Lines Matching full:h1

180 my ($h0,$h1,$h2)=("%r14","%rbx","%r10");
196 	mulq	$h1			# h1*r0
201 	mulq	$h1			# h1*s1
202 	 mov	$h2,$h1			# borrow $h1
206 	imulq	$s1,$h1			# h2*s1
207 	add	$h1,$d2
208 	 mov	$d1,$h1
212 	add	$d2,$h1
222 	adc	\$0,$h1
330 	mov	8($ctx),$h1
342 	adc	8($inp),$h1
358 	mov	$h1,8($ctx)
419 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
440 	mov	$r1,$h1
472 	mov	$h1,%rax
485 	mov	$h1,$d1
530 	mov	$h1,%rax
536 	mov	$h1,$d1
570 	mov	$h1,%rax
576 	mov	$h1,$d1
649 	mov	$d2#d,$h1#d
655 	shr	\$12,$h1
658 	adc	$d2,$h1
663 	add	$d1,$h1
673 	adc	\$0,$h1
682 	adc	8($inp),$h1
695 	mov	$h1,$r0
696 	mov	$h1,$r1
701 	shr	\$14,$h1
706 	and	\$0x3ffffff,$h1		# h[3]
713 	vmovd	%rdx#d,$H1
715 	vmovd	$h1#d,$H3
722 	mov	$h1,8($ctx)
731 	mov	$h1#d,12($ctx)
776 	mov	8($ctx),$h1
788 	adc	8($inp),$h1
800 	mov	$h1,$d1
801 	mov	$h1,$d2
806 	shr	\$14,$h1
811 	and	\$0x3ffffff,$h1		# h[3]
815 	vmovd	%rdx#d,$H1
817 	vmovd	$h1#d,$H3
845 	vmovd		4*1($ctx),$H1
970 	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
971 	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
972 	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
973 	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
974 	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
980 	vpmuludq	$T1,$D4,$D1		# d1 = h1*r0
989 	  vmovdqa	$H1,0x10(%r11)				#
990 	vpmuludq	$T3,$H2,$H1		# h3*r1
992 	vpaddq		$H1,$D4,$D4		# d4 += h3*r1
995 	vpmuludq	$T1,$H2,$H1		# h1*r1
998 	vpaddq		$H1,$D2,$D2		# d2 += h1*r1
1006 	vpmuludq	$T1,$H3,$H1		# h1*r2
1008 	vpaddq		$H1,$D3,$D3		# d3 += h1*r2
1018 	vpmuludq	$T1,$H2,$H1		# h1*r3
1020 	vpaddq		$H1,$D4,$D4		# d4 += h1*r3
1023 	vpmuludq	$T3,$H3,$H1		# h3*s3
1026 	vpaddq		$H1,$D1,$D1		# d1 += h3*s3
1031 	 vmovdqu	16*1($inp),$H1				#
1038 	 vpsrldq	\$6,$H1,$H3				#
1040 	vpmuludq	$T1,$H4,$T0		# h1*s4
1041 	 vpunpckhqdq	$H1,$H0,$H4		# 4
1044 	vpaddq		$T0,$D0,$D0		# d0 += h1*s4
1046 	vpunpcklqdq	$H1,$H0,$H0		# 0:1
1051 	vpsrlq		\$26,$H0,$H1
1054 	vpand		$MASK,$H1,$H1		# 1
1062 	vpaddq		0x10(%r11),$H1,$H1
1075 	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1076 	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1077 	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1078 	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1079 	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1082 	vpmuludq	$H1,$T4,$T1		# h1*r0
1100 	vpmuludq	$H1,$T2,$T1		# h1*r1
1102 	vpaddq		$T1,$D2,$D2		# d2 += h1*r1
1107 	vpmuludq	$H1,$T3,$T1		# h1*r2
1109 	vpaddq		$T1,$D3,$D3		# d3 += h1*r2
1117 	 vpmuludq	$H1,$T2,$T1		# h1*r3
1121 	vpaddq		$T1,$D4,$D4		# d4 += h1*r3
1142 	vpmuludq	$H1,$T4,$H0
1145 	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
1173 	vpaddq		$D0,$D1,$H1		# h0 -> h1
1178 	vpsrlq		\$26,$H1,$D1
1179 	vpand		$MASK,$H1,$H1
1180 	vpaddq		$D1,$H2,$H2		# h1 -> h2
1192 	vpaddq		$D0,$H1,$H1		# h0 -> h1
1210 	vpaddq		$H1,$T1,$T1
1217 	vmovdqa		$H1,0x10(%r11)
1221 	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1222 	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1223 	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1224 	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1225 	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1230 	vpmuludq	$T1,$D4,$D1		# d1 = h1*r0
1237 	vpmuludq	$T2,$H2,$H1		# h2*r1
1238 	vpaddq		$H1,$D3,$D3		# d3 += h2*r1
1240 	vpmuludq	$T1,$H2,$H0		# h1*r1
1241 	vpaddq		$H0,$D2,$D2		# d2 += h1*r1
1248 	vpmuludq	$T2,$H4,$H1		# h2*r2
1249 	vpaddq		$H1,$D4,$D4		# d4 += h2*r2
1250 	vpmuludq	$T1,$H4,$H0		# h1*r2
1251 	vpaddq		$H0,$D3,$D3		# d3 += h1*r2
1255 	vpmuludq	$T4,$H2,$H1		# h4*s2
1256 	vpaddq		$H1,$D1,$D1		# d1 += h4*s2
1261 	vpmuludq	$T1,$H3,$H0		# h1*r3
1262 	vpaddq		$H0,$D4,$D4		# d4 += h1*r3
1266 	vpmuludq	$T4,$H4,$H1		# h4*s3
1267 	vpaddq		$H1,$D2,$D2		# d2 += h4*s3
1276 	vpmuludq	$T4,$H3,$H1		# h4*s4
1277 	vpaddq		$H1,$D3,$D3		# h3 = d3 + h4*s4
1280 	vpmuludq	$T2,$H3,$H1		# h2*s4
1281 	vpaddq		$H1,$D1,$D1		# h1 = d1 + h2*s4
1282 	vpmuludq	$T1,$H3,$H3		# h1*s4
1283 	vpaddq		$H3,$D0,$D0		# h0 = d0 + h1*s4
1288 	vmovdqu		16*1($inp),$H1
1291 	vpsrldq		\$6,$H1,$H3
1292 	vpunpckhqdq	$H1,$H0,$H4		# 4
1293 	vpunpcklqdq	$H1,$H0,$H0		# 0:1
1297 	vpsrlq		\$26,$H0,$H1
1300 	vpand		$MASK,$H1,$H1		# 1
1308 	vpaddq		0x10(%r11),$H1,$H1
1318 	vpmuludq	$H1,$T4,$T1		# h1*r0
1319 	vpaddq		$T1,$D1,$D1		# d1 += h1*r0
1334 	vpmuludq	$H1,$T2,$T0		# h1*r1
1335 	vpaddq		$T0,$D2,$D2		# d2 += h1*r1
1344 	vpmuludq	$H1,$T4,$T0		# h1*r2
1345 	vpaddq		$T0,$D3,$D3		# d3 += h1*r2
1355 	vpmuludq	$H1,$T3,$T0		# h1*r3
1356 	vpaddq		$T0,$D4,$D4		# d4 += h1*r3
1376 	vpmuludq	$H1,$T3,$T3		# h1*s4
1377 	vpaddq		$T3,$D0,$D0		# d0 += h1*s4
1403 	vpaddq		$H0,$D1,$D1		# h0 -> h1
1408 	vpsrlq		\$26,$D1,$H1
1410 	vpaddq		$H1,$D2,$D2		# h1 -> h2
1422 	vpaddq		$H0,$D1,$D1		# h0 -> h1
1483 	add	%rax,%r9	# h1
1516 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1571 	mov	$d2#d,$h1#d
1577 	shr	\$12,$h1
1580 	adc	$d2,$h1
1585 	add	$d1,$h1
1595 	adc	\$0,$h1
1605 	adc	8($inp),$h1
1623 	mov	$h1,$r0
1624 	mov	$h1,$r1
1629 	shr	\$14,$h1
1634 	and	\$0x3ffffff,$h1		# h[3]
1641 	vmovd	%rdx#d,%x#$H1
1643 	vmovd	$h1#d,%x#$H3
1650 	mov	$h1,8($ctx)
1659 	mov	$h1#d,12($ctx)
1704 	mov	8($ctx),$h1
1717 	adc	8($inp),$h1
1733 	mov	$h1,$d1
1734 	mov	$h1,$d2
1739 	shr	\$14,$h1
1744 	and	\$0x3ffffff,$h1		# h[3]
1748 	vmovd	%rdx#d,%x#$H1
1750 	vmovd	$h1#d,%x#$H3
1789 	vmovd		4*1($ctx),%x#$H1
1906 	vpaddq		$H1,$T1,$H1
1914 	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1915 	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1916 	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1917 	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1918 	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1923 	# d4 = h2*r2   + h4*r0 + h3*r1             + h1*r3   + h0*r4
1924 	# d3 = h2*r1   + h3*r0           + h1*r2   + h0*r3   + h4*5*r4
1925 	# d2 = h2*r0           + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1926 	# d1 = h2*5*r4 + h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3
1927 	# d0 = h2*5*r3 + h0*r0 + h4*5*r1 + h3*5*r2           + h1*5*r4
1936 	vpmuludq	$H1,$T1,$H2		# h1*r1, borrow $H2 as temp
1938 	vpaddq		$H2,$D2,$D2		# d2 += h1*r1
1946 	vpmuludq	$H1,$T0,$H2		# h1*r0
1948 	vpaddq		$H2,$D1,$D1		# d1 += h1*r0
1962 	vpmuludq	$H1,$T2,$T4		# h1*r2
1964 	vpaddq		$T4,$D3,$D3		# d3 += h1*r2
1969 	vpmuludq	$H1,$H2,$T4		# h1*r3
1972 	vpaddq		$T4,$D4,$D4		# d4 += h1*r3
1988 	vpmuludq	$H1,$S4,$H0		# h1*s4
1991 	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
2002 	vpaddq		$D0,$D1,$H1		# h0 -> h1
2009 	vpsrlq		\$26,$H1,$D1
2010 	vpand		$MASK,$H1,$H1
2011 	vpaddq		$D1,$H2,$H2		# h1 -> h2
2029 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2056 	vpaddq		$H1,$T1,$H1
2071 	vpmuludq	$H1,$T1,$H2		# h1*r1
2073 	vpaddq		$H2,$D2,$D2		# d2 += h1*r1
2080 	vpmuludq	$H1,$T0,$H2		# h1*r0
2083 	vpaddq		$H2,$D1,$D1		# d1 += h1*r0
2094 	vpmuludq	$H1,$T2,$T4		# h1*r2
2096 	vpaddq		$T4,$D3,$D3		# d3 += h1*r2
2099 	vpmuludq	$H1,$H2,$T4		# h1*r3
2101 	vpaddq		$T4,$D4,$D4		# d4 += h1*r3
2113 	vpmuludq	$H1,$S4,$H0		# h1*s4
2116 	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
2152 	vpaddq		$D0,$D1,$H1		# h0 -> h1
2157 	vpsrlq		\$26,$H1,$D1
2158 	vpand		$MASK,$H1,$H1
2159 	vpaddq		$D1,$H2,$H2		# h1 -> h2
2171 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2178 	vmovd		%x#$H1,`4*1-48-64`($ctx)
2213 map(s/%y/%z/,($H0,$H1,$H2,$H3,$H4));
2454 	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
2455 	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
2456 	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
2457 	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
2458 	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
2463 	# d3 = h2*r1   + h0*r3 + h1*r2   + h3*r0 + h4*5*r4
2464 	# d4 = h2*r2   + h0*r4 + h1*r3   + h3*r1 + h4*r0
2465 	# d0 = h2*5*r3 + h0*r0 + h1*5*r4         + h3*5*r2 + h4*5*r1
2466 	# d1 = h2*5*r4 + h0*r1           + h1*r0 + h3*5*r3 + h4*5*r2
2467 	# d2 = h2*r0           + h0*r2   + h1*r1 + h3*5*r4 + h4*5*r3
2478 	 vpaddq		$H1,$T1,$H1		# accumulate input
2494 	vpmuludq	$H1,$R2,$M3
2495 	vpmuludq	$H1,$R3,$M4
2496 	vpmuludq	$H1,$S4,$M0
2498 	vpaddq		$M3,$D3,$D3		# d3 += h1*r2
2499 	vpaddq		$M4,$D4,$D4		# d4 += h1*r3
2500 	vpaddq		$M0,$D0,$D0		# d0 += h1*s4
2508 	vpmuludq	$H1,$R0,$M1
2509 	vpmuludq	$H1,$R1,$M2
2512 	vpaddq		$M1,$D1,$D1		# d1 += h1*r0
2513 	vpaddq		$M2,$D2,$D2		# d2 += h1*r1
2530 	vpaddq		$M1,$D1,$H1		# h1 = d2 + h4*s2
2547 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2554 	vpsrlq		\$26,$H1,$D1
2555 	vpandq		$MASK,$H1,$H1
2556 	vpaddq		$D1,$H2,$H2		# h1 -> h2
2573 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2620 	 vpaddq		$H1,$T1,$H1		# accumulate input
2635 	vpmuludq	$H1,$R2,$M3
2636 	vpmuludq	$H1,$R3,$M4
2637 	vpmuludq	$H1,$S4,$M0
2639 	vpaddq		$M3,$D3,$D3		# d3 += h1*r2
2640 	vpaddq		$M4,$D4,$D4		# d4 += h1*r3
2641 	vpaddq		$M0,$D0,$D0		# d0 += h1*s4
2647 	vpmuludq	$H1,$R0,$M1
2648 	vpmuludq	$H1,$R1,$M2
2651 	vpaddq		$M1,$D1,$D1		# d1 += h1*r0
2652 	vpaddq		$M2,$D2,$D2		# d2 += h1*r1
2670 	vpaddq		$M1,$D1,$H1		# h1 = d2 + h4*s2
2680 	vpermq		\$0xb1,$H1,$D1
2685 	vpaddq		$D1,$H1,$H1
2692 	vpermq		\$0x2,$H1,$D1
2697 	vpaddq		$D1,$H1,$H1
2703 	vextracti64x4	\$0x1,$H1,%y#$D1
2708 	vpaddq		$D1,$H1,${H1}{%k3}{z}
2712 map(s/%z/%y/,($H0,$H1,$H2,$H3,$H4, $D0,$D1,$D2,$D3,$D4, $MASK));
2728 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2733 	vpsrlq		\$26,$H1,$D1
2734 	vpand		$MASK,$H1,$H1
2737 	vpaddq		$D1,$H2,$H2		# h1 -> h2
2755 	vpaddq		$D0,$H1,$H1		# h0 -> h1
2769 	vmovd		%x#$H1,`4*1-48-64`($ctx)
2899 my ($H0,$H1,$H2,$r2r1r0,$r1r0s2,$r0s2s1,$Dlo,$Dhi) = map("%ymm$_",(0..5,16,17));
2964 	vpermq		\$0b01010101,$Dlo,${H1}{%k7}{z}
2973 	vpmadd52luq	$r1r0s2,$H1,$Dlo
2974 	vpmadd52huq	$r1r0s2,$H1,$Dhi
3022 my ($H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2) = map("%ymm$_",(0..5,16,17));
3048 	vmovq		8($ctx),%x#$H1
3100 	vmovdqa		$R1,$H1
3126 	vpmadd52luq	$H1,$S2,$D0lo
3127 	vpmadd52huq	$H1,$S2,$D0hi
3128 	vpmadd52luq	$H1,$R0,$D1lo
3129 	vpmadd52huq	$H1,$R0,$D1hi
3130 	vpmadd52luq	$H1,$R1,$D2lo
3131 	vpmadd52huq	$H1,$R1,$D2hi
3144 	vpandq		$mask44,$D1lo,$H1
3162 	vpaddq		$tmp,$H1,$H1
3167 	vpunpcklqdq	$R1,$H1,$R1		# 1,2
3168 	vpbroadcastq	%x#$H1,%x#$H1		# 2,2
3186 	vinserti128	\$1,%x#$R1,$H1,$R1	# 1,2,3,4
3199 	vmovq		8($ctx),%x#$H1
3263 	vpaddq		$T1,$H1,$H1
3290 	vpmadd52luq	$H1,$S2,$D0lo
3291 	vpmadd52huq	$H1,$S2,$D0hi
3292 	vpmadd52luq	$H1,$R0,$D1lo
3293 	vpmadd52huq	$H1,$R0,$D1hi
3294 	vpmadd52luq	$H1,$R1,$D2lo
3295 	vpmadd52huq	$H1,$R1,$D2hi
3310 	vpandq		$mask44,$D1lo,$H1
3334 	vpaddq		$tmp,$H1,$H1
3352 	vpaddq		$T1,$H1,$H1
3374 	vpmadd52luq	$H1,$S2,$D0lo
3375 	vpmadd52huq	$H1,$S2,$D0hi
3376 	vpmadd52luq	$H1,$R0,$D1lo
3377 	vpmadd52huq	$H1,$R0,$D1hi
3378 	vpmadd52luq	$H1,$R1,$D2lo
3379 	vpmadd52huq	$H1,$R1,$D2hi
3389 	vpsrldq		\$8,$D1hi,$H1
3395 	vpaddq		$H1,$D1hi,$D1hi
3402 	vpermq		\$0x2,$D1hi,$H1
3408 	vpaddq		$H1,$D1hi,${D1hi}{%k1}{z}
3423 	vpandq		$mask44,$D1lo,$H1
3441 	vpaddq		$tmp,$H1,$H1
3448 	vmovq		%x#$H1,8($ctx)
3463 my ($H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2) = map("%ymm$_",(0..5,16,17));
3485 	vmovq	8($ctx),%x#$H1
3575 map(s/%y/%z/, $H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2);
3628 	vpaddq		$T1,$H1,$H1
3655 	vpmadd52luq	$H1,$S2,$D0lo
3656 	vpmadd52huq	$H1,$S2,$D0hi
3657 	vpmadd52luq	$H1,$R0,$D1lo
3658 	vpmadd52huq	$H1,$R0,$D1hi
3659 	vpmadd52luq	$H1,$R1,$D2lo
3660 	vpmadd52huq	$H1,$R1,$D2hi
3675 	vpandq		$mask44,$D1lo,$H1
3699 	vpaddq		$tmp,$H1,$H1
3707 	vpaddq		$T1,$H1,$H1
3729 	vpmadd52luq	$H1,$SS2,$D0lo
3730 	vpmadd52huq	$H1,$SS2,$D0hi
3731 	vpmadd52luq	$H1,$RR0,$D1lo
3732 	vpmadd52huq	$H1,$RR0,$D1hi
3733 	vpmadd52luq	$H1,$RR1,$D2lo
3734 	vpmadd52huq	$H1,$RR1,$D2hi
3744 	vpsrldq		\$8,$D1hi,$H1
3750 	vpaddq		$H1,$D1hi,$D1hi
3757 	vpermq		\$0x2,$D1hi,$H1
3763 	vpaddq		$H1,$D1hi,$D1hi
3770 	vextracti64x4	\$1,$D1hi,%y#$H1
3775 map(s/%z/%y/, $H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2);
3783 	vpaddq		$H1,$D1hi,${D1hi}{%k1}{z}
3798 	vpandq		$mask44,$D1lo,$H1
3816 	vpaddq		$tmp,$H1,$H1
3821 	vmovq		%x#$H1,8($ctx)