Compiling the following code snippet with "g++ -fopenmp -O -Wall" I get a hosed error message: ============================ int i; void foo() { #pragma omp parallel { int j; i = j; } } ============================ bug.cc: In function 'void _Z3foov.omp_fn.0(void*)': bug.cc:8: warning: 'j' is used uninitialized in this function Because of the ".omp_fn.0" at the end of the function name the demangler cannot process it. Jakub, this is because of your patch http://gcc.gnu.org/ml/gcc-patches/2005-10/msg01644.html Should we regard the ".omp_fn.0" suffix as an extension to the name mangling scheme and teach the demangler how to handle those things? This problem will probably also come up with debuggers etc.
actually there is no demangling in the C++ front-end. What is usually is giving is just a friendly version of the string (there is a way to do this but I don't know how). Now the demangler might need to be expanded to support this (but that is a different issue from the C++ front-end issue).
No, this isn't a failure of the demangler. Nor do ever expect these functions to ever be demangled. That they contain the original function name at all is merely an aid to us the compiler developer. I don't see that this is any better or worse, from the user's perspective, than just "__omp_fn.N", which is what we had before the patch. Unless you can come up with some convincing user usage scenario, I'm not going to address this problem at all.
What we perhaps could do is hack up C++ lang_decl_name etc., so that for the artificially created .omp_fn* functions the diagnostic machinery would instead print the name of the original function, perhaps annotated with #pragma omp region of ... or something similar. Not sure if we can use for that just if (DECL_ARTIFICIAL (current_function_decl) && name starts with create_omp_child_function_name prefix) or if we would need to add some special attribute or bit (and if the latter, if it is worth doing).
Now it says "In function ‘<built-in>’". Jakub, could we do better here?
Now, `_Z3foov._omp_fn.0' is demangled as `foo() [clone ._omp_fn.0]'. GCC could use that instead of just '<built-in>'.