Fix family of _mmX_alignr_epiX functions

[marxin]

The function is supposed to return the first argument for IMM8 == 8.

Can be verified by both Clang and GCC for the following C++ code snippet:

#include <immintrin.h>
#include <stdint.h>

#include <iostream>

template<class T> inline void log(const __m256i& value)
{
    const size_t n = sizeof(__m256i) / sizeof(T);
    T buffer[n];
    _mm256_storeu_si256((__m256i*)buffer, value);
    for (int i = 0; i < n; i++)
        std::cout << std::hex << buffer[i] << " ";
    std::cout << std::endl;
}

int main()
{
    const auto a = _mm256_setr_epi8(
        1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
        25, 26, 27, 28, 29, 30, 31, 32
        );
    const auto b = _mm256_setr_epi8(
        -1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16, -17, -18, -19,
        -20, -21, -22, -23, -24, -25, -26, -27, -28, -29, -30, -31, -32
        );
    std::cout << "a: ";
    log<uint64_t>(a); 
    std::cout << "b: ";
    log<uint64_t>(b);
    const auto r = _mm256_alignr_epi8(a, b, 16);
    std::cout << "_mm256_alignr_epi8(a, b, 16):\n   ";
    log<uint64_t>(r);

    return 0;
}

✦ ❯ g++ stdarch.cc -march=native && ./a.out 
a: 807060504030201 100f0e0d0c0b0a09 1817161514131211 201f1e1d1c1b1a19 
b: f8f9fafbfcfdfeff f0f1f2f3f4f5f6f7 e8e9eaebecedeeef e0e1e2e3e4e5e6e7 
_mm256_alignr_epi8(a, b, 16):
   807060504030201 100f0e0d0c0b0a09 1817161514131211 201f1e1d1c1b1a19 

✦ ❯ clang++ stdarch.cc -march=native && ./a.out 
a: 807060504030201 100f0e0d0c0b0a09 1817161514131211 201f1e1d1c1b1a19 
b: f8f9fafbfcfdfeff f0f1f2f3f4f5f6f7 e8e9eaebecedeeef e0e1e2e3e4e5e6e7 
_mm256_alignr_epi8(a, b, 16):
   807060504030201 100f0e0d0c0b0a09 1817161514131211 201f1e1d1c1b1a19 

[rustbot]

r? @Amanieu

rustbot has assigned @Amanieu.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[marxin]

CC: @alexcrichton

[Amanieu]

Could you check all the other alignr intrinsics to see if they have the same issue?

[marxin]

Oh, a good point. Yes, there is at least one more function that behaves also wrongly. I will get to it over the weekend.

[marxin]

All right, so the situation is more complicated, and for the purpose of comparison (in between GCC/Clang C++ and Rust stdarch), I wrote the following script that utilizes all the non-mask functions from the alignr family with all intermediate masks up to 256.

gen.py.txt

I noticed the following issues:

• _mm256_alignr_epi8, _mm512_alignr_epi8 - bad handling of IMM8 == 16 and 32 (that was my original motivation for this PR)
• _mm_alignr_epi32, _mm256_alignr_epi32, _mm_alignr_epi64, _mm256_alignr_epi64 - these intrinsic have a limited number of bits that are used from the mask (please see the Intel Intrinsics Guide).
• mm512_alignr_epi32, mm512_alignr_epi64 - also have a limit on the number of bits, but it's enough (no code change needed) + docs update mentions the maximum shift
• _mm512_alignr_epi8 is a bit weird (as opposed to _mm512_alignr_epi32, _mm512_alignr_epi64): it does not concatenate the 2 512-bit vectors first, so I updated the docs