Use table.fill in externref-xform (TODO)

[lukaslihotzki]

No description provided.

[daxpedda]

First off: this is amazing, thank you!

As far as I know the reference-type proposal is not a requirements for the atomic proposal, so either we will have to adjust the documentation or gate it.

I would prefer to gate it, which should be possible.

Will pull in alexcrichton when we are ready to review this, I don't feel comfortable enough with Wasm's instruction set to approve this.

[lukaslihotzki]

I don't understand where gating makes sense. The table.set and table.fill instructions are both newly defined in the reference-types proposal, so I don't see any additional requirements. If externref_stack == 0, neither of those instructions is used anyway.

[daxpedda]

Ah, apologies! For some reason I thought we are in threads-xform 🤣.
@alexcrichton if you wouldn't mind taking a look at this.

[alexcrichton]

Yes I believe that this is correct but I'm also not sure whether it's worth doing. For small sizes it may be better to have the currently unrolled loop perf-wise and it may only be advantageous for larger sizes to use table.fill. I'm not entirely sure though in that I haven't benchmarked any wasm engine to see the difference.

[daxpedda]

I would have argued that loop unrolling is a compilers job when it's applying optimizations, on the other hand this is not something I really know much about.

@lukaslihotzki would you mind researching this a bit:

• Benchmark it.
• Find out if the LLVM or wasm-opt unrolls table.fill.
• See how it would make sense to proceed here in your opinion.

Thank you!