GitLab

The fix in commit v3.2.3~3^2 (Fix assertion failure on unmatched foreach
in function, 2015-05-18) broke handling of unmatched non-loop blocks
because it assumed all function blockers removed during error unwinding
were for loops, essentially switching the set of mishandled cases.

The purpose of the loop block push/pop operations is to define a scope
matching the lifetime of the loop function blockers.  Since our function
blockers already have the proper lifetime, simply move the push/pop
operations to their constructor/destructor.

Extend the RunCMake.Syntax test with a case covering this.

The CMake 3.2 release branch does not have 'cmAlgorithms.h' and simply
provides 'cmHasLiteralSuffix' in 'cmStandardIncludes.h' instead.

Zack Galbreath authored May 27, 2015 and Brad King committed May 27, 2015

The regex update in commit v3.1.2~4^2 (ctest_build: Update GNU make
error message matching, 2015-01-28) broke matching of the messages it
updated.  Remove the escape character before the square brackets added
by that commit.

The lexical scope counting added by commit v3.2.0-rc1~332^2~1 (Track
nested loop levels in CMake language with a stack of counters,
2014-11-18) forgot to account for scopes popped by error messages about
unclosed scopes.  Teach the error handler to pop the lexical scope it
reports as unclosed.  Re-order the lexical scope RAII object to be
inside the variable scope RAII object scope so that the lexical scope
is fully closed before we check assertions about variable scopes.

Extend the RunCMake.Syntax test with a case covering this.

The part of the MS C Runtime library that applications need to
distribute has been renamed from "msvcr*.dll" to "vcruntime*.dll"
starting with VS 2015.  See the Visual C++ Team Blog:

 Introducing the Universal CRT
 http://blogs.msdn.com/b/vcblog/archive/2015/03/03/introducing-the-universal-crt.aspx

When given the command line

  tool a\ b c

mingw32-make incorrectly passes "a b" and "c" to the tool.  When given
the command line

  tool a\ b "c"

mingw32-make correctly passes "a\", "b", and "c" to the tool.

Since commit v3.1.0-rc1~861^2 (MSVC: Add properties to configure
compiler PDB files, 2014-02-24) we pass the compiler pdb option to
MS-style compiler tools as "/Fd<dir>\" but mingw32-make may consume
the backslash as escaping a following space as described above.
Workaround this problem by changing the backslash to a forward
slash as had been used prior to the above commit.

The use of "cmake -E touch_nocreate" added in commit v3.2.1~4^2
(Makefile: Fix multiple custom command outputs regression, 2015-03-06)
caused builds to fail when one of the outputs is intentionally not
created.  This was fixed by our parent commit by making touch_nocreate
succeed when the file is missing.  Add a test case covering it.

For the Watcom WMake generator, check for the SYMBOLIC source file
property separately on each output.  The mark is needed on outputs that
are not really created to tell 'wmake' not to complain that it is
missing.  The mark is also needed on outputs that are created or 'wmake'
will not consider them out of date when they exist.
Inspired-by: Ben Boeckel <ben.boeckel@kitware.com>

The GNU 3.3 optimizer causes bad behavior in liblzma, so disable it.

Somehow the XL compiler optimizations create incorrect behavior in
liblzma and lead to crashes or truncated output during compression.

Tim Kientzle authored Sep 25, 2014 and Brad King committed Apr 09, 2015

Port upstream LibArchive commit "compute string pointers after
concatenation" (2014-09-25) and commit "Move variables to top of
function for non-C99 compilers" (2014-11-15) to our CMake copy.
Otherwise we may compute a pointer to memory that is about to be freed
and then compute a bad size to give to CryptGenRandom.
Inspired-by: Tim Kientzle <kientzle@gmail.com>

Our Windows API header checks are conditioned on if(NOT UNIX) but
libarchive checks HAVE_WINDOWS_H anyway so the result leaks into
the cache and influenes the curl build.  Set the check results to
false explicitly for curl when not on Windows to tolerate this.
Reported-by: Vyacheslav Karpukhin <Vyacheslav.Karpukhin@jetbrains.com>

Replace an old hack of the form 'if("${VAR}" MATCHES "^${VAR}$")'
with the much simpler 'if(NOT DEFINED ${VAR})'.  This was exposed
by a CMP0054 warning.

Betsy McPhail authored Mar 23, 2015 and Brad King committed Mar 25, 2015

The file lock functionality added in commit v3.2.0-rc1~297^2~1 (file:
Add LOCK subcommand to do file and directory locking, 2014-11-26) forgot
to close the lock file descriptors.  Eventually it was possible to run
out of file descriptors and locks could not longer be acquired.  Fix
this by closing the file descriptor or handle when we are done with it.
Also set the member back to the initial value from the constructor
to leave everything in a consistent state (useful for debugging).

Co-Author: Ruslan Baratov <ruslan_baratov@yahoo.com>

Protect against variables named "CSHARP", "PERL", "PYTHON", or "UNKNOWN"
when CMP0054 is not set to NEW.
Reported-by: Tuukka Pasanen <tuukka@iocaste.ilmi.fi>

Simply check for whether the local generator has a parent instead of
depending on a string comparison of directory names.

Move the test cases from Tests/CMakeTests/ConfigureFileTest.cmake.in
over to use the RunCMake.configure_file infrastructure.  This does much
more robust verification of CMake output for each test case, and would
have caught the regression fixed in our parent commit.

The unknown argument warning added by commit v3.2.0-rc1~452^2
(configure_file: Warn about unknown arguments, 2014-10-31) failed to
account for options handled by the NewLineStyle member instead of
directly in the main loop.  Simply whitelist them for now.

In commit v3.2.0-rc1~272^2~2 (Makefile: Fix rebuild with multiple custom
command outputs, 2014-12-05) we changed the generated makefile pattern
for multiple outputs from

  out1: depends...
          commands...
  out2: out1

to

  out1 out2: depends...
          commands...

This was based on the incorrect assumption that make tools would treat
this as a combined output rule and run the command(s) exactly once for
them.  It turns out that instead this new pattern is equivalent to

  out1: depends...
          commands...
  out2: depends...
          commands...

so the commands may be run more than once.

Some documents suggest using a "dedicated witness" stamp file:

  stamp: depends...
          rm -f stamp
          touch stamp.tmp
          commands...
          mv stamp.tmp stamp
  out1 out2: stamp

However, if the commands fail the error message will refer to the stamp
instead of any of the real outputs, which may be confusing to readers.
Also, this approach seems to have the same behavior of the original
approach that motiviated the above commit: multiple invocations are
needed to bring consumers of the outputs up to date.

Instead we can return to the original approach but add an explicit
touch to each extra output rule:

  out1: depends...
          commands...
  out2: out1
          touch -c out2

This causes make tools to recognize that all outputs have changed and
therefore to execute any commands that consume them.