use zero-extension in the loongarch64 crc32 implementation

[folkertdev]

Using sign extension caused some issues in miri which actually checks the range of the input.

[folkertdev]

r? sayantn

[sayantn]

Should the signatures be signed or unsigned? Because typically crypto always uses unsigned

View changes since the review

[folkertdev]

Clang uses signed, see also discussion at #2137 (comment)

[sayantn]

I think you meant to use simd_test here instead of target_feature

View changes since the review

[folkertdev]

Should be good now, we can't run this on arm at the moment due to the feature detection macro not supporting a feature we need.

[heiher]

Would it make sense to handle this in Miri instead? Casting to an unsigned type isn't free, and crc.w.{b,h}.w only looks at the relevant low bits, ignoring the upper bits.

fn crc32_test(a: i8, b:i8) -> i32 {
    let r = crc_w_b_w(a, -1);
    crc_w_b_w(b, r)
}

b4800:   02fffc06    li.d            $a2, -1
b4804:   00241884    crc.w.b.w       $a0, $a0, $a2
b4808:   002410a4    crc.w.b.w       $a0, $a1, $a0
b480c:   4c000020    ret

->

b4800:   0343fc84    andi            $a0, $a0, 0xff
b4804:   02fffc06    li.d            $a2, -1
b4808:   00241884    crc.w.b.w       $a0, $a0, $a2
b480c:   0343fca5    andi            $a1, $a1, 0xff
b4810:   002410a4    crc.w.b.w       $a0, $a1, $a0
b4814:   4c000020    ret

[heiher]

Would it make sense to handle this in Miri instead?

rust-lang/miri#5102

[folkertdev]

I'd expect LLVM to optimize the mask out when the instruction that uses the result won't look at the upper bits anyway?