bn/asm/ppc.pl

230 #.set r7,7
275 # Freely use registers r5,r6,r7,r8,r9,r10,r11 as follows:
278 # r7,r8 are the results of the 32x32 giving 64 bit multiply.
297 	$UMULL		r7,r5,r6
300 	addc		r7,r7,r7		# compute (r7,r8)=2*(r7,r8)
305 	addc		r10,r7,r10		# now add to temp result.
311 	$UMULL		r7,r6,r6
313 	addc		r11,r7,r11
318 	$UMULL		r7,r5,r6
321 	addc		r7,r7,r7
325 	addc		r11,r7,r11
331 	$UMULL		r7,r5,r6
333 	addc		r7,r7,r7
337 	addc		r9,r7,r9
343 	$UMULL		r7,r5,r6
346 	addc		r7,r7,r7
349 	addc		r9,r7,r9
354 	$UMULL		r7,r6,r6
356 	addc		r10,r7,r10
361 	$UMULL		r7,r5,r6
363 	addc		r7,r7,r7
367 	addc		r10,r7,r10
373 	$UMULL		r7,r5,r6
375 	addc		r7,r7,r7
379 	addc		r11,r7,r11
384 	$UMULL		r7,r6,r6
386 	addc		r9,r7,r9
415 # Freely use registers r5,r6,r7,r8,r9,r10,r11 as follows:
418 # r7,r8 are the results of the 32x32 giving 64 bit multiply.
435 	$UMULL		r7,r5,r6
438 	addc		r10,r7,r10		#add the two register number
439 	adde		r11,r8,r0 		# (r8,r7) to the three register
442 	addc		r10,r7,r10		#add the two register number
443 	adde		r11,r8,r11 		# (r8,r7) to the three register
449 	$UMULL		r7,r6,r6
451 	addc		r11,r7,r11
456 	$UMULL		r7,r5,r6
459 	addc		r11,r7,r11
463 	addc		r11,r7,r11
470 	$UMULL		r7,r5,r6
473 	addc		r9,r7,r9
477 	addc		r9,r7,r9
483 	$UMULL		r7,r5,r6
486 	addc		r9,r7,r9
490 	addc		r9,r7,r9
496 	$UMULL		r7,r6,r6
499 	addc		r10,r7,r10
504 	$UMULL		r7,r5,r6
507 	addc		r10,r7,r10
511 	addc		r10,r7,r10
517 	$UMULL		r7,r5,r6
520 	addc		r10,r7,r10
524 	addc		r10,r7,r10
530 	$UMULL		r7,r5,r6
533 	addc		r11,r7,r11
537 	addc		r11,r7,r11
543 	$UMULL		r7,r5,r6
546 	addc		r11,r7,r11
550 	addc		r11,r7,r11
556 	$UMULL		r7,r5,r6
559 	addc		r11,r7,r11
563 	addc		r11,r7,r11
568 	$UMULL		r7,r6,r6
570 	addc		r9,r7,r9
575 	$UMULL		r7,r5,r6
578 	addc		r9,r7,r9
582 	addc		r9,r7,r9
588 	$UMULL		r7,r5,r6
591 	addc		r9,r7,r9
595 	addc		r9,r7,r9
601 	$UMULL		r7,r5,r6
603 	addc		r9,r7,r9
606 	addc		r9,r7,r9
612 	$UMULL		r7,r5,r6
615 	addc		r10,r7,r10
618 	addc		r10,r7,r10
624 	$UMULL		r7,r5,r6
627 	addc		r10,r7,r10
630 	addc		r10,r7,r10
636 	$UMULL		r7,r5,r6
638 	addc		r10,r7,r10
641 	addc		r10,r7,r10
647 	$UMULL		r7,r5,r6
650 	addc		r10,r7,r10
653 	addc		r10,r7,r10
658 	$UMULL		r7,r6,r6
660 	addc		r11,r7,r11
665 	$UMULL		r7,r5,r6
667 	addc		r11,r7,r11
670 	addc		r11,r7,r11
676 	$UMULL		r7,r5,r6
678 	addc		r11,r7,r11
682 	addc		r11,r7,r11
688 	$UMULL		r7,r5,r6
690 	addc		r11,r7,r11
693 	addc		r11,r7,r11
699 	$UMULL		r7,r5,r6
702 	addc		r9,r7,r9
705 	addc		r9,r7,r9
711 	$UMULL		r7,r5,r6
713 	addc		r9,r7,r9
716 	addc		r9,r7,r9
722 	$UMULL		r7,r5,r6
724 	addc		r9,r7,r9
727 	addc		r9,r7,r9
732 	$UMULL		r7,r6,r6
734 	addc		r10,r7,r10
739 	$UMULL		r7,r5,r6
741 	addc		r10,r7,r10
744 	addc		r10,r7,r10
750 	$UMULL		r7,r5,r6
752 	addc		r10,r7,r10
755 	addc		r10,r7,r10
761 	$UMULL		r7,r5,r6
763 	addc		r11,r7,r11
766 	addc		r11,r7,r11
772 	$UMULL		r7,r5,r6
774 	addc		r11,r7,r11
777 	addc		r11,r7,r11
782 	$UMULL		r7,r6,r6
784 	addc		r9,r7,r9
789 	$UMULL		r7,r5,r6
791 	addc		r9,r7,r9
794 	addc		r9,r7,r9
801 	$UMULL		r7,r5,r6
803 	addc		r10,r7,r10
806 	addc		r10,r7,r10
811 	$UMULL		r7,r6,r6
813 	addc		r11,r7,r11
841 # r6, r7 are the 2 BN_ULONGs being multiplied.
848 	$LD	r7,`0*$BNSZ`(r5)
849 	$UMULL	r10,r6,r7
850 	$UMULH	r11,r6,r7
853 	$LD	r7,`1*$BNSZ`(r5)
854 	$UMULL	r8,r6,r7
855 	$UMULH	r9,r6,r7
861 	$LD	r7, `0*$BNSZ`(r5)
862 	$UMULL	r8,r6,r7
863 	$UMULH	r9,r6,r7
870 	$UMULL	r8,r6,r7
871 	$UMULH	r9,r6,r7
877 	$LD	r7,`1*$BNSZ`(r5)
878 	$UMULL	r8,r6,r7
879 	$UMULH	r9,r6,r7
885 	$LD	r7,`2*$BNSZ`(r5)
886 	$UMULL	r8,r6,r7
887 	$UMULH	r9,r6,r7
893 	$LD	r7,`3*$BNSZ`(r5)
894 	$UMULL	r8,r6,r7
895 	$UMULH	r9,r6,r7
901 	$LD	r7,`2*$BNSZ`(r5)
902 	$UMULL	r8,r6,r7
903 	$UMULH	r9,r6,r7
909 	$LD	r7,`1*$BNSZ`(r5)
910 	$UMULL	r8,r6,r7
911 	$UMULH	r9,r6,r7
917 	$LD	r7,`0*$BNSZ`(r5)
918 	$UMULL	r8,r6,r7
919 	$UMULH	r9,r6,r7
925 	$LD	r7,`1*$BNSZ`(r5)
926 	$UMULL	r8,r6,r7
927 	$UMULH	r9,r6,r7
933 	$LD	r7,`2*$BNSZ`(r5)
934 	$UMULL	r8,r6,r7
935 	$UMULH	r9,r6,r7
941 	$LD	r7,`3*$BNSZ`(r5)
942 	$UMULL	r8,r6,r7
943 	$UMULH	r9,r6,r7
950 	$UMULL	r8,r6,r7
951 	$UMULH	r9,r6,r7
957 	$LD	r7,`2*$BNSZ`(r5)
958 	$UMULL	r8,r6,r7
959 	$UMULH	r9,r6,r7
965 	$LD	r7,`3*$BNSZ`(r5)
966 	$UMULL	r8,r6,r7
967 	$UMULH	r9,r6,r7
995 # r6, r7 are the 2 BN_ULONGs being multiplied.
1003 	$LD	r7,`0*$BNSZ`(r5)	#b[0]
1004 	$UMULL	r10,r6,r7
1005 	$UMULH	r11,r6,r7
1008 	$LD	r7,`1*$BNSZ`(r5)
1009 	$UMULL	r8,r6,r7
1010 	$UMULH	r9,r6,r7
1016 	$LD	r7,`0*$BNSZ`(r5)
1017 	$UMULL	r8,r6,r7
1018 	$UMULH	r9,r6,r7
1025 	$UMULL	r8,r6,r7
1026 	$UMULH	r9,r6,r7
1032 	$LD	r7,`1*$BNSZ`(r5)
1033 	$UMULL	r8,r6,r7
1034 	$UMULH	r9,r6,r7
1040 	$LD	r7,`2*$BNSZ`(r5)
1041 	$UMULL	r8,r6,r7
1042 	$UMULH	r9,r6,r7
1048 	$LD	r7,`3*$BNSZ`(r5)
1049 	$UMULL	r8,r6,r7
1050 	$UMULH	r9,r6,r7
1056 	$LD	r7,`2*$BNSZ`(r5)
1057 	$UMULL	r8,r6,r7
1058 	$UMULH	r9,r6,r7
1065 	$LD	r7,`1*$BNSZ`(r5)
1066 	$UMULL	r8,r6,r7
1067 	$UMULH	r9,r6,r7
1073 	$LD	r7,`0*$BNSZ`(r5)
1074 	$UMULL	r8,r6,r7
1075 	$UMULH	r9,r6,r7
1082 	$UMULL	r8,r6,r7
1083 	$UMULH	r9,r6,r7
1089 	$LD	r7,`1*$BNSZ`(r5)
1090 	$UMULL	r8,r6,r7
1091 	$UMULH	r9,r6,r7
1097 	$LD	r7,`2*$BNSZ`(r5)
1098 	$UMULL	r8,r6,r7
1099 	$UMULH	r9,r6,r7
1105 	$LD	r7,`3*$BNSZ`(r5)
1106 	$UMULL	r8,r6,r7
1107 	$UMULH	r9,r6,r7
1113 	$LD	r7,`4*$BNSZ`(r5)
1114 	$UMULL	r8,r6,r7
1115 	$UMULH	r9,r6,r7
1121 	$LD	r7,`5*$BNSZ`(r5)
1122 	$UMULL	r8,r6,r7
1123 	$UMULH	r9,r6,r7
1129 	$LD	r7,`4*$BNSZ`(r5)
1130 	$UMULL	r8,r6,r7
1131 	$UMULH	r9,r6,r7
1137 	$LD	r7,`3*$BNSZ`(r5)
1138 	$UMULL	r8,r6,r7
1139 	$UMULH	r9,r6,r7
1145 	$LD	r7,`2*$BNSZ`(r5)
1146 	$UMULL	r8,r6,r7
1147 	$UMULH	r9,r6,r7
1153 	$LD	r7,`1*$BNSZ`(r5)
1154 	$UMULL	r8,r6,r7
1155 	$UMULH	r9,r6,r7
1161 	$LD	r7,`0*$BNSZ`(r5)
1162 	$UMULL	r8,r6,r7
1163 	$UMULH	r9,r6,r7
1170 	$UMULL	r8,r6,r7
1171 	$UMULH	r9,r6,r7
1177 	$LD	r7,`1*$BNSZ`(r5)
1178 	$UMULL	r8,r6,r7
1179 	$UMULH	r9,r6,r7
1185 	$LD	r7,`2*$BNSZ`(r5)
1186 	$UMULL	r8,r6,r7
1187 	$UMULH	r9,r6,r7
1193 	$LD	r7,`3*$BNSZ`(r5)
1194 	$UMULL	r8,r6,r7
1195 	$UMULH	r9,r6,r7
1201 	$LD	r7,`4*$BNSZ`(r5)
1202 	$UMULL	r8,r6,r7
1203 	$UMULH	r9,r6,r7
1209 	$LD	r7,`5*$BNSZ`(r5)
1210 	$UMULL	r8,r6,r7
1211 	$UMULH	r9,r6,r7
1217 	$LD	r7,`6*$BNSZ`(r5)
1218 	$UMULL	r8,r6,r7
1219 	$UMULH	r9,r6,r7
1225 	$LD	r7,`7*$BNSZ`(r5)
1226 	$UMULL	r8,r6,r7
1227 	$UMULH	r9,r6,r7
1233 	$LD	r7,`6*$BNSZ`(r5)
1234 	$UMULL	r8,r6,r7
1235 	$UMULH	r9,r6,r7
1241 	$LD	r7,`5*$BNSZ`(r5)
1242 	$UMULL	r8,r6,r7
1243 	$UMULH	r9,r6,r7
1249 	$LD	r7,`4*$BNSZ`(r5)
1250 	$UMULL	r8,r6,r7
1251 	$UMULH	r9,r6,r7
1257 	$LD	r7,`3*$BNSZ`(r5)
1258 	$UMULL	r8,r6,r7
1259 	$UMULH	r9,r6,r7
1265 	$LD	r7,`2*$BNSZ`(r5)
1266 	$UMULL	r8,r6,r7
1267 	$UMULH	r9,r6,r7
1273 	$LD	r7,`1*$BNSZ`(r5)
1274 	$UMULL	r8,r6,r7
1275 	$UMULH	r9,r6,r7
1281 	$LD	r7,`0*$BNSZ`(r5)
1282 	$UMULL	r8,r6,r7
1283 	$UMULH	r9,r6,r7
1289 	$LD	r7,`1*$BNSZ`(r5)
1290 	$UMULL	r8,r6,r7
1291 	$UMULH	r9,r6,r7
1297 	$LD	r7,`2*$BNSZ`(r5)
1298 	$UMULL	r8,r6,r7
1299 	$UMULH	r9,r6,r7
1305 	$LD	r7,`3*$BNSZ`(r5)
1306 	$UMULL	r8,r6,r7
1307 	$UMULH	r9,r6,r7
1313 	$LD	r7,`4*$BNSZ`(r5)
1314 	$UMULL	r8,r6,r7
1315 	$UMULH	r9,r6,r7
1321 	$LD	r7,`5*$BNSZ`(r5)
1322 	$UMULL	r8,r6,r7
1323 	$UMULH	r9,r6,r7
1329 	$LD	r7,`6*$BNSZ`(r5)
1330 	$UMULL	r8,r6,r7
1331 	$UMULH	r9,r6,r7
1337 	$LD	r7,`7*$BNSZ`(r5)
1338 	$UMULL	r8,r6,r7
1339 	$UMULH	r9,r6,r7
1346 	$UMULL	r8,r6,r7
1347 	$UMULH	r9,r6,r7
1353 	$LD	r7,`6*$BNSZ`(r5)
1354 	$UMULL	r8,r6,r7
1355 	$UMULH	r9,r6,r7
1361 	$LD	r7,`5*$BNSZ`(r5)
1362 	$UMULL	r8,r6,r7
1363 	$UMULH	r9,r6,r7
1369 	$LD	r7,`4*$BNSZ`(r5)
1370 	$UMULL	r8,r6,r7
1371 	$UMULH	r9,r6,r7
1377 	$LD	r7,`3*$BNSZ`(r5)
1378 	$UMULL	r8,r6,r7
1379 	$UMULH	r9,r6,r7
1385 	$LD	r7,`2*$BNSZ`(r5)
1386 	$UMULL	r8,r6,r7
1387 	$UMULH	r9,r6,r7
1393 	$LD	r7,`3*$BNSZ`(r5)
1394 	$UMULL	r8,r6,r7
1395 	$UMULH	r9,r6,r7
1401 	$LD	r7,`4*$BNSZ`(r5)
1402 	$UMULL	r8,r6,r7
1403 	$UMULH	r9,r6,r7
1409 	$LD	r7,`5*$BNSZ`(r5)
1410 	$UMULL	r8,r6,r7
1411 	$UMULH	r9,r6,r7
1417 	$LD	r7,`6*$BNSZ`(r5)
1418 	$UMULL	r8,r6,r7
1419 	$UMULH	r9,r6,r7
1425 	$LD	r7,`7*$BNSZ`(r5)
1426 	$UMULL	r8,r6,r7
1427 	$UMULH	r9,r6,r7
1434 	$UMULL	r8,r6,r7
1435 	$UMULH	r9,r6,r7
1441 	$LD	r7,`6*$BNSZ`(r5)
1442 	$UMULL	r8,r6,r7
1443 	$UMULH	r9,r6,r7
1449 	$LD	r7,`5*$BNSZ`(r5)
1450 	$UMULL	r8,r6,r7
1451 	$UMULH	r9,r6,r7
1457 	$LD	r7,`4*$BNSZ`(r5)
1458 	$UMULL	r8,r6,r7
1459 	$UMULH	r9,r6,r7
1465 	$LD	r7,`5*$BNSZ`(r5)
1466 	$UMULL	r8,r6,r7
1467 	$UMULH	r9,r6,r7
1473 	$LD	r7,`6*$BNSZ`(r5)
1474 	$UMULL	r8,r6,r7
1475 	$UMULH	r9,r6,r7
1481 	$LD	r7,`7*$BNSZ`(r5)
1482 	$UMULL	r8,r6,r7
1483 	$UMULH	r9,r6,r7
1490 	$UMULL	r8,r6,r7
1491 	$UMULH	r9,r6,r7
1497 	$LD	r7,`6*$BNSZ`(r5)
1498 	$UMULL	r8,r6,r7
1499 	$UMULH	r9,r6,r7
1505 	$LD	r7,`7*$BNSZ`(r5)
1506 	$UMULL	r8,r6,r7
1507 	$UMULH	r9,r6,r7
1544 	subfc.	r7,r0,r6        # If r6 is 0 then result is 0.
1553 	$LDU	r7,$BNSZ(r4)
1555 	subfe	r6,r8,r7	# r6 = r7+carry bit + onescomplement(r8)
1556 				# if carry = 1 this is r7-r8. Else it
1557 				# is r7-r8 -1 as we need.
1602 	$LDU	r7,$BNSZ(r4)
1604 	adde	r8,r7,r8
1642 	$CNTLZ.	r7,r5			#r7 = num leading 0s in d.
1644 	subf	r8,r7,r8		#r8 = BN_num_bits_word(d)
1651 Lppcasm_div3:				#r7 = BN_BITS2-i. so r7=i
1652 	cmpi	0,0,r7,0		# is (i == 0)?
1654 	$SHL	r3,r3,r7		# h = (h<< i)
1656 	$SHL	r5,r5,r7		# d<<=i
1658 	$SHL	r4,r4,r7		# l <<=i
1684 	$SHRI	r7,r10,`$BITS/2`	#r7= (t &BN_MASK2H), sort of...
1685 	addic.	r7,r7,0			#test if r7 == 0. used below.
1687 					# r7 = (t<<BN_BITS4)|((l&BN_MASK2h)>>BN_BITS4)
1689 	$SHLI	r7,r10,`$BITS/2`	# r7 = (t<<BN_BITS4)
1690 	or	r7,r7,r11		# r7|=((l&BN_MASK2h)>>BN_BITS4)
1691 	$UCMP	cr1,r6,r7		# compare (tl <= r7)
1749 #	r7,r8 = product.
1761 	$UMULL	r7,r6,r6
1763 	$STU	r7,$BNSZ(r3)
1791 	rlwinm.	r7,r5,30,2,31		# num >> 2
1793 	mtctr	r7
1900 	rlwinm.	r7,r5,30,2,31		# num >> 2
1902 	mtctr	r7