chore: bump spin version to 1.5.0

[jprendes]

Bump spin 1.4.1 -> 1.5.0.

[jprendes]

@cpuguy83 @devigned PTAL

[Mossaka]

lgtm! But it looks like the CI failed

[jprendes]

IIUC, there's a problem building a dependency introduced by the new LLM feature of spin.

[Mossaka]

Is it becuase the spin image is still using spin SDK v1.4.1?

Nvm, I realized that it is the shim build that failed.

[jprendes]

Ok, I digged a bit into it.
What's failing is building llama-cpp in aarch64 musl.
It builds fine with gnu ABI adding the -Ctarget-feature=+fp16 (what upstream spin does).
But it fails with musl ABI.
There are a few diferences, including the gcc version that cross uses for those targets.

The main issue is that llama-cpp is using the vld1q_u8_x4 and vld1q_s8_x4 functions, which should be SIMD intrinsics exposed by the compiler.

According to this issue xmrig/xmrig#2882, that intrinsic is not present in gcc <= 11.

I am wondering what the right fix is, since spin doesn't provide a feature flag to disable llm.

[jprendes]

Ok, the solution I found is less than idea, and it's polyfilling the missing functions in the cross container.
I think this is an issue to be fixed upstream, this is just an ugly workaround.
I created fermyon/spin#1786

[Mossaka]

lgtm!

minor comment: add a readme on polyfill folder to describe what this does and when shall we remove it.

This can be done in a future PR.

[Mossaka]

Need a rebase

[jprendes]

minor comment: add a readme on polyfill folder to describe what this does and when shall we remove it.

Done.

If there's an update to the compiler in cross-rs, the polyfill will become a no-op as it does a feature test before acting.
I left a link to the upstream issue in spin.
There are many different solutions, off the top of my head I can think of:

• a feature gate in spin
• a compiler update in cross
• replacing cross with something like xx, similar to what we do in runwasi (although cross is really nice, but would also allow us to build with systemd and seccomp support here)

[Mossaka]

Could you please sign your commit @jprendes ?

[jprendes]

sorry, changed laptops and hadn't set up git correctly!
it's signed now :-)