Windows build issue

[eddelbuettel]

Reported on rcpp-devel in this post

@yixuan Any thoughts?

[dcdillon]

I must be an idiot but have browsed all the docs for this issue and I can't find a usable error message. Can someone clarify what the error message actually is?

[eddelbuettel]

If you follow the first link in the email by Francois Rousset you get to the CRAN tests results for Eigen which flashes a bright red for Windows and this link. Now that's of course just R CMD check and once again not the actual error.

Which we still need to reproduce.

[dcdillon]

Gotcha. Thanks
On May 22, 2016 3:13 PM, "Dirk Eddelbuettel" notifications@github.com
wrote:

If you follow the first link in the email by Francois Rousset you get to
the CRAN tests results for Eigen which flashes a bright red for Windows and this
link
https://www.r-project.org/nosvn/R.check/r-devel-windows-ix86+x86_64/RcppEigen-00check.html.
Now that's of course just R CMD check and once again not the actual
error.

Which we still need to reproduce.

—
You are receiving this because you commented.
Reply to this email directly or view it on GitHub
#34 (comment)

[kevinushey]

FWIW, I can't reproduce the R CMD check failure on my Windows machine (32bit R, using R-patched 3.3.0, r70654; dev version of RcppEigen) :-/

> sessionInfo()
R version 3.3.0 Patched (2016-05-20 r70654)
Platform: i386-w64-mingw32/i386 (32-bit)
Running under: Windows >= 8 x64 (build 9200)

[yixuan]

Hi everyone, now I can confirm that this issue CAN be reproduced when -O3 is defined AND -DNDEBUG is NOT defined in the compilation flag.

[eddelbuettel]

@yixuan you are very, very impressive.

Now where do we go from here? I vaguely recall that -DNDEBUG was recommended already. Behaviour that varies with/without certain optimization flags is more commonly compiler-related.

[aadler]

Lurking here, but core R for Windows (which mainly means recommended packages) overwrites -O3 with -O2 for C++ (lines 74 and 77 of R_HOME/src/gnuwin32/fixed/etc/Makeconf) and has done so for a very long time—not just for the gcc 4.9.3 based Rtools. As there may be other package issues with -O3, would suggesting -O2 for the time being be appropriate?

[kevinushey]

Here's the (mostly unhelpful) output from WinDbg:

(2e40.2550): Access violation - code c0000005 (first chance)
First chance exceptions are reported before any exception handling.
This exception may be expected and handled.
*** ERROR: Symbol file could not be found.  Defaulted to export symbols for C:\Users\kevin\R\win-library\3.3\RcppEigen\libs\i386\RcppEigen.dll - 
eax=2c28d000 ebx=70daf5ce ecx=6e2c28d0 edx=00000002 esi=000001ce edi=6e2c28d0
eip=6e2b5685 esp=0141b668 ebp=00000000 iopl=0         nv up ei pl zr na pe nc
cs=0023  ss=002b  ds=002b  es=002b  fs=0053  gs=002b             efl=00010246
RcppEigen!ZN5Eigen8internal11gebp_kernelIddiLi2ELi2ELb0ELb0EEclEPdiPKdS5_iiidiiiiS3_+0x275:
6e2b5685 dd00            fld     qword ptr [eax] ds:002b:2c28d000=????????????????

 # ChildEBP RetAddr  
WARNING: Stack unwind information not available. Following frames may be wrong.
00 00000000 00000000 RcppEigen!ZN5Eigen8internal11gebp_kernelIddiLi2ELi2ELb0ELb0EEclEPdiPKdS5_iiidiiiiS3_+0x275

The demangled name (in case someone wanted to dive in)

C:\Users\kevin>c++filt -n _ZN5Eigen8internal11gebp_kernelIddiLi2ELi2ELb0ELb0EEclEPdiPKdS5_iiidiiii
Eigen::internal::gebp_kernel<double, double, int, 2, 2, false, false>::operator()(double*, int, double const*, double const*, int, int, int, double, int, int, int, int)

IMHO the best solution is for us to just force -O2 for compilation.

[eddelbuettel]

I just got this email but it was probably tickled by someone else :)

Date: Mon, 23 May 2016 16:47:06 +0200                                                                                                                                                                               
From: Uwe.Ligges@R-Project.org                                                                                                                                                                                      
To: edd@debian.org                                                                                                                                                                                                  
Cc: Uwe.Ligges@R-Project.org                                                                                                                                                                                        
Subject: winbuilder: Package RcppEigen_666.3.2.8.1.tar.gz has been checked and built                                                                                                                                

Dear package maintainer,

this notification has been generated automatically.
Your package RcppEigen_666.3.2.8.1.tar.gz has been built (if working) and checked for Windows.
Please check the log files and (if working) the binary package at:
http://win-builder.r-project.org/QlY633B1xgv5
The files will be removed after roughly 72 hours.
Installation time in seconds: 61
Check time in seconds: 364
Status: 2 NOTEs
R Under development (unstable) (2016-05-23 r70656)

All the best,
Uwe Ligges
(CRAN maintainer of binary packages for Windows)

[f-rousset]

Hi,

yes I submitted to win-builder RcppEigen_666.3.2.8.1.tar.gz, which is RcppEigen_0.3.2.8.1.tar.gz with -O2 and -DNDEBUG added in Makevars.win, to check whether it would pass the tests that the current version does not. I forgot to change the maintainer email address, so you can see that this works.

Thanks to everyone for caring about this issue.

F.

[kevinushey]

Another option is to force -fno-omit-frame-pointer in Makevars.win as this seems to sidestep the compilation issue when -O3 and -UNDEBUG are both applied (perhaps it indirectly disables some optimization that is breaking here?)

[f-rousset]

I just tried -fno-omit-frame-pointer on win-builder and it indeed works.

(For definiteness: I modified he RcppEigen``Makevars.win to contain

ifeq "${R_ARCH}" "/i386"
PKG_CXXFLAGS= -I../inst/include -fno-omit-frame-pointer
else
PKG_CXXFLAGS = -I../inst/include 
endif

)

[eddelbuettel]

You can add it unconditionally. It won't do harm for x86_64.

I can't quite recall where but we used it before methinks. Or maybe I just kept it around in ~/.R/Makevars.

[aadler]

@f-rousset Have you tried -O2 without -DNDEBUG?

[f-rousset]

@aadler -O2 without -DNDEBUG fails to pass the tests for RcppEigen. It fails in the same way as the current RcppEigen version, as far as the win-builder logs show.

[aadler]

@f-rousset Thanks. For what it is worth, in the hidden machinations of R for windows (R_HOME/src/gnuwin32/fixed/etc/Makeconf, otherwise known as "Here Be Monsters") has "R_XTRA_CPPFLAGS = -I"$(R_HOME)/include" -DNDEBUG" on line 165.

[f-rousset]

I have now had an exchange with Uwe Ligges about the compilation flags on win-builder and his latest response can be summarized as "ARGH!". We can expect that they will be fixed back to the expected -DNDEBUG and -O2.

[eddelbuettel]

Very nice -- thanks so much for reporting back and getting it settled upstream at winbuilder!

[eddelbuettel]

And I just got email from Uwe (or rather his bot) that a new RcppEigen has been built.

[yixuan]

Great! Thank everyone who provided help.

[eddelbuettel]

And now got email from Uwe himself, so I think we can close this.

Kudos to @f-rousset for following this through.