Segfault from C stack overflow for mips64el and risc64 architectures

[tillea]

Summary:

As per Debian bug report rstanarm does not build for mips64el and risc64 architectures.

Description:

As you can see at the end of the build log for mips64el the package fails to build. The relevant part of the log is probably

g++ -std=gnu++17 -I"/usr/share/R/include" -DNDEBUG -I"../inst/include" -I"/usr/lib/R/site-library/StanHeaders/include/src" -DBOOST_DISABLE_ASSERTS -DEIGEN_NO_DEBUG -DBOOST_MATH_OVERFLOW_ERROR_POLICY=errno_on_error -      DUSE_STANC3 -D_HAS_AUTO_PTR_ETC=0 -I'/usr/lib/R/site-library/StanHeaders/include' -I'/usr/lib/R/site-library/rstan/include' -I'/usr/lib/R/site-library/BH/include' -I'/usr/lib/R/site-library/Rcpp/include' -I'/usr/lib/R/site-  library/RcppEigen/include' -I'/usr/lib/R/site-library/RcppParallel/include'    -I/usr/include -DTBB_INTERFACE_NEW -I'/usr/lib/R/site-library/RcppParallel/include' -D_REENTRANT -DSTAN_THREADS -DTBB_INTERFACE_NEW    -fpic  -g -O2 -ffile-prefix-map=/<<BUILDDIR>>/r-base-4.3.2=. -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2  -c stan_files/count.cc -o stan_files/count.o
In file included from /usr/include/boost/math/special_functions/detail/bessel_jy.hpp:14,
                 from /usr/include/boost/math/special_functions/bessel.hpp:20,
                 from /usr/lib/R/site-library/StanHeaders/include/stan/math/prim/fun/bessel_first_kind.hpp:6,
                 from /usr/lib/R/site-library/StanHeaders/include/stan/math/prim/fun.hpp:31,
                 from /usr/lib/R/site-library/StanHeaders/include/stan/math/prim.hpp:14,
                 from /usr/lib/R/site-library/StanHeaders/include/src/stan/io/dump.hpp:7,
                 from /usr/lib/R/site-library/rstan/include/rstan/stan_fit.hpp:43,
                 from /usr/lib/R/site-library/rstan/include/rstan/rstaninc.hpp:4,
                 from stan_files/count.hpp:18,
                 from stan_files/count.cc:3:
/usr/include/boost/math/special_functions/gamma.hpp: In instantiation of ‘boost::math::detail::upper_incomplete_gamma_fract<T>::result_type boost::math::detail::upper_incomplete_gamma_fract<T>::operator()() [with T =     double; result_type = std::pair<double, double>]’:
/usr/include/boost/math/tools/fraction.hpp:217:20:   required from ‘typename boost::math::tools::detail::fraction_traits<Gen>::result_type boost::math::tools::continued_fraction_a(Gen&, const U&, uintmax_t&) [with Gen =  boost::math::detail::upper_incomplete_gamma_fract<double>; U = double; typename detail::fraction_traits<Gen>::result_type = double; uintmax_t = long unsigned int]’
/usr/include/boost/math/tools/fraction.hpp:252:31:   required from ‘typename boost::math::tools::detail::fraction_traits<Gen>::result_type boost::math::tools::continued_fraction_a(Gen&, const U&) [with Gen = boost::math::detail::upper_incomplete_gamma_fract<double>; U = double; typename detail::fraction_traits<Gen>::result_type = double]’
/usr/include/boost/math/special_functions/gamma.hpp:314:68:   required from ‘T boost::math::detail::upper_gamma_fraction(T, T, T) [with T = double]’
/usr/include/boost/math/special_functions/gamma.hpp:1176:44:   required from ‘T boost::math::detail::gamma_incomplete_imp(T, T, bool, bool, const Policy&, T*) [with T = double; Policy = boost::math::policies::            policy<boost::math::policies::pole_error<boost::math::policies::errno_on_error>, boost::math::policies::promote_float<false>, boost::math::policies::promote_double<false>, boost::math::policies::default_policy, boost::math:: policies::default_policy, boost::math::policies::default_policy, boost::math::policies::default_policy, boost::math::policies::default_policy, boost::math::policies::default_policy, boost::math::policies::default_policy,     boost::math::policies::default_policy, boost::math::policies::default_policy, boost::math::policies::default_policy>]’
/usr/include/boost/math/special_functions/gamma.hpp:2130:35:   required from ‘boost::math::tools::promote_args_t<RT1, RT2> boost::math::gamma_p(RT1, RT2, const Policy&) [with RT1 = double; RT2 = double; Policy =          policies::policy<policies::overflow_error<boost::math::policies::errno_on_error>, policies::pole_error<boost::math::policies::errno_on_error>, policies::promote_double<false>, policies::digits2<0>, policies::default_policy,  policies::default_policy, policies::default_policy, policies::default_policy, policies::default_policy, policies::default_policy, policies::default_policy, policies::default_policy, policies::default_policy>; tools::         promote_args_t<RT1, RT2> = double]’
/usr/lib/R/site-library/StanHeaders/include/stan/math/prim/fun/gamma_p.hpp:76:30:   required from here
/usr/include/boost/math/special_functions/gamma.hpp:299:16: note: the ABI for returning a value with C++17 empty bases but otherwise an aggregate with only one or two floating-point fields was changed in GCC 12.1
  299 |    result_type operator()()
      |                ^~~~~~~~
"/usr/lib/R/bin/Rscript" -e "source(file.path('..', 'tools', 'make_cc.R')); make_cc(commandArgs(TRUE))" stan_files/jm.stan
code for methods in class "Rcpp_model_base" was not checked for suspicious field assignments (recommended package 'codetools' not available?)
code for methods in class "Rcpp_model_base" was not checked for suspicious field assignments (recommended package 'codetools' not available?)
code for methods in class "Rcpp_stan_fit" was not checked for suspicious field assignments (recommended package 'codetools' not available?)
code for methods in class "Rcpp_stan_fit" was not checked for suspicious field assignments (recommended package 'codetools' not available?)
Error: segfault from C stack overflow
Execution halted
make[1]: *** [Makevars:24: stan_files/jm.cc] Error 1

Reproducible Steps:

I've just got a hint about some dockerfile you could build with --platform=linux/mips64le or --platform=linux/risc64

FROM docker.io/debian:sid-slim

WORKDIR /build

RUN apt-get update && apt-get install -y --no-install-recommends devscripts curl debian-keyring
RUN dget http://ftp.debian.org/debian/pool/main/r/r-cran-rstanarm/r-cran-rstanarm_2.32.1-1.dsc
WORKDIR /build/r-cran-rstanarm-2.32.1
RUN apt-get -y build-dep .
RUN dpkg-buildpackage -uc -us -b

Please note: I have not tested this way to use docker (just copied what others tried with some other package).

RStanARM Version:

2.32.1-1

R Version:

4.3.2

Operating System:

Debian (sid)

[bgoodri]

Thanks. We don't test Stan on less common archs but maybe @SteveBronder has some idea what we need to change in StanHeaders?

[SteveBronder]

From the line here

/usr/include/boost/math/special_functions/gamma.hpp:299:16: note: the ABI for returning a value with C++17 empty bases but otherwise an aggregate with only one or two floating-point fields was changed in GCC 12.1
  299 |    result_type operator()()

Is this an issue with Boost's BH R package?

[bgoodri]

Possibly an issue with Boost. Those Debian packages are using the system boost (which seems to be 1.83) rather than the one in the BH package (which is 1.84) but I doubt that version difference is relevant to gamma.hpp which has barely changed in the last couple of years.

[WardBrian]

The command that's failing is make stan_files/jm.cc, which as far as I can tell just runs stanc3 and then the Rccp expose method.

Since it's jm.stan (a model I have been using to benchmark stanc3 since I know it's a bear) my guess is that the stack overflow is happening inside the JS VM running stanc. @tillea - do you know if this is running in quickjs or V8?

[tillea]

Am Wed, Apr 10, 2024 at 07:24:39AM -0700 schrieb Brian Ward:

@tillea - do you know if this is running in `quickjs` or `V8`?

Unfortunately I have no idea. In the log I linked to when opening this issue I find r-cran-v8 is mentioned but only r-cran-quickjsr_1.1.3-1_riscv64.deb is really installed. Maybe that's a hint - but I have no idea about those details. Hope that helps, Andreas.

[WardBrian]

Patching to install V8 is what I'd try next. RStan should automatically switch over, and hopefully that resolves it. I don't know enough about debian packing to know how to try this myself, but I am able to run that docker file locally so if you can think of a command which I could add to do so, I can try it.

[bgoodri]

Indeed, V8 parses complicated Stan programs more reliably than quickjs does, but V8 is more difficult to build for normal users, so rstan depends on quickjs but uses V8 if it happens to be installed. On Debian, or any binary linux repository, V8 should already be packaged by someone.

[WardBrian]

I did some manual hacking around inside docker (moved V8 up from Suggests, manually ran apt install r-cran-v8) and I'm trying to re-run dpkg-buildpackage now.

[tillea]

Am Wed, Apr 10, 2024 at 08:34:33AM -0700 schrieb Brian Ward:

I did some manual hacking around inside docker (moved `V8` up from `Suggests`, manually ran `apt install r-cran-v8`) and I'm trying to re-run `dpkg-buildpackage` now.

I'll happily add r-cran-v8 to the Build-Depends if this might solve the problem. Kind regards, Andreas.

[WardBrian]

It is still running, but it has already progressed past the point it was previously failing

[WardBrian]

The build using r-cran-v8 succeeded

[tillea]

Am Wed, Apr 10, 2024 at 12:36:53PM -0700 schrieb Brian Ward:

The build using `r-cran-v8` succeeded

Succeeds in Debian as well. Thanks a lot, Andreas.

[WardBrian]

@tillea - glad to hear it!

I noticed you are also the debian packager for r-cran-rstan -- I would recommend adding r-cran-v8 to the depends for that package as well, as it would benefit users to have that packaged, as @bgoodri mentions

[tillea]

Am Thu, Apr 11, 2024 at 07:55:34AM -0700 schrieb Brian Ward:

I noticed you are also the debian packager for `r-cran-rstan` -- I would recommend adding `r-cran-v8` to the depends for that package as well, as it would benefit users to have that packaged, as @bgoodri mentions

Just uploaded. Thanks a lot for the hint, Andreas.