Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

cmake: add a proper test to determine if explicit linking against -lc is required #238

Merged
merged 2 commits into from
Feb 10, 2024

Conversation

chris-se
Copy link
Contributor

@chris-se chris-se commented Feb 8, 2024

This pull request hopefully contains the proper way to fix the -lc issue raised in #187, without breaking MinGW as in #233.

This was tested on Linux and Windows/MinGW, but not on OpenBSD (I don't have a system running). However, since the check is operating-system agnostic, it should hopefully work there as well.

@chris-se chris-se mentioned this pull request Feb 8, 2024
@chris-se
Copy link
Contributor Author

chris-se commented Feb 8, 2024

Note: the windows-mingw test in this PR fails because the test execution doesn't work correctly, probably a minor CI setup issue. (See https://github.com/tbeu/matio/actions/runs/7831127379/job/21366733842?pr=238) Compare this to a different PR where the MinGW build wasn't yet fixed, and it failed already at the build stage in the windows-mingw test. (See https://github.com/tbeu/matio/actions/runs/7827524982/job/21355503568?pr=237)

cmake/compilerOptions.cmake Outdated Show resolved Hide resolved
@tbeu tbeu requested a review from MaartenBent February 8, 2024 14:40
@tbeu
Copy link
Owner

tbeu commented Feb 8, 2024

@seanm Can you please cross-check in OpenBSD? Thanks!

@tbeu
Copy link
Owner

tbeu commented Feb 8, 2024

Note: the windows-mingw test in this PR fails because the test execution doesn't work correctly, probably a minor CI setup issue. (See https://github.com/tbeu/matio/actions/runs/7831127379/job/21366733842?pr=238) Compare this to a different PR where the MinGW build wasn't yet fixed, and it failed already at the build stage in the windows-mingw test. (See https://github.com/tbeu/matio/actions/runs/7827524982/job/21355503568?pr=237)

Yes, my bad. Can you fix it on the fly. $MATDUMP is just different.

@tbeu tbeu force-pushed the master branch 3 times, most recently from 6911ba8 to d78a523 Compare February 8, 2024 20:31
@tbeu
Copy link
Owner

tbeu commented Feb 8, 2024

There now is a OpenBSD CI available. Feel free to rebase.

@seanm
Copy link
Contributor

seanm commented Feb 8, 2024

There now is a OpenBSD CI available. Feel free to rebase.

Oh, nice! I didn't think github had that. Where can I RTFM on that?

@tbeu
Copy link
Owner

tbeu commented Feb 8, 2024

What excatly? OpenBSD CI or rebase button?

@seanm
Copy link
Contributor

seanm commented Feb 8, 2024

The former. Thanks for sharing.

Copy link
Collaborator

@MaartenBent MaartenBent left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks good. But I would prefer message(VERBOSE ...).
This status messages should not be relevant for end users.

CMake also has some macros to handle the CMAKE_REQUIRED_ variables, that you could use instead of manually resetting them: https://cmake.org/cmake/help/latest/module/CMakePushCheckState.html

@tbeu tbeu linked an issue Feb 9, 2024 that may be closed by this pull request
@chris-se chris-se force-pushed the proper-libc-link-test branch from dd46722 to 58e52d6 Compare February 9, 2024 07:51
@chris-se
Copy link
Contributor Author

chris-se commented Feb 9, 2024

I've now updated this PR to include all of the previous feedback, including using VERBOSE for specific messages, as well as reworking some other things. (+ rebasing on current master)

The OpenBSD CI has now run successfully, and the CMake checks work as intended:

  -- Performing Test HAVE_LINK_NO_UNDEFINED_IMPLICIT_LIBC
  -- Performing Test HAVE_LINK_NO_UNDEFINED_IMPLICIT_LIBC - Failed
  -- Performing Test HAVE_LINK_NO_UNDEFINED_EXPLICIT_LIBC
  -- Performing Test HAVE_LINK_NO_UNDEFINED_EXPLICIT_LIBC - Success

This means that -lc is added on OpenBSD because that is needed there, but it's not added on other platforms, where it either isn't required (Linux) or even breaks things (MinGW).

@tbeu tbeu requested a review from MaartenBent February 9, 2024 08:23
@tbeu tbeu force-pushed the master branch 12 times, most recently from a91ab85 to 9893d4c Compare February 9, 2024 21:13
@tbeu tbeu force-pushed the master branch 4 times, most recently from f249429 to 8442fbc Compare February 9, 2024 21:46
Christian Seiler added 2 commits February 10, 2024 18:02
Issue tbeu#187 appears to indicate that
OpenBSD requires an explicit specification of -lc on the linker line
if -Wl,--no-undefined is used.

Unfortunately, -lc is not available on MinGW (due to the C runtime
working different on Windows), so adding it whenever -lm is also linked
against does not work.

This commit reworks the entire CMake linker flag logic:

 - Remove dependency on CMake >= 3.17.0 for some checks -- we need to
   do a manual check_c_source_compiles() anyway (due to
   check_linker_flag() not triggering the potential error when -lc is
   not supplied), so we can drop that dependency.

 - Check for -Wl,--no-undefined both without and with an explicit -lc
   on the command line. If either version works, assume we can pass
   GNU-style linker options. If only the explicit version works, also
   pass -lc explicitly. (If both work don't, because even if -lc works,
   such as on a standard Linux, one would typically not pass it
   explicitly while compiling software.)

Fixes tbeu#233
@tbeu tbeu force-pushed the proper-libc-link-test branch from 58e52d6 to d39f93f Compare February 10, 2024 17:02
@tbeu tbeu merged commit b0b3140 into tbeu:master Feb 10, 2024
5 of 6 checks passed
@tbeu
Copy link
Owner

tbeu commented Feb 10, 2024

Thanks for fixing! Very appreciated.

@tbeu
Copy link
Owner

tbeu commented Feb 10, 2024

I've just noticed that the merge breaks the linking of matdump.exe on my Cygwin setup.

The relevant configure output is:

-- Performing Test HAVE_LINK_NO_UNDEFINED_IMPLICIT_LIBC
-- Performing Test HAVE_LINK_NO_UNDEFINED_IMPLICIT_LIBC - Success
-- Performing Test HAVE_LINK_NO_UNDEFINED_EXPLICIT_LIBC
-- Performing Test HAVE_LINK_NO_UNDEFINED_EXPLICIT_LIBC - Success
-- Performing Test HAVE_LINK_RETAIN_SYMBOLS_FILE
-- Performing Test HAVE_LINK_RETAIN_SYMBOLS_FILE - Success
-- Performing Test HAVE_LINK_UNDEFINED_ERROR
-- Performing Test HAVE_LINK_UNDEFINED_ERROR - Success
-- Using GNU-style linker flags

The link command is

/usr/bin/cc -O3 -DNDEBUG -Wl,--enable-auto-import CMakeFiles/matdump.dir/tools/matdump.c.o CMakeFiles/matdump.dir/snprintf/snprintf.c.o -o matdump.exe -Wl,--out-implib,libmatdump.dll.a -Wl,--major-image-version,0,--minor-image-version,0  libmatio.a -lm "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libhdf5.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libszaec.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libaec.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libzlib.lib" -lshlwapi 

which results in about thousands lines of error messages.

The previous link command was

/usr/bin/cc -O3 -DNDEBUG -Wl,--enable-auto-import CMakeFiles/matdump.dir/tools/matdump.c.o CMakeFiles/matdump.dir/snprintf/snprintf.c.o -o matdump.exe -Wl,--out-implib,libmatdump.dll.a -Wl,--major-image-version,0,--minor-image-version,0  libmatio.a -lm -lc "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libhdf5.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libszaec.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libaec.lib" "/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/lib/libzlib.lib" -lshlwapi 

with -lc being the only difference.

I tried to setup a Cygwin CI yesterday but failed.

@MaartenBent
Copy link
Collaborator

So cygwin does need libc, even though the implicit check succeeds.

It can't find printf. Can you modify the test code to use printf instead of (or in addition to) malloc and free?

@MaartenBent
Copy link
Collaborator

Also are you linking with msvc hdf5 libraries? I've never used cygwin myself, but mixing gcc and msvc is usually not a good idea.

@tbeu
Copy link
Owner

tbeu commented Feb 11, 2024

I tried a lot - even forcing -lc, but it only works if the merge is reverted.

Yes, I also wondered why it links with MSVC static HDF5 libs.

@chris-se
Copy link
Contributor Author

Most of the messages for undefined symbols you posted come from linking against MSVC static HDF5 libs, which obviously don't use the same C runtime as Cygwin, which is why those fail. (What I don't get is why they succeed if you revert the patch? They shouldn't work at all.)

My guess is that -lc isn't actually what's causing this here, but something probably related to linking against static MSVC libraries (perhaps the header of the wrong C runtime is used during compilation?). I haven't used cygwin in more than a decade, but I can take a look. Could you describe your setup a bit more? Did you just install cygwin + the official Windows binaries of HDF5?

@chris-se
Copy link
Contributor Author

So interestingly, in the constellation you describe (HDF5 installed via the official binary installer) and building from within Cygwin, I get the following error already compiling in cygwin before the CMake change:

In file included from /cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/include/hdf5.h:21,
                 from /home/christian/old.matio/src/matio_private.h:76,
                 from /home/christian/old.matio/src/endian.c:31:
/cygdrive/c/Program Files/HDF_Group/HDF5/1.14.3/include/H5public.h:285:19: error: conflicting types for ‘ssize_t’; have ‘long long int’
  285 | typedef long long ssize_t;
      |                   ^~~~~~~
In file included from /home/christian/old.matio/src/matio.h:37,
                 from /home/christian/old.matio/src/matio_private.h:36,
                 from /home/christian/old.matio/src/endian.c:31:
/usr/include/stdio.h:81:18: note: previous declaration of ‘ssize_t’ with type ‘ssize_t’ {aka ‘long int’}
   81 | typedef _ssize_t ssize_t;
      |                  ^~~~~~~

However, when I install the hdf5-devel Cygwin package, I have the following behavior:

  • If I just call CMake without additional options, it still finds the HDF5 version installed in C:\Program Files, not the HDF5 version that is part of Cygwin.
  • If I specify -DHDF5_ROOT=/usr when invoking CMake (build directory has to be clean, if there's already a CMake configuration there, this won't work), then everything compiles successfully on Cygwin, both with and without my patch, and ./matdump ../share/test_file_hdf5.mat appears to work as expected.

Since my commit only touches the build system, but not the source files, could it be that the object files from your previous build were still seen as current (because when using make instead of e.g. ninja, compiler flag changes, such as different include paths, aren't detected, and don't cause a rebuild), which is why you don't see the ssize_t error in your case (because it only tried to re-link, not re-compile). And that because you now installed the official HDF5 installer on your system (but previously hadn't) and CMake detects the Visual Studio one over the Cygwin on, that this broke the Cygwin build, and not my change? And that because you didn't start with a fresh build directory, you didn't see the issues? (Only guessing here.)

@tbeu
Copy link
Owner

tbeu commented Feb 15, 2024

OK, after reinstalling libhdf5-devel in CygWin, unsinstalling the MSVC static libs and setting -DMATIO_SHARED=ON I can successfully link in CygWin. (It would also work if -lc is set, but not necessary.) CygWin does not install the static hdf5 libs, which is the reason why configuration with -DMATIO_SHARED=OFF fails: Could NOT find HDF5 (missing: HDF5_LIBRARIES) (found version "1.12.3")

Finally, we can leave this as is. Thanks again for fixing and investigating.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

MinGW Build Regression
4 participants