This is part of a goal to reduce release artifacts to ≤50 files so the list of
files is not truncated in Gitlab pages. MD5 is an obsolete checksum format that
has been deprecated for many years. This commit removes these checksums from
being produced in the deployment flow, leaving SHA256 checksums as the only
option. This is consistent with Gitlab’s Generic Package Registry that provides
an alternative way of seeing file checksums and also only offers SHA256.
This change reduces the number of release artifacts from 56 to 46.
Now that nothing is written to the sibling `debug` and `source` directories,
there does not seem to be an advantage to retaining this level in the hierarchy.
rename CMake-produced DEB t *-cmake to disambiguate
Now that the Autotools-produced DEBs are consolidated into tarballs, someone
looking at a flat list of all the release artifacts could incorrectly conclude
that Graphviz-${GV_VERSION}-Linux.deb is an alternative complete package to the
tarballs. Renaming this makes it clearer that it originates from the CMake build
system and is incomplete.
Note that this also removes the `os` hierarchy in the packaging directories for
these artifacts. All DEB artifacts are now produced without any
`os`/`debug`/`source` subdirectory.
This drops the total number of Graphviz release artifacts from 63 to 58. This is
still not below the 50 artifact ceiling we need to hit, but this is progress.
Note that this removes the `os`/`debug`/`source` hierarchy in the packaging
directories for these artifacts, now that all three of these are packaged into a
single tarball. The `os` hierarchy remains for the CMake-produced DEB for these
platforms.
This drops the total number of Graphviz release artifacts from 168 to 63. This
is still not below the 50 artifact ceiling we need to hit, but this is
significant progress.
rename CMake-produced RPM to *-cmake to disambiguate
Now that the Autotools-produced RPMs are consolidated into tarballs, someone
looking at a flat list of all the release artifacts could incorrectly conclude
that Graphviz-${GV_VERSION}-Linux.rpm is an alternative complete package to the
tarballs. Renaming this makes it clearer that it originates from the CMake build
system and is incomplete.
Note that this also removes the `os` hierarchy in the packaging directories for
these artifacts. All RPM artifacts are now produced without any
`os`/`debug`/`source` subdirectory.
This drops the total number of Graphviz release artifacts from 176 to 168. This
is still not below the 50 artifact ceiling we need to hit, but this is progress.
Note that this removes the `os`/`debug`/`source` hierarchy in the packaging
directories for these artifacts, now that all three of these are packaged into a
single tarball. The `os` hierarchy remains for the CMake-produced RPM for these
platforms.
This drops the total number of Graphviz release artifacts from 340 to 176. This
is still not below the 50 artifact ceiling we need to hit, but this is
significant progress.
This #define was not conditional (it was always set to `free`) and was not used
consistently (there were still direct calls to `free` in some places that should
have used this macro), incorrectly giving the impression that it could be safely
swapped for another function.
gts_surface_foreach_edge: fix mismatch of calling convention in callback
The compiler said about this code:
delaunay.c: In function ‘edgeStats’:
delaunay.c:245:34: warning: cast between incompatible function types from
‘void (*)(GtsSegment *, estats *)’ {aka ‘void (*)(struct _GtsSegment *,
struct <anonymous> *)’} to ‘gint (*)(void *, void *)’ {aka
‘int (*)(void *, void *)’} [-Wcast-function-type]
gts_surface_foreach_edge (s, (GtsFunc) cnt_edge, sp);
^
This is not quite as bad as the previous instance of this fixed. However, it
seems to have been relying on a coincident return value. E.g. on x86-64 this
code relies on 0 ending up in `RAX` at the end of the callback to indicate
iteration should continue, despite `cnt_edge` having no declared return value.
delaunay_remove_holes: fix mismatch of calling convention in callback
The compiler said about this code:
delaunay.c: In function ‘delaunay_remove_holes’:
delaunay.c:45:9: warning: cast between incompatible function types from
‘gboolean (*)(GtsTriangle *)’ {aka ‘int (*)(struct _GtsTriangle *)’}
to ‘gint (*)(void *, void *)’ {aka ‘int (*)(void *, void *)’}
[-Wcast-function-type]
(GtsFunc) triangle_is_hole, NULL);
This warning is not spurious. In particular, the mismatch in the number of
arguments passed to the callback means the callback function has a different
calling convention than expected by the code calling it. The result of this can
be stack corruption or incorrect interpretation of function arguments.
In practice, all major native calling conventions use registers for both
function types, so this is somewhat benign. However, this likely caused problems
on stack-based environments like WASM and JITs.
Each Graphviz version has a notion of “collection,” which traditionally could be
the values “stable” or “development.” The semantically versioned releases that
are published to the website are considered stable. The inter-release packages
that the deploy.py script still uploads to the Gitlab package repository are
considered development. gen_version.py knows how to determine the collection of
a version during build.
Additionally .gitlab-ci.yml knew of another collection value, “experimental,” of
which gen_version.py was unaware. It would manually tweak the collection to this
for certain non-mainstream builds.
deploy.py knows nothing of this notion of collections.
The combined effect of the above has led to some negative outcomes:
1. Published releases include a mixture of files labelled “stable” and files
labelled “experimental.” It is not clear to the general public what either
of these labels mean or how they differ.
2. In recent times, “development” versions are no more or less stable than
“stable” versions. The terminology here unfortunately has become
inaccurate.
3. The directory hierarchy for published releases has included an extra level
for collection. If deployment worked as intended, this would be redundant
as every release artifact would be “stable.” However, as discussed above,
the deployment script actually publishes “experimental” artifacts as well.
To reduce confusion and simplify build and packaging going forwards, this change
removes the notion of collection. The generated COLLECTION file is no longer
produced and users should have no need to be aware of this concept.
gen_version.py still internally knows the idea of collections to distinguish
published release version numbers from inter-release version numbers. But this
is no longer exposed to the rest of the build system and packaging.
The notion of “experimental” as a collection value is completely removed.
standardize on '/usr/bin/env bash' shebang for CI scripts
This is slightly more portable (e.g. on FreeBSD Bash does not live at
/bin/bash). The previous shebang was fine on all the operating systems we
currently support, but this adds robustness as we expand CI coverage.
Commit d64b75d2704bd3ed47b3f74c0686e19ca74e65fc explicitly enabled Lefty in
several places in preparation for it becoming disabled by default. However, it
accidentally missed the out-of-source build.
While it is probably not a good idea to have a GVC config that is 100000 bytes,
I see no reason it should be arbitrarily rejected. In future, it might be a good
idea to make `gvconfig_plugin_install_from_config` accept a file pointer instead
of a string and read/parse the config file incrementally. If this were done,
then there would probably be no issue at all with large configs.
gvconfig.c: realign alternative for 'glob' with POSIX
If the `glob` system function¹ is available, it is used in this code. If not (on
Windows), an alternative static function is provided. The POSIX spec mandates
only three members of the `glob_t` struct, `gl_pathc`, `gl_pathv`, and
`gl_offs`. This commit realigns the members with their types in the POSIX spec.
This squashes some -Wsign-compare warnings on POSIX platforms.
fix: exclude cmd/smyrna when Smyrna is disabled during build
Similar to the situation just fixed with Lefty, when Smyrna was disabled
(`--with-smyrna=no` passed to the build system), several Smyrna related
artifacts would still be built and installed. We now exclude the entire Smyrna
subdirectory, which should be less error prone.
fix: exclude cmd/lefty when Lefty is disabled during build
When Lefty was disabled (`--enable-lefty` not passed to the build system),
several Lefty related artifacts would still be built and installed. E.g. the man
page. We now exclude the entire Lefty subdirectory, which should be less error
prone.
enable persistent Chocolatey cache in Windows CI jobs
The intention of this change is to accelerate Windows CI jobs, currently one of
the limiting factors on our CI throughput. This change takes a conservative
approach, giving each relevant job its own cache. In theory these jobs could all
share a cache, but the Gitlab docs¹ imply this is only safe if only a single job
has a `push` policy. To do this effectively, we would probably have to create a
new preliminary CI job that only built the cache. Then all other jobs would
become `pull` consumers of this cache. Perhaps an improvement to make in future.
projects / graphviz / log