Proposal to add integer abs()

[jan-wassenberg]

abs() for integers would be helpful for JPEG XL - we predict pixels from their neighbours and compute the absolute error.

On x86 and ARM, INT32_MIN etc. are unchanged - negating them leads to the same result.

Without this operation, I think we could do u32x4_min(x, neg(x)), where neg(x) is either ~x + 1 or _mm_sign_epi32(x, something_negative), or whatever wasm_i32x4_neg uses, so about 3 instructions. By contrast, abs() can be a single instruction on x86 (_mm_abs_epi8..32) and ARM (vabsq_s8..32).

Has this been considered/discussed already?

[Maratyszcza]

There is PR #128 that propose integer absolute value instructions.

[jan-wassenberg]

Oh, nice! Sorry about the duplication, I only checked for other issues, not PRs.

[dtig]

Closing as duplicate of #128 which is now merged.