Some expect_ad calls fail to compile

[peterwicksstringfield]

Description

I think that there is a problem with expect_ad, preventing certain expect_ad calls which would otherwise work as intended from compiling.

This test case for example:

TEST(test_unit_math_test_ad, test_ad_show_problem) {
  auto f = [](auto& x, auto& y) { return x + y(0); };
  Eigen::VectorXd v(3);
  v << 1, 2, 3;
  stan::test::expect_ad(f, 0, v);
}

Produces this compilation error:

[...]
./test/unit/math/test_ad.hpp:493:21: error: no matching function for call to ‘expect_all_throw(const test_unit_math_test_ad_test_ad_show_problem_Test::TestBody()::<lambda(auto:50&, auto:51&)>&, int&, const Eigen::Matrix<double, -1, 1>&)’
     expect_all_throw(f, x1, x2);
[...]

expect_ad here delegates to expect_ad_vv(tols, f, int, T2), and expect_ad_vv(tols, f, int, T2) tries to call expect_all_throw(f, int, T2), but expect_all_throw only has the following signatures:

void expect_all_throw(const F& f, const Eigen::VectorXd& x) {
void expect_all_throw(const F& f, double x1) {
void expect_all_throw(const F& f, double x1, double x2) {
void expect_all_throw(const F& f, double x1, double x2, double x3) {

I think expect_all_throw is only supposed to be called on the serialized form of the functor f?

After trying to call expect_all_throw, and assuming except_all_throw does not catch an exception, expect_ad_vv(tols, f, int, T2) computes the value at the given arguments, and then again at the same arguments with the int converted to a double. After that it delegates to expect_ad_vv(tols, f, T1, T2) which serializes f and delegates to expect_ad_helper, which itself calls expect_all_throw on the serialized form of f.

Example

Above.

Expected Output

expect_ad compiles in this case, test coverage is not reduced.

Current Version:

v3.0.0

[peterwicksstringfield]

Ah. Well now I see why that exception check was there. Ticket updated. I'm going to take another look at this tomorrow. I do think we can get things like "except_ad(f, 0, v)" to compile, but it will take a bit more work.

[bob-carpenter]

Let me know if you need help. That case of expect_ad(F, double, VectorXd) is intended to work.

…

On Dec 22, 2019, at 1:47 AM, Peter Wicks Stringfield ***@***.***> wrote: Ah. Well now I see why that exception check was there. Ticket updated. I'm going to take another look at this tomorrow. I do think we can get things like "except_ad(f, 0, v)" to compile, but it will take a bit more work. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.