x86_64 adcx/adox

bignum.cc

#include <cstdint>
#include <array>
#include <utility>
#include <x86intrin.h>

using Big = std::array<uint64_t, 4>;
/*
unsigned char sum(Big & __restrict__ a, const Big & __restrict__ b) {
    unsigned char c = 0;
    for (std::size_t i = a.size(); i > 0; --i) {
        c = _addcarry_u64(c, a[i - 1], b[i - 1], (unsigned long long*)&a[i - 1]);
    }
    return c;
}
*/
unsigned char sum(Big & __restrict__ a, const Big & __restrict__ b) {
    unsigned char c;
    uint64_t t;
    uint64_t *pa = &a[0];
    const uint64_t *pb = &b[0];
    uint64_t count = a.size();
    asm("   xor %1, %1 \n"
        " 1: \n"
        "   mov     -8(%4, %%rcx, 8), %0 \n"
        "   adc     -8(%5, %%rcx, 8), %0 \n"
        "   mov     %0, -8(%4, %%rcx, 8) \n"
        "   loop 1b \n"
        "   setc %1"
         : "=&r" (t), "=&r" (c), "+c" (count), "+m" (*(uint64_t(*)[a.size()])pa)
         : "r" (pa), "r" (pb), "m" (*(uint64_t(*)[b.size()])pb)
         : "cc"
    );
    return c;
}

std::pair<unsigned char, unsigned char> sum2(Big & __restrict__ a, const Big & __restrict__ b, Big & __restrict__ c, const Big & __restrict__ d) {
    unsigned char c0, c1;
    uint64_t t;
    uint64_t *pa = &a[0];
    const uint64_t *pb = &b[0];
    uint64_t *pc = &c[0];
    const uint64_t *pd = &d[0];
    uint64_t count = a.size();
    asm("   xor %1, %1 \n"
        "   xor %2, %2 \n"
        " 1: \n"
        "   mov     -8(%6, %%rcx, 8), %0 \n"
        "   adcx    -8(%7, %%rcx, 8), %0 \n"
        "   mov     %0, -8(%6, %%rcx, 8) \n"
        "   mov     -8(%8, %%rcx, 8), %0 \n"
        "   adox    -8(%9, %%rcx, 8), %0 \n"
        "   mov     %0, -8(%8, %%rcx, 8) \n"
        "   loop 1b \n"
        "   setc %1 \n"
        "   seto %2"
         : "=&r" (t), "=&r" (c0), "=&r" (c1), "+c" (count), "+m" (*(uint64_t(*)[a.size()])pa), "+m" (*(uint64_t(*)[c.size()])pc)
         : "r" (pa), "r" (pb), "r" (pc), "r" (pd), "m" (*(uint64_t(*)[b.size()])pb), "m" (*(uint64_t(*)[d.size()])pd)
         : "cc"
    );
    return std::make_pair(c0, c1);
}

#include <cstdio>
int main() {
    Big a{~0LLU,~0LLU,~0LLU,~0LLU};
    Big b{2,3,4,6};
    Big c{~0LLU,~0LLU,~0LLU,~0LLU};
    Big d{4,6,8,12};
    // unsigned char c0 = sum(a, b);
    // unsigned char c1 = sum(c, d);
    auto [c0, c1] = sum2(a, b, c, d);
    //auto [c2, c3] = sum2(a, b, c, d);
    printf("%x ", c0);
    for(std::size_t i = 0; i < a.size(); ++i) {
        printf("%016lx ", a[i]);
    }
    printf("\n");
    printf("%x ", c1);
    for(std::size_t i = 0; i < c.size(); ++i) {
        printf("%016lx ", c[i]);
    }
    printf("\n");
}