Lines Matching full:t4
419 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
892 vpunpckhqdq $T1,$T0,$T4 # 4
896 vpsrlq \$40,$T4,$T4 # 4
904 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
985 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
988 vpmuludq 0x20(%rsp),$T4,$H0 # h4*s1
1011 vpmuludq $T4,$H4,$H0 # h4*s2
1022 vpmuludq $T4,$H3,$H0 # h4*s3
1034 vpmuludq $T4,$H4,$T4 # h4*s4
1037 vpaddq $T4,$D3,$D3 # d3 += h4*s4
1039 vpmuludq 0x70(%rsp),$T0,$T4 # h0*r4
1042 vpaddq $T4,$D4,$D4 # d4 += h0*r4
1043 vmovdqa -0x90(%r11),$T4 # r0^4
1081 vpmuludq $H0,$T4,$T0 # h0*r0
1082 vpmuludq $H1,$T4,$T1 # h1*r0
1086 vpmuludq $H2,$T4,$T0 # h2*r0
1087 vpmuludq $H3,$T4,$T1 # h3*r0
1090 vpmuludq $H4,$T4,$T4 # h4*r0
1092 vpaddq $T4,$D4,$D4
1105 vmovdqa -0x50(%r11),$T4 # s2^4
1112 vpmuludq $H4,$T4,$T0 # h4*s2
1116 vpmuludq $H3,$T4,$T4 # h3*s2
1118 vpaddq $T4,$D0,$D0 # d0 += h3*s2
1120 vmovdqa -0x10(%r11),$T4 # s4^4
1133 vpmuludq $H2,$T4,$H2 # h2*s4
1134 vpmuludq $H3,$T4,$H3 # h3*s4
1137 vpmuludq $H4,$T4,$H4 # h4*s4
1142 vpmuludq $H1,$T4,$H0
1143 vpunpckhqdq $T1,$T0,$T4 # 4
1150 #vpsrlq \$40,$T4,$T4 # 4
1151 vpsrldq \$`40/8`,$T4,$T4 # 4
1157 vpand 0(%rcx),$T4,$T4 # .Lmask24
1161 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1212 vpaddq $H4,$T4,$T4
1232 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
1244 vpmuludq $T4,$H3,$H3 # h4*s1
1255 vpmuludq $T4,$H2,$H1 # h4*s2
1266 vpmuludq $T4,$H4,$H1 # h4*s3
1276 vpmuludq $T4,$H3,$H1 # h4*s4
1306 vpshufd \$0x32,`16*0-64`($ctx),$T4 # r0^n, 34xx -> x3x4
1316 vpmuludq $H0,$T4,$T0 # h0*r0
1318 vpmuludq $H1,$T4,$T1 # h1*r0
1320 vpmuludq $H2,$T4,$T0 # h2*r0
1323 vpmuludq $H3,$T4,$T1 # h3*r0
1325 vpmuludq $H4,$T4,$T4 # h4*r0
1326 vpaddq $T4,$D4,$D4 # d4 += h4*r0
1333 vpshufd \$0x32,`16*3-64`($ctx),$T4 # r2
1342 vpmuludq $H2,$T4,$T1 # h2*r2
1344 vpmuludq $H1,$T4,$T0 # h1*r2
1347 vpmuludq $H0,$T4,$T4 # h0*r2
1348 vpaddq $T4,$D2,$D2 # d2 += h0*r2
1351 vpshufd \$0x32,`16*6-64`($ctx),$T4 # s3
1360 vpmuludq $H4,$T4,$T1 # h4*s3
1363 vpmuludq $H3,$T4,$T0 # h3*s3
1365 vpmuludq $H2,$T4,$T4 # h2*s3
1366 vpaddq $T4,$D0,$D0 # d0 += h2*s3
1383 vpsrldq \$8,$D4,$T4
1389 vpaddq $T4,$D4,$D4
1516 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1837 vmovdqu `16*2-64`($ctx),%x#$T4
1847 vpermd $T4,$T0,$T4
1852 vmovdqa $T4,0x40-0x90(%rax)
1875 vpunpckhqdq $T1,$T0,$T4 # 4
1882 vpsrlq \$40,$T4,$T4 # 4
1887 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1910 vpaddq $H4,$T4,$H4
1935 vpmuludq $H0,$T1,$T4 # h0*r1
1937 vpaddq $T4,$D1,$D1 # d1 += h0*r1
1939 vpmuludq $H3,$T1,$T4 # h3*r1
1941 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1945 vpmuludq $H0,$T0,$T4 # h0*r0
1947 vpaddq $T4,$D0,$D0 # d0 += h0*r0
1949 vpmuludq $H3,$T0,$T4 # h3*r0
1952 vpaddq $T4,$D3,$D3 # d3 += h3*r0
1956 vpmuludq $H3,$T1,$T4 # h3*s2
1959 vpaddq $T4,$D0,$D0 # d0 += h3*s2
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
1974 vpmuludq $H3,$T3,$T4 # h3*s3
1977 vpaddq $T4,$D1,$D1 # d1 += h3*s3
1979 vpunpckhqdq $T1,$T0,$T4 # 4
2031 vpsrlq \$40,$T4,$T4 # 4
2040 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
2060 vpaddq $H4,$T4,$H4
2070 vpmuludq $H0,$T1,$T4 # h0*r1
2072 vpaddq $T4,$D1,$D1 # d1 += h0*r1
2074 vpmuludq $H3,$T1,$T4 # h3*r1
2076 vpaddq $T4,$D4,$D4 # d4 += h3*r1
2079 vpmuludq $H0,$T0,$T4 # h0*r0
2081 vpaddq $T4,$D0,$D0 # d0 += h0*r0
2084 vpmuludq $H3,$T0,$T4 # h3*r0
2086 vpaddq $T4,$D3,$D3 # d3 += h3*r0
2089 vpmuludq $H3,$T1,$T4 # h3*s2
2091 vpaddq $T4,$D0,$D0 # d0 += h3*s2
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
2103 vpmuludq $H3,$T3,$T4 # h3*s3
2105 vpaddq $T4,$D1,$D1 # d1 += h3*s3
2124 vpsrldq \$8,$H4,$T4
2129 vpaddq $T4,$H4,$H4
2133 vpermq \$0x2,$H4,$T4
2138 vpaddq $T4,$H4,$H4
2211 map(s/%y/%z/,($T4,$T0,$T1,$T2,$T3)); # switch to %zmm domain
2258 vmovdqu `16*8-64`($ctx),%x#$T4 # ... ${S4}
2274 vpermd $T4,$T2,$S4
2314 vpsrlq \$32,$R4,$T4
2332 vpmuludq $T4,$S4,$M3
2333 vpmuludq $T4,$R0,$M4
2334 vpmuludq $T4,$S1,$M0
2335 vpmuludq $T4,$S2,$M1
2336 vpmuludq $T4,$S3,$M2
2346 vmovdqu64 16*4($inp),%z#$T4
2387 vpunpcklqdq $T4,$T3,$T0 # transpose input
2388 vpunpckhqdq $T4,$T3,$T4
2423 vpsllq \$12,$T4,$T3
2426 vpsrlq \$14,$T4,$T3
2427 vpsrlq \$40,$T4,$T4 # 4
2432 #vporq $PADBIT,$T4,$T4 # padbit, yes, always
2476 vporq $PADBIT,$T4,$T4 # padbit, yes, always
2480 vpaddq $H4,$T4,$H4
2483 vmovdqu64 16*4($inp),$T4
2503 vpunpcklqdq $T4,$T3,$T0 # transpose input
2504 vpunpckhqdq $T4,$T3,$T4
2537 vpsllq \$12,$T4,$T3
2569 vpsrlq \$14,$T4,$T3
2575 vpsrlq \$40,$T4,$T4 # 4
2584 #vporq $PADBIT,$T4,$T4 # padbit, yes, always
2619 vporq $PADBIT,$T4,$T4 # padbit, yes, always
2622 vpaddq $H4,$T4,$H4
2711 map(s/%z/%y/,($T0,$T1,$T2,$T3,$T4, $PADBIT));
2721 vpunpckhqdq $T1,$T0,$T4 # 4
2742 vpsrlq \$40,$T4,$T4 # 4
2760 vpor 32(%rcx),$T4,$T4 # padbit, yes, always