Drop separate `TrackMemory` (aka `CURLDEBUG`) debug feature.
After recent changes (thread-safety,
193cb00ce9, and updates leading up to
it), `TrackMemory` is unlikely to cause build or runtime issues.
To simplify builds and debug options, enable `TrackMemory`
unconditionally for debug-enabled (aka `DEBUGBUILD`) builds. Before
this patch, this was already the default, with an option to disable
it, or enable it in non-debug-enabled builds.
Note, in practice these two debug options already went hand in hand. It
was not possible to toggle them separately for a long time due to bugs,
before 59dc9f7e69 (2024-05-28) fixed it.
This patch also removes/deprecates separate knobs and feature flags for
`TrackMemory`:
- autotools: `--enable-curldebug`/`--disable-curldebug`
- cmake: `-DENABLE_CURLDEBUG=ON`/`OFF`
- C macro: `CURLDEBUG`
- libcurl: `CURL_VERSION_CURLDEBUG` symbol deprecated in favor
of `CURL_VERSION_DEBUG`. They always return the same value after this
patch.
Also:
- drop `TrackMemory` from `curl -V` output.
- rename internal `CURLDEBUG` macro to `CURL_MEMDEBUG` internally.
To avoid confusion with `DEBUGBUILD`, but to keep guarding
`TrackMemory`-related internals for readability.
- runtests: bind `TrackMemory` to debug feature. Keep it a separate
test feature requirement, for clarity.
- CI: drop test builds for combinations of the two options.
- GHA/linux: no longer disable TrackMemory in the TSAN job.
Ref: https://github.com/curl/curl/pull/20328#issuecomment-3754528407Closes#20331
To simplify the directory layout.
- OS400 and vms support move from `packages` to `projects`.
- Windows README and `generate.bat` files move from `projects`
to `projects/Windows`.
Closes#20271
Turns out in practice the internal cache name contains the arch where it
matters (arm or dual-arch local builds), which is part of the cache key
string. Drop `runner.arch` to avoid two arch strings in the key.
Cache keys without an explicit arch designator mean intel.
Revert ff78af5752#20234Closes#20249
`mbedtls-arm` and `mbedtls-prev` were matching `contains()` expressions
looking for `mbedtls`. It caused an unnecessary cache restore and an
redundant mbedtls build on bumps, and made the build flavors require
different local directory names to avoid this accidental collision.
Also drop `-threadsafe` from internal names. All local builds are.
Follow-up to 8806035344#20240
Follow-up to 3a305831d1#19077Closes#20248
All libresll jobs, wolfssl-all, and one mbedtls job.
As noted earlier, arm jobs run faster than intel ones, especially
valgrind steps that run almost twice as fast. Package install runs
slower, but this is offset by faster build and test steps, even in
non-valgrind jobs.
Follow-up to ff78af5752#20234
Follow-up to 2b0d8dcc16#20231Closes#20246
`libressl-filc` was matching `contains()` expressions looking for
`libressl`, causing build confusion and failure in specific cases.
Rename `libressl` to `libressl-c` to avoid this. Also rename the cache
id for consistency, though not necessary for the fix.
Bug: https://github.com/curl/curl/actions/runs/20860412340/job/59938315276
Follow-up to c262481873#19407
Cherry-picked from #20234Closes#20240
More cost-effective and seems as fast or faster than Intel.
In particular, valgrind seems to be almost 2x fast. So fast the job pair
could fit under 10 minutes if merged again (but would be the longest in
GHA/Linux.)
Installing packages is slightly slower. The package repo is Ubuntu's
which is slower than the Azure mirror used on Intel (unless Azure is
broken, which happened a lot last year).
To add to more jobs, the locally built deps also need to be migrated.
Also:
- add workaround for failing sshd server on Linux arm runners, caused by
world-writable `HOME` directory.
Closes#20231
It also means that now all torture test-pairs use the same shallowness,
while earlier FTP used 20.
Also:
- This made macOS torture jobs slow enough to split them into 3 jobs.
Follow-up to 2cbcde90e7#20153Closes#20172
To avoid building libcurl in both static and shared flavor by default.
It results in 1.5-2.x speed-up for the curl build step in most jobs.
Saving a total of 6-7 minutes. In the Cygwin job alone it saves 1-1.5m.
Also:
- enable static + shared in a Windows job to keep testing this combo.
Follow-up to ff958fc4b2#20159Closes#20167
To make CI turnaround time shorter, by cutting the longest running jobs.
After this patch all jobs should finish around 10-11 minutes. Down from
15-16 minutes before this patch.
Suggested-by: Stefan Eissing
The fuzzing workflow is now the slowest (with a 7-minute startup time
needed to build deps from source on each run), followed by macOS
and Windows torture tests (both split in two now). Without fuzzing, it's
under 10 minutes.
Notes:
- an extra cost with job-pairs is installing prereqs,
configuring/building curl and tests twice. GitHub doesn't support
making a matrix job a prereq for another workflow that may fix this:
https://github.com/orgs/community/discussions/42335
This overhead is significant on Windows: 11m20 -> 9m20 + 8m40
- job-pairs are annoying to maintain and keep in sync.
- splitting tests into halves is a manual process and needs to be
revisited from time to time. Possibly something to automate with
a runtests option, e.g. with `1 of 50%` and `2 of 50%`?
Also:
- split torture tests in two equal pieces, replacing the `FTP` + `!FTP`
split used earlier.
Related perf improvements from today:
- make scan-build 2x fast:
ff958fc4b2#20159
- drop build-only Testi386 fuzz workflow (from curl), saving 9m per run:
0106023c1fhttps://github.com/curl/curl-fuzzer/pull/236Closes#20153
To run more pytest sshd tests, and for more static analysis.
Also:
- drop redundant option from `openssl libssh2 ...` config.
- GHA/linux: enable pytest in the LTO job (to test libssh2).
- avoid both with local builds of OpenSSL-forks, due to crypto lib
mixups causing a mixture of build error, crashes, test failures.
Follow-up to eb39fee40b#19934Closes#19943
- cmake: automatically disable typecheck when running clang-tidy,
to avoid possible interference, and to improve performance.
- INSTALL-CMAKE: document both this, and unity=off for clang-tidy.
- GHA/linux: disable for some static analyzers CI jobs to avoid possible
interference.
- GHA/linux: disable in Fil-C job to improve build performance.
Follow-up to 9e6f1c5efb#19637
Follow-up to fd2ca2399e#17955Closes#19941
To detect mistakes made in the runtests framework that reduce
the number of test runs. Before this patch it could go undetected with
a green CI.
The minimum thresholds will need light maintenance going forward (either
bumping them periodically, or adjust if some may fell below minimums for
justified reasons). We may also make minimums tighter or looser, or more
job-specific.
Latest number of test runs for each job can be seen at Test Clutch:
https://testclutch.curl.se/static/reports/feature-matrix.html
Also:
- GHA: set minimums.
Assisted-by: Dan Fandrich
Follow-up to f2a75a14dd
Follow-up to bb1391f943#19510Closes#19942
Only the `test-ci` build target sets `--buildinfo` automatically,
since 985f39c0ce. It needs to be set
manually for other targets used in CI, such as `test-torture`,
to enable the `buildinfo.txt` dump in the runtests step.
For Test Clutch. In an attempt to re-sync `targetarch` with the rest of
macOS jobs on the feature matrix page:
https://testclutch.curl.se/static/reports/feature-matrix.html
Before this patch and possibly since the breaking update It's `aarch64e`
for torture jobs and `aarch64` for the rest
(stricly speaking `aarch64e` is the correct value for all macOS jobs, but
autotools and cmake report arm64/aarch64 without the `e`.)
Regression from 985f39c0ce#18147Closes#19601
It also means that all supported OpenSSL versions and forks support
TLSv1.3 after this patch.
It reduces `openssl.c` size by more than 10%, or 400 LOC.
Ref: #18822Closes#18330
They were disabled since these jobs ran in Zuul. The tests are 320, 321,
322, 323, 324. Of which, 323 runs in CI, the rest needs `gnutls-serv`
with SRP enabled, which is not available in current jobs and no longer
offered by Ubuntu's `gnutls-bin` package. 324 doesn't appear as
a skipped test, 323 seems to be running fine, the rest are logged as
skipped. This suggests it's safe to drop the exceptions.
Closes#19413
Build and cache LibreSSL locally with Fil-C and enable it in the Fil-C
job.
Also:
- disable test 776 in the Fil-C job. It fails consistently, and due to
flakiness seen earlier its result is disabled. In this job it seems to
be adding 1 to 9 minues to the test run step and fails consistently.
- include Fil-C version in the LibreSSL cache key to prepare for Fil-C
ABI changes.
- GHA/linux: fully quote `tflags` values to avoid breaking YAML.
Tested and confirmed working with OpenSSL too, but ended up with
LibreSSL for faster, smaller builds.
Closes#19407
Requirements for Fil-C:
- not to accidentally pick up system headers. E.g. from `/usr/include`
on Linux. It can happen when any dependency is auto-detected on this
header path. This makes Fil-C find the wrong system headers, which
in turn breaks the configuration step in subtle ways (with CMake) and
less subtle ways (autotools). Then CMake ends up running into an error
while compiling.
- build all dependencies with Fil-C too.
(this patch doesn't build any dependencies yet.)
- "unity" mode disabled. It should work, but needs a lot of memory and
slower than a standard compiler, or a Fil-C non-unity build.
- x86_64 Linux host platform when using the pre-built toolchain.
Observations on a minimal, static build made with no dependencies and
Fil-C 0.674 (based on clang 20.1.8).
- curl tool sizes:
- cmake, default, w/o -O: 30 MB (gcc 14.2.0: 1.7 MB)
- cmake, default, w/o -O, stripped: 29.6 MB (gcc: 1.4 MB)
- cmake, Release, -O3: 7.2 MB (gcc: 1 MB)
- cmake, Release, -O3, stripped: 6.8 MB (gcc: 0.93 MB)
- autotools, default, -O2: 7 MB
- libcurl.a size is 32 MB (cmake, default, w/o -O) (gcc: 2.7 MB)
- build times 3-3.5x longer (compared to system gcc 14.2.0):
- all runtests available pass OK.
- all pytests skipped due to missing features/dependencies.
- shared libcurl builds also work (cmake, default: 25 MB libcurl.so and
5.75 MB (5.6 stripped) curl tool)
- autotools works fine too, with dependencies disabled or set to avoid
`/usr/include`.
Closes#19391
