This parameter cannot be removed without breaking API. Doing this does not seem
worth it just to squash a warning. Additionally this function may use the `gvc`
parameter in future.
The ctype.h `is*` functions are typically implemented as macros using a lookup
table. This is fine, but on Cygwin the compiler is somewhat picky about this:
lib/ast/fmtesc.c:116:59: warning: array subscript has type char
[-Wchar-subscripts]
116 | (c == '#' && (b == f || isspace(*(b - 1)))
| ^~~~~~~~
I expect this release to be more turbulent than others in recent times, simply
because of the greater-than-average number of changes since the last release,
both in aggregate as well as specifically to the release process itself.
use 'sed' instead of 'rm;git-checkout' to fix CRLF line endings in Cygwin CI
Only selective artifacts in the repository are sensitive to line endings. E.g.
C/C++ and Python code is happily parsed by the compiler/interpreter regardless
of which line ending style it uses.
This should hopefully accelerate the Cygwin CI tasks slightly. The assumption is
that running `sed` on a few files should be faster than deleting all files and
recreating them.
pic plugin: replace 'pic_string' with 'gvputs_nonascii'
This finishes the process of de-duping `fig_string`, `mp_string`, and
`pic_string`. This change makes this plugin more thread safe (no static buffer
is no longer used), and improves its efficiency (bytes are now written directly
into the output instead of a temporary buffer, requiring dynamic allocation and
two copies).
mp plugin: replace 'mp_string' with 'gvputs_nonascii'
This partially de-dupes some code (`pic_string` is a duplicate of `mp_string`),
makes this plugin more thread safe (no static buffer is no longer used), and
improves its efficiency (bytes are now written directly into the output instead
of a temporary buffer, requiring dynamic allocation and two copies).
fig plugin: replace 'fig_string' with 'gvputs_nonascii'
This partially de-dupes some code (`mp_string` and `pic_string` are duplicates
of `fig_string`), makes this plugin more thread safe (no static buffer is no
longer used), and improves its efficiency (bytes are now written directly into
the output instead of a temporary buffer, requiring dynamic allocation and two
copies).
GDI+ plugin: rewrite 'FileStream' reference counting to be thread safe
Based on Microsoft sample code,¹ it seems like the expectation is that consumers
of the `IStream` API may expect to concurrently manipulate reference counts.
This change guards against this by making the referencing counting code use
lock-free atomics, based on the same Microsoft sample.
Note that this is essentially proactively addressing a latent issue as it is, in
general, unsafe to use Graphviz libraries or plugins in a multithreaded
environment right now.
GDI+ plugin: use 'override' instead of 'virtual' on 'FileStream' methods
This C++11 mechanism is an improvement as it causes compilation to error out if
any of these methods were not already declared virtual in a parent class. That
is, it provides a sanity check that these overrides are really doing what they
intend.
This variable affects `find_file` and `find_path`.¹ `find_file` is not used in
Graphviz and `find_path` already defaults to the correct values for the current
platform.² This change avoids, among other things, hard-coding of the target
architecture as x86-64.
Related to #1973.
¹ see https://cmake.org/cmake/help/latest/variable/CMAKE_INCLUDE_PATH.html#variable:CMAKE_INCLUDE_PATH
² https://cmake.org/cmake/help/latest/command/find_path.html#command:find_path
The code base extensively uses `getenv` without checking the `HAVE_GETENV` macro
this detection supplies. This is not a problem because `getenv` is provided as
far back as C89, and can be guaranteed to exist on all platforms for which
Graphviz currently compiles, as we require C99.
The if-style used in CMake files had no consistency. Meanwhile, the expression
in a CMake `endif` is ignored. We learned in the process of constructing the
prior commit that this is supported right back to CMake 3.1. So this commit
standardizes on no-space, no-expression if-style.
Commit 354ed2c644452f300b316d166386bd40d8fec76f, among others, introduced the
use of `$<TARGET_OBJECTS>` in `target_link_libraries` calls. The CMake docs¹
claim this is only officially supported from 3.21 onwards, despite working in
some prior versions. CMake provides binaries going back many versions² so it is
possible to manually bisect exactly what this wording in the CMake docs means.
Running this process yields the answer that CMake 3.9 is the first version that
does not error out when encountering `$<TARGET_OBJECTS>` in a
`target_link_libraries` expression.
Note that the Graphviz CMake files also use some features like `endif` without
an expression, which claims to only be supported from CMake 3.14 onwards.³
Despite this, versions as far back as CMake 3.1 support this syntax just fine.
¹ https://cmake.org/cmake/help/latest/command/target_link_libraries.html#linking-object-libraries-via-target-objects
² https://cmake.org/files/
³ https://cmake.org/cmake/help/v3.13/command/endif.html vs
https://cmake.org/cmake/help/v3.14/command/endif.html
edgepaint: remove 'LIBRARY' declaration from the MSVC defs file
Similar to the prior commit and 2b92772d480f321bda6615140efa546abc5c66e9, this
is a compatibility change that allows MinGW to also understand this file
correctly.
Magnus Jacobsson [Sun, 28 Nov 2021 12:32:28 +0000 (13:32 +0100)]
pathplan: don't override library name in .def file
Similarly as reported in
https://gitlab.com/graphviz/graphviz/-/issues/2156, the export name
override breaks the MinGW and Cygwin autotools builds. By not
specifying the export name, each toolchain should use its appropriate
default, which should at least work, even if it's not consistent
across toolchains.
projects / graphviz / log