crash/segfault when reusing a model

Summary:

R reliably crashes when re-using a pre-compiled/cached model.

Description:

R crashes the 2nd time a model is run, possibly related to #221.

Reproducible Steps:

library(rstan)
stan_demo(1)  # runs fine
stan_demo(1)  # crash here

Current Output:

*** caught segfault ***
address 0x1c, cause 'memory not mapped'

Traceback:
 1: FUN(X[[i]], ...)
 2: lapply(name, function(id) {    v <- .Internal(getSymbolInfo(as.character(id), PACKAGE, as.logical(withRegistrationInfo)))    if (is.null(v)) {        msg <- paste("no such symbol", id)        if (length(pkgName) && nzchar(pkgName))             msg <- paste(msg, "in package", pkgName)        stop(msg, domain = NA)    }    names(v) <- c("name", "address", "package", "numParameters")[seq_along(v)]    v})
 3: getNativeSymbolInfo(name, PACKAGE)
 4: doTryCatch(return(expr), name, parentenv, handler)
 5: tryCatchOne(expr, names, parentenv, handlers[[1L]])
 6: tryCatchList(expr, classes, parentenv, handlers)
 7: tryCatch(getNativeSymbolInfo(name, PACKAGE), error = function(e) e)
 8: Module(module, mustStart = TRUE)
 9: .getModulePointer(x)
10: <S4 object of class "Module">$stan_fit4model658c7fe3d509_ideo_interactions
11: eval(expr, envir, enclos)
12: eval(call("$", mod, paste("stan_fit4", model_cppname, sep = "")))
13: object@mk_cppmodule(object)
14: .local(object, ...)
15: (function (object, ...) {    standardGeneric("sampling")})(object = <S4 object of class "stanmodel">, data = <environment>)
16: (function (object, ...) {    standardGeneric("sampling")})(object = <S4 object of class "stanmodel">, data = <environment>)
17: do.call(sampling, args = dots)
18: stan_demo(1)

Expected Output:

N/A

RStan Version:

> packageVersion("rstan")
[1] ‘2.14.1’

R Version:

> R.version.string
[1] "R version 3.3.2 (2016-10-31)"

Operating System:

OS X 10.11.6

Reproduce the failure with the two consecutive stan_demo(1) calls and inspect the Module/.getModulePointer and sampling paths shown in the traceback. Compare the first and second model runs, using the reported rstan 2.14.1 and R 3.3.2 environment; done means reusing the cached model no longer produces a segfault.