On Nov 20, 2011, at 1:55 AM, Ian Lance Taylor wrote: > Amittai Aviram <amittai.aviram@xxxxxxxx> writes: > >> I have been creating a modified version of GCC's libgomp library to support a special version of OpenMP. In order for it to work properly, at the end of the main function in an OpenMP program, I have to insert a call to a special clean-up function (say, "clean_up()"), just before the "return" statement. (I define my clean_up function in my version of libgomp.) I would like to modify GCC so that it inserts the call to clean_u automatically, just before the code representing the return statement, such as the RET instruction in x86 assembly code. How might I go about this? >> >> Alternatively, it might work if GCC inserted the function call into the boilerplate code of __exit. How might I enable GCC to do that? > > Would it work to just use atexit? If that works, perhaps you could just > add a global constructor which calls atexit with your cleanup function. > > Ian Thanks a lot, Ian! That works! However, I had to do a bit of extra hacking. At first, for some reason, my clean-up function was being called twice. Since it did some clean-up operations that depended on data structures in a mapped file, and then unmapped the file afterwards, the second call was causing a segmentation fault (dereferencing a null pointer). I got around this problem through a simple hack: my library now has a boolean variable set to false on startup; the clean-up function checks the value and returns if the variable is true; and it sets it to true at the end, before returning. However, I am still wondering why the code gets run twice, requiring this hack. This is a multiprocess program, where I fork a bunch of child processes in a "start" function and I join them back in the clean-up function. The "start" function calls atexit to register the clean-up function--after it forks the child processes. I see no reason why the child processes would ever register the clean-up function themselves, since they exit before they can come back into the "start" code that has the "atexit" statement. However, just for good measure, I also test for the process ID; if it's not 0, the clean-up function returns without doing anything. But this did not prevent the clean-up code from running twice. Amittai Aviram PhD Student in Computer Science Yale University 646 483 2639 amittai.aviram@xxxxxxxx http://www.amittai.com
