sha.cpp

mysql-5.0.33源码包

// ASMR3 = ASMR2 but f is (x&y)|(z&(x|y))
//               which is (w&x)|(y&(w|x))

#define ASMR3(v,w,x,y,z,i,W1,W2,W3,W4) \
    AS2(    mov   edi, [esp + W1 * 4]           )   \
    AS2(    mov   esi, x                        )   \
    AS2(    xor   edi, [esp + W2 * 4]           )   \
    AS2(     or   esi, w                        )   \
    AS2(    xor   edi, [esp + W3 * 4]           )   \
    AS2(    and   esi, y                        )   \
    AS2(    xor   edi, [esp + W4 * 4]           )   \
    AS2(    movd  mm0, esi                      )   \
    AS2(    rol   edi, 1                        )   \
    AS2(    mov   esi, x                        )   \
    AS2(    mov   [esp + W1 * 4], edi           )   \
    AS2(    and   esi, w                        )   \
    AS2(    lea     z, [edi + z + 0x8F1BBCDC]   )   \
    AS2(    movd  edi, mm0                      )   \
    AS2(     or   esi, edi                      )   \
    AS2(    mov   edi, v                        )   \
    AS2(    rol   edi, 5                        )   \
    AS2(    add     z, esi                      )   \
    AS2(    rol     w, 30                       )   \
    AS2(    add     z, edi                      )


// ASMR4 = ASMR2 but different constant

#define ASMR4(v,w,x,y,z,i,W1,W2,W3,W4) \
    AS2(    mov   edi, [esp + W1 * 4]           )   \
    AS2(    mov   esi, x                        )   \
    AS2(    xor   edi, [esp + W2 * 4]           )   \
    AS2(    xor   esi, y                        )   \
    AS2(    xor   edi, [esp + W3 * 4]           )   \
    AS2(    xor   esi, w                        )   \
    AS2(    xor   edi, [esp + W4 * 4]           )   \
    AS2(    rol   edi, 1                        )   \
    AS2(    add     z, esi                      )   \
    AS2(    mov   [esp + W1 * 4], edi           )   \
    AS2(    lea     z, [edi + z + 0xCA62C1D6]   )   \
    AS2(    mov   edi, v                        )   \
    AS2(    rol   edi, 5                        )   \
    AS2(    rol     w, 30                       )   \
    AS2(    add     z, edi                      )


#ifdef _MSC_VER
    __declspec(naked) 
#endif
void SHA::AsmTransform(const byte* data, word32 times)
{
#ifdef __GNUC__
    #define AS1(x)    asm(#x);
    #define AS2(x, y) asm(#x ", " #y);

    #define PROLOG()  \
        asm(".intel_syntax noprefix"); \
        AS2(    movd  mm3, edi                      )   \
        AS2(    movd  mm4, ebx                      )   \
        AS2(    movd  mm5, esi                      )   \
        AS2(    movd  mm6, ebp                      )   \
        AS2(    mov   ecx, DWORD PTR [ebp +  8]     )   \
        AS2(    mov   edi, DWORD PTR [ebp + 12]     )   \
        AS2(    mov   eax, DWORD PTR [ebp + 16]     )

    #define EPILOG()  \
        AS2(    movd  ebp, mm6                  )   \
        AS2(    movd  esi, mm5                  )   \
        AS2(    movd  ebx, mm4                  )   \
        AS2(    mov   esp, ebp                  )   \
        AS2(    movd  edi, mm3                  )   \
        AS1(    emms                            )   \
        asm(".att_syntax");
#else
    #define AS1(x)    __asm x
    #define AS2(x, y) __asm x, y

    #define PROLOG() \
        AS1(    push  ebp                           )   \
        AS2(    mov   ebp, esp                      )   \
        AS2(    movd  mm3, edi                      )   \
        AS2(    movd  mm4, ebx                      )   \
        AS2(    movd  mm5, esi                      )   \
        AS2(    movd  mm6, ebp                      )   \
        AS2(    mov   edi, data                     )   \
        AS2(    mov   eax, times                    )

    #define EPILOG() \
        AS2(    movd  ebp, mm6                  )   \
        AS2(    movd  esi, mm5                  )   \
        AS2(    movd  ebx, mm4                  )   \
        AS2(    movd  edi, mm3                  )   \
        AS2(    mov   esp, ebp                  )   \
        AS1(    pop   ebp                       )   \
        AS1(    emms   )                            \
        AS1(    ret 8  )   
#endif

    PROLOG()

    AS2(    mov   esi, ecx              )

    #ifdef OLD_GCC_OFFSET
        AS2(    add   esi, 20               )   // digest_[0]
    #else
        AS2(    add   esi, 16               )   // digest_[0]
    #endif

    AS2(    movd  mm2, eax              )   // store times_
    AS2(    movd  mm1, esi              )   // store digest_

    AS2(    sub   esp, 68               )   // make room on stack

AS1( loopStart: )

    // byte reverse 16 words of input, 4 at a time, put on stack for W[]

    // part 1
    AS2(    mov   eax, [edi]        )
    AS2(    mov   ebx, [edi +  4]   )
    AS2(    mov   ecx, [edi +  8]   )
    AS2(    mov   edx, [edi + 12]   )

    AS1(    bswap eax   )
    AS1(    bswap ebx   )
    AS1(    bswap ecx   )
    AS1(    bswap edx   )

    AS2(    mov   [esp],      eax   )
    AS2(    mov   [esp +  4], ebx   )
    AS2(    mov   [esp +  8], ecx   )
    AS2(    mov   [esp + 12], edx   )

    // part 2
    AS2(    mov   eax, [edi + 16]   )
    AS2(    mov   ebx, [edi + 20]   )
    AS2(    mov   ecx, [edi + 24]   )
    AS2(    mov   edx, [edi + 28]   )

    AS1(    bswap eax   )
    AS1(    bswap ebx   )
    AS1(    bswap ecx   )
    AS1(    bswap edx   )

    AS2(    mov   [esp + 16], eax   )
    AS2(    mov   [esp + 20], ebx   )
    AS2(    mov   [esp + 24], ecx   )
    AS2(    mov   [esp + 28], edx   )


    // part 3
    AS2(    mov   eax, [edi + 32]   )
    AS2(    mov   ebx, [edi + 36]   )
    AS2(    mov   ecx, [edi + 40]   )
    AS2(    mov   edx, [edi + 44]   )

    AS1(    bswap eax   )
    AS1(    bswap ebx   )
    AS1(    bswap ecx   )
    AS1(    bswap edx   )

    AS2(    mov   [esp + 32], eax   )
    AS2(    mov   [esp + 36], ebx   )
    AS2(    mov   [esp + 40], ecx   )
    AS2(    mov   [esp + 44], edx   )


    // part 4
    AS2(    mov   eax, [edi + 48]   )
    AS2(    mov   ebx, [edi + 52]   )
    AS2(    mov   ecx, [edi + 56]   )
    AS2(    mov   edx, [edi + 60]   )

    AS1(    bswap eax   )
    AS1(    bswap ebx   )
    AS1(    bswap ecx   )
    AS1(    bswap edx   )

    AS2(    mov   [esp + 48], eax   )
    AS2(    mov   [esp + 52], ebx   )
    AS2(    mov   [esp + 56], ecx   )
    AS2(    mov   [esp + 60], edx   )

    AS2(    mov   [esp + 64], edi   )   // store edi for end

    // read from digest_
    AS2(    mov   eax, [esi]            )   // a1
    AS2(    mov   ebx, [esi +  4]       )   // b1
    AS2(    mov   ecx, [esi +  8]       )   // c1
    AS2(    mov   edx, [esi + 12]       )   // d1
    AS2(    mov   ebp, [esi + 16]       )   // e1


    ASMR0(eax, ebx, ecx, edx, ebp,  0)
    ASMR0(ebp, eax, ebx, ecx, edx,  1)
    ASMR0(edx, ebp, eax, ebx, ecx,  2)
    ASMR0(ecx, edx, ebp, eax, ebx,  3)
    ASMR0(ebx, ecx, edx, ebp, eax,  4)
    ASMR0(eax, ebx, ecx, edx, ebp,  5)
    ASMR0(ebp, eax, ebx, ecx, edx,  6)
    ASMR0(edx, ebp, eax, ebx, ecx,  7)
    ASMR0(ecx, edx, ebp, eax, ebx,  8)
    ASMR0(ebx, ecx, edx, ebp, eax,  9)
    ASMR0(eax, ebx, ecx, edx, ebp, 10)
    ASMR0(ebp, eax, ebx, ecx, edx, 11)
    ASMR0(edx, ebp, eax, ebx, ecx, 12)
    ASMR0(ecx, edx, ebp, eax, ebx, 13)
    ASMR0(ebx, ecx, edx, ebp, eax, 14)
    ASMR0(eax, ebx, ecx, edx, ebp, 15)

    ASMR1(ebp, eax, ebx, ecx, edx, 16,  0,  2,  8, 13)
    ASMR1(edx, ebp, eax, ebx, ecx, 17,  1,  3,  9, 14)
    ASMR1(ecx, edx, ebp, eax, ebx, 18,  2,  4, 10, 15)
    ASMR1(ebx, ecx, edx, ebp, eax, 19,  3,  5, 11,  0)

    ASMR2(eax, ebx, ecx, edx, ebp, 20,  4,  6, 12,  1)
    ASMR2(ebp, eax, ebx, ecx, edx, 21,  5,  7, 13,  2)
    ASMR2(edx, ebp, eax, ebx, ecx, 22,  6,  8, 14,  3)
    ASMR2(ecx, edx, ebp, eax, ebx, 23,  7,  9, 15,  4)
    ASMR2(ebx, ecx, edx, ebp, eax, 24,  8, 10,  0,  5)
    ASMR2(eax, ebx, ecx, edx, ebp, 25,  9, 11,  1,  6)
    ASMR2(ebp, eax, ebx, ecx, edx, 26, 10, 12,  2,  7)
    ASMR2(edx, ebp, eax, ebx, ecx, 27, 11, 13,  3,  8)
    ASMR2(ecx, edx, ebp, eax, ebx, 28, 12, 14,  4,  9)
    ASMR2(ebx, ecx, edx, ebp, eax, 29, 13, 15,  5, 10)
    ASMR2(eax, ebx, ecx, edx, ebp, 30, 14,  0,  6, 11)
    ASMR2(ebp, eax, ebx, ecx, edx, 31, 15,  1,  7, 12)
    ASMR2(edx, ebp, eax, ebx, ecx, 32,  0,  2,  8, 13)
    ASMR2(ecx, edx, ebp, eax, ebx, 33,  1,  3,  9, 14)
    ASMR2(ebx, ecx, edx, ebp, eax, 34,  2,  4, 10, 15)
    ASMR2(eax, ebx, ecx, edx, ebp, 35,  3,  5, 11,  0)
    ASMR2(ebp, eax, ebx, ecx, edx, 36,  4,  6, 12,  1)
    ASMR2(edx, ebp, eax, ebx, ecx, 37,  5,  7, 13,  2)
    ASMR2(ecx, edx, ebp, eax, ebx, 38,  6,  8, 14,  3)
    ASMR2(ebx, ecx, edx, ebp, eax, 39,  7,  9, 15,  4)


    ASMR3(eax, ebx, ecx, edx, ebp, 40,  8, 10,  0,  5)
    ASMR3(ebp, eax, ebx, ecx, edx, 41,  9, 11,  1,  6)
    ASMR3(edx, ebp, eax, ebx, ecx, 42, 10, 12,  2,  7)
    ASMR3(ecx, edx, ebp, eax, ebx, 43, 11, 13,  3,  8)
    ASMR3(ebx, ecx, edx, ebp, eax, 44, 12, 14,  4,  9)
    ASMR3(eax, ebx, ecx, edx, ebp, 45, 13, 15,  5, 10)
    ASMR3(ebp, eax, ebx, ecx, edx, 46, 14,  0,  6, 11)
    ASMR3(edx, ebp, eax, ebx, ecx, 47, 15,  1,  7, 12)
    ASMR3(ecx, edx, ebp, eax, ebx, 48,  0,  2,  8, 13)
    ASMR3(ebx, ecx, edx, ebp, eax, 49,  1,  3,  9, 14)
    ASMR3(eax, ebx, ecx, edx, ebp, 50,  2,  4, 10, 15)
    ASMR3(ebp, eax, ebx, ecx, edx, 51,  3,  5, 11,  0)
    ASMR3(edx, ebp, eax, ebx, ecx, 52,  4,  6, 12,  1)
    ASMR3(ecx, edx, ebp, eax, ebx, 53,  5,  7, 13,  2)
    ASMR3(ebx, ecx, edx, ebp, eax, 54,  6,  8, 14,  3)
    ASMR3(eax, ebx, ecx, edx, ebp, 55,  7,  9, 15,  4)
    ASMR3(ebp, eax, ebx, ecx, edx, 56,  8, 10,  0,  5)
    ASMR3(edx, ebp, eax, ebx, ecx, 57,  9, 11,  1,  6)
    ASMR3(ecx, edx, ebp, eax, ebx, 58, 10, 12,  2,  7)
    ASMR3(ebx, ecx, edx, ebp, eax, 59, 11, 13,  3,  8)

    ASMR4(eax, ebx, ecx, edx, ebp, 60, 12, 14,  4,  9)
    ASMR4(ebp, eax, ebx, ecx, edx, 61, 13, 15,  5, 10)
    ASMR4(edx, ebp, eax, ebx, ecx, 62, 14,  0,  6, 11)
    ASMR4(ecx, edx, ebp, eax, ebx, 63, 15,  1,  7, 12)
    ASMR4(ebx, ecx, edx, ebp, eax, 64,  0,  2,  8, 13)
    ASMR4(eax, ebx, ecx, edx, ebp, 65,  1,  3,  9, 14)
    ASMR4(ebp, eax, ebx, ecx, edx, 66,  2,  4, 10, 15)
    ASMR4(edx, ebp, eax, ebx, ecx, 67,  3,  5, 11,  0)
    ASMR4(ecx, edx, ebp, eax, ebx, 68,  4,  6, 12,  1)
    ASMR4(ebx, ecx, edx, ebp, eax, 69,  5,  7, 13,  2)
    ASMR4(eax, ebx, ecx, edx, ebp, 70,  6,  8, 14,  3)
    ASMR4(ebp, eax, ebx, ecx, edx, 71,  7,  9, 15,  4)
    ASMR4(edx, ebp, eax, ebx, ecx, 72,  8, 10,  0,  5)
    ASMR4(ecx, edx, ebp, eax, ebx, 73,  9, 11,  1,  6)
    ASMR4(ebx, ecx, edx, ebp, eax, 74, 10, 12,  2,  7)
    ASMR4(eax, ebx, ecx, edx, ebp, 75, 11, 13,  3,  8)
    ASMR4(ebp, eax, ebx, ecx, edx, 76, 12, 14,  4,  9)
    ASMR4(edx, ebp, eax, ebx, ecx, 77, 13, 15,  5, 10)
    ASMR4(ecx, edx, ebp, eax, ebx, 78, 14,  0,  6, 11)
    ASMR4(ebx, ecx, edx, ebp, eax, 79, 15,  1,  7, 12)


    AS2(    movd  esi, mm1              )   // digest_

    AS2(    add   [esi],      eax       )   // write out
    AS2(    add   [esi +  4], ebx       )
    AS2(    add   [esi +  8], ecx       )
    AS2(    add   [esi + 12], edx       )
    AS2(    add   [esi + 16], ebp       )

    // setup next round
    AS2(    movd  ebp, mm2              )   // times
 
    AS2(    mov   edi, DWORD PTR [esp + 64] )   // data
    
    AS2(    add   edi, 64               )   // next round of data
    AS2(    mov   [esp + 64], edi       )   // restore
    
    AS1(    dec   ebp                   )
    AS2(    movd  mm2, ebp              )
    AS1(    jnz   loopStart             )


    EPILOG()
}


#endif // DO_SHA_ASM

} // namespace