Replace combinational-loop analysis with sparse SSA #4772

Workflow file for this run

.github/workflows/codspeed.yml at cba3429

	name: CodSpeed

	on:
	push:
	branches:
	- "master"
	pull_request:
	# `workflow_dispatch` allows CodSpeed to trigger backtest
	# performance analysis in order to generate initial data.
	workflow_dispatch:

	jobs:
	benchmarks:
	strategy:
	fail-fast: false
	matrix:
	include:
	# veryl-tests (parse/format/analyze) is LTO-sensitive, so keep the
	# fat-LTO `bench` default.
	- package: veryl-tests
	build_args: ""
	# The simulator's hot loop is JIT-compiled native code that LTO
	# barely touches; build it lto=false (release-verylup) to skip the
	# fat-LTO link that OOM-kills the 2-core runner via wasmtime+cranelift.
	- package: veryl-simulator
	build_args: "--profile release-verylup"

	name: Run benchmarks
	runs-on: ubuntu-latest
	env:
	RUSTC_WRAPPER: sccache
	CARGO_INCREMENTAL: "0"
	# ~0.5G holds one full build closure; 1G adds a generation of headroom.
	# Bigger only hoards stale object versions and pushes the repo over the
	# 30GB cache quota, evicting the rust-cache generations that let cargo
	# skip dependency compilation entirely.
	SCCACHE_CACHE_SIZE: "1G"
	steps:
	- uses: actions/checkout@v6

	- name: Setup rust toolchain, cache and cargo-codspeed binary
	uses: moonrepo/setup-rust@v1
	with:
	channel: stable
	cache: false
	bins: cargo-codspeed

	# rust-cache caches ~/.cargo and target/ (dependency compilation) so cargo
	# skips deps; sccache (content-addressed) then serves the workspace crates,
	# which rust-cache alone cannot reuse across runs (cargo's mtime fingerprints
	# break after checkout). The key is distinct because building under sccache
	# changes every crate's fingerprint. Master saves; PRs restore read-only.
	- uses: Swatinem/rust-cache@v2
	continue-on-error: true
	with:
	shared-key: "${{ runner.os }}-codspeed-${{ matrix.package }}-sccache"
	save-if: ${{ github.ref == 'refs/heads/master' }}
	- uses: mozilla-actions/sccache-action@v0.0.10
	with:
	# stats stay in the job log via the explicit step below; keep them off
	# the run-summary annotations
	disable_annotations: true
	- uses: actions/cache/restore@v4
	continue-on-error: true
	with:
	path: ~/.cache/sccache
	key: sccache-codspeed-${{ matrix.package }}-${{ github.sha }}
	restore-keys: sccache-codspeed-${{ matrix.package }}-

	- name: Build the benchmark target(s)
	run: cargo codspeed build -p ${{ matrix.package }} --bench benchmark ${{ matrix.build_args }}

	- name: sccache stats
	if: always()
	run: sccache --show-stats

	- uses: actions/cache/save@v4
	if: github.ref == 'refs/heads/master'
	continue-on-error: true
	with:
	path: ~/.cache/sccache
	key: sccache-codspeed-${{ matrix.package }}-${{ github.sha }}

	- name: Run the benchmarks
	uses: CodSpeedHQ/action@v4
	with:
	run: cargo codspeed run -p ${{ matrix.package }}
	mode: "simulation"
	token: ${{ secrets.CODSPEED_TOKEN }}

	# heliodor 1-core Linux boot benchmark (CodSpeed walltime). Runs on a CodSpeed
	# Macro Runner (codspeed-macro, ARM64) — walltime is not available on standard
	# GitHub-hosted runners. Needs a C compiler (the simulator cc backend).
	# Run locally with: cargo bench -p veryl-simulator --bench heliodor_boot
	boot-benchmark:
	name: Run heliodor boot benchmark (walltime)
	runs-on: codspeed-macro
	env:
	CARGO_INCREMENTAL: "0"
	steps:
	- uses: actions/checkout@v6
	with:
	submodules: recursive

	- name: Setup rust toolchain, cache and cargo-codspeed binary
	uses: moonrepo/setup-rust@v1
	with:
	channel: stable
	cache: false
	bins: cargo-codspeed

	# rust-cache handles dependency compilation (cargo skips deps on a hit).
	# sccache is intentionally not used here: on this Macro Runner its per-sha
	# cache re-uploads ~0.5GB every master run (~270s) for only marginal
	# workspace-crate reuse, which rust-cache already covers via the deps.
	- uses: Swatinem/rust-cache@v2
	continue-on-error: true
	with:
	shared-key: "${{ runner.os }}-${{ runner.arch }}-codspeed-boot"
	save-if: ${{ github.ref == 'refs/heads/master' }}

	# The simulator stores compiled aot_c .so files content-addressed under
	# ~/.cache/veryl/aot_c; without this cache every run pays the full gcc
	# compile of the emitted C (~10 min on this runner). The lookup inside
	# the dir is content-addressed, so the cache key only needs to pick the
	# newest snapshot: restore by prefix, save under the unique commit sha.
	- uses: actions/cache/restore@v4
	id: aot-cache
	continue-on-error: true
	with:
	path: ~/.cache/veryl/aot_c
	key: ${{ runner.os }}-${{ runner.arch }}-aot-c-${{ github.sha }}
	restore-keys: ${{ runner.os }}-${{ runner.arch }}-aot-c-

	# release-verylup (lto=false) avoids the slow fat-LTO link the `bench`
	# profile forces; the JIT'd simulation loop is unaffected by LTO. Matches
	# the simulator build in the benchmarks job above.
	- name: Build the benchmark target
	run: cargo codspeed build -p veryl-simulator --bench heliodor_boot --profile release-verylup

	- name: Run the benchmark
	uses: CodSpeedHQ/action@v4
	with:
	run: cargo codspeed run -p veryl-simulator
	mode: "walltime"
	token: ${{ secrets.CODSPEED_TOKEN }}

	# Save only on master (like rust-cache above): PR-scoped caches are
	# invisible to other refs and just eat the 10GB quota.
	- uses: actions/cache/save@v4
	if: github.ref == 'refs/heads/master'
	continue-on-error: true
	with:
	path: ~/.cache/veryl/aot_c
	key: ${{ steps.aot-cache.outputs.cache-primary-key }}

Navigation Menu

Uh oh!

Uh oh!

Replace combinational-loop analysis with sparse SSA #4772

Workflow file

Replace combinational-loop analysis with sparse SSA #4772

Uh oh!

Workflow file for this run