The cached tar held the 3rdparty work trees too, and its extraction fails on
the dangling symlink 3rdparty/zstd/tests/cli-tests/bin/unzstd, which Git Bash
tar can't create. Cache .git/modules alone and let git submodule update check
the work trees out from it, as a clean checkout does.
Co-Authored-By: Claude <noreply@anthropic.com>
Push builds of main ran one after another, so a merge waited for the
builds of every earlier merge before main got a result for its tree; a
fix of a red main sat behind a build of the tree it fixes. The newest
push now cancels the build in progress, like pull request and dispatched
builds already do. The merge queue's baselines already skip cancelled
main builds and use the nearest completed ones.
Co-Authored-By: Claude <noreply@anthropic.com>
A new select input of the Build workflow limits the Linux and Windows jobs
to the test executables and tests whose build inputs the merged pull
requests change. Push builds of main always build and run everything and
save the test graph of their tree; dispatched runs with select restore the
latest graph and fall back to a full run without one.
tools/select_tests.py dump walks the ninja graph of every test executable
(order-only dependencies included, so prx libraries loaded at run time are
in it) plus the header dependencies of its objects, the CMake files that
register each test, the source files its command line names, and its
fixtures. select picks the tests whose closure, registration or arguments
contain a changed file, and every test that is not a native executable of
the tree; files no build reads select the tests under their directory;
changes to the root build, cmake/, 3rdparty/, the workflow or the tool run
everything. Tests added since the graph was written always run. ctest gets
the selection as -I indices through the tool, which keeps the cmd shell on
Windows and stays under its command line limit.
Co-Authored-By: Claude <noreply@anthropic.com>
The runner cannot create directories under /var/cache/anyps5 (mkdir: Permission denied), so the install step failed
and the linux job built nothing.
Co-Authored-By: Claude <noreply@anthropic.com>
The self-hosted runners share one 32-thread host and were fixed at one job each, so
a build with many ccache misses compiled them one after another (686 misses took 37 min).
Pick jobs from the free CPUs and available memory (1 to 8, ctest at most 4) and pass a
load limit to ninja so a busy host backs off.
Co-Authored-By: Claude <noreply@anthropic.com>
Git for Windows 2.56 (runner image 20261002) moved from mingw64 to ucrt64,
so the fixed SSL_CERT_FILE path no longer exists and the WinLibs CMake
fails the FFmpeg download with status 60 on every Windows build. Take the
bundle from whichever of ucrt64 and mingw64 exists.
Co-Authored-By: Claude <noreply@anthropic.com>
The eligibility job enabled the relinker matrix for a pull request only
when repos.listCommits on main with the author's login returned a
commit. That filter misses authors whose commits in main are not
matched to their login. On 2026-10-13 it found no commit for 26 authors
of open pull requests who have 1 to 44 pull requests merged into main,
so their pull requests skip the relinker, against CONTRIBUTING.md ("if
its author already has a commit in main").
When listCommits finds nothing, the job now also searches for the
author's merged pull requests into main, with the query pr-intake.yml
uses. Incomplete search results count as none. A failed search, such as
the search API's own rate limit, skips the relinker with a warning
instead of failing the job. Both requests now run only when the pull
request changes a relinker file, so other pull requests make neither.
test_build_eligibility.py runs the step's script in node with a mocked
github object and skips when node is missing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A push to main ran every job: relinker on three hosts, register context
on four and write watch on two, next to the full Linux and Windows
builds. The eligibility gate now reads the files of the push (compare
before...after) and applies the same path rules as pull requests; when
it cannot (a first push, 300 or more files, an API error) every job runs
as before.
Whenever the full builds run (push and manual builds), the Linux and
Windows relinker jobs repeat tests the full builds already run: the
relinker-only ctest lists (150 and 145 tests) are subsets of the full
Linux and Windows ones (1052 and 1069) in run 38397922579. The relinker
matrix now keeps only macOS then; automatic pull request builds, which
have no full build, still run all three.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3rdparty/mbedtls (#3897) has its own framework submodule. The PR build
fetched it in its merge step, but push builds on main skipped that step
and failed at configure.
Co-Authored-By: Claude <noreply@anthropic.com>
Build SDL's Wayland backend next to X11 and ask SDL for "wayland,x11"
on Linux, so the window runs natively under a Wayland compositor instead
of through XWayland. SDL loads libwayland, xkbcommon and libdecor at run
time, so systems without them, or sessions without WAYLAND_DISPLAY, keep
using X11. SDL_VIDEODRIVER still overrides the choice.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since #5263 the Linux job restores a lavapipe shader cache keyed by the
runner's /proc/cpuinfo flags. llvmpipe compiles shaders with LLVM's host
CPU *name* (from vendor, family and model) plus the detected feature bits,
and keys its cache only by the feature bits, so two runners with the same
flags but different models share cached code. A hull shader JIT-compiled
on a znver4/znver5 or Ice Lake class host uses GFNI (vgf2p8affineqb);
loaded on a Zen 3 host with the same flags it raises SIGILL, which is what
agc_driver_tessellation has been hitting on main and on dispatched builds
since the first main build after #5234 saved that hull variant.
Hash vendor_id, cpu family, model and flags into the key, so a runner only
restores code compiled for its own CPU model, and print the model name so a
future mismatch can be read off the log.
The cache from #5107 was keyed by Mesa version only, so a runner restored
llvmpipe code compiled on another runner's CPU and about 85 agc tests died
with SIGILL (Illegal instruction) on main.
Co-Authored-By: Claude <noreply@anthropic.com>
The bundled SDL builds its PipeWire, PulseAudio and ALSA drivers only
when their development files are found, and otherwise keeps just the
silent dummy and disk drivers without an error. The Linux CI and
release jobs install none of them, so the v0.1.1 Linux libraries
(libSceAudioOut, libSceAgcDriver, libfmod) contain no audible driver.
Fail configuration on Linux when SDL found none of the three, install
the Debian development packages in the Linux build and release jobs,
and list the packages in the build instructions.
No job saved its ccache, so every build restored a day-old cache and
recompiled about half the tree (971/2038 hits). Main push builds now
save it, and they no longer cancel each other, so one finishes and
refreshes it. Pull request builds still only restore, so PR code never
writes the shared cache.
macOS has no counterpart to userfaultfd write-protect or
MEM_WRITE_WATCH, so GuestWriteWatch had no backend there and the driver
re-read every guest buffer on each use.
The macOS backend write-protects watched pages. The first write to a
page faults, the SIGBUS/SIGSEGV handler records the page and restores
its protection, and Collect reports the recorded pages and protects
them again. A page is protected before its record is cleared, under the
lock the handler takes, so a write that lands in between is recorded.
The lock hands out turns in order: an unfair one starved writers under
Rosetta.
The protection is real, which adds three entry points:
- GuestWriteWatchBeginHostWrite/EndHostWrite lift it around a system
call that writes guest memory, which would otherwise fail with
EFAULT. GuestArena::HostWrite calls them, as it calls the Windows
equivalent.
- GuestWriteWatchProtectionChanged tells the watch the protection given
to watched pages by mprotect or a new mapping.
- GuestWriteWatchProtection reports that protection, since a watched
page is read-only between a collection and the next write.
Register keeps the recorded protection of a page it already watches:
read from the kernel it would be the watch's own read-only protection,
and the page would never be watched again.
The tests build without the prx libraries and run natively and through
Rosetta.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GitHub refreshes merge_commit_sha only when the pull request's branch
changes, so a manual build could test the pull request against a main
from days earlier. Each job now checks out main as it is when the build
starts and merges the pull request head into it.
The Linux job reports "100% tests passed out of 371" with 20 agc_driver execution tests skipped
(llvmpipe reports subgroup size 8, so the tests that need a wider wave exit 77), and the Windows
job reports "100% tests passed, 0 tests failed out of 368" with 143 skipped - the runner has no
Vulkan driver and ANYPS5_REQUIRE_VULKAN is not set there, so 143 of the 147 GPU execution tests
silently do not run. ctest counts a skip as a pass, so neither job notices when the set grows.
Add tools/check_skips.py. It reads the file ctest writes with --output-junit and fails when a test
skips for a reason that is not expected:
- linux: a skip is expected only when the test's own output reports that the device's subgroups
are too narrow for it, so adding a wave32 test does not need the rule edited;
- windows: any agc_driver test, since the runner has no driver for them, while a skip of anything
else fails.
Both jobs now pass --output-junit and run the check after ctest. The JUnit file needs no
redirection, so the console output stays visible and no pipe can hide a failing ctest.
Verified here with the CI command and the CI device (lavapipe, subgroup 8): the real run reports
20 skipped, all expected; a 21st wave32 test skipping for the same reason passes; a skip whose
reason is something else exits 1; a file with 143 agc_driver skips passes with --job windows and
one non-agc skip in it exits 1.
