Use simd-lite

[Licenser]

Use simd-lite for ARM support.

[sunnygleason]

@Licenser it's awesome that this works! The developer opinions aren't a deal-breaker, I think we can fix that in future revisions anyway.

[sunnygleason]

@Licenser was just looking at the test failures, I think you might want to use this impl of vtstq. I did an integration earlier against your intrinsics work, and this is what I needed (which is why I was complaining about the private fields):

#[inline]
fn test_u8(a: u8, b: u8) -> u8 {
    if a & b != 0 {
        0xFF
    } else {
        0x00
    }
}

#[inline]
pub unsafe fn vtstq_u8(a: uint8x16_t, b: uint8x16_t) -> uint8x16_t {
    uint8x16_t(
        test_u8(a.0, b.0),
        test_u8(a.1, b.1),
        test_u8(a.2, b.2),
        test_u8(a.3, b.3),
        test_u8(a.4, b.4),
        test_u8(a.5, b.5),
        test_u8(a.6, b.6),
        test_u8(a.7, b.7),
        test_u8(a.8, b.8),
        test_u8(a.9, b.9),
        test_u8(a.10, b.10),
        test_u8(a.11, b.11),
        test_u8(a.12, b.12),
        test_u8(a.13, b.13),
        test_u8(a.14, b.14),
        test_u8(a.15, b.15),
    )
}

#[inline]
fn test_s8(a: i8, b: i8) -> i8 {
    if a & b != 0 {
        -1
    } else {
        0x00
    }
}

#[inline]
pub unsafe fn vtstq_s8(a: int8x16_t, b: int8x16_t) -> int8x16_t {
    int8x16_t(
        test_s8(a.0, b.0),
        test_s8(a.1, b.1),
        test_s8(a.2, b.2),
        test_s8(a.3, b.3),
        test_s8(a.4, b.4),
        test_s8(a.5, b.5),
        test_s8(a.6, b.6),
        test_s8(a.7, b.7),
        test_s8(a.8, b.8),
        test_s8(a.9, b.9),
        test_s8(a.10, b.10),
        test_s8(a.11, b.11),
        test_s8(a.12, b.12),
        test_s8(a.13, b.13),
        test_s8(a.14, b.14),
        test_s8(a.15, b.15),
    )
}

[Licenser]

/// Compare bitwise test bits nonzero
#[inline]
#[target_feature(enable = "neon")]
#[cfg_attr(target_arch = "arm", target_feature(enable = "v7"))]
#[cfg_attr(test, assert_instr(cmtst))]
// even gcc compiles this to ldr: https://clang.godbolt.org/z/1bvH2x
// #[cfg_attr(test, assert_instr(ld1))]
pub unsafe fn vtstq_u8(a: uint8x16_t, b: uint8x16_t) -> uint8x16_t {
    vcgtq_u8(vandq_u8(a, b), vdupq_n_u8(0))
}

Ja I think that should be the same as vcgtq_u8(vandq_u8(a, b), vdupq_n_u8(0))

vandq_u8(a, b) should be a & b (vectorized) and vcgtq_u8(..., vdupq_n_u8(0)) should be != 0 (or >= ) vectorized (I dind't find a not-equal intrinsic but perhaps I just missed it?)

sadly not even clang compiles vtstq_u8 to a own intrinsic: https://godbolt.org/z/lwHshr

[sunnygleason]

I'm crying 😂

[Licenser]

I know it's ugly :( but it's correcter then mem::transmute, it's the load intrinsic that's made to well ... populate a register.

[Licenser]

ideally we want an neon equivalent of _mm256_setr_epi8

[Licenser]

hmm I got an idea! We could use a trait to add a ::new() method how about that?

[sunnygleason]

hmm I got an idea! We could use a trait to add a ::new() method how about that?

Yes I love that idea! Only problem is can't add from outside the crate?

[Licenser]

If we control the trait we don't need to control the type :D and it compiles to a vld1 instruciton one way or the other so ¯_(ツ)_/¯ doens't hurt to do the as *const dance in there - we take the hit complexity but it's nice for the user.

[Licenser]

ahhh damn we can't have a trait with different arities for functions but we could pass in the slice, that's mildly better I think.

[sunnygleason]

crazily enough I think mem::transmute made the most sense 😂
right now, I'm up for whatever -- this is close enough

[sunnygleason]

interesting looks like they do compile down to the same: https://rust.godbolt.org/z/0ORCzh

[Licenser]

:D yay! I think the ::new is a lot nicer.

[sunnygleason]

sorry I left this on the wrong PR, oops!

@Licenser did you find the bug? I think I might have it:

8875cff#diff-e3cc34fc9e09211fc01983167aaf86d7R47

(signed versus unsigned subtraction in utf8check, we had an error in the previous intrinsics)

PS. ignore all the intrinsics changes, that's just me bisecting the problem

[Licenser]

Wow, well done! I think that did it!!! Once this all works we should give it a sweep to see if we can simplify some of those intriniscs but most importantly it seems to actually work :D good bye std::arch!

[sunnygleason]

@Licenser this looks good to me! Are there big changes you'd want to make before releasing the simd-lite crate,are we getting close?

[Licenser]

I really really would like to see if we can get rid of the copied code in simd-lite before releasing it wherever possible. And were not we need to make sure it's properly attributed and includes the respective copyrights.

[sunnygleason]

Merging, beautiful work @Licenser !