explicitly make exported functions visible

[jeroen]

Marks exported Rcpp functions as visible, on supported platforms. The attribute_visible macro is defined in <R_ext/Visibility.h>. If the compiler supports visibility it is defined as

 __attribute__ ((visibility ("default")))

And otherwise it is blank. The idea is that we can compile packages with -fvisibility=hidden as defined in make variable $(C_VISIBILITY) on supported platforms. So pkg authors can do:

PKG_CXXFLAGS=$(C_VISIBILITY)

Limiting visibility to exported functions reduces the chances of symbol conflicts like we saw in ropensci/hunspell#21. Here both rstudio and the hunspell package include a copy of libhunspell but different versions. Apparently their C++ classes start conflicting under certain circumstances. Hiding the libhunspell symbols from the dll solved the problem.

[codecov-io]

Codecov Report

Merging #720 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #720   +/-   ##
=======================================
  Coverage   93.15%   93.15%           
=======================================
  Files          66       66           
  Lines        3432     3432           
=======================================
  Hits         3197     3197           
  Misses        235      235

Impacted Files	Coverage Δ
inst/include/RcppCommon.h	100% <ø> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5ca4f6a...f873b97. Read the comment docs.

[eddelbuettel]

Interesting. I was so far only aware of a user-space clash between Boost and RStudio which I had hit with anytime and (the new, non CRAN) RcppMLPACK. This got fixed in the RStudio dailies a little while ago thanks to @kevinushey

I am always a little wary of making a global change but this one may be sensible. Maybe to wait till I get home though.

Alternatively could we change this to defining and inly if another variable is defined? I seem to have

C_VISIBILITY = -fvisibility=hidden
F77_VISIBILITY = -fvisibility=hiddenb

so should we condition on that?

[jeroen]

I think the conditioning is already done by R itself. Header R_ext/Visibility.h has:

#ifdef HAVE_VISIBILITY_ATTRIBUTE
# define attribute_visible __attribute__ ((visibility ("default")))
# define attribute_hidden __attribute__ ((visibility ("hidden")))
#else
# define attribute_visible
# define attribute_hidden
#endif

Where HAVE_VISIBILITY_ATTRIBUTE is defined in Rconfig.h if the compiler supports it. When HAVE_VISIBILITY_ATTRIBUTE is not defined, both C_VISIBILITY and attribute_visible are blank.

[eddelbuettel]

Thanks for that, very helpful. I very vaguely recalled it but did not have time to hunt.

[jjallaire]

Seems like a good change to me, I'll merge it. I can't forsee any problems with revdeps but we'll see on the next pass!