wasm-opt crashes with a bus error on a 91 kB Wasm file

[samestep]

Issue

I've put my input file prospero-opt-grad.wat in a GitHub Gist here: https://gist.github.com/samestep/ca2333b48a0aac67dab4bba4e51be92d

I'm on a 2020 MacBook Pro with M1 chip. I downloaded binaryen-version_122-arm64-macos.tar.gz from GitHub Releases, and attempted to run it on this Wasm file as follows:

wasm-opt prospero-opt-grad.wat -O3 -o prospero-opt-grad-opt.wasm

This ran for 9 seconds and then crashed with the following error output:

zsh: bus error  wasm-opt prospero-opt-grad.wat -O3 -o prospero-opt-grad-opt.wasm

Context

This came up in samestep/floretta#7 while I was experimenting with automatic differentiation of Wasm for an example computation graph with 8k nodes to see how well it would work on a single large function without control flow.

In particular, I started with prospero.wat which I generated from prospero.vm using the code in that PR I linked, then ran it through wasm-opt like this:

wasm-opt prospero.wat -O3 -o prospero-opt.wasm

That worked fine. Then I took prospero-opt.wasm and ran it through my Floretta tool at commit 919c361, like this:

cargo run prospero-opt.wasm --reverse --export main backprop --wat -o prospero-opt-grad.wat

That also worked fine. Finally I ran prospero-opt-grad.wat through wasm-opt, which fails as detailed above. Also, to clarify, the "91 kB" number in this issue title comes from running this command:

wasm-tools parse prospero-opt-grad.wat -o prospero-opt-grad.wasm && wc -c prospero-opt-grad.wasm

The same wasm-opt error still happens when running on prospero-opt-grad.wasm instead of prospero-opt-grad.wat.

[kripken]

Adding -all fixes the error. This module does need features enabled due to use of at least multivalue.

The crash looks like an OOM - we accumulate all errors before printing, and it seems like there are just so many we run out of memory. I didn't check specifically why, but it might be the incredibly-deeply nested call $f32_mins - we may be trying to print out each nesting, leading to quadratic time.

We do need to accumulate errors before printing, for determinism, but perhaps we can stop after a fixed number, to avoid this...