When using msvcdeps, header dependencies are not detected reliably for
generated source files. The root cause is a bug in versions of MSVC
prior to VS2019 16.0 in which it emits lower-case path prefixes when
resolving include paths relative to the containing file. Absolute paths
and paths relative to include directories passed in the MSVC command
line are, in contrast, case-correct.
Such a file-relative include directive with an incorrect lower-case
prefix derails waf's node hash signature handling and fails silently.
This change uses ant_glob() with the ignorecase keyword argument to
find the file on the filesystem with the correct case. The prior
case-correction code has been superseded and was removed.
See the following Visual Studio bug report for details on the issue:
https://developercommunity.visualstudio.com/content/problem/233871/showincludes-lowercases-some-path-segments.html
Make path_to_node() only accept a path as a string instead of also as a
list. That requires joining the list of path components in the relative
path case before calling path_to_node(). Also use path.pop(0) to remove
the first path component instead of copying the remainder of the path
using a slice operator.
Rework how msvcdeps' cached_nodes lock is used so acquiring the lock is
only necessary on a cache miss. Also use a "with" context manager to
simplify management of the lock lifecycle.
ant_matcher() converts an ANT glob pattern to an equivalent regex
pattern. This commit adds support for escaping parenthesis in the
input pattern so they don't end up being treated as a regex capture
group.
Also add a unit test to verify ant_glob()'s ability to handle special
characters in the input pattern.
Previously one could explicitly state to use PySide2 or PyQt4 but not PyQt5 which was picked just by default. In this way the option can override local configurations and also this prevents to have mixed tools versions if we are sure we need PyQt5.
This patch corrects an error in the exec_response_command exception
handler which always assumed that the execution's stdout would be bound
to the the WafError exception object.
However, this assumption is only true when the execution completes with
a non-zero status code. For other exceptions, the stdout attribute is
not bound.
Now, when stdout is not available, the WafError msg will be used
instead.
The order of the lines in a doxyfile are important. This patch uses an
ordered dictionary to keep the keys of the doxyfile in the same order.
This is particularly important for doxyfiles that contain @INCLUDE
lines. In such cases, if the dictionary is not ordered, the @INCLUDE
line can end up in the middle of the generated doxyfile and thus
override all entries that were seen before it.
Currently PDBs are only installed if the /DEBUG flag appears in the
current toolchain's LINKFLAGS attribute. This patch expands support
so that /DEBUG:FULL and /DEBUG:FASTLINK also cause PDBs to be
installed.
mac-o symbols are prefixed with an underscore. when specifying multiple
sub-regexes (e.g. 'sym1|sym2|sym3'), only the first will be matched
(since the expansion turns into '(?P<symbol>_?sym1|sym2|sym3)'). here,
this is remedied by wrapping the symbol regex in a paren group.
This patch attempts to detects if, when we are running from within an
MSYS2 environement (MSYSTEM is set) we are also executing inside an
MSYS2 provided version of python. It does this by assuming that if we
are not in a cygwin environment and we are building on windows, If the
value of sys.executable is /usr/bin or /bin or /usr/local/bin (somethign
unixy) then we are running in an MSYS2 python interpreter and shoudl
compensate for msys2 root paths. Otherwise we shouldn't be doing extra
path manipulation.
Not all tools executed by tasks support the '@argsfile' syntax for
shunting commandline arguments to a file. This means that if such
commands are shunted to a file early, he command will not work. On
windows the rc.exe command is such an example, but some tools on linux
have similar limitations. In the posix case, we artifically limit our
commandline size because it is difficult/variable to caluclate what the
actual limit is (it is partially dependent on environment size). This
could artifically cause commands to fail due to commandline length when
they otherwise wouldn't.
This patch fixes this issue by adding the 'allow_argsfile' flag to the
task. This way certain task instances will be able to specify if they
are compatible with the '@argsfile' syntax or not.
This patch addresses the bug described in issue #2225 where in using
posix paths and an empty PREFIX value can result in files being
installed to the root of the drive specified by destdir instead of to
the desired prefix value. This is a bug in the assumption that user
specified paths that are strings will contain directory separators that
match the target operating system.
The previous patches to workaround
http://support.microsoft.com/kb/830473 drastically over estimated the
number of characters in commands by treating the repr() version of the
command array as a reasonable estimator of commandline length. This
caused commands attempt to write argsfiles before they should have.
The new calculation calculates the number characters in the command
array and adds the number of spaces that would be added by ' '.join(cmd)
this provides a much closer estimate of the commandline length.
This also limits the CLI-length on non windows platforms to 200kB. This
prevents us hitting the much larger argument limits on Linux/BSD/MacOS
platforms.
Most of the ant_globs used are explicitly and knowingly on build directory
(ie. javadoc, jar re) so the warning is quite spurious. The only one that
may be in doubt is the source re one: I added also here because if you use
a code generator (ie. protoc) then it is also correct to glob on builds and
the warning is misleading.
