6314       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6315       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6316       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6317       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6318       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6319       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6320       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6321       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6322       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6323       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6324       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6325       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6326       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6327       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6328       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6329       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6330       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6332       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6338       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
6340       "ld1 {v2.s}[0], [x1], #4\n"  in Transform()
6345       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6346       "sxtl v0.4s, v0.4h\n"  in Transform()
6347       "sxtl2 v3.4s, v2.8h\n"  in Transform()
6348       "sxtl v2.4s, v2.4h\n"  in Transform()
6349       "scvtf v0.4s, v0.4s\n"  in Transform()
6350       "scvtf v1.4s, v1.4s\n"  in Transform()
6351       "scvtf v2.4s, v2.4s\n"  in Transform()
6352       "scvtf v3.4s, v3.4s\n"  in Transform()
6353       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6354       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6355       "fmul v2.4s, v2.4s, v11.4s\n"  in Transform()
6356       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
6357       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6358       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6359       "fadd v2.4s, v2.4s, v10.4s\n"  in Transform()
6360       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
6361       "fadd v0.4s, v0.4s, v2.4s\n"  in Transform()
6362       "fadd v1.4s, v1.4s, v3.4s\n"  in Transform()
6363       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6364       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6365       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6366       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6367       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6368       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6369       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6370       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6372       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
6373       "st1 {v1.2s}, [%x[output]], #8\n"  in Transform()
6407       "dup v8.4s, w0\n"  in Transform()
6409       "dup v9.4s, w0\n"  in Transform()
6411       "dup v10.4s, w0\n"  in Transform()
6413       "dup v11.4s, w0\n"  in Transform()
6415       "dup v12.4s, w0\n"  in Transform()
6417       "dup v13.4s, w0\n"  in Transform()
6419       "dup v14.4s, w0\n"  in Transform()
6429       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6430       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6436       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6437       "sxtl v2.4s, v1.4h\n"  in Transform()
6438       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6439       "sxtl v6.4s, v5.4h\n"  in Transform()
6440       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6441       "sxtl v0.4s, v0.4h\n"  in Transform()
6442       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6443       "sxtl v4.4s, v4.4h\n"  in Transform()
6444       "scvtf v0.4s, v0.4s\n"  in Transform()
6445       "scvtf v1.4s, v1.4s\n"  in Transform()
6446       "scvtf v2.4s, v2.4s\n"  in Transform()
6447       "scvtf v3.4s, v3.4s\n"  in Transform()
6448       "scvtf v4.4s, v4.4s\n"  in Transform()
6449       "scvtf v5.4s, v5.4s\n"  in Transform()
6450       "scvtf v6.4s, v6.4s\n"  in Transform()
6451       "scvtf v7.4s, v7.4s\n"  in Transform()
6452       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6453       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6454       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6455       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6456       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6457       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6458       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6459       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6460       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6461       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6462       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6463       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6464       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6465       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6466       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6467       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6468       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6469       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6470       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6471       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6472       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6473       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6474       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6475       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6476       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6477       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6478       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6479       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6480       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6481       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6482       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6483       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6484       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6485       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6486       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6487       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6489       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6495       "ld1 {v0.s}[0], [%x[input]], #4\n"  in Transform()
6498       "ld1 {v2.s}[0], [x1], #4\n"  in Transform()
6504       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6505       "sxtl v0.4s, v0.4h\n"  in Transform()
6506       "sxtl2 v3.4s, v2.8h\n"  in Transform()
6507       "sxtl v2.4s, v2.4h\n"  in Transform()
6508       "scvtf v0.4s, v0.4s\n"  in Transform()
6509       "scvtf v1.4s, v1.4s\n"  in Transform()
6510       "scvtf v2.4s, v2.4s\n"  in Transform()
6511       "scvtf v3.4s, v3.4s\n"  in Transform()
6512       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6513       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6514       "fmul v2.4s, v2.4s, v11.4s\n"  in Transform()
6515       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
6516       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6517       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6518       "fadd v2.4s, v2.4s, v10.4s\n"  in Transform()
6519       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
6520       "fadd v0.4s, v0.4s, v2.4s\n"  in Transform()
6521       "fadd v1.4s, v1.4s, v3.4s\n"  in Transform()
6522       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6523       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6524       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6525       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6526       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6527       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6528       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6529       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6531       "st1 {v0.4s}, [%x[output]], #16\n"  in Transform()
6532       "st1 {v1.2s}, [%x[output]], #8\n"  in Transform()
6533       "st1 {v1.s}[2], [%x[output]], #4\n"  in Transform()
6567       "dup v8.4s, w0\n"  in Transform()
6569       "dup v9.4s, w0\n"  in Transform()
6571       "dup v10.4s, w0\n"  in Transform()
6573       "dup v11.4s, w0\n"  in Transform()
6575       "dup v12.4s, w0\n"  in Transform()
6577       "dup v13.4s, w0\n"  in Transform()
6579       "dup v14.4s, w0\n"  in Transform()
6589       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6590       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6596       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6597       "sxtl v2.4s, v1.4h\n"  in Transform()
6598       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6599       "sxtl v6.4s, v5.4h\n"  in Transform()
6600       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6601       "sxtl v0.4s, v0.4h\n"  in Transform()
6602       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6603       "sxtl v4.4s, v4.4h\n"  in Transform()
6604       "scvtf v0.4s, v0.4s\n"  in Transform()
6605       "scvtf v1.4s, v1.4s\n"  in Transform()
6606       "scvtf v2.4s, v2.4s\n"  in Transform()
6607       "scvtf v3.4s, v3.4s\n"  in Transform()
6608       "scvtf v4.4s, v4.4s\n"  in Transform()
6609       "scvtf v5.4s, v5.4s\n"  in Transform()
6610       "scvtf v6.4s, v6.4s\n"  in Transform()
6611       "scvtf v7.4s, v7.4s\n"  in Transform()
6612       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6613       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6614       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6615       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6616       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6617       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6618       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6619       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6620       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6621       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6622       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6623       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6624       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6625       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6626       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6627       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6628       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6629       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6630       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6631       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6632       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6633       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6634       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6635       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6636       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6637       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6638       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6639       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6640       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6641       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6642       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6643       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6644       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6645       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6646       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6647       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6649       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6655       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
6656       "ld1 {v2.2s}, [x1], #8\n"  in Transform()
6660       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6661       "sxtl v0.4s, v0.4h\n"  in Transform()
6662       "sxtl2 v3.4s, v2.8h\n"  in Transform()
6663       "sxtl v2.4s, v2.4h\n"  in Transform()
6664       "scvtf v0.4s, v0.4s\n"  in Transform()
6665       "scvtf v1.4s, v1.4s\n"  in Transform()
6666       "scvtf v2.4s, v2.4s\n"  in Transform()
6667       "scvtf v3.4s, v3.4s\n"  in Transform()
6668       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6669       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6670       "fmul v2.4s, v2.4s, v11.4s\n"  in Transform()
6671       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
6672       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6673       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6674       "fadd v2.4s, v2.4s, v10.4s\n"  in Transform()
6675       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
6676       "fadd v0.4s, v0.4s, v2.4s\n"  in Transform()
6677       "fadd v1.4s, v1.4s, v3.4s\n"  in Transform()
6678       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6679       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6680       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6681       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6682       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6683       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6684       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6685       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6687       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
6721       "dup v8.4s, w0\n"  in Transform()
6723       "dup v9.4s, w0\n"  in Transform()
6725       "dup v10.4s, w0\n"  in Transform()
6727       "dup v11.4s, w0\n"  in Transform()
6729       "dup v12.4s, w0\n"  in Transform()
6731       "dup v13.4s, w0\n"  in Transform()
6733       "dup v14.4s, w0\n"  in Transform()
6743       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6744       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6750       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6751       "sxtl v2.4s, v1.4h\n"  in Transform()
6752       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6753       "sxtl v6.4s, v5.4h\n"  in Transform()
6754       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6755       "sxtl v0.4s, v0.4h\n"  in Transform()
6756       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6757       "sxtl v4.4s, v4.4h\n"  in Transform()
6758       "scvtf v0.4s, v0.4s\n"  in Transform()
6759       "scvtf v1.4s, v1.4s\n"  in Transform()
6760       "scvtf v2.4s, v2.4s\n"  in Transform()
6761       "scvtf v3.4s, v3.4s\n"  in Transform()
6762       "scvtf v4.4s, v4.4s\n"  in Transform()
6763       "scvtf v5.4s, v5.4s\n"  in Transform()
6764       "scvtf v6.4s, v6.4s\n"  in Transform()
6765       "scvtf v7.4s, v7.4s\n"  in Transform()
6766       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6767       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6768       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6769       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6770       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6771       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6772       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6773       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6774       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6775       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6776       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6777       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6778       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6779       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6780       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6781       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6782       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6783       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6784       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6785       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6786       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6787       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6788       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6789       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6790       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6791       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6792       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6793       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6794       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6795       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6796       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6797       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6798       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6799       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6800       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6801       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6803       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6809       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
6811       "ld1 {v3.2s}, [x1], #8\n"  in Transform()
6818       "sxtl v2.4s, v1.4h\n"  in Transform()
6819       "sxtl v5.4s, v4.4h\n"  in Transform()
6820       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6821       "sxtl v0.4s, v0.4h\n"  in Transform()
6822       "sxtl2 v4.4s, v3.8h\n"  in Transform()
6823       "sxtl v3.4s, v3.4h\n"  in Transform()
6824       "scvtf v0.4s, v0.4s\n"  in Transform()
6825       "scvtf v1.4s, v1.4s\n"  in Transform()
6826       "scvtf v2.4s, v2.4s\n"  in Transform()
6827       "scvtf v3.4s, v3.4s\n"  in Transform()
6828       "scvtf v4.4s, v4.4s\n"  in Transform()
6829       "scvtf v5.4s, v5.4s\n"  in Transform()
6830       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6831       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6832       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6833       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
6834       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6835       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6836       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6837       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6838       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6839       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
6840       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6841       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6842       "fadd v0.4s, v0.4s, v3.4s\n"  in Transform()
6843       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
6844       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
6845       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6846       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6847       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6848       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6849       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6850       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6851       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6852       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6853       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6854       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6855       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6856       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6858       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
6859       "st1 {v2.s}[0], [%x[output]], #4\n"  in Transform()
6893       "dup v8.4s, w0\n"  in Transform()
6895       "dup v9.4s, w0\n"  in Transform()
6897       "dup v10.4s, w0\n"  in Transform()
6899       "dup v11.4s, w0\n"  in Transform()
6901       "dup v12.4s, w0\n"  in Transform()
6903       "dup v13.4s, w0\n"  in Transform()
6905       "dup v14.4s, w0\n"  in Transform()
6915       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
6916       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
6922       "sxtl2 v3.4s, v1.8h\n"  in Transform()
6923       "sxtl v2.4s, v1.4h\n"  in Transform()
6924       "sxtl2 v7.4s, v5.8h\n"  in Transform()
6925       "sxtl v6.4s, v5.4h\n"  in Transform()
6926       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6927       "sxtl v0.4s, v0.4h\n"  in Transform()
6928       "sxtl2 v5.4s, v4.8h\n"  in Transform()
6929       "sxtl v4.4s, v4.4h\n"  in Transform()
6930       "scvtf v0.4s, v0.4s\n"  in Transform()
6931       "scvtf v1.4s, v1.4s\n"  in Transform()
6932       "scvtf v2.4s, v2.4s\n"  in Transform()
6933       "scvtf v3.4s, v3.4s\n"  in Transform()
6934       "scvtf v4.4s, v4.4s\n"  in Transform()
6935       "scvtf v5.4s, v5.4s\n"  in Transform()
6936       "scvtf v6.4s, v6.4s\n"  in Transform()
6937       "scvtf v7.4s, v7.4s\n"  in Transform()
6938       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
6939       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
6940       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
6941       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
6942       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
6943       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
6944       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
6945       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
6946       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
6947       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
6948       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
6949       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
6950       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
6951       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
6952       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
6953       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
6954       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
6955       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
6956       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
6957       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
6958       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
6959       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
6960       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
6961       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
6962       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
6963       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
6964       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
6965       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
6966       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
6967       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
6968       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
6969       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
6970       "fcvtzs v0.4s, v0.4s\n"  in Transform()
6971       "fcvtzs v1.4s, v1.4s\n"  in Transform()
6972       "fcvtzs v2.4s, v2.4s\n"  in Transform()
6973       "fcvtzs v3.4s, v3.4s\n"  in Transform()
6975       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
6981       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
6983       "ld1 {v3.2s}, [x1], #8\n"  in Transform()
6990       "sxtl v2.4s, v1.4h\n"  in Transform()
6991       "sxtl v5.4s, v4.4h\n"  in Transform()
6992       "sxtl2 v1.4s, v0.8h\n"  in Transform()
6993       "sxtl v0.4s, v0.4h\n"  in Transform()
6994       "sxtl2 v4.4s, v3.8h\n"  in Transform()
6995       "sxtl v3.4s, v3.4h\n"  in Transform()
6996       "scvtf v0.4s, v0.4s\n"  in Transform()
6997       "scvtf v1.4s, v1.4s\n"  in Transform()
6998       "scvtf v2.4s, v2.4s\n"  in Transform()
6999       "scvtf v3.4s, v3.4s\n"  in Transform()
7000       "scvtf v4.4s, v4.4s\n"  in Transform()
7001       "scvtf v5.4s, v5.4s\n"  in Transform()
7002       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7003       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7004       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7005       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
7006       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7007       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7008       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7009       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7010       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7011       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
7012       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7013       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7014       "fadd v0.4s, v0.4s, v3.4s\n"  in Transform()
7015       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
7016       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
7017       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7018       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7019       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7020       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7021       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7022       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7023       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7024       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7025       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7026       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7027       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7028       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7030       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
7031       "st1 {v2.2s}, [%x[output]], #8\n"  in Transform()
7065       "dup v8.4s, w0\n"  in Transform()
7067       "dup v9.4s, w0\n"  in Transform()
7069       "dup v10.4s, w0\n"  in Transform()
7071       "dup v11.4s, w0\n"  in Transform()
7073       "dup v12.4s, w0\n"  in Transform()
7075       "dup v13.4s, w0\n"  in Transform()
7077       "dup v14.4s, w0\n"  in Transform()
7087       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
7088       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
7094       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7095       "sxtl v2.4s, v1.4h\n"  in Transform()
7096       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7097       "sxtl v6.4s, v5.4h\n"  in Transform()
7098       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7099       "sxtl v0.4s, v0.4h\n"  in Transform()
7100       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7101       "sxtl v4.4s, v4.4h\n"  in Transform()
7102       "scvtf v0.4s, v0.4s\n"  in Transform()
7103       "scvtf v1.4s, v1.4s\n"  in Transform()
7104       "scvtf v2.4s, v2.4s\n"  in Transform()
7105       "scvtf v3.4s, v3.4s\n"  in Transform()
7106       "scvtf v4.4s, v4.4s\n"  in Transform()
7107       "scvtf v5.4s, v5.4s\n"  in Transform()
7108       "scvtf v6.4s, v6.4s\n"  in Transform()
7109       "scvtf v7.4s, v7.4s\n"  in Transform()
7110       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7111       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7112       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7113       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7114       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7115       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7116       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7117       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7118       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7119       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7120       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7121       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7122       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7123       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7124       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7125       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7126       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7127       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7128       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7129       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7130       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7131       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7132       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7133       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7134       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7135       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7136       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7137       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7138       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7139       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7140       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7141       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7142       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7143       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7144       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7145       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7147       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
7153       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
7156       "ld1 {v3.2s}, [x1], #8\n"  in Transform()
7164       "sxtl v2.4s, v1.4h\n"  in Transform()
7165       "sxtl v5.4s, v4.4h\n"  in Transform()
7166       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7167       "sxtl v0.4s, v0.4h\n"  in Transform()
7168       "sxtl2 v4.4s, v3.8h\n"  in Transform()
7169       "sxtl v3.4s, v3.4h\n"  in Transform()
7170       "scvtf v0.4s, v0.4s\n"  in Transform()
7171       "scvtf v1.4s, v1.4s\n"  in Transform()
7172       "scvtf v2.4s, v2.4s\n"  in Transform()
7173       "scvtf v3.4s, v3.4s\n"  in Transform()
7174       "scvtf v4.4s, v4.4s\n"  in Transform()
7175       "scvtf v5.4s, v5.4s\n"  in Transform()
7176       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7177       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7178       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7179       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
7180       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7181       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7182       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7183       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7184       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7185       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
7186       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7187       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7188       "fadd v0.4s, v0.4s, v3.4s\n"  in Transform()
7189       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
7190       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
7191       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7192       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7193       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7194       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7195       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7196       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7197       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7198       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7199       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7200       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7201       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7202       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7204       "st1 {v0.4s, v1.4s}, [%x[output]], #32\n"  in Transform()
7205       "st1 {v2.2s}, [%x[output]], #8\n"  in Transform()
7206       "st1 {v2.s}[2], [%x[output]], #4\n"  in Transform()
7240       "dup v8.4s, w0\n"  in Transform()
7242       "dup v9.4s, w0\n"  in Transform()
7244       "dup v10.4s, w0\n"  in Transform()
7246       "dup v11.4s, w0\n"  in Transform()
7248       "dup v12.4s, w0\n"  in Transform()
7250       "dup v13.4s, w0\n"  in Transform()
7252       "dup v14.4s, w0\n"  in Transform()
7262       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
7263       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
7269       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7270       "sxtl v2.4s, v1.4h\n"  in Transform()
7271       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7272       "sxtl v6.4s, v5.4h\n"  in Transform()
7273       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7274       "sxtl v0.4s, v0.4h\n"  in Transform()
7275       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7276       "sxtl v4.4s, v4.4h\n"  in Transform()
7277       "scvtf v0.4s, v0.4s\n"  in Transform()
7278       "scvtf v1.4s, v1.4s\n"  in Transform()
7279       "scvtf v2.4s, v2.4s\n"  in Transform()
7280       "scvtf v3.4s, v3.4s\n"  in Transform()
7281       "scvtf v4.4s, v4.4s\n"  in Transform()
7282       "scvtf v5.4s, v5.4s\n"  in Transform()
7283       "scvtf v6.4s, v6.4s\n"  in Transform()
7284       "scvtf v7.4s, v7.4s\n"  in Transform()
7285       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7286       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7287       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7288       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7289       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7290       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7291       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7292       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7293       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7294       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7295       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7296       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7297       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7298       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7299       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7300       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7301       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7302       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7303       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7304       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7305       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7306       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7307       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7308       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7309       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7310       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7311       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7312       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7313       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7314       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7315       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7316       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7317       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7318       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7319       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7320       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7322       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
7328       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
7329       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
7330       "ld1 {v3.2s}, [x1], #8\n"  in Transform()
7331       "ld1 {v3.s}[2], [x1], #4\n"  in Transform()
7337       "sxtl v2.4s, v1.4h\n"  in Transform()
7338       "sxtl v5.4s, v4.4h\n"  in Transform()
7339       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7340       "sxtl v0.4s, v0.4h\n"  in Transform()
7341       "sxtl2 v4.4s, v3.8h\n"  in Transform()
7342       "sxtl v3.4s, v3.4h\n"  in Transform()
7343       "scvtf v0.4s, v0.4s\n"  in Transform()
7344       "scvtf v1.4s, v1.4s\n"  in Transform()
7345       "scvtf v2.4s, v2.4s\n"  in Transform()
7346       "scvtf v3.4s, v3.4s\n"  in Transform()
7347       "scvtf v4.4s, v4.4s\n"  in Transform()
7348       "scvtf v5.4s, v5.4s\n"  in Transform()
7349       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7350       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7351       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7352       "fmul v3.4s, v3.4s, v11.4s\n"  in Transform()
7353       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7354       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7355       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7356       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7357       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7358       "fadd v3.4s, v3.4s, v10.4s\n"  in Transform()
7359       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7360       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7361       "fadd v0.4s, v0.4s, v3.4s\n"  in Transform()
7362       "fadd v1.4s, v1.4s, v4.4s\n"  in Transform()
7363       "fadd v2.4s, v2.4s, v5.4s\n"  in Transform()
7364       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7365       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7366       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7367       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7368       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7369       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7370       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7371       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7372       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7373       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7374       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7375       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7377       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
7411       "dup v8.4s, w0\n"  in Transform()
7413       "dup v9.4s, w0\n"  in Transform()
7415       "dup v10.4s, w0\n"  in Transform()
7417       "dup v11.4s, w0\n"  in Transform()
7419       "dup v12.4s, w0\n"  in Transform()
7421       "dup v13.4s, w0\n"  in Transform()
7423       "dup v14.4s, w0\n"  in Transform()
7433       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
7434       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
7440       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7441       "sxtl v2.4s, v1.4h\n"  in Transform()
7442       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7443       "sxtl v6.4s, v5.4h\n"  in Transform()
7444       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7445       "sxtl v0.4s, v0.4h\n"  in Transform()
7446       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7447       "sxtl v4.4s, v4.4h\n"  in Transform()
7448       "scvtf v0.4s, v0.4s\n"  in Transform()
7449       "scvtf v1.4s, v1.4s\n"  in Transform()
7450       "scvtf v2.4s, v2.4s\n"  in Transform()
7451       "scvtf v3.4s, v3.4s\n"  in Transform()
7452       "scvtf v4.4s, v4.4s\n"  in Transform()
7453       "scvtf v5.4s, v5.4s\n"  in Transform()
7454       "scvtf v6.4s, v6.4s\n"  in Transform()
7455       "scvtf v7.4s, v7.4s\n"  in Transform()
7456       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7457       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7458       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7459       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7460       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7461       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7462       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7463       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7464       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7465       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7466       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7467       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7468       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7469       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7470       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7471       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7472       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7473       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7474       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7475       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7476       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7477       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7478       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7479       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7480       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7481       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7482       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7483       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7484       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7485       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7486       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7487       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7488       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7489       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7490       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7491       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7493       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
7499       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
7500       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
7502       "ld1 {v4.2s}, [x1], #8\n"  in Transform()
7503       "ld1 {v4.s}[2], [x1], #4\n"  in Transform()
7510       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7511       "sxtl v2.4s, v1.4h\n"  in Transform()
7512       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7513       "sxtl v6.4s, v5.4h\n"  in Transform()
7514       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7515       "sxtl v0.4s, v0.4h\n"  in Transform()
7516       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7517       "sxtl v4.4s, v4.4h\n"  in Transform()
7518       "scvtf v0.4s, v0.4s\n"  in Transform()
7519       "scvtf v1.4s, v1.4s\n"  in Transform()
7520       "scvtf v2.4s, v2.4s\n"  in Transform()
7521       "scvtf v3.4s, v3.4s\n"  in Transform()
7522       "scvtf v4.4s, v4.4s\n"  in Transform()
7523       "scvtf v5.4s, v5.4s\n"  in Transform()
7524       "scvtf v6.4s, v6.4s\n"  in Transform()
7525       "scvtf v7.4s, v7.4s\n"  in Transform()
7526       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7527       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7528       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7529       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7530       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7531       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7532       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7533       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7534       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7535       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7536       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7537       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7538       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7539       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7540       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7541       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7542       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7543       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7544       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7545       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7546       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7547       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7548       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7549       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7550       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7551       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7552       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7553       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7554       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7555       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7556       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7557       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7558       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7559       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7560       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7561       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7563       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
7564       "st1 {v3.s}[0], [%x[output]], #4\n"  in Transform()
7598       "dup v8.4s, w0\n"  in Transform()
7600       "dup v9.4s, w0\n"  in Transform()
7602       "dup v10.4s, w0\n"  in Transform()
7604       "dup v11.4s, w0\n"  in Transform()
7606       "dup v12.4s, w0\n"  in Transform()
7608       "dup v13.4s, w0\n"  in Transform()
7610       "dup v14.4s, w0\n"  in Transform()
7620       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
7621       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
7627       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7628       "sxtl v2.4s, v1.4h\n"  in Transform()
7629       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7630       "sxtl v6.4s, v5.4h\n"  in Transform()
7631       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7632       "sxtl v0.4s, v0.4h\n"  in Transform()
7633       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7634       "sxtl v4.4s, v4.4h\n"  in Transform()
7635       "scvtf v0.4s, v0.4s\n"  in Transform()
7636       "scvtf v1.4s, v1.4s\n"  in Transform()
7637       "scvtf v2.4s, v2.4s\n"  in Transform()
7638       "scvtf v3.4s, v3.4s\n"  in Transform()
7639       "scvtf v4.4s, v4.4s\n"  in Transform()
7640       "scvtf v5.4s, v5.4s\n"  in Transform()
7641       "scvtf v6.4s, v6.4s\n"  in Transform()
7642       "scvtf v7.4s, v7.4s\n"  in Transform()
7643       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7644       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7645       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7646       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7647       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7648       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7649       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7650       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7651       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7652       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7653       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7654       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7655       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7656       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7657       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7658       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7659       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7660       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7661       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7662       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7663       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7664       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7665       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7666       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7667       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7668       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7669       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7670       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7671       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7672       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7673       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7674       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7675       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7676       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7677       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7678       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7680       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
7686       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
7687       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
7689       "ld1 {v4.2s}, [x1], #8\n"  in Transform()
7690       "ld1 {v4.s}[2], [x1], #4\n"  in Transform()
7697       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7698       "sxtl v2.4s, v1.4h\n"  in Transform()
7699       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7700       "sxtl v6.4s, v5.4h\n"  in Transform()
7701       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7702       "sxtl v0.4s, v0.4h\n"  in Transform()
7703       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7704       "sxtl v4.4s, v4.4h\n"  in Transform()
7705       "scvtf v0.4s, v0.4s\n"  in Transform()
7706       "scvtf v1.4s, v1.4s\n"  in Transform()
7707       "scvtf v2.4s, v2.4s\n"  in Transform()
7708       "scvtf v3.4s, v3.4s\n"  in Transform()
7709       "scvtf v4.4s, v4.4s\n"  in Transform()
7710       "scvtf v5.4s, v5.4s\n"  in Transform()
7711       "scvtf v6.4s, v6.4s\n"  in Transform()
7712       "scvtf v7.4s, v7.4s\n"  in Transform()
7713       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7714       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7715       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7716       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7717       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7718       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7719       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7720       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7721       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7722       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7723       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7724       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7725       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7726       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7727       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7728       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7729       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7730       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7731       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7732       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7733       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7734       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7735       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7736       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7737       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7738       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7739       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7740       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7741       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7742       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7743       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7744       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7745       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7746       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7747       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7748       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7750       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
7751       "st1 {v3.2s}, [%x[output]], #8\n"  in Transform()
7785       "dup v8.4s, w0\n"  in Transform()
7787       "dup v9.4s, w0\n"  in Transform()
7789       "dup v10.4s, w0\n"  in Transform()
7791       "dup v11.4s, w0\n"  in Transform()
7793       "dup v12.4s, w0\n"  in Transform()
7795       "dup v13.4s, w0\n"  in Transform()
7797       "dup v14.4s, w0\n"  in Transform()
7807       "ld1 {v0.4s}, [%x[input]], #16\n"  in Transform()
7808       "ld1 {v4.4s}, [x1], #16\n"  in Transform()
7814       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7815       "sxtl v2.4s, v1.4h\n"  in Transform()
7816       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7817       "sxtl v6.4s, v5.4h\n"  in Transform()
7818       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7819       "sxtl v0.4s, v0.4h\n"  in Transform()
7820       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7821       "sxtl v4.4s, v4.4h\n"  in Transform()
7822       "scvtf v0.4s, v0.4s\n"  in Transform()
7823       "scvtf v1.4s, v1.4s\n"  in Transform()
7824       "scvtf v2.4s, v2.4s\n"  in Transform()
7825       "scvtf v3.4s, v3.4s\n"  in Transform()
7826       "scvtf v4.4s, v4.4s\n"  in Transform()
7827       "scvtf v5.4s, v5.4s\n"  in Transform()
7828       "scvtf v6.4s, v6.4s\n"  in Transform()
7829       "scvtf v7.4s, v7.4s\n"  in Transform()
7830       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7831       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7832       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7833       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7834       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7835       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7836       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7837       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7838       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7839       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7840       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7841       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7842       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7843       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7844       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7845       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7846       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7847       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7848       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7849       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7850       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7851       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7852       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7853       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7854       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7855       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7856       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7857       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7858       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7859       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7860       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7861       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7862       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7863       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7864       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7865       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7867       "st1 {v0.4s, v1.4s, v2.4s, v3.4s}, [%x[output]], #64\n"  in Transform()
7873       "ld1 {v0.2s}, [%x[input]], #8\n"  in Transform()
7874       "ld1 {v0.s}[2], [%x[input]], #4\n"  in Transform()
7877       "ld1 {v4.2s}, [x1], #8\n"  in Transform()
7878       "ld1 {v4.s}[2], [x1], #4\n"  in Transform()
7886       "sxtl2 v3.4s, v1.8h\n"  in Transform()
7887       "sxtl v2.4s, v1.4h\n"  in Transform()
7888       "sxtl2 v7.4s, v5.8h\n"  in Transform()
7889       "sxtl v6.4s, v5.4h\n"  in Transform()
7890       "sxtl2 v1.4s, v0.8h\n"  in Transform()
7891       "sxtl v0.4s, v0.4h\n"  in Transform()
7892       "sxtl2 v5.4s, v4.8h\n"  in Transform()
7893       "sxtl v4.4s, v4.4h\n"  in Transform()
7894       "scvtf v0.4s, v0.4s\n"  in Transform()
7895       "scvtf v1.4s, v1.4s\n"  in Transform()
7896       "scvtf v2.4s, v2.4s\n"  in Transform()
7897       "scvtf v3.4s, v3.4s\n"  in Transform()
7898       "scvtf v4.4s, v4.4s\n"  in Transform()
7899       "scvtf v5.4s, v5.4s\n"  in Transform()
7900       "scvtf v6.4s, v6.4s\n"  in Transform()
7901       "scvtf v7.4s, v7.4s\n"  in Transform()
7902       "fmul v0.4s, v0.4s, v9.4s\n"  in Transform()
7903       "fmul v1.4s, v1.4s, v9.4s\n"  in Transform()
7904       "fmul v2.4s, v2.4s, v9.4s\n"  in Transform()
7905       "fmul v3.4s, v3.4s, v9.4s\n"  in Transform()
7906       "fmul v4.4s, v4.4s, v11.4s\n"  in Transform()
7907       "fmul v5.4s, v5.4s, v11.4s\n"  in Transform()
7908       "fmul v6.4s, v6.4s, v11.4s\n"  in Transform()
7909       "fmul v7.4s, v7.4s, v11.4s\n"  in Transform()
7910       "fadd v0.4s, v0.4s, v8.4s\n"  in Transform()
7911       "fadd v1.4s, v1.4s, v8.4s\n"  in Transform()
7912       "fadd v2.4s, v2.4s, v8.4s\n"  in Transform()
7913       "fadd v3.4s, v3.4s, v8.4s\n"  in Transform()
7914       "fadd v4.4s, v4.4s, v10.4s\n"  in Transform()
7915       "fadd v5.4s, v5.4s, v10.4s\n"  in Transform()
7916       "fadd v6.4s, v6.4s, v10.4s\n"  in Transform()
7917       "fadd v7.4s, v7.4s, v10.4s\n"  in Transform()
7918       "fadd v0.4s, v0.4s, v4.4s\n"  in Transform()
7919       "fadd v1.4s, v1.4s, v5.4s\n"  in Transform()
7920       "fadd v2.4s, v2.4s, v6.4s\n"  in Transform()
7921       "fadd v3.4s, v3.4s, v7.4s\n"  in Transform()
7922       "fsub v0.4s, v0.4s, v12.4s\n"  in Transform()
7923       "fsub v1.4s, v1.4s, v12.4s\n"  in Transform()
7924       "fsub v2.4s, v2.4s, v12.4s\n"  in Transform()
7925       "fsub v3.4s, v3.4s, v12.4s\n"  in Transform()
7926       "fmul v0.4s, v0.4s, v13.4s\n"  in Transform()
7927       "fmul v1.4s, v1.4s, v13.4s\n"  in Transform()
7928       "fmul v2.4s, v2.4s, v13.4s\n"  in Transform()
7929       "fmul v3.4s, v3.4s, v13.4s\n"  in Transform()
7930       "fadd v0.4s, v0.4s, v14.4s\n"  in Transform()
7931       "fadd v1.4s, v1.4s, v14.4s\n"  in Transform()
7932       "fadd v2.4s, v2.4s, v14.4s\n"  in Transform()
7933       "fadd v3.4s, v3.4s, v14.4s\n"  in Transform()
7934       "fcvtzs v0.4s, v0.4s\n"  in Transform()
7935       "fcvtzs v1.4s, v1.4s\n"  in Transform()
7936       "fcvtzs v2.4s, v2.4s\n"  in Transform()
7937       "fcvtzs v3.4s, v3.4s\n"  in Transform()
7939       "st1 {v0.4s, v1.4s, v2.4s}, [%x[output]], #48\n"  in Transform()
7940       "st1 {v3.2s}, [%x[output]], #8\n"  in Transform()
7941       "st1 {v3.s}[2], [%x[output]], #4\n"  in Transform()