[sse4a] implement all non-immediate-mode intrinsics

[gnzlbg]

No description provided.

[gnzlbg]

Does somebody have access to an AMD machine and can check whether the tests pass there (using --nocapture to verify that sse4a is detected to be enabled)?

[alexcrichton]

I don't personally have access but they look good to me, thanks!

[gnzlbg]

We should figure out a way to test AMD intrinsics.

…

On Fri 22. Dec 2017 at 17:14, Alex Crichton ***@***.***> wrote: Merged #249 <#249>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#249 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA3Npgm7fGshK-n3pqmBxW-rBgY81QuEks5tC9VygaJpZM4RLIYH> .

[goodmanjonathan]

I get assertion failures for the _mm_stream_ss and _mm_stream_sd tests:

running 1725 tests
test runtime::macros::tests::test_macros ... ok
test runtime::x86::tests::compare_with_cupid ... ok
test runtime::x86::tests::dump ... sse: true
sse2: true
sse3: true
ssse3: false
sse4.1: false
sse4.2: false
sse4a: true

<==== snip ====>

test x86::i686::sse4a::assert__mm_extract_si64_extrq ... ok
test x86::i686::sse4a::assert__mm_insert_si64_insertq ... ok
test x86::i686::sse4a::assert__mm_stream_sd_movntsd ... ok
test x86::i686::sse4a::assert__mm_stream_ss_movntss ... ok
test x86::i686::sse4a::tests::_mm_extract_si64 ... ok
test x86::i686::sse4a::tests::_mm_insert_si64 ... ok
test x86::i686::sse4a::tests::_mm_stream_sd ... thread 'x86::i686::sse4a::tests::_mm_stream_sd' panicked at 'assertion failed: (left == right)
left: 3.0,
right: 4.0', src/x86/i686/sse4a.rs:128:8
note: Run with RUST_BACKTRACE=1 for a backtrace.
FAILED
test x86::i686::sse4a::tests::_mm_stream_ss ... thread 'x86::i686::sse4a::tests::_mm_stream_ss' panicked at 'assertion failed: (left == right)
left: 5.0,
right: 8.0', src/x86/i686/sse4a.rs:150:8
FAILED

<==== snip ====>

If I'm correct, it looks like the first f32/f64 in a f32x4/f64x2 is the least significant f32/f64, not the last f32/f64, as confirmed by this equivalent C code:

#include <assert.h>
#include <x86intrin.h>

__attribute__((target("sse4a"))) void f(void) {
    double d[] = { 1.0, 2.0 };
    __m128d v1 = { 3.0, 4.0 };
    _mm_stream_sd(d, v1);
    assert(d[0] == 3.0); // not 4.0
    assert(d[1] == 2.0);

    float f[] = { 1.0, 2.0, 3.0, 4.0 };
    __m128 v2 = { 5.0, 6.0, 7.0, 8.0 };
    _mm_stream_ss(f, v2);
    assert(f[0] == 5.0); // not 8.0
    assert(f[1] == 2.0);
    assert(f[2] == 3.0);
    assert(f[3] == 4.0);
}

int main(void) {
    f();
}

[gnzlbg]

@goodmanjonathan thanks for looking into this, are you willing to submit a patch that fixes the tests and the docs of the intrinsics?

[goodmanjonathan]

I'd be happy to!