Evaluate: Peformance in various WebAssembly engines

[alexcrichton]

For the initial proposal an implementation was made in Wasmtime and Cranelift. The primary motivation of this proposal is performance which means that other engines must also be evaluated as well. Other engines should be evaluted both before/after this proposal to see if the same performance gap that was exposed by Wasmtime is additionally exposed in other engines as well. This will help motivate various instructions as "broadly applicable to all engines" vs "just this one engine needs to improve".

[alexcrichton]

I've collated some benchmarks at https://github.com/alexcrichton/wasm-benchmark-i128 and here's some results of WebAssembly as-is today with no modifications. All results here are reported as percent slowdown relative to native, so smaller is better.

The x86_64 computer here is Intel(R) Core(TM) i9-14900K and the aarch64 computer is a macOS M2 Pro.

engine	blind-sig	lehmer	mul-bignum	fib_10000	shl	shr	div-bignum
Node 22.6.0 x86_64	556%	250%	337%	138%	24%	66%	342%
Wasmtime 23.0.2 x86_64	638%	357%	388%	129%	8%	22%	388%
Node 22.6.0 aarch64	486%	105%	233%	82%	-4%	1%	233%
Wasmtime 23.0.2 aarch64	328%	113%	196%	80%	9%	8%	222%
Firefox 129.0 aarch64	446%	110%	232%	83%	15%	18%	255%

I don't personally have other hardware to test on at this time. Testing in the spidermonkey shell or d8 will require some more work in terms of setup scripts and integration so I also haven't done that yet.

I'll also note that this is just an initial set of benchmarks, and the set of benchmarks should probably grow over time.

[alexcrichton]

@jakobkummerow protoyped this proposal in V8 and got these numbers, which are quite similar to the improvements I saw in Wasmtime.