30 my @vtmp=map("v$_",(0..3));
150 	tbl	@vtmp[0].16b, {$dat.16b}, $MaskV.16b
152 	&mul_matrix(@vtmp[0], $TAHMatV, $TALMatV, $vtmp4);
154 	eor @vtmp[1].16b, @vtmp[1].16b, @vtmp[1].16b
155 	aese @vtmp[0].16b,@vtmp[1].16b
157 	&mul_matrix(@vtmp[0], $ATAHMatV, $ATALMatV, $vtmp4);
159 	mov	$dat.16b,@vtmp[0].16b
162 	ushr	@vtmp[0].4s,$dat.4s,32-2
163 	ushr	@vtmp[1].4s,$dat.4s,32-10
164 	ushr	@vtmp[2].4s,$dat.4s,32-18
165 	ushr	@vtmp[3].4s,$dat.4s,32-24
166 	sli	@vtmp[0].4s,$dat.4s,2
167 	sli	@vtmp[1].4s,$dat.4s,10
168 	sli	@vtmp[2].4s,$dat.4s,18
169 	sli	@vtmp[3].4s,$dat.4s,24
170 	eor	$vtmp4.16b,@vtmp[0].16b,$dat.16b
171 	eor	$vtmp4.16b,$vtmp4.16b,$vtmp[1].16b
172 	eor	$dat.16b,@vtmp[2].16b,@vtmp[3].16b
184 	tbl	@vtmp[0].16b, {$dat.16b}, $MaskV.16b
185 	tbl	@vtmp[1].16b, {$datx.16b}, $MaskV.16b
187 	&mul_matrix(@vtmp[0], $TAHMatV, $TALMatV, $vtmp4);
188 	&mul_matrix(@vtmp[1], $TAHMatV, $TALMatV, $vtmp4);
191 	aese @vtmp[0].16b,$vtmp5.16b
192 	aese @vtmp[1].16b,$vtmp5.16b
194 	&mul_matrix(@vtmp[0], $ATAHMatV, $ATALMatV,$vtmp4);
195 	&mul_matrix(@vtmp[1], $ATAHMatV, $ATALMatV,$vtmp4);
197 	mov	$dat.16b,@vtmp[0].16b
198 	mov	$datx.16b,@vtmp[1].16b
201 	ushr	@vtmp[0].4s,$dat.4s,32-2
203 	ushr	@vtmp[1].4s,$dat.4s,32-10
204 	ushr	@vtmp[2].4s,$dat.4s,32-18
205 	ushr	@vtmp[3].4s,$dat.4s,32-24
206 	sli	@vtmp[0].4s,$dat.4s,2
208 	sli	@vtmp[1].4s,$dat.4s,10
209 	sli	@vtmp[2].4s,$dat.4s,18
210 	sli	@vtmp[3].4s,$dat.4s,24
211 	eor	$vtmp4.16b,@vtmp[0].16b,$dat.16b
212 	eor	$vtmp4.16b,$vtmp4.16b,@vtmp[1].16b
213 	eor	$dat.16b,@vtmp[2].16b,@vtmp[3].16b
215 	ushr	@vtmp[1].4s,$datx.4s,32-10
216 	ushr	@vtmp[2].4s,$datx.4s,32-18
217 	ushr	@vtmp[3].4s,$datx.4s,32-24
218 	sli	@vtmp[1].4s,$datx.4s,10
219 	sli	@vtmp[2].4s,$datx.4s,18
220 	sli	@vtmp[3].4s,$datx.4s,24
222 	eor	$vtmp4.16b,$vtmp4.16b,@vtmp[1].16b
223 	eor	$datx.16b,@vtmp[2].16b,@vtmp[3].16b
233 	mov	@vtmp[3].s[0],$word
235 	tbl	@vtmp[0].16b, {@vtmp[3].16b}, $MaskV.16b
237 	&mul_matrix(@vtmp[0], $TAHMatV, $TALMatV, @vtmp[2]);
239 	eor @vtmp[1].16b, @vtmp[1].16b, @vtmp[1].16b
240 	aese @vtmp[0].16b,@vtmp[1].16b
242 	&mul_matrix(@vtmp[0], $ATAHMatV, $ATALMatV, @vtmp[2]);
245 	mov	$wtmp0,@vtmp[0].s[0]
355 	eor	@vtmp[0].16b,@data[1].16b,$rk0.16b
356 	eor	@vtmp[1].16b,@datax[1].16b,$rk0.16b
357 	eor	$rk0.16b,$rka.16b,@vtmp[0].16b
358 	eor	$rk1.16b,$rkb.16b,@vtmp[1].16b
382 	eor	@vtmp[0].16b,@data[3].16b,$rk0.16b
383 	eor	@vtmp[1].16b,@datax[3].16b,$rk0.16b
384 	eor	$rk0.16b,$rka.16b,@vtmp[0].16b
385 	eor	$rk1.16b,$rkb.16b,@vtmp[1].16b
447 	&rev32(@vtmp[3],@data[0]);
448 	&rev32(@vtmp[2],@data[1]);
449 	&rev32(@vtmp[1],@data[2]);
450 	&rev32(@vtmp[0],@data[3]);
464 	&rev32(@vtmp[3],@data[0]);
465 	&rev32(@vtmp[2],@data[1]);
466 	&rev32(@vtmp[1],@data[2]);
467 	&rev32(@vtmp[0],@data[3]);
526 	&rbit(@vtmp[2],$src,$std);
529 	shl  $des.16b, @vtmp[2].16b, #1
530 	ext  @vtmp[1].16b, @vtmp[2].16b, @vtmp[2].16b,#15
531 	ushr @vtmp[1].16b, @vtmp[1].16b, #7
532 	mul  @vtmp[1].16b, @vtmp[1].16b, @vtmp[0].16b
533 	eor  $des.16b, $des.16b, @vtmp[1].16b
593 	movi	@vtmp[0].16b,#64
606 	tbl	@vtmp[0].16b, {@data[0].16b}, $MaskV.16b
608 	&mul_matrix(@vtmp[0], $TAHMatV, $TALMatV, @vtmp[2]);
610 	eor @vtmp[1].16b, @vtmp[1].16b, @vtmp[1].16b
611 	aese @vtmp[0].16b,@vtmp[1].16b
613 	&mul_matrix(@vtmp[0], $ATAHMatV, $ATALMatV, @vtmp[2]);
615 	mov	$wtmp,@vtmp[0].s[0]
763 	st4	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
779 	st4	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
805 	st4	{@vtmp[0].s-@vtmp[3].s}[0],[$outp],#16
806 	st4	{@vtmp[0].s-@vtmp[3].s}[1],[$outp]
817 	st4	{@vtmp[0].s-@vtmp[3].s}[0],[$outp],#16
818 	st4	{@vtmp[0].s-@vtmp[3].s}[1],[$outp],#16
819 	st4	{@vtmp[0].s-@vtmp[3].s}[2],[$outp]
925 	&transpose(@vtmp,@datax);
932 	eor	@vtmp[0].16b,@vtmp[0].16b,$ivec1.16b
934 	eor	@vtmp[1].16b,@vtmp[1].16b,@datax[0].16b
935 	eor	@vtmp[2].16b,@vtmp[2].16b,@datax[1].16b
936 	eor	@vtmp[3].16b,$vtmp[3].16b,@datax[2].16b
943 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
963 	&transpose(@vtmp,@datax);
965 	eor	@vtmp[0].16b,@vtmp[0].16b,$ivec1.16b
966 	eor	@vtmp[1].16b,@vtmp[1].16b,@data[0].16b
968 	eor	@vtmp[2].16b,@vtmp[2].16b,@data[1].16b
969 	eor	@vtmp[3].16b,$vtmp[3].16b,@data[2].16b
970 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
1005 	&transpose(@vtmp,@datax);
1007 	eor	@vtmp[0].16b,@vtmp[0].16b,$ivec1.16b
1008 	eor	@vtmp[1].16b,@vtmp[1].16b,@data[0].16b
1009 	st1	{@vtmp[0].4s,@vtmp[1].4s},[$outp],#32
1024 	&transpose(@vtmp,@datax);
1026 	eor	@vtmp[0].16b,@vtmp[0].16b,$ivec1.16b
1027 	eor	@vtmp[1].16b,@vtmp[1].16b,@data[0].16b
1028 	eor	@vtmp[2].16b,@vtmp[2].16b,@data[1].16b
1029 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s},[$outp],#48
1100 	eor	@vtmp[0].16b,@vtmp[0].16b,@vtmpx[0].16b
1101 	eor	@vtmp[1].16b,@vtmp[1].16b,@vtmpx[1].16b
1102 	eor	@vtmp[2].16b,@vtmp[2].16b,@vtmpx[2].16b
1103 	eor	@vtmp[3].16b,@vtmp[3].16b,@vtmpx[3].16b
1104 	st4	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
1123 	eor	@vtmp[0].16b,@vtmp[0].16b,@vtmpx[0].16b
1124 	eor	@vtmp[1].16b,@vtmp[1].16b,@vtmpx[1].16b
1125 	eor	@vtmp[2].16b,@vtmp[2].16b,@vtmpx[2].16b
1126 	eor	@vtmp[3].16b,@vtmp[3].16b,@vtmpx[3].16b
1131 	st4	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
1163 	eor	@vtmp[0].16b,@vtmp[0].16b,@vtmpx[0].16b
1164 	eor	@vtmp[1].16b,@vtmp[1].16b,@vtmpx[1].16b
1165 	eor	@vtmp[2].16b,@vtmp[2].16b,@vtmpx[2].16b
1166 	eor	@vtmp[3].16b,@vtmp[3].16b,@vtmpx[3].16b
1167 	st4	{@vtmp[0].s,@vtmp[1].s,@vtmp[2].s,@vtmp[3].s}[0],[$outp],#16
1168 	st4	{@vtmp[0].s,@vtmp[1].s,@vtmp[2].s,@vtmp[3].s}[1],[$outp],#16
1177 	eor	@vtmp[0].16b,@vtmp[0].16b,@vtmpx[0].16b
1178 	eor	@vtmp[1].16b,@vtmp[1].16b,@vtmpx[1].16b
1179 	eor	@vtmp[2].16b,@vtmp[2].16b,@vtmpx[2].16b
1180 	eor	@vtmp[3].16b,@vtmp[3].16b,@vtmpx[3].16b
1181 	st4	{@vtmp[0].s,@vtmp[1].s,@vtmp[2].s,@vtmp[3].s}[0],[$outp],#16
1182 	st4	{@vtmp[0].s,@vtmp[1].s,@vtmp[2].s,@vtmp[3].s}[1],[$outp],#16
1183 	st4	{@vtmp[0].s,@vtmp[1].s,@vtmp[2].s,@vtmp[3].s}[2],[$outp],#16
1321 	&transpose(@data,@vtmp);
1322 	&transpose(@datax,@vtmp);
1326 	&transpose(@vtmp,@datax);
1329 	eor @vtmp[0].16b, @vtmp[0].16b, @tweak[0].16b
1330 	eor @vtmp[1].16b, @vtmp[1].16b, @tweak[1].16b
1331 	eor @vtmp[2].16b, @vtmp[2].16b, @tweak[2].16b
1332 	eor @vtmp[3].16b, @vtmp[3].16b, @tweak[3].16b
1340 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
1364 	&transpose(@data,@vtmp);
1368 	&transpose(@vtmp,@data);
1370 	eor @vtmp[0].16b, @vtmp[0].16b, @tweak[0].16b
1371 	eor @vtmp[1].16b, @vtmp[1].16b, @tweak[1].16b
1372 	eor @vtmp[2].16b, @vtmp[2].16b, @tweak[2].16b
1373 	eor @vtmp[3].16b, @vtmp[3].16b, @tweak[3].16b
1374 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s,@vtmp[3].4s},[$outp],#64
1413 	&transpose(@data,@vtmp);
1417 	&transpose(@vtmp,@data);
1419 	eor @vtmp[0].16b, @vtmp[0].16b, @tweak[0].16b
1420 	eor @vtmp[1].16b, @vtmp[1].16b, @tweak[1].16b
1421 	st1	{@vtmp[0].4s,@vtmp[1].4s},[$outp],#32
1439 	&transpose(@data,@vtmp);
1443 	&transpose(@vtmp,@data);
1445 	eor @vtmp[0].16b, @vtmp[0].16b, @tweak[0].16b
1446 	eor @vtmp[1].16b, @vtmp[1].16b, @tweak[1].16b
1447 	eor @vtmp[2].16b, @vtmp[2].16b, @tweak[2].16b
1448 	st1	{@vtmp[0].4s,@vtmp[1].4s,@vtmp[2].4s},[$outp],#48
1483 	mov @vtmp[0].16B,@tweak[1].16b
1485 	mov @tweak[2].16B,@vtmp[0].16b

Last Index update Fri Nov 01 00:06:59 UTC 2024

Dedicated to & Maintained by Room-11