skc: high memory usage during clang invocation causes OOM on 4GB CI

[mbouaziz]

Problem

When skc compiles a binary with -O2, it peaks at ~4.3GB total memory because it holds ~3.4GB RSS while spawning clang -cc1 -O2 which adds ~0.9GB on top. This causes OOM kills on CI environments with 4GB RAM (e.g. CircleCI medium).

Measured Memory Profile

Captured building the skjs binary (skargo build --release --bin skjs) inside skiplabs/skip:latest:

Time	Process	RSS
T+0s	skc starts	~200MB
T+10s	skc (compilation)	~1.8GB
T+15s	skc (peak, pre-clang)	~3.4GB
T+15s	skc spawns clang -cc1	skc: 3.4GB + clang: ~270MB
T+25s	PEAK	skc: 3.4GB + clang: ~940MB = ~4.3GB total
T+35s	done	released

Key observations

• skc holds 3.4GB RSS the entire time clang runs
• clang -cc1 grows to ~940MB during LLVM codegen at -O2
• The two processes overlap: skc doesn't free memory before clang finishes

Suggested Fix

Since skc's job is done once it writes the .ll file, it could:

1. exec() into clang — replace the skc process entirely, reducing peak to max(3.4GB, 0.9GB) = 3.4GB
2. Free compilation state before invoking clang
3. Fork/exec clang and exit — let clang run independently

Option 1 would be the most effective.

Reproduction

git clone https://github.com/SkipLabs/skjs
git clone --recurse-submodules https://github.com/SkipLabs/skip

# This reliably OOMs:
docker run -d --name skc-oom-test --memory=4g --memory-swap=4g \
  -v "$PWD/skjs:/work/skjs" -v "$PWD/skip:/work/skip" \
  -w /work/skjs skiplabs/skip:latest \
  skargo build --release

docker inspect skc-oom-test --format '{{.State.OOMKilled}}'  # true = OOM

Current Workaround

skjs CI uses skargo build --release --bin skjs instead of skargo build --release to avoid building the test binary (which causes a second 4GB spike). This brings peak from ~4.3GB to ~3.9GB.

Environment

• Docker image: skiplabs/skip:latest (skc 0.1.0, LLVM 20)
• CI: CircleCI medium executor (4GB RAM)