Building R-devel from Source on Snow Leopard

Hi Simon, Tony, and anybode else who attempts to complie 2.9.2 for
themselves on Snow Leopard,

The configure line which eventually led to R configuring, building,  
and
running was this:

./configure --prefix=/usr/local --with-x --disable-R-framework -- 
enable-threads=posix --without-aqua --with-lapack=" /usr/local/lib/ 
liblapack.a /usr/local/lib/libptf77blas.a /usr/local/lib/ 
libptcblas.a -lpthread /usr/local/lib/libatlas.a " --with-blas=" / 
usr/local/lib/liblapack.a /usr/local/lib/libptf77blas.a /usr/local/ 
lib/libptcblas.a -lpthread /usr/local/lib/libatlas.a " CFLAGS="-O3 - 
fopenmp -mtune=native -m64" CXXFLAGS="-O3 -fopenmp -mtune=native - 
m64" FFLAGS="-O3 -fopenmp -mtune=native -m64" LDFLAGS="-m64 -L/sw/ 
lib -L/usr/X11R6/lib -L/usr/lib/gcc/i686-apple-darwin9/4.2.1/x86_64 - 
lgfortran" F77="gfortran -arch x86_64" CPPFLAGS="-I/sw/include"  
FC="gfortran -arch x86_64" FCFLAGS="-O3 -fopenmp -mtune=native -m64"

...dropping the --with-blas and --with-lapack items also lets R  
compile fine
(but see benchmarks below). Note also that I have libreadline and  
libiconv
installed in the /sw tree using 64-bit Fink for Snow  
Leopard...you'll need
64bit versions of readline and iconv to compile R; check out
/Developer/SDKs/MacOSX10.6.sdk/usr/lib maybe.

The one oddity is that when I pass

--with-blas="-L/usr/local/lib -llapack -lptf77blas -lptcblas - 
lpthread -latlas"
(and --with-lapack the same, or one and not the other)

...the build fails in the manner described in my last message  
(lapack.so,
compiled by R, ends up not containing the functions it should),  
while if I
explicitly list the libraries:

--with-blas=" /usr/local/lib/liblapack.a /usr/local/lib/ 
libptf77blas.a /usr/local/lib/libptcblas.a -lpthread /usr/local/lib/ 
libatlas.a

...it builds successfully. Not sure what this portends.

Also, I have done some benchmarking to see how much faster R is (on  
my Mac
Pro, dual quad-core Xeon @ 2.8GHz, 6G RAM) when I go through all this
trouble:

(benchmarked using http://r.research.att.com/benchmarks/R-benchmark-25.R 
)

nightly build
(http://r.research.att.com/R-2.9-branch-leopard-universal.tar.gz,  
built Sept
5 2009, run with --arch x86_64):
Overall mean (sum of I, II and III trimmed means/3)_ (sec):   
0.835276859437347

my build, no BLAS or LAPACK:
Overall mean (sum of I, II and III trimmed means/3)_ (sec):   
1.83643467491601 (This build did consistently better on non-matrix  
math, FAR worse on the matrix math)

my build, BLAS, no LAPACK:
Overall mean (sum of I, II and III trimmed means/3)_ (sec):   
0.760949143043198 (outperforms nightly on most metrics,  
underperforms on none, outperforms no-blas on all)

my build, LAPACK & BLAS:
Overall mean (sum of I, II and III trimmed means/3)_ (sec):   
0.70660451743888 (outperforms everything else on most metrics,  
underperforms on none)

So it seems that letting ATLAS compile your BLAS and LAPACK (Atlas  
3.9.14
does this), plus adding optimization flags, leads to a 15%  
improvement over
the nightly build (when using this specific benchmarking system, ymmv
depending on what functions you actually use R for).

--Adam

On Fri, 4 Sep 2009, Adam D. I. Kramer wrote:

Simon,

	Thanks for your comments; responses below.

On Fri, 4 Sep 2009, Simon Urbanek wrote:

Adam,
On Sep 4, 2009, at 5:52 PM, Adam D. I. Kramer wrote:

Hi Tony and Simon,

	I was having this trouble today/yesterday and have been following
along on this thread. The solution I found was to add this to my
./configure:
LDFLAGS="-m64 -L/usr/lib/gcc/i686-apple-darwin9/4.2.1/x86_64 - 
lgfortran"
and
F77="gfortran -arch x86_64"

You should not touch LDFLAGS and besides F77 you'll also need to  
set FC.

From configure --help:

LDFLAGS     linker flags, e.g. -L<lib dir> if you have libraries in a
            nonstandard directory <lib dir>

...the gfortran library lives in a nonstandard directory, which is  
why I put
the -L there... so I then moved -lgfortran to LIBS='-lgfortran' and  
things
broke in the same way, as did moving the -L...-lgfortran there.   
Both of
these need to be passed to the linker when mixing C/Fortran code.

It also throws the error "Maybe check LDFLAGS for paths to Fortran
libraries?" which is what tipped me off to the fact that I should  
provide a
link to the Fortran libraries in LDFLAGS.  If this is incorrect,  
please let
me know where to put them instead (and consider updating the  
configure
script).

...this gets me through the "can't make mixed C/Fortran code"  
errors. However, through one hoop of fire and I'm confronted by  
another:
checking for iconv.h... yes
checking for iconv... no
checking for iconvlist... no
configure: error: --with-iconv=yes (default) and a suitable iconv  
is not
available
...this strikes me as odd. The relevant lines from the config.log  
file are:
configure:39212: gcc -std=gnu99 -o conftest -O3 -fopenmp - 
mtune=native -m64 -m64 -L/usr/lib/gcc/i686-apple-darwin9/4.2.1/ 
x86_64 -lgfortran conftest.c -lm >&5
conftest.c: In function 'main':
conftest.c:189: warning: passing argument 1 of 'libiconvlist' from
incompatible pointer type
Undefined symbols:
"_libiconvlist", referenced from:
   _main in cc2XgkAg.o
ld: symbol(s) not found
collect2: ld returned 1 exit status
However:
swiss:R-2.9.2$ nm /usr/lib/libiconv.dylib | grep _libiconvlist
00008807 T _libiconvlist
...I'm not sure where else configure could be trying to find  
libiconv,

check /usr/local/lib that is the most common problem area ...

I had, and there was no libiconv there.

I eventually solved this problem by upgrading fink to the 64bit  
version and
installing libiconv and adding CPPFLAGS="-I/sw/include" and adding - 
L/sw/lib
to LDFLAGS, forcing it to link against the new library. It seems  
like the
Apple libraries should serve the purpose, but for whatever reason  
they did
not.

My current problem is that, during make, R complains about the  
lapack.so
file it has compiled from my lapack libraries (I configured with
--with-lapack="-L/usr/local -llapack -lptf77blas -lcblas -lpthread - 
latlas"
and a similar --with-blas):

Warning in solve.default(rgb) :
unable to load shared library '/Volumes/Tubby2/nobackup/R-2.9.2/ 
modules//lapack.so':
dlopen(/Volumes/Tubby2/nobackup/R-2.9.2/modules//lapack.so, 6):  
Symbol not found: _cblas_cdotc_sub
Referenced from: /Volumes/Tubby2/nobackup/R-2.9.2/modules//lapack.so
Expected in: flat namespace
in /Volumes/Tubby2/nobackup/R-2.9.2/modules//lapack.so
Error in solve.default(rgb) : lapack routines cannot be loaded
Error: unable to load R code in package 'grDevices'
Execution halted

...that said, I understand that --with-lapack is "not recommended"  
in the
admin manual, so I'm pretty much on my own here, but I'd be  
interested if
you have an idea or two about what might be going on.  
_cblas_cdotc_sub is
indeed in my lapack.a file (but not the lapack.so file R refers  
to), and I
can use _cblas_cdotc if I write some simple c code and link it  
against
lapack using the same flags as noted above.

My interest, really, is in whether ATLAS's LAPACK-tuning system  
(which is
pretty new) actually leads to speed improvements in R.  So, I don't
desperately need to use it, but am still concerned that a compile  
that used
to run smoothly no longer does.

--Adam

Cheers,
Simon

but
running "locate libiconv" shows versions in
/Developer/SDKs/MacOSX10.6.sdk/usr/lib/ also, which also show  
_libiconvlist
when I nm them.
Any suggestions? I'm not sure how to further debug this...the  
symbols are
there.
Cordially,
Adam

Hi Simon,
I think that you might have resolved my issue. I did not specify  
the arch,
so I will try that and let everyone know if that works.
Cheers,
--Tony
On Thu, Sep 3, 2009 at 4:11 PM, Simon Urbanek
<simon.urbanek at r-project.org>wrote:

Tony,
On Sep 3, 2009, at 6:14 PM, Tony Chiang wrote:
So I am trying to configure and build the latest of R-devel on  
one of the

new Macbook Pro's running Snow leopard. I have installed the  
latest
X-Code
tools (downloaded from the Apple Site) and have gcc installed:
gcc --versioni686-apple-darwin10-gcc-4.2.1 (GCC) 4.2.1 (Apple  
Inc. build
5646)
So as I am trying to configure, I get this error message:

What is the exact configure command you're running? I suspect  
that your
problem is a 32-bit vs 64-bit mismatch, because SL changed the  
gcc default
to 64-bit whereas the gfortran default is 32-bit, so you have  
to make sure
your archs match (depending on which you actually want to  
build) - ideally
you should specify -arch for all compilers (as the CRAN binary  
does - also
see the FAQ).
Cheers,
Simon
...cut...

checking for Fortran 77 libraries of gfortran...  -L/usr/local/ 
lib

-L/usr/local/lib/gcc/i686-apple-darwin8/4.2.3
-L/usr/local/lib/gcc/i686-apple-darwin8/4.2.3/../../.. - 
lgfortranbegin
-lgfortran
checking how to get verbose linking output from gcc - 
std=gnu99... -v
checking for C libraries of gcc -std=gnu99...  -lcrt1.10.6.o
-L/usr/local/lib -L/usr/lib/gcc/i686-apple-darwin10/4.2.1/ 
x86_64
-L/usr/lib/i686-apple-darwin10/4.2.1
-L/usr/lib/gcc/i686-apple-darwin10/4.2.1

-L/usr/lib/gcc/i686-apple-darwin10/4.2.1/../../../i686-apple- 
darwin10/4.2.1
-L/usr/lib/gcc/i686-apple-darwin10/4.2.1/../../.. -lSystem
checking for dummy main to link with Fortran 77 libraries...  
rm:
conftest.dSYM: is a directory
none
checking for Fortran 77 name-mangling scheme... rm:  
conftest.dSYM: is a
directory
unknown
configure: WARNING: unknown Fortran name-mangling scheme
checking whether gfortran appends underscores to external  
names...
unknown
configure: error: cannot use Fortran

I have not seen any messages quite like this on the mailing  
list. What is
pretty strange is the conftest.dSYM needs to be removed (?)  
for some
reason
but is a directory. Any ideas or suggestions on how to resolve  
this?
Best,
--Tony

    [[alternative HTML version deleted]]