Lines Matching refs:D4
419 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
880 vmovdqu `16*3`($ctx),$D4 # preload r0^2
911 vpshufd \$0xEE,$D4,$D3 # 34xx -> 3434
912 vpshufd \$0x44,$D4,$D0 # xx12 -> 1212
915 vpshufd \$0xEE,$D1,$D4
918 vmovdqa $D4,-0x80(%r11)
925 vpshufd \$0xEE,$D0,$D4
928 vmovdqa $D4,-0x60(%r11)
935 vpshufd \$0xEE,$D2,$D4
938 vmovdqa $D4,-0x40(%r11)
945 vpshufd \$0xEE,$D1,$D4
947 vmovdqa $D4,-0x20(%r11)
950 vmovdqa 0x00(%rsp),$D4 # preload r0^2
977 # and $D4 is preloaded with r0^2...
979 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
980 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
982 vpmuludq $T2,$D4,$D2 # d3 = h2*r0
984 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
985 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
992 vpaddq $H1,$D4,$D4 # d4 += h3*r1
1005 vpaddq $H0,$D4,$D4 # d4 += h2*r2
1020 vpaddq $H1,$D4,$D4 # d4 += h1*r3
1042 vpaddq $T4,$D4,$D4 # d4 += h0*r4
1092 vpaddq $T4,$D4,$D4
1099 vpaddq $T0,$D4,$D4 # d4 += h3*r1
1108 vpaddq $T0,$D4,$D4 # d4 += h2*r2
1121 vpaddq $T1,$D4,$D4 # d4 += h1*r3
1144 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
1153 vmovdqa 0x00(%rsp),$D4 # preload r0^2
1204 vpshufd \$0x10,$D4,$D4 # r0^n, xx12 -> x1x2
1227 vpmuludq $T2,$D4,$D2 # d2 = h2*r0
1228 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
1230 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
1231 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
1232 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
1235 vpaddq $H0,$D4,$D4 # d4 += h3*r1
1249 vpaddq $H1,$D4,$D4 # d4 += h2*r2
1262 vpaddq $H0,$D4,$D4 # d4 += h1*r3
1275 vpaddq $H2,$D4,$D4 # h4 = d4 + h0*r4
1326 vpaddq $T4,$D4,$D4 # d4 += h4*r0
1329 vpaddq $T0,$D4,$D4 # d4 += h3*r1
1343 vpaddq $T1,$D4,$D4 # d4 += h2*r2
1356 vpaddq $T0,$D4,$D4 # d4 += h1*r3
1369 vpaddq $T2,$D4,$D4 # d4 += h0*r4
1383 vpsrldq \$8,$D4,$T4
1389 vpaddq $T4,$D4,$D4
1399 vpaddq $H3,$D4,$D4 # h3 -> h4
1405 vpsrlq \$26,$D4,$H4
1406 vpand $MASK,$D4,$D4
1426 vpaddq $H3,$D4,$D4 # h3 -> h4
1432 vmovd $D4,`4*4-48-64`($ctx)
1516 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1844 vmovdqu `16*7-64`($ctx),%x#$D4
1857 vpermd $D4,$T0,$D4
1861 vmovdqa $D4,0xe0-0x90(%rax)
1931 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
1941 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1953 vpaddq $H2,$D4,$D4 # d4 += h4*r0
1972 vpaddq $T4,$D4,$D4 # d4 += h1*r3
1990 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
2004 vpsrlq \$26,$H4,$D4
2013 vpaddq $D4,$H0,$H0
2014 vpsllq \$2,$D4,$D4
2015 vpaddq $D4,$H0,$H0 # h4 -> h0
2066 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
2076 vpaddq $T4,$D4,$D4 # d4 += h3*r1
2087 vpaddq $H2,$D4,$D4 # d4 += h4*r0
2101 vpaddq $T4,$D4,$D4 # d4 += h1*r3
2115 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
2154 vpsrlq \$26,$H4,$D4
2161 vpaddq $D4,$H0,$H0
2162 vpsllq \$2,$D4,$D4
2163 vpaddq $D4,$H0,$H0 # h4 -> h0
2212 map(s/%y/%z/,($D0,$D1,$D2,$D3,$D4));
2257 vmovdqu `16*7-64`($ctx),%x#$D4 # ... ${R4}
2272 vpermd $D4,$T2,$R4
2294 vpmuludq $T0,$R4,$D4 # d4 = r0'*r4
2307 vpaddq $M4,$D4,$D4 # d4 += r1'*r3
2318 vpaddq $M4,$D4,$D4 # d4 += r2'*r2
2328 vpaddq $M4,$D4,$D4 # d4 += r3'*r1
2338 vpaddq $M4,$D4,$D4 # d4 += r2'*r0
2354 vpaddq $M3,$D4,$D4 # d3 -> d4
2360 vpsrlq \$26,$D4,$M4
2361 vpandq $MASK,$D4,$D4
2381 vpaddq $M3,$D4,$D4 # d3 -> d4
2385 # $D0-$D4, ...
2409 vpermd $D4,$M0,${R4}{%k1}
2471 vpmuludq $H2,$R2,$D4 # d4 = h2*r2
2490 vpaddq $M4,$D4,$D4 # d4 += h0*r4
2499 vpaddq $M4,$D4,$D4 # d4 += h1*r3
2511 vpaddq $M4,$D4,$D4 # d4 += h3*r1
2521 vpaddq $M4,$D4,$D4 # d4 += h4*r0
2541 vpaddq $H3,$D4,$H4 # h3 -> h4
2551 vpsrlq \$26,$H4,$D4
2558 vpaddq $D4,$H0,$H0
2559 vpsllq \$2,$D4,$D4
2560 vpaddq $D4,$H0,$H0 # h4 -> h0
2613 vpmuludq $H2,$R2,$D4 # d4 = h2*r2
2630 vpaddq $M4,$D4,$D4 # d4 += h0*r4
2640 vpaddq $M4,$D4,$D4 # d4 += h1*r3
2650 vpaddq $M4,$D4,$D4 # d4 += h3*r1
2661 vpaddq $M4,$D4,$D4 # d4 += h4*r0
2678 vpermq \$0xb1,$D4,$H4
2683 vpaddq $D4,$H4,$H4
2690 vpermq \$0x2,$H4,$D4
2695 vpaddq $D4,$H4,$H4
2701 vextracti64x4 \$0x1,$H4,%y#$D4
2706 vpaddq $D4,$H4,${H4}{%k3}{z} # it's passed to .Ltail_avx2
2712 map(s/%z/%y/,($H0,$H1,$H2,$H3,$H4, $D0,$D1,$D2,$D3,$D4, $MASK));
2730 vpsrlq \$26,$H4,$D4
2739 vpaddq $D4,$H0,$H0
2740 vpsllq \$2,$D4,$D4
2743 vpaddq $D4,$H0,$H0 # h4 -> h0