An update on the windows toolchain developments
On Mon, Aug 31, 2015 at 9:22 AM, Avraham Adler <avraham.adler at gmail.com> wrote:
Hello, Jeroen. Regarding the Rf_error, does the issue still exist in v4 of GCC 4.9.2 and is a regression in 4.9.3, or has it never been solved in 4.9.2 either? Regarding the Rcpp error, wouldn't throwing no-sse also prevents the use of almost all vecotrizations (sse2-4.1 and avx/avx2), slowing results down noticeably? Avi On Mon, Aug 31, 2015 at 9:14 AM, Jeroen Ooms <jeroenooms at gmail.com> wrote:
A new version of the toolchain based on gcc 4.9.3 and corresponding build of R-devel are available from [1]. See the readme.txt [2] for details. The installer should work out of the box and make it easy for Windows users to compile and test R packages with the new toolchain. The bad news: unfortunately it appears that the sjlj problem on win64 (triggered by Rf_error) is still not completely solved in the latest mingw-w64. Therefore I restored -fno-asynchronous-unwind-tables in CXXFLAGS in the latest build of R. This workaround allows us to move forward while we investigate this problem. Even with -fno-asynchronous-unwind-tables there are still two Rcpp unit tests that crash on win64: test.Module.member and test.Module.property. These tests segfault when an exception is raised within a module. The problem still exists in the latest Rcpp version from Github. It is unclear if this is related to the sjlj bug above, or a separate problem in Rcpp / R-devel. The issue can be worked around by changing Rcpp's Makevars.win to: PKG_CXXFLAGS=-mno-sse With this fix, Rcpp is able to execute all unit tests on both win32 and win64. All tests pass, except for an issue with dates (also only on win64) that is reported at [3]. [1] http://www.stat.ucla.edu/~jeroen/mingw-w64/ [2] http://www.stat.ucla.edu/~jeroen/mingw-w64/readme.txt [3] https://github.com/RcppCore/Rcpp/issues/359
_______________________________________________ R-SIG-windows mailing list R-SIG-windows at r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-windows
Is this error [1] dependent on using RToolsTest (for which I get
package ?RToolTest? is not available (for R version 3.2.2 Patched) )?
In my test compilation of R and Rcpp using the 4.9.2 toolchain, given:
#include <Rcpp.h>
using namespace Rcpp;
// [[Rcpp::export]]
SEXP ctor_mdy() {
Date dt = Date(12,31,2005);
return wrap(dt);
}
// [[Rcpp::export]]
SEXP ctor_ymd() {
Date dt = Date(2005,12,31);
return wrap(dt);
}
I get
ctor_mdy()
[1] "2005-12-31"
ctor_ymd()
[1] "2005-12-31" If it matters, EOPTS in MkRules.local were -march=ivybridge -mtune=ivybridge --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256 -pipe .R/Makevars CXXFLAGS= -std=gnu++11 -O2 -march=ivybridge -mtune=ivybridge --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256 -pipe Thank you, Avi [1] https://github.com/RcppCore/Rcpp/issues/359