Lines Matching refs:pDst
151 #MACRO swizzle pDst, pSrc, tmp1, tmp2
152 # pDst and pSrc are modified
155 my ($pDst, $pSrc, $cnt, $d0)=@_;
160 mov $d0#w, ($pDst)
162 mov $d0#w, (+64*1)($pDst)
164 mov $d0#w, (+64*2)($pDst)
166 mov $d0#w, (+64*3)($pDst)
168 lea 64*4($pDst), $pDst
177 #MACRO unswizzle pDst, pSrc, tmp*3
180 my ($pDst, $pSrc, $cnt, $d0, $d1)=@_;
198 mov $d0, (+8*0)($pDst)
199 mov $d1, (+8*1)($pDst)
201 lea 8*2($pDst), $pDst
365 #MULADD_256x512 MACRO pDst, pA, pB, OP, TMP, X7, X6, X5, X4, X3, X2, X1, X0
367 # Inputs: pDst: Destination (768 bits, 12 qwords)
376 my ($pDst, $pA, $pB, $OP, $TMP, $X)=@_;
380 &MULSTEP_512_ADD($X, "(+8*0)($pDst)", $pB, $pA, $OP, $TMP);
386 &MULSTEP_512($X, "(+8*1)($pDst)", $pB, $OP, $TMP);
392 &MULSTEP_512($X, "(+8*2)($pDst)", $pB, $OP, $TMP);
398 &MULSTEP_512($X, "(+8*3)($pDst)", $pB, $OP, $TMP);
618 #MUL_512x512 MACRO pDst, pA, pB, x7, x6, x5, x4, x3, x2, x1, x0, tmp*2
620 # Inputs: pDst: Destination (1024 bits, 16 qwords)
627 my ($pDst
628 my ($pDst, $pDst_o) = ($pDst =~ m/([^+]*)\+?(.*)?/);
636 mov %rax, (+$pDst_o+8*0)($pDst)
654 &MULSTEP_512(\@X, "(+$pDst_o+8*$i)($pDst)", $pB, $OP, $TMP);
659 mov $X[0], (+$pDst_o+8*8)($pDst)
660 mov $X[1], (+$pDst_o+8*9)($pDst)
661 mov $X[2], (+$pDst_o+8*10)($pDst)
662 mov $X[3], (+$pDst_o+8*11)($pDst)
663 mov $X[4], (+$pDst_o+8*12)($pDst)
664 mov $X[5], (+$pDst_o+8*13)($pDst)
665 mov $X[6], (+$pDst_o+8*14)($pDst)
666 mov $X[7], (+$pDst_o+8*15)($pDst)
711 my ($pDst, $pA, $x, $A, $tmp, $x7, $x6, $pDst_o)=@_;
712 my ($pDst, $pDst_o) = ($pDst =~ m/([^+]*)\+?(.*)?/);
722 mov %rax, (+$pDst_o+8*1)($pDst)
736 mov $X[0], (+$pDst_o+8*2)($pDst)
748 mov $X[1], (+$pDst_o+8*3)($pDst)
757 mov $X[2], (+$pDst_o+8*4)($pDst)
802 mov $X[3], (+$pDst_o+8*5)($pDst)
811 mov $X[4], (+$pDst_o+8*6)($pDst)
849 mov $X[5], (+$pDst_o+8*7)($pDst)
858 mov $x7, (+$pDst_o+8*8)($pDst)
887 mov $X[1], (+$pDst_o+8*9)($pDst)
896 mov $X[2], (+$pDst_o+8*10)($pDst)
917 mov $X[5], (+$pDst_o+8*11)($pDst)
926 mov $X[1], (+$pDst_o+8*12)($pDst)
939 mov $X[2], (+$pDst_o+8*13)($pDst)
941 mov %rdx, (+$pDst_o+8*14)($pDst)
944 mov (+$pDst_o+8*1)($pDst), $X[0]
945 mov (+$pDst_o+8*2)($pDst), $X[1]
946 mov (+$pDst_o+8*3)($pDst), $X[2]
947 mov (+$pDst_o+8*4)($pDst), $X[3]
948 mov (+$pDst_o+8*5)($pDst), $X[4]
949 mov (+$pDst_o+8*6)($pDst), $X[5]
966 mov %rax, (+$pDst_o+8*0)($pDst)
977 mov $X[0], (+$pDst_o+8*1)($pDst)
978 mov $X[1], (+$pDst_o+8*2)($pDst)
989 mov $X[2], (+$pDst_o+8*3)($pDst)
990 mov $X[3], (+$pDst_o+8*4)($pDst)
997 mov $X[4], (+$pDst_o+8*5)($pDst)
998 mov $X[5], (+$pDst_o+8*6)($pDst)
1003 mov (+$pDst_o+8*7)($pDst), $X[0]
1004 mov (+$pDst_o+8*8)($pDst), $X[1]
1005 mov (+$pDst_o+8*9)($pDst), $X[2]
1006 mov (+$pDst_o+8*10)($pDst), $X[3]
1007 mov (+$pDst_o+8*11)($pDst), $X[4]
1008 mov (+$pDst_o+8*12)($pDst), $X[5]
1009 mov (+$pDst_o+8*13)($pDst), $x6
1010 mov (+$pDst_o+8*14)($pDst), $x7
1038 mov $X[0], (+$pDst_o+8*7)($pDst)
1039 mov $X[1], (+$pDst_o+8*8)($pDst)
1050 mov $X[2], (+$pDst_o+8*9)($pDst)
1051 mov $X[3], (+$pDst_o+8*10)($pDst)
1060 mov $X[4], (+$pDst_o+8*11)($pDst)
1061 mov $X[5], (+$pDst_o+8*12)($pDst)
1067 mov $x6, (+$pDst_o+8*13)($pDst)
1068 mov $x7, (+$pDst_o+8*14)($pDst)
1069 mov $A, (+$pDst_o+8*15)($pDst)
1278 # Call mod_mul_a1(pDst, pSrc1, pSrc2, pM, pData)
