Higher memory use in CI with 1.2.0 as base compiler

[straight-shoota]

I have created a branch to prepare master for 1.3-dev (https://github.com/crystal-lang/crystal/tree/ci/1.3-dev) including updates to 1.2.0 as base compiler for CI.
But there are errors in CI: https://app.circleci.com/pipelines/github/crystal-lang/crystal/7044/workflows/0dffd249-0e5a-44dd-a6f1-38476f2022b8
Looks like LLVM runs out of memory. It's inconsistent (which is not unlikely for such a failure). Do we need more memory on our CircleCI runners?

[beta-ziliani]

Does that mean 1.2.0 is consuming more memory? What could be the reason for that?

[straight-shoota]

For all I know, 1.2.0 should maybe even use less memory due to #11167 and #11188.

[asterite]

There's more code and more specs, so LLVM will need more memory for that.

[beta-ziliani]

We're comparing the same set of specs, what changes is the base compiler, so it's about memory usage of the compiled specs with 1.1 vs 1.2

[asterite]

Ah, you are right. But there has been many changes to the compiler, right? It's hard to track which of these introduces more memory allocations, either on the semantic pass or the codegen pass.

[straight-shoota]

We'll have to look into that.

As an intermediary solution until we know what's causing the memory issues, we can just use a CI runner with more memory to get in business again.

[beta-ziliani]

Yes, the point is knowing if there's something we should worry about. If our programs consume a bit more memory that's OK, but if they consume a lot more, then we should fix it.

[straight-shoota]

I've added configuration for resource_class: large (with 8 GB memory instead of 6 GB) to #11317
CI workflow run: https://app.circleci.com/pipelines/github/crystal-lang/crystal/7058/workflows/8c38d619-9477-497e-b35f-d86a471f6f0d (Update: all green 🟢).

Related: crystal-lang/distribution-scripts#140