Update relaxed simd instructions

[yurydelendik]

Changes wasm-tools to implement current https://github.com/WebAssembly/relaxed-simd/blob/main/proposals/relaxed-simd/Overview.md

• Update instructions names and codes
• Add new instructions i16x8.relaxed_q15mulr_s, i16x8.dot_i8x16_i7x16_s, i32x4.dot_i8x16_i7x16_add_s, and f32x4.relaxed_dot_bf16x8_add_f32x4

[alexcrichton]

Thanks for this!

Out of curiosity, I believe your main motivation for this is updating the text parser for the usage in Firefox, is that right? If so, do you find it onerous/difficult/annoying to update all the other crates in the workspace? I find we have a lot here so updating little things can be a surprising amount of effort and I'm always hoping we can find more central locations to define all of this. Recently I'm wondering if we should try to deduplicate all the instruction encodings/names into one place to avoid having to name them in so many places. Anyway just curious, always happy to merge these PRs!

[alexcrichton]

Oh also, would you like a publish?

[yurydelendik]

If so, do you find it onerous/difficult/annoying to update all the other crates in the workspace?
Recently I'm wondering if we should try to deduplicate all the instruction encodings/names into one place to avoid having to name them in so many places.

It is not a big deal to update in multiple place, as long as testing will check typos and discrepancies in implementations.

Oh also, would you like a publish?

Yes, at your earliest convenience. I anticipate a reasonable delay up to 2 weeks between version release.

[alexcrichton]

@yurydelendik I'm not following the the relaxed-simd spec much so you may know the answer to this. With WebAssembly/testsuite#62 I was testing out updating the wasm-tools submodule to the testsuite to include relaxed-simd to ensure that all tests pass. Currently the only error is:

failed test: tests/testsuite/proposals/relaxed-simd/relaxed_fma_fms.wast

Caused by:
    unknown operator or unexpected token
         --> tests/testsuite/proposals/relaxed-simd/relaxed_fma_fms.wast:5:78
          |
        5 |     (func (export "f32x4.relaxed_fms") (param v128 v128 v128) (result v128) (f32x4.relaxed_fms (local.get 0) (local.get 1) (local.get 2)))
          |                                                                              ^

It looks like fms operations were reimplemented as fnma but there aren't any spec tests for fnma, so I'm wondering if you know what the current state of this is and how to reconcile this.

[alexcrichton]

Also I'm happy to publish whenever, I'm just waiting for a day with no fuzz-bugs to publish. (should be today, checking on oss-fuzz now)

[yurydelendik]

It looks like fms operations were reimplemented as fnma but there aren't any spec tests for fnma

The difference between f32x4.relaxed_fms and f32x4.relaxed_fnma is operands order. The existing test will be changed to use third operand as an accumulator (in fms and fma it was the first one). The old f32x4.relaxed_fma test will be failing as well. This change was made to keep signature similar to dot instructions.

Also I'm happy to publish whenever, I'm just waiting for a day with no fuzz-bugs to publish.

No problem. It is sort of a breaking change, and probably has not good tests. Disabling tests for fma/fms can be an option.

[alexcrichton]

Ok sounds good, I'll handle that accordingly and publish these changes today.