sys/src/libsec/amd64/sha1block.s

   1 /* x = (wp[off-f] ^ wp[off-8] ^ wp[off-14] ^ wp[off-16]) <<< 1;
   2  * wp[off] = x;
   3  * x += A <<< 5;
   4  * E += 0xca62c1d6 + x;
   5  * x = FN(B,C,D);
   6  * E += x;
   7  * B >>> 2
   8  */
   9 #define BSWAPDI BYTE $0x0f; BYTE $0xcf;
  10
  11 #define BODY(off,FN,V,A,B,C,D,E)\
  12         MOVL (off-64)(BP),DI;\
  13         XORL (off-56)(BP),DI;\
  14         XORL (off-32)(BP),DI;\
  15         XORL (off-12)(BP),DI;\
  16         ROLL $1,DI;\
  17         MOVL DI,off(BP);\
  18         LEAL V(DI)(E*1),E;\
  19         MOVL A,DI;\
  20         ROLL $5,DI;\
  21         ADDL DI,E;\
  22         FN(B,C,D)\
  23         ADDL DI,E;\
  24         RORL $2,B;\
  25
  26 #define BODY0(off,FN,V,A,B,C,D,E)\
  27         MOVLQZX off(BX),DI;\
  28         BSWAPDI;\
  29         MOVL DI,off(BP);\
  30         LEAL V(DI)(E*1),E;\
  31         MOVL A,DI;\
  32         ROLL $5,DI;\
  33         ADDL DI,E;\
  34         FN(B,C,D)\
  35         ADDL DI,E;\
  36         RORL $2,B;\
  37
  38 /*
  39  * fn1 = (((C^D)&B)^D);
  40  */
  41 #define FN1(B,C,D)\
  42         MOVL C,DI;\
  43         XORL D,DI;\
  44         ANDL B,DI;\
  45         XORL D,DI;\
  46
  47 /*
  48  * fn24 = B ^ C ^ D
  49  */
  50 #define FN24(B,C,D)\
  51         MOVL B,DI;\
  52         XORL C,DI;\
  53         XORL D,DI;\
  54
  55 /*
  56  * fn3 = ((B ^ C) & (D ^= B)) ^ B
  57  * D ^= B to restore D
  58  */
  59 #define FN3(B,C,D)\
  60         MOVL B,DI;\
  61         XORL C,DI;\
  62         XORL B,D;\
  63         ANDL D,DI;\
  64         XORL B,DI;\
  65         XORL B,D;\
  66
  67 /*
  68  * stack offsets
  69  * void sha1block(uchar *DATA, int LEN, ulong *STATE)
  70  */
  71 #define DATA    0
  72 #define LEN     8
  73 #define STATE   16
  74
  75 /*
  76  * stack offsets for locals
  77  * ulong w[80];
  78  * uchar *edata;
  79  * ulong *w15, *w40, *w60, *w80;
  80  * register local
  81  * ulong *wp = BP
  82  * ulong a = eax, b = ebx, c = ecx, d = edx, e = esi
  83  * ulong tmp = edi
  84  */
  85 #define Rpdata  R8
  86 #define WARRAY  (-8-(80*4))
  87 #define TMP1    (-16-(80*4))
  88 #define TMP2    (-24-(80*4))
  89 #define W15     (-32-(80*4))
  90 #define W40     (-40-(80*4))
  91 #define W60     (-48-(80*4))
  92 #define W80     (-56-(80*4))
  93 #define EDATA   (-64-(80*4))
  94
  95 TEXT    _sha1block+0(SB),$384
  96
  97         MOVQ RARG, Rpdata
  98         MOVLQZX len+LEN(FP),BX
  99         ADDQ BX, RARG
 100         MOVQ RARG,edata+EDATA(SP)
 101
 102         LEAQ aw15+(WARRAY+15*4)(SP),DI
 103         MOVQ DI,w15+W15(SP)
 104         LEAQ aw40+(WARRAY+40*4)(SP),DX
 105         MOVQ DX,w40+W40(SP)
 106         LEAQ aw60+(WARRAY+60*4)(SP),CX
 107         MOVQ CX,w60+W60(SP)
 108         LEAQ aw80+(WARRAY+80*4)(SP),DI
 109         MOVQ DI,w80+W80(SP)
 110
 111 mainloop:
 112         LEAQ warray+WARRAY(SP),BP
 113
 114         MOVQ state+STATE(FP),DI
 115         MOVL (DI),AX
 116         MOVL 4(DI),BX
 117         MOVL BX,tmp1+TMP1(SP)
 118         MOVL 8(DI),CX
 119         MOVL 12(DI),DX
 120         MOVL 16(DI),SI
 121
 122         MOVQ Rpdata,BX
 123
 124 loop1:
 125         BODY0(0,FN1,0x5a827999,AX,tmp1+TMP1(SP),CX,DX,SI)
 126         MOVL SI,tmp2+TMP2(SP)
 127         BODY0(4,FN1,0x5a827999,SI,AX,tmp1+TMP1(SP),CX,DX)
 128         MOVL tmp1+TMP1(SP),SI
 129         BODY0(8,FN1,0x5a827999,DX,tmp2+TMP2(SP),AX,SI,CX)
 130         BODY0(12,FN1,0x5a827999,CX,DX,tmp2+TMP2(SP),AX,SI)
 131         MOVL SI,tmp1+TMP1(SP)
 132         BODY0(16,FN1,0x5a827999,SI,CX,DX,tmp2+TMP2(SP),AX)
 133         MOVL tmp2+TMP2(SP),SI
 134
 135         ADDQ $20,BX
 136         ADDQ $20,BP
 137         CMPQ BP,w15+W15(SP)
 138         JCS loop1
 139
 140         BODY0(0,FN1,0x5a827999,AX,tmp1+TMP1(SP),CX,DX,SI)
 141         ADDQ $4,BX
 142         MOVQ BX,R8
 143         MOVQ tmp1+TMP1(SP),BX
 144
 145         BODY(4,FN1,0x5a827999,SI,AX,BX,CX,DX)
 146         BODY(8,FN1,0x5a827999,DX,SI,AX,BX,CX)
 147         BODY(12,FN1,0x5a827999,CX,DX,SI,AX,BX)
 148         BODY(16,FN1,0x5a827999,BX,CX,DX,SI,AX)
 149
 150         ADDQ $20,BP
 151
 152 loop2:
 153         BODY(0,FN24,0x6ed9eba1,AX,BX,CX,DX,SI)
 154         BODY(4,FN24,0x6ed9eba1,SI,AX,BX,CX,DX)
 155         BODY(8,FN24,0x6ed9eba1,DX,SI,AX,BX,CX)
 156         BODY(12,FN24,0x6ed9eba1,CX,DX,SI,AX,BX)
 157         BODY(16,FN24,0x6ed9eba1,BX,CX,DX,SI,AX)
 158
 159         ADDQ $20,BP
 160         CMPQ BP,w40+W40(SP)
 161         JCS loop2
 162
 163 loop3:
 164         BODY(0,FN3,0x8f1bbcdc,AX,BX,CX,DX,SI)
 165         BODY(4,FN3,0x8f1bbcdc,SI,AX,BX,CX,DX)
 166         BODY(8,FN3,0x8f1bbcdc,DX,SI,AX,BX,CX)
 167         BODY(12,FN3,0x8f1bbcdc,CX,DX,SI,AX,BX)
 168         BODY(16,FN3,0x8f1bbcdc,BX,CX,DX,SI,AX)
 169
 170         ADDQ $20,BP
 171         CMPQ BP,w60+W60(SP)
 172         JCS loop3
 173
 174 loop4:
 175         BODY(0,FN24,0xca62c1d6,AX,BX,CX,DX,SI)
 176         BODY(4,FN24,0xca62c1d6,SI,AX,BX,CX,DX)
 177         BODY(8,FN24,0xca62c1d6,DX,SI,AX,BX,CX)
 178         BODY(12,FN24,0xca62c1d6,CX,DX,SI,AX,BX)
 179         BODY(16,FN24,0xca62c1d6,BX,CX,DX,SI,AX)
 180
 181         ADDQ $20,BP
 182         CMPQ BP,w80+W80(SP)
 183         JCS loop4
 184
 185         MOVQ state+STATE(FP),DI
 186         ADDL AX,0(DI)
 187         ADDL BX,4(DI)
 188         ADDL CX,8(DI)
 189         ADDL DX,12(DI)
 190         ADDL SI,16(DI)
 191
 192         MOVQ edata+EDATA(SP),DI
 193         CMPQ Rpdata,DI
 194         JCS mainloop
 195
 196         RET
 197         END