Cache key isolation: Include the build flavour (ccwarn/asan/gcov) in the
build ccache key, so the coverage (gcov), asan, and RTLMeter
(ccwarn-off) configurations no longer share and thrash the single
'<runs-on>-<cc>' bucket, nor warm-start from each other's (near-useless)
caches.
Prune stale entries instead of a fixed size cap: After building/testing,
evict ccache entries this run did not touch (--evict-older-than, based
on the run's own duration). This bounds the saved cache to the current
working set, so CCACHE_MAXSIZE and CCACHE_LIMIT_MULTIPLE are no longer
needed and are removed (ccache falls back to its 5G default ceiling, but
is realistically much smaller).
ccache configuration:
- Drop CCACHE_COMPRESS (already the default in ccache 4.x).
- Add CCACHE_COMPILERCHECK=content, so a changed compiler mtime on a rebuilt
ephemeral runner image does not invalidate the whole restored cache.
- Hoist all ccache-related variables (CACHE_BASE_KEY, CCACHE_*, CXX) to the
workflow-level env.
Remove dead ccache scaffolding:
- reusable-lint-py: the job never compiles or caches; drop its ccache env.
- reusable-rtlmeter-run: remove the disabled persistent-cache step and its
storage variables; keep CCACHE_DISABLE=1, since caching must stay off to
keep RTLMeter timings representative.
Store test job ccache as artifacts for reuse
Code coverage and RTLMeter complete (concluding 'skipped') on branch
pushes where they do no work, and each completion fires the Pages
workflow_run trigger. Rebuilding and redeploying master's pages for
these no-op events is wasteful.
Gate the build job so a workflow_run whose upstream conclusion is
'skipped' does nothing; deploy and notify need build, so they
cascade-skip.
Route such no-op runs to their own throwaway concurrency group. In the
shared 'pages' group they would, as the newest queued run, cancel a
pending real run and then deploy nothing (concurrency keeps only the
in-progress and newest pending run). Real runs keep sharing 'pages' and
serialize as before.
Add retry to deploy step to reduce flakyness due to the Pages back end.
Fix trigger on workflow file changes.
The Windows build never actually used ccache; the '.ccache' directory
only ever held the win_flex_bison install, and cmake/MSVC did no compiler
caching. Clean this up:
- Drop the CCACHE_* environment variables from the job (meaningless here).
- Store win_flex_bison in a dedicated 'win_flex_bison' directory instead of
the misleadingly named '.ccache', and cache just that (static key, as it
is a fixed third-party tool).
- Set WIN_FLEX_BISON inside ci-win-compile.ps1 as an absolute path, rather
than in the workflow, making the script self-contained.
- Use the runner's processor count for build parallelism instead of a fixed
'-j 3', in both ci-win-compile.ps1 and ci-win-test.ps1.
- Remove the verilator-win.zip archive and upload steps; nothing consumes
that artifact.
- Build Verilator with the Ninja generator from an MSVC developer shell
instead of MSBuild. MSBuild only parallelizes across projects, and
Verilator is a single project, so it compiled serially; Ninja keeps all
cores busy. Also compile with /Od: this job only checks that Verilator
builds and can verilate an example, so an optimized binary is not needed
and the otherwise-dominant optimizer time is wasted.
- Merge reusable-rtlmeter-build.yml into reusable-build.yml, selected by
an 'install' flag: archive the installed Verilator for RTLMeter, or
the whole repository tree for the test jobs.
- Align reusable-test.yml input arguments with reusable-build.yml
- Pass build/test options as booleans instead of 0/1.
- Inline the single-entry build matrices, and hoist matrix keys that are
constant across a job.
- Split the monolithic ci-script.bash into ci-build.bash and
ci-test.bash, with shared setup in ci-common.bash.
- Pass job parameters as explicit command-line options rather than
environment variables (--compiler, --prefix, --suite, --asan,
--gcov, --reloc, --ccwarn).
- Detect the host OS and distribution at runtime (uname,
/etc/os-release) instead of the legacy Travis-era CI_OS_NAME,
CI_RUNS_ON and CI_BUILD_STAGE_NAME variables.
- Remove dead code: FreeBSD support, Windows wavediff install, the
unused CI_COMMIT variable, and various redundant operations.
- Inline the single-entry build-job matrices.
- Rename osx to macOS
- Refuse to run these ci scripts outside GitHub Actions to protect
developer checkouts.
PR #7776 removed Ubuntu 22.04 clang from the test matrix, which was the
only one testing the relocation of an installation. Add back under
Ubuntu 24.04 clang.
Restrict VL_RESTORER to be usable only with trivially copyable types.
Introduce VL_RESTORER_COPY and VL_RESTORER_CLEAR, which are more
efficient versions usable for non trivially copyable types.
VL_RESTORER_COPY semantically behaves the same as VL_RESTORER, but only
does one copy at initialization, and on scope exit restores via a move.
VL_RESTORER_CLEAR swaps the variable with an new one constructed via
the no-args constructor (e.g. empty collection), which does not require
any copying at any point.
Static assertions enforce picking one of the new flavours when copying
might be expensive.
Fetching a thread-local is relatively expensive. Random initializing
wides used to do it once or twice per word, and on short runs for
large designs can be noticeably expensive, so fetch once per variable
instead. Also remove unused VL_RAND_RESET_{Q,W} functions.
Rewrite module inlining decision to be based on a bipartite Module/Cell
graph, similar to V3InlineCFuncs. Preserved all old heuristics, but
added 2 new ones:
- If a module, and all the sub-hierarchy below it, is less than 10% the
total flattened size of the design, then flatten the contents of that
module (but the module itself is not necessarily inlined).
- If the flattened size of all instances of a module is less than 20% of
the total flattened size of the design, then inline all instances of
that module.
These are both relative to the total size of the design, so they
auto-scale with complexity. The net effect is that large shared
instances are preserved, but their contents are flattened out. E.g. in a
multi-core CPU this would keep the cores non-inlined but flatten out
most everything else. This still enables V3Combining and sharing those
later, but avoids potentially big overheads e.g. with small widely used
library modules.
Empirically this yields less generated C++ than the previous version
(due to removing lots of small functions), and can improve performance
10-20% while still having meaningful combining relative to the size of
the design.
Geza Lore
Wilson Snyder
Pawel Klopotek
Eryk Szpotański
Dragon-Git
Pawel Kojma
Jakub Michalski
github action
Sergey Chusov
dependabot[bot]
Joshua Leahy
Nick Brereton
Matthew Ballance
Yilou Wang