R CMD check warning about compiler warning flags
Hi, See inline below.
On Dec 22, 2017, at 9:12 AM, Martin Maechler <maechler at stat.math.ethz.ch> wrote:
Duncan Murdoch <murdoch.duncan at gmail.com> on Thu, 21 Dec 2017 14:23:13 -0500 writes:
On 21/12/2017 1:02 PM, Winston Chang wrote:
On recent builds of R-devel, R CMD check gives a WARNING when some compiler warning flags are detected, such as -Werror, because they are non-portable. This appears to have been added in this commit: https://github.com/wch/r-source/commit/2e80059
That is not the canonical R sources.
Yes, that is obvious. The main page for that repository says it is a mirror of the R sources, right at the top. I know that because I put the message there, and because I see it every time I visit the repository. If you have a good way of pointing people to the changes made in a commit with the canonical R sources, please let us know. I and many others would be happy to use it.
The usual way is just to refer to the revision number, i.e. "This appears to have been added in rev 73909".
People who don't have the sources checked out can see the diff on your Github mirror using
https://github.com/wch/r-source/search?q="trunk at 73909"&type=Commits
and following the listed search hit. (Thanks to Thierry Onkelinx for helping me with this.) This only works for commits to the trunk. I guessed that something like
https://github.com/wch/r-source/search?q="R-3-4-branch at 73937"&type=Commits
would work if the commit was to the 3.4 branch, but apparently not. I don't know how to find those commits. Presumably there's a way, but I don't know it.
Another possibility is that someone could set up (or already has?) one of the web viewers (WebSVN, etc.) for the real repository. That would be better for those of us who are SVN users, but probably harder for Git users.
Duncan Murdoch
As you know I had setup (the first few versions of) the svn at https://svn.r-project.org/ at the time, I wanted to keep that machine protected as much as possible and had decided not to install any other apache modules and svn - niceties just such that the server would run minimal services and hence would be minimally vulnerable. The times have changed though and I will look into adding WebSVN to svn.r-project.org as one of the first things in 2018. Martin Maechler
Martin, Just to play a bit of a devil's advocate here, WebSVN has not been updated/maintained since June of 2011, so is a number of years old at this point. That should raise some reasonable concerns over bugs, security issues and related matters. To put that in perspective, the last update to WebSVN was around the time that R 2.13.1 was released. That general pattern, of SVN clients not being actively maintained, seems to be consistent across a number of the web based (and even browser plugin based) SVN clients, which may in turn, be an indication of the general shift to Git in recent years, much as the shift from CVS to SVN occurred years ago. In researching other SVN clients, the few that still seem to be actively maintained are typically dedicated desktop clients/GUIs, that in some cases are closed source and/or are OS specific. One of the few web based SVN clients that still seems to be actively maintained is Trac (https://trac.edgewall.org), however it is under a modified BSD license, which may raise some issues and from what I can tell, may be more complicated in terms of set up, since it also supports bug tracking, wiki and other functionality. As with any such application, there would be ongoing maintenance issues as well. I am not advocating any particular solution. I just want to point out potential issues with WebSVN and raise for discussion, what options may make sense to consider, based upon how many folks might actually use such "live" web based functionality versus remote SVN access via a desktop client, which is certainly an option, since they won't have write access anyway. Also, a number of the Git desktop clients also support Git SVN (https://git-scm.com/docs/git-svn), for compatibility, which offers Git users an alternative to using a dedicated desktop SVN client. Regards, Marc
And your description seems wrong: there is now an _optional_ check controlled by an environment variable, primarily for CRAN checks.
The check is "optional", but not for packages submitted to CRAN.
I'm working on a package where these compiler warning flags are present in a Makefile generated by a configure script -- that is, the configure script detects whether the compiler supports these flags, and if so, puts them in the Makefile. (The configure script is for a third-party C library which is in a subdirectory of src/.) Because the flags are added only if the system supports them, there shouldn't be any worries about portability in practice.
Please read the explanation in the manual: there are serious concerns about such flags which have bitten CRAN users several times. To take your example, you cannot know what -Werror does on all compilers (past, present or future) where it is supported (and -W flags do do different things on different compilers). On current gcc it does -Werror Make all warnings into errors. and so its effect depends on what other flags are used (people typically use -Wall, and most new versions of both gcc and clang add more warnings to -Wall -- I read this week exactly such a discussion about the interaction of -Werror with -Wtautological-constant-compare as part of -Wall in clang trunk).
Is there a way to get R CMD check to not raise warnings in cases like this? I know I could modify the C library's configure.ac (which is used to generate the configure script) but I'd prefer to leave the library's code untouched if possible.
You don't need to (and most likely should not) use the C[XX]FLAGS it generates ... just use the flags which R passes to the package to use.
It turns out that there isn't even a risk of these compiler flags being used -- I learned from of my colleagues that the troublesome compiler flags, like -Werror, never actually appear in the Makefile. The configure script prints out those compiler flags out when it checks for them, but in the end it creates a Makefile with the CFLAGS inherited from R. So there's no chance that the library would be compiled using those flags (unless R passed them along). His suggested workaround is to silence the output of the configure script. That also hides some useful information, but it does work for this issue. -Winston