[Github] Add initial version of precommit checks

[boomanaiden154]

This patch adds an initial version of the LLVM precommit checks. These checks should be reasonably performant (full test cycle in ~30 minutes with a fully warm cache, ~80 minutes with a completely cold cache) and should catch regressions within LLVM.

This is mainly intended to test the scalability of the current design and to start eliminating issues before we begin to scale to other subprojects.

[llvmbot]

@llvm/pr-subscribers-github-workflow

Author: Aiden Grossman (boomanaiden154)

Changes

This patch adds an initial version of the LLVM precommit checks. These checks should be reasonably performant (full test cycle in ~30 minutes with a fully warm cache, ~80 minutes with a completely cold cache) and should catch regressions within LLVM.

This is mainly intended to test the scalability of the current design and to start eliminating issues before we begin to scale to other subprojects.

---

Full diff: https://github.com/llvm/llvm-project/pull/80951.diff

1 Files Affected:

• (added) .github/workflows/precommit.yaml (+47)

diff --git a/.github/workflows/precommit.yaml b/.github/workflows/precommit.yaml
new file mode 100644
index 0000000000000..8285a19b684f3
--- /dev/null
+++ b/.github/workflows/precommit.yaml
@@ -0,0 +1,47 @@
+name: "Precommit tests"
+
+permissions:
+  contents: read
+
+on:
+  pull_request:
+    branches:
+      - main
+    paths:
+      - '.github/workflows/precommit.yaml'
+      - 'llvm/**'
+
+jobs:
+  build-llvm:
+    name: "Build and test LLVM"
+    runs-on: ubuntu-latest
+    container:
+      image: ghcr.io/llvm/ci-ubuntu-22.04:latest
+    steps:
+      - name: Fetch LLVM sources
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
+      - name: Setup ccache
+        uses: hendrikmuhs/ccache-action@v1
+        with:
+          max-size: 500M
+          variant: sccache
+          key: precommit-linux
+      - name: Configure LLVM
+        run: |
+          cmake -B llvm-build -GNinja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER_LAUNCHER=sccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
+            -DLLVM_ENABLE_ASSERTIONS=ON \
+            -DCMAKE_C_COMPILER=clang \
+            -DCMAKE_CXX_COMPILER=clang++ \
+            -DLLVM_LIT_ARGS="-v --no-progress-bar" \
+            ./llvm
+      - name: Build LLVM
+        run: |
+          ninja -C llvm-build llvm-test-depends
+      - name: Check LLVM
+        run: |
+          ninja -C llvm-build check-llvm

[boomanaiden154]

I think this is in a workable enough state at this point given the recent container work. After this (and fixing any following issues), I plan on working on the following:

• Enabling more subprojects using logic from the existing premerge checks to select what to build to test (as per the RFC discussion).
• Introducing a windows docker container/workflow to get testing coverage back there in hopefully a reasonably performant manner, potentially only running after Linux tests have passed to reduce job load.
• Potentially moving to label-triggered tests so we could potentially make this required at some point and could enable opt-in tests.

Probably a couple other misc things in there. This will help us figure out system scalability however and get things rolling in those directions.

[tstellar]

Before this gets committed, we need to make sure we get consensus on the RFC thread. I'm not sure we are there yet.

[tstellar]

Do you think we'll be able to use the same file for Windows testing? If not, I would call this "Linux Precommit Tests" and change the filename too.

[boomanaiden154]

I'm planning on a separate file per OS. Github doesn't have native runs-on support for containers on Windows, so we'll have to run the docker commands explicitly. The container image would need to be different either way, which means we would at least need separate jobs. I'm hoping to make sure common parts are pulled out into something common, but the OSes are different (at least as exposed to GHA for the way I want to use it).