Build model methods with -fno-gnu-unique on Linux with GCC, so that they can be loaded together with rstan

NEWS.md

@@ -322,7 +322,7 @@ without running its cleanup, for example a future worker interrupted from the
 parent session. Previously they kept running as orphans. (#1086)
 * `$init_model_methods()` and `$expose_functions()` now work in the same R
 session as rstan or brms, whichever loads first, when RcppParallel 6.2.0 or
-later is installed. (#1270)
+later is installed. (#1177, #1270)
 * A function exposed with `$expose_functions()` that returns a tuple no longer
 occasionally crashes R or returns a list with wrong elements. (#1001)
 * `$expose_functions(global = TRUE)` now also keeps the functions in the

R/expose.R

@@ -112,7 +112,10 @@ rcpp_source_stan <- function(code, env, verbose = FALSE, ...) {
     make_args
   )
   check_sundials_fpic(flags[8], verbose)
-  cxxflags <- flags[1]
+  # With GCC, Stan Math's autodiff stack is a unique symbol that the loader
+  # binds to rstan's, or any other Stan Math build in the session, and R
+  # crashes (#1177). Hidden visibility keeps it out of the exported symbols.
+  cxxflags <- paste(flags[1], "$(CXX_VISIBILITY)")
   cppflags <- flags[2]
   libs <- paste(flags[3:7], collapse = " ")
   cmdstanr_includes <- system.file(

tests/testthat/test-model-methods.R

@@ -124,6 +124,24 @@ test_that("Methods return correct values", {
   expect_equal(unconstrained_variables, c(0.1))
 })
 
+test_that("model methods don't export Stan Math's autodiff stack", {
+  skip_on_os("windows")
+  # On Linux the loader would bind rstan's copy to it (#1177)
+  instance <- paste0(
+    "_ZN4stan4math22AutodiffStackSingletonINS0_9vari_baseE",
+    "NS0_15chainable_allocEE9instance_E"
+  )
+  fit$init_model_methods()
+  dlls <- Filter(
+    function(dll) startsWith(dll[["name"]], "sourceCpp"),
+    getLoadedDLLs()
+  )
+  expect_gt(length(dlls), 0)
+  for (dll in dlls) {
+    expect_false(is.loaded(instance, PACKAGE = dll[["name"]]))
+  }
+})
+
 test_that("Model methods environments are independent", {
   data_list_2 <- data_list
   data_list_2$N <- 20