[Rcpp-devel] Optimization flag may be needed in Makevars.win for 4.9.2 toolchain

[Avraham Adler]

The -march is local, of course. The question is should Rcpp always
enforce at least -O2 or the like? From your answer I guess not.

The reason why not to simply have an optimization flag in the local
makevars is that, at least in my experience, some code is faster with
O2 and some with O3, and by having one in makevars, I am overriding
the cases where the package author (who knows better) is passing a
different flag. Of course one can just add that flag and install Rcpp
by itself.

Thanks again,

Avi

On Tue, Aug 25, 2015 at 5:50 PM, Dirk Eddelbuettel <edd at debian.org> wrote:
>
> On 25 August 2015 at 17:39, Avraham Adler wrote:
> | I've been testing a GCC 4.9.2-based toolchain for Windows for the last
> | week or so, and today fell afoul of the error uncovered here
> | <https://stat.ethz.ch/pipermail/r-devel/2015-March/070814.html>. I've
> | been passing different optimization flags, and Rcpp compiled properly
> | under every one EXCEPT when a -march (ivybridge) flag was passed
> | without an optimization flag. In that case, I received the
> | "collect2.exe: error: ld returned 5 exit status" error as discussed on
> | R-devel. Passing a -mtune flag without an -O flag did not cause an
> | error.
> |
> | Downloading the Rcpp source code and manually adding -O2 to Rcpp's
> | Makevars.win, re-tar/gzipping it and installing the local version
> | allowed Rcpp to compile properly.
> |
> | I don't know if it is better to tell people to change their local
> | Makevars under HOME/.R to pass -O2 or -O3 (that would probably be a
> | call by you, Dirk, as you would know best if O3 would help or hinder),
> | or if it is better to have Rcpp enforce it, but it should be noted
> | somewhere once a 4.9.2-based Windows is eventually released.
>
> CRAN disallows non-portable flags such as -march.
>
> So yes, ~/.R/Makevars it is.
>
> Dirk
>
> --
> http://dirk.eddelbuettel.com | @eddelbuettel | edd at debian.org