All commands accepting file paths should normalize the slashes so that
the string-represented names can be compared reliably. The commands
add_library and add_executable have done this for years. We taught
add_custom_command to normalize its OUTPUT names in commit a75a0a14
(Normalize add_custom_command OUTPUT names, 2010-12-15). We handled a
special case of the DEPENDS option in commit 7befc007 (Handle trailing
slashes on add_custom_command DEPENDS, 2011-01-26).
Teach both add_custom_command and add_custom_target to normalize slashes
of DEPENDS files up front. This approach subsumes the above-mentioned
special case so remove the one line added for it but keep its test.
Extend the CustomCommand test to check that slash count mismatches
between custom command OUTPUT and DEPENDS can still be linked correctly.
A search-and-replace in commit 8d32d229 (make commands lower case by
default, 2007-10-10) accidentally changed the variable reference
CMAKE_INCLUDE_DIRECTORIES_BEFORE to CMAKE_include_directories_BEFORE.
Change it back.
The Cray Fortran compiler started using module init symbols in version 7.3.2.
Starting in commit 71287734 (Teach FortranC interface for Intel, PGI, and gcc
4.2, 2009-08-05) we provide C versions of the module init symbols so that the
detection executable can link when the C versions of the module-mangled symbols
are picked up.
If no C module-mangled symbol matches then we cannot let the C module init
symbol appear because it will be duplicated by the Fortran copy that provides
the module-mangled symbol. This was first handled for the PathScale compiler
in commit 21faaa5d (FortranCInterface: Fix PathScale detection, 2010-01-22) and
commit 46858720 (FortranCInterface: Fix PathScale detection again, 2010-02-16).
Handle it now for the Cray compiler too.
The curses dialog (ccmake) allows variables to be specified on the
command line. If any of these variables is used during any configure
iteration or during generate we must not warn about it.
The Qt dialog (cmake-gui) allows variables to be added and removed in
the GUI interactively. If a variable is added, removed, and then added
again we must still warn if it is unused.
Set target property RUNTIME_OUTPUT_DIRECTORY explicitly on ProcessFwd9x
and EncodeExecutable so that we know exactly where the executables will
exist on disk.
At some point in the past VS 2010 failed some tests with custom commands when
relative paths were not used. It seems that those problems have been fixed.
However, the relative paths apparently are appended to the current working
directoy before vs accesses the file. So, with a long path, relative paths
cause it to create a combined path that is too long.
86cb17b Pass include directories with response files to GNU on Windows
9a0b9bc Optionally pass include directories with response files
6e8a67f Generate target-wide flags before individual build rules
d099546 Factor old-style -D flags out from -I flag generation
e6c2701 ProcessorCount: Use ERROR_QUIET with execute_process (#11302)
4dd2ec2 ProcessorCount: Test fails if count is 0 (#11302)
6dd74d5 ProcessorCount: Add support for remaining platforms (#11302)
c159836 ProcessorCount test: more output, do not fail. (#11302)
6259bc4 Compare ProcessorCount to SystemInformation count. (#11302)
3430955 Add ProcessorCount support for QNX via pidin. (#11302)
4d6418f If getconf returns empty output, try cpuinfo. (#11302)
9cc8ad9 Add correct module notice header.
abbaa12 Add module ProcessorCount.cmake (#11302)
Also, comment out all "debugging" calls to message() that helped
us interpret the output on other platforms when running on the
dashboard clients.
Using ERROR_QUIET avoids unnecessary stderr output while calling
external tools to determine the processor count. If there's an
error parsing the output, we set the count to 0 anyhow.
Also, the test will fail on a CMake dashboard run if the count
comes back equal to 0.
Now that the code is "done"-ish, remove the debugging output.
Expect no output on stdout or stderr when calling the
ProcessorCount function from now on.
The parent commit 46c0a583 (Enable Java test more carefully on Apple,
2011-03-18) failed to restore the exclusion of Xcode when enabling the
Java test that was originally removed by commit c8f39193 (Avoid problem
reading jni.h on Macs, 2010-10-25). The Xcode generator does not work
with the current Java support at all.
This check if the function exists and the prototype we want to use is
correct. There are still functions which have different prototypes on
different UNIX systems.
The CTEST_RUN_Java option added by commit c8f39193 (Avoid problem
reading jni.h on Macs, 2010-10-25) was a quick hack to disable the Java
test on Mac machines after an update from Apple created a broken jni.h
symlink. Remove the option and instead test whether jni.h exists as a
readable file before reading it. This restores the original Java test
enabling logic but makes it robust to the broken symlink.
Some samples of things that got unnoticed by our nightly builds:
$ JAVA_HOME= mvn
Warning: JAVA_HOME environment variable is not set.
...
$ mvn
[INFO] Scanning for projects...
[INFO] ------------------------------------------------------------------------
[ERROR] BUILD FAILURE
...
The GNU 4.x toolchain on MinGW (and therefore MSYS) allows compiler
options to be passed via response files. Use this to pass include
directory -I options. This allows the include file search path to be
very long despite shell and mingw32-make command line length limits.
Create platform option CMAKE_<lang>_USE_RESPONSE_FILE_FOR_INCLUDES to
enable use of response files for passing the list of include directories
to compiler command lines.
This switches the internal generation order but does not affect the
results. The new order ensures that any internal state changed by
generating target-wide flags is known when the individual rules that use
those flags are generated.
Adjust whitespace to make the output of "--help-module FindMPI" look
good. Also separate the comment containing the copyright and license
notice so it does not appear in the documentation.