Evaluate: Performance on other architectures

[alexcrichton]

This proposal should be evaluated on architectures other than x86_64 and numbers/methodology should be recorded in this repository's proposal documentation. This should at least include aarch64 and should ideally include riscv64 and a 32-bit architecture too.

[alexcrichton]

Using https://github.com/alexcrichton/wasm-benchmark-i128 as a source of benchmarks I've analyzed a prototype of i64.{add,sub,mul}128. Numbers here are slowdowns-relative-to-native to lower is better. Negative numbers mean that the wasm is faster than native (take that with a grain of salt these are very-micro micro-benchmarks). This was evaluated with Wasmtime 25 and using Rust nightly-2024-08-20 to compile the wasm module (roughly). The x86_64 machine was Linux with Intel(R) Core(TM) i9-14900K and the aarch64 machine is a macOS M2 pro.

benchmark	x86_64 before	x86_64 after	aarch64 before	aarch64 after
blind-sig	616%	76%	325%	78%
lehmer	362%	13%	111%	-28%
mul-bignum	386%	28%	187%	10%
fib_10000	120%	9%	72%	-2%
shl	94%	94%	35%	30%
shr	101%	102%	35%	35%
div-bignum	353%	168%	218%	111%
sort-signed	81%	95%	79%	79%
sort-unsigned	100%	95%	77%	78%

This to me shows that aarch64 is, in cases where overflow-related instructions can be used, going to see similar performance improvements as x86_64. I unfortunately do not personally have other hardware to test on at this time.