[Rcpp-devel] RcppEigen fails on win-builder i386, and a possibly related issue.

[François Rousset]

Le 22/05/2016 à 23:59, Dirk Eddelbuettel a écrit :
> On 22 May 2016 at 23:31, François Rousset wrote:
> | Le 22/05/2016 à 22:51, Dan Dillon a écrit :
> |
> |
> |     Do you have the same version of mingw as cran?  If not , that would be a
> |     good test
> |
> | I have installed the Rtools 3.3 as described there :
> | https://github.com/rwinlib/r-base
> | which is linked from the "R-devel-experimental" download page
> | https://cran.r-project.org/bin/windows/base/rdevel.html
> | so I am using gcc 4.9.3, and win-builder too.
> |
> | The full win-builder logs for 'minimalTest' are accessible (for ca. 72h) here :
> | http://win-builder.r-project.org/u84hzB80Fol0/
>
> There is almost nothing there. It builds, and then on i386 we get a run-time
> error and it shows is
>
>    > mini(14)
>    Trivial code: swZ.selfadjointView<Lower>().rankUpdate(swZ):
>
> from http://win-builder.r-project.org/u84hzB80Fol0/examples_and_tests/minimalTest-Ex_i386.Rout
> and the summary in http://win-builder.r-project.org/u84hzB80Fol0/00check.log
>
> Can you expand on it?  Decompose the expressesion?  Compare it to alternate
> versions of the same expression?
I don't know how to decompose it further in more elementary steps. I can 
surely compute a cross-product in a different way.
However, having submitted 131 versions of packages to win-builder in the 
last week in order to produce the small example, and given how slow 
using win-builder makes any debugging attempt, I wish someone would be 
able to reproduce the problem on another machine, or else to give me 
more specific advice as to what alternative versions of the code would 
be most useful to test, because I feel have reached the point where my 
limited understanding of Eigen and RcppEigen's internals makes the blind 
testing exercise too inefficient.

F.