Skip to content

[mlir][Sol][Yul] Canonicalize the Yul module before its lowering #910

[mlir][Sol][Yul] Canonicalize the Yul module before its lowering

[mlir][Sol][Yul] Canonicalize the Yul module before its lowering #910

name: Regression Tests
on:
pull_request:
# The shared build-llvm action only saves ccache on non-pull_request events:
# push-to-main runs warm the cache that PR runs restore, and double as
# post-merge regression signal.
push:
branches: [main]
schedule:
# Keepalive: GitHub evicts caches untouched for 7 days; a weekly warmup
# covers quiet weeks so PR builds don't go cold.
- cron: '0 4 * * 1'
workflow_dispatch:
inputs:
commits-to-test:
description: "Number of commits to test"
required: true
default: '1'
# For more information about the supported sanitizers in LLVM, see `LLVM_USE_SANITIZER` option in:
# https://www.llvm.org/docs/CMake.html
llvm-sanitizer:
required: false
default: 'Address'
type: string
description: 'A sanitizer to build LLVM with. Possible values are Address, Memory, MemoryWithOrigins, Undefined, Thread, DataFlow, and Address;Undefined'
enable-valgrind:
required: false
default: 'false'
type: string
description: 'Run LLVM regression tests under valgrind.'
defaults:
run:
shell: bash
permissions:
contents: read
packages: read # pull the private ghcr.io/nomicfoundation/solx-ci-runner image
concurrency:
group: ${{ github.repository_id }}-${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
# Never cancel main-branch runs: their ccache save at the end of the job is
# what keeps PR builds warm.
cancel-in-progress: ${{ github.event_name == 'pull_request' }}
jobs:
target-machine:
runs-on: ubuntu-26.04
outputs:
default: ${{ steps.default.outputs.machine }}
steps:
- name: Check for default target
id: default
shell: bash -ex {0}
# steps.*.outputs.* is this job's own step outputs (extensibility hook),
# not external input.
run: | # zizmor: ignore[template-injection]
if [[ "${{ join(steps.*.outputs.*) }}" == "" ]]; then
echo "machine=default" | tee -a "${GITHUB_OUTPUT}"
fi
prepare-test-matrix:
needs: target-machine
runs-on: ubuntu-26.04
outputs:
matrix-formatting: ${{ steps.prepare.outputs.matrix-formatting }}
matrix-tests: ${{ steps.prepare.outputs.matrix-tests }}
fetch_depth: ${{ steps.extract_branch.outputs.fetch_depth }}
steps:
# On push/schedule events there is neither a PR commit count nor a
# dispatch input; default to testing the single HEAD commit.
- name: Extract branch name
id: extract_branch
# commits-to-test is a workflow_dispatch input (attacker-controllable);
# route it and base_ref through env so neither is expanded into the
# shell / arithmetic context.
env:
BASE_REF: ${{ github.base_ref || github.event.repository.default_branch }}
COMMITS: ${{ github.event.pull_request.commits || github.event.inputs.commits-to-test || 1 }}
run: |
echo "head_ref=${GITHUB_HEAD_REF:-${GITHUB_REF#refs/heads/}}" | tee -a "${GITHUB_OUTPUT}"
echo "base_ref=${BASE_REF}" | tee -a "${GITHUB_OUTPUT}"
echo "fetch_depth=$(( COMMITS + 1 ))" | tee -a "${GITHUB_OUTPUT}"
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
persist-credentials: false
ref: ${{ steps.extract_branch.outputs.head_ref }}
fetch-depth: ${{ github.event.pull_request.commits || github.event.inputs.commits-to-test || 1 }}
repository: ${{ github.event.pull_request.head.repo.full_name || github.repository }}
- name: Prepare commits
id: prepare
shell: bash -ex {0}
env:
TEST_COMMITS_LIMIT: 25 # Limit the number of commits to test
COMMITS: ${{ github.event.pull_request.commits || github.event.inputs.commits-to-test || 1 }}
run: | # zizmor: ignore[template-injection] toJSON(needs.*) is controlled job data, not external input
COMMITS_TO_TEST=$(git log -n "${COMMITS}" --pretty='"%H"' | head -n "${TEST_COMMITS_LIMIT}")
echo "matrix-formatting={ \"commit\": [${COMMITS_TO_TEST//$'\n'/, }] }" | tee -a "${GITHUB_OUTPUT}"
TARGETS=$(echo '${{ toJSON(needs.target-machine.outputs) }}' | jq '.[]')
echo "matrix-tests={ \"commit\": [${COMMITS_TO_TEST//$'\n'/, }], \"target\": [${TARGETS//$'\n'/, }] }" | tee -a "${GITHUB_OUTPUT}"
check-formatting:
if: github.event_name == 'pull_request'
runs-on: ubuntu-26.04
permissions:
contents: read
packages: read
pull-requests: write # code-format-helper posts review comments on the PR
container:
image: ghcr.io/nomicfoundation/solx-ci-runner@sha256:7cc8aba2983f6feb71f046876611b48e9649e72acdfcdbcad756e2030c3a356a
needs: prepare-test-matrix
strategy:
fail-fast: false
max-parallel: 3
matrix: ${{ fromJson(needs.prepare-test-matrix.outputs.matrix-formatting) }}
steps:
# Shallow like the regression job: the diff below only needs
# `matrix.commit` and its parent, and a full LLVM-history clone
# (fetch-depth: 0) took ~9 minutes per matrix job.
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
persist-credentials: false
ref: ${{ matrix.commit }}
fetch-depth: ${{ needs.prepare-test-matrix.outputs.fetch_depth }}
repository: ${{ github.event.pull_request.head.repo.full_name || github.repository }}
- name: Mark everything as safe for Git
run: git config --global --add safe.directory '*'
- name: Define previous commit
id: previous-commit
run: echo "sha=$(git rev-parse -q '${{ matrix.commit }}^' 2>/dev/null || echo '${{ github.event.pull_request.base.sha }}')" >> "$GITHUB_OUTPUT" # zizmor: ignore[template-injection] matrix.commit and base.sha are git SHAs, not external input
- name: Get changed files
id: changed-files
uses: tj-actions/changed-files@9426d40962ed5378910ee2e21d5f8c6fcbf2dd96 # v47.0.6
with:
separator: ","
sha: ${{ matrix.commit }}
base_sha: ${{ steps.previous-commit.outputs.sha }}
# Job-local venv: the image's system interpreter is Debian-managed (its
# packages collide with these pins) and ships no pip. GITHUB_PATH makes
# the venv's python3/darker win in the formatter step below.
- name: Install formatting requirements
run: |
python3 -m venv "${RUNNER_TEMP}/venv-format"
"${RUNNER_TEMP}/venv-format/bin/pip" install -r ./llvm/utils/git/requirements_formatting_solx_llvm.txt
echo "${RUNNER_TEMP}/venv-format/bin" >> "${GITHUB_PATH}"
- name: Run code formatter
# start-rev/end-rev are git SHAs, changed-files a controlled comma list,
# PR number an integer — all from prior steps, not shell-injectable.
run: | # zizmor: ignore[template-injection]
set -x
if [ '${{ github.event.pull_request.head.repo.fork }}' = 'true' ]; then
TOKEN_PARAM=""
else
TOKEN_PARAM="--token ${{ secrets.GITHUB_TOKEN }}"
fi
python3 ./llvm/utils/git/code-format-helper-solx-llvm.py ${TOKEN_PARAM} \
--issue-number ${{ github.event.pull_request.number }} \
--start-rev ${{ steps.previous-commit.outputs.sha }} \
--end-rev ${{ matrix.commit }} \
--changed-files ${{ steps.changed-files.outputs.all_changed_files }}
regression-tests:
runs-on: ubuntu-26.04
# Backstop: a thrashed VM hangs its job indefinitely — the starved
# runner agent can't even act on cancellation. Sized for a cold
# ccache-less build with MLIR enabled, which approaches the limit.
timeout-minutes: 360
container:
image: ghcr.io/nomicfoundation/solx-ci-runner@sha256:7cc8aba2983f6feb71f046876611b48e9649e72acdfcdbcad756e2030c3a356a
# Cap container memory below the 16 GB VM (and disallow swap) so a
# memory-hungry test gets OOM-killed inside the container — lit records
# it as a named crash and continues — instead of starving the runner
# agent on the host, which kills the whole job as "lost communication".
options: --memory 13g --memory-swap 13g
# Host tooling bind-mounted so the free-disk-space action can reclaim
# ~24 GB before the LLVM build fills the runner's ~14 GB free budget.
# Keep in sync with solx's test.yaml / cache-warmup.yaml.
volumes:
- /usr/share/dotnet:/mnt/free-disk-space/dotnet
- /usr/local/lib/android:/mnt/free-disk-space/android
- /opt/ghc:/mnt/free-disk-space/ghc
- /usr/local/.ghcup:/mnt/free-disk-space/ghcup
- /opt/hostedtoolcache/CodeQL:/mnt/free-disk-space/codeql
needs: prepare-test-matrix
strategy:
fail-fast: false
max-parallel: 3
matrix: ${{ fromJson(needs.prepare-test-matrix.outputs.matrix-tests) }}
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
persist-credentials: false
repository: NomicFoundation/solx
ref: 'main'
# Local ref into the solx checkout above (like build-llvm below), so the
# action version always matches the checked-out solx-dev sources.
- name: Free disk space
uses: ./.github/actions/free-disk-space-linux
- name: Clean up LLVM directory
run: rm -rf solx-llvm
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
persist-credentials: false
ref: ${{ matrix.commit }}
path: "solx-llvm"
fetch-depth: ${{ needs.prepare-test-matrix.outputs.fetch_depth }}
repository: ${{ github.event.pull_request.head.repo.full_name || github.repository }}
- name: Mark everything as safe for Git
run: git config --global --add safe.directory '*'
- name: Define previous commit
if: github.event_name == 'pull_request'
working-directory: solx-llvm
id: previous-commit
run: echo "sha=$(git rev-parse -q '${{ matrix.commit }}^' 2>/dev/null || echo '${{ github.event.pull_request.base.sha }}')" >> "$GITHUB_OUTPUT" # zizmor: ignore[template-injection] matrix.commit and base.sha are git SHAs, not external input
- name: Build LLVM
uses: ./.github/actions/build-llvm
with:
enable-tests: true
enable-assertions: true
# The Sol/Yul dialects are developed in this fork, and solx CI
# builds LLVM with MLIR everywhere — so MLIR is part of every
# regression build, not an opt-in.
enable-mlir: true
# This tree changes every PR, so the whole-artifact cache never hits
# usefully — and a hit would skip the build, leaving no build-final
# directory for the lit and clang-tidy steps below. ccache (inside
# the action) is the only useful cache layer here.
enable-artifact-cache: 'false'
sanitizer: ${{ inputs.llvm-sanitizer }}
enable-valgrind: ${{ inputs.enable-valgrind }}
# The shared action only saves ccache on non-PR events, so PR reruns
# would rebuild cold (~2 h). Save under its restore-keys prefix:
# ccache-action prepends `ccache-` to every restore-key, so the leading
# `ccache-` below is required for the restore to prefix-match this
# entry, and the `-mlir-utils-tests` suffix must track what the action
# derives from the build-llvm inputs above. Placed before the lit step
# (which can still take down the runner) and on `always()` so even a
# failed build banks its partial ccache. Restores pick the newest
# matching entry, so only job 0 saves; more entries per run would be
# pure quota churn.
- name: Save PR-scoped ccache
if: always() && github.event_name == 'pull_request' && strategy.job-index == 0
continue-on-error: true
uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
with:
path: ${{ runner.temp }}/ccache-llvm
key: ccache-llvm-${{ runner.os }}-${{ runner.arch }}-mlir-utils-tests-end-pr${{ github.run_id }}-${{ github.run_attempt }}
# `verify-llvm` is a special target that runs all the regression tests
# it includes `check-llvm` and special codegen tests
# called via `verify-llvm-codegen-opt` and `verify-llvm-codegen-llc` targets
- name: Running Lit tests
run: |
# Crash tests (`not --crash`, gtest death tests, lit self-tests) pipe
# cores to the HOST's systemd-coredump — outside this container's
# cgroup and memory cap — which OOMed the 16 GB VM. Piped patterns
# ignore RLIMIT_CORE except the magic value of exactly 1 byte, which
# aborts the pipe handler (fs/coredump.c; the same trick LLVM's
# Process::PreventCoreFiles uses). `ulimit -c` scales its argument
# by 1024, so it cannot express that byte.
echo "core_pattern: $(cat /proc/sys/kernel/core_pattern)"
prlimit --core=1:1 --pid $$ || echo "::warning::prlimit failed; piped core dumps remain enabled"
prlimit --core --pid $$
# Per-test timeout needs psutil, which older images lack; without it
# lit only loses timeout support, so don't fail the run over it.
if python3 -c 'import psutil' 2>/dev/null; then
export LIT_OPTS='--timeout=600'
else
echo "psutil unavailable; running without per-test timeouts"
fi
ninja -C './target-llvm/build-final' verify-llvm -v
ninja -C './target-llvm/build-final' check-lld-unit -v
# MLIR's lit + unit tests, covering the Sol/Yul dialect and
# SolToYul/CAPI changes this fork carries.
ninja -C './target-llvm/build-final' check-mlir -v
# build-llvm configures with -DCMAKE_EXPORT_COMPILE_COMMANDS=Off;
# re-running cmake on the configured build dir flips it back On without
# touching any build rule, so ninja rebuilds nothing.
- name: Export compile commands
if: github.event_name == 'pull_request'
run: cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=On './target-llvm/build-final'
- name: clang-tidy
if: github.event_name == 'pull_request'
working-directory: solx-llvm
run: | # zizmor: ignore[template-injection] previous-commit.outputs.sha is a git SHA, not external input
CHANGES=$(git diff -U0 ${{ steps.previous-commit.outputs.sha }})
echo "Running clang-tidy on the following changes: ${CHANGES}"
echo "${CHANGES}" | \
./clang-tools-extra/clang-tidy/tool/clang-tidy-diff_Zegar.py \
-p1 -clang-tidy-binary "$(which clang-tidy)" \
-path ../target-llvm/build-final/compile_commands.json