stanc --O flag fails to build simple Stan model

[andrewfowlie]

Summary:

Compilation error on simple Stan model only when using --O stanc flag.

Description:

The --O flag generates a model header including a stan::model::assign call that doesn't appear to have a matching implementation.

Reproducible Steps:

For the simple model bug.stan

parameters {
  vector[1] x;
}
transformed parameters {
  vector[1] theta = 1 * x;
}
model {
    target += theta[1];  # using theta here seems necessary for the bug
}

the build

make bug   # works

works as expected, whereas

rm bug  # start from fresh
make bug STANCFLAGS=--O  # fails

fails with (partial) output

g++ -std=c++17 -pthread -D_REENTRANT -Wno-sign-compare -Wno-ignored-attributes -Wno-class-memaccess      -I stan/lib/stan_math/lib/tbb_2020.3/include    -O3 -I src -I stan/src -I stan/lib/rapidjson_1.1.0/ -I lib/CLI11-1.9.1/ -I stan/lib/stan_math/ -I stan/lib/stan_math/lib/eigen_3.4.0 -I stan/lib/stan_math/lib/boost_1.84.0 -I stan/lib/stan_math/lib/sundials_6.1.1/include -I stan/lib/stan_math/lib/sundials_6.1.1/src/sundials    -DBOOST_DISABLE_ASSERTS          -c -Wno-ignored-attributes   -x c++ -o bug.o bug.hpp
bug.hpp: In instantiation of ‘stan::scalar_type_t<T2> bug_model_namespace::bug_model::log_prob_impl(VecR&, VecI&, std::ostream*) const [with bool propto__ = false; bool jacobian__ = false; VecR = Eigen::Matrix<stan::math::var_value<double>, -1, 1>; VecI = Eigen::Matrix<int, -1, 1>; stan::require_vector_like_t<VecR>* <anonymous> = 0; stan::require_vector_like_vt<std::is_integral, VecI>* <anonymous> = 0; stan::require_st_var<VecR>* <anonymous> = 0; stan::scalar_type_t<T2> = stan::math::var_value<double>; std::ostream = std::basic_ostream<char>]’:
bug.hpp:371:0:   required from ‘T_ bug_model_namespace::bug_model::log_prob(Eigen::Matrix<T_a, -1, 1>&, std::ostream*) const [with bool propto__ = false; bool jacobian__ = false; T_ = stan::math::var_value<double>; std::ostream = std::basic_ostream<char>]’
stan/src/stan/model/model_base_crtp.hpp:98:0:   required from ‘stan::math::var stan::model::model_base_crtp<M>::log_prob(Eigen::Matrix<stan::math::var_value<double>, -1, 1>&, std::ostream*) const [with M = bug_model_namespace::bug_model; stan::math::var = stan::math::var_value<double>; std::ostream = std::basic_ostream<char>]’
stan/src/stan/model/model_base_crtp.hpp:96:0:   required from here
bug.hpp:129: error: no matching function for call to ‘assign(Eigen::Matrix<stan::math::var_value<double>, -1, 1>&, stan::math::var_value<Eigen::Matrix<double, -1, 1> >, const char [30])’
  129 |       stan::model::assign(lcm_sym7__, stan::math::multiply(1, x),
      |       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  130 |         "assigning variable lcm_sym7__");
      | 
In file included from stan/src/stan/model/indexing.hpp:9,
                 from stan/src/stan/model/model_header.hpp:13:

Expected Output:

I don't know whether my Stan program contains some undefined behavior such that it should never compile, or if this is a bug with the --O flag.

I understand that the --O flag enables experimental optimization features, that may or may not yield performance benefits, but I didn't expect it to fail to build such a simple program.

If it's just the case that we should expect that --O sometimes fails to build valid Stan programs, please make that clearer in the documents.

Current Version:

v2.35.0

[WardBrian]

We will investigate, but I do want to say that yes, regular users should basically never use this flag. We renamed it to —Oexperimental for a reason, and if not for worrying about people already using it we probably would have removed it

[andrewfowlie]

Thanks Brian! In that case, one way forward:

• confirming that it is a bug only in —O and not in other optimization levels
• telling in the docs that —O it isn’t guaranteed to build valid Stan programs

This will put more people off using it!

[WardBrian]

I did some analysis of the cause and opened the above issues