~/bin$ CC=clang CFLAGS=-g -O0 CXX=clang++ ~/src/R-devel/configure --enable-memory-profiling --enable-R-shlib --disable-byte-compiled-packages
Looking at R's config.log carefully, I eventually spotted a difference depending on compilation with -O0 (when the symbol is not found) and -O2 (when everything is fine). The essence of the test is
~/tmp$ clang -O0 conftest.c
/tmp/conftest-hn9uU6.o: In function `main':
conftest.c:(.text+0x9): undefined reference to `clock_gettime'
clang: error: linker command failed with exit code 1 (use -v to see invocation)
~/tmp$ clang -O2 conftest.c
for either gcc or clang, in R-2-15 or R-devel. conftest.c is, at the end of the day
#include <time.h>
int
main ()
{
#ifndef clock_gettime
char *p = (char *) clock_gettime;
#endif
;
return 0;
}
with -O2 succeeding because, well, the 'check' is optimized out (right?). This is from R-devel/configure.ac:1835
R_CHECK_FUNCS([clock_gettime timespec_get], [#include <time.h>])
if test "${ac_cv_have_decl_clock_gettime}" = "yes"; then
AC_CHECK_LIB(rt, clock_gettime)
fi
I'm not completely sure of this...
Martin
From the error message, it happens when "testing if installed package can be
loaded" and trying to load Rcpp.so through dyn.load. I am not very sure how
dyn.load works (and thus please ignore me if I am talking non-sense), but
from the help page it seems that it loads additional shared object into the
executing R image. On my machine, R was built as a shared library, and ldd
libR.so shows among other things librt. In other words, librt was loaded when
R was started. However building Rcpp only builds libraries rather than
executables, and thus there shouldn't be errors when building Rcpp while
loading Rcpp may cause errors unless librt was already loaded or linked to
libRcpp.so
Another reason that I suspect that this is not a problem with compiler but
how R is built is that, try the following program,
#include <time.h>
int main () { clock_gettime(0,0); }
with either g++ or clang++, without explicit linking to -lrt, the same
undefined symbol error occurs. For instance, on my RHEL, with g++ 4.4, $ g++
-o foo foo.cpp /tmp/ccKB2Ftv.o: In function `main': foo.cpp:(.text+0xf):
undefined reference to `clock_gettime' collect2: ld returned 1 exit status
Change main to say, foo, and build a shared or static library without linking
to librt causes no problems at all (though loading that library will cause
problems as usual).
If so far what I am talking is not non-sense, then I suggest you find your
libR.so library, and check if it indeed does not link to librt. If that is
the case, we need to find out how to build R to trigger the linking to librt
(build R with profiling, which is the default?)
Anyway, as Dirk said, linking to librt is harmless on Linux nonetheless, so
that is perhaps the easiest 'fix'.
Best,
Yan Zhou
On Dec 24, 2012, at 1:11 PM, Dirk Eddelbuettel <edd at debian.org> wrote:
On 23 December 2012 at 18:31, Martin Morgan wrote: | When trying to install
Rcpp 0.10.2 with | | $ R --version | R Under development (unstable)
(2012-12-18 r61372) -- "Unsuffered Consequences" | $ clang++ --version |
Ubuntu clang version 3.0-6ubuntu3 (tags/RELEASE_30/final) (based on LLVM
3.0) | Target: x86_64-pc-linux-gnu | Thread model: posix
I just upgraded my r-devel version today; the one I used to test 0.10.2
was probably a week or at the most two weeks old. I did quite a few R CMD
check runs; they all passed -- but I only used g++.
Methinks I could have caught that if I had used clang.
| I get | | $ R CMD INSTALL Rcpp_0.10.2.tar.gz | ... | clang++
-I/home/mtmorgan/bin/R-devel/include -DNDEBUG -I../inst/include/ |
-I/usr/local/include -fpic -g -O2 -c Timer.cpp -o Timer.o | ... | **
testing if installed package can be loaded | Error in dyn.load(file,
DLLpath = DLLpath, ...) : | unable to load shared object |
'/home/mtmorgan/R/x86_64-unknown-linux-gnu-library/2.16/Rcpp/libs/Rcpp.so':
| /home/mtmorgan/R/x86_64-unknown-linux-gnu-library/2.16/Rcpp/libs/Rcpp.so:
| undefined symbol: clock_gettime | Error: loading failed | | From man
clock_gettime I have | | ... | Link with -lrt. | | and adding
PKG_LIBS += -lrt to src/Makevars addresses the problem for me, but | I'm
not sure that is a universal solution? Also, I had an identical problem
with | another package (Matrix, since corrected) so there is perhaps
something | a-typical about my configuration.
I don't think so. The internal-to-Rcpp Timer class is new in 0.10.2. And
Yan is (as usual) correct, -lrt is AFACR specific to Linux. I needed that
in the past when using the kernel-based high-res timer.
So this is something we need to figure out. So we seem to know that
r-devel and clang need it. Can you check if r-release needs it? With g++,
neither R versions seems to need it.
OTOH I don't think adding -lrt does any harm so maybe we'll add "in
general". Might run this by Simon too.... This needs more work.
Thanks so much for the heads-up!
Dirk
| | Martin | -- | Computational Biology / Fred Hutchinson Cancer Research
Center | 1100 Fairview Ave. N. | PO Box 19024 Seattle, WA 98109 | |
Location: Arnold Building M1 B861 | Phone: (206) 667-2793 |