#8199: Get rid of HEAP_ALLOCED ----------------------------+---------------------------------------------- Reporter: ezyang | Owner: ezyang Type: feature | Status: new request | Milestone: 7.10.1 Priority: normal | Version: 7.7 Component: | Keywords: Compiler | Architecture: Unknown/Multiple Resolution: | Difficulty: Project (more than a week) Operating System: | Blocked By: 5435 Unknown/Multiple | Related Tickets: Type of failure: | None/Unknown | Test Case: | Blocking: | ----------------------------+---------------------------------------------- Comment (by ezyang): I discovered an interesting flaw in the current design while debugging a segfault today. The problem goes like this: currently, on a per-module basis we generate an extra C file responsible for registering staticclosureinds with the initialization loop which will run at the beginning of the file. The implicit assumption being made here, of course, is that a statically linked module will be loaded atomically together, all of the object files being brought in for the ride. However, this is NOT TRUE when split objects are being used! In this case, it is quite possible for the object file containing the initialization function to *not* be loaded, while another split object which uses an indirection does get loaded. In this case, the static closure never gets initialized, and we'll get a segfault if we try to use it. How can we fix this? One possibility is, when splitting objects, to ensure if the staticclosureinds section is nonzero, then we have a reference to the initialization function somewhere else in the resulting object file, so it gets loaded in. However, it's not even clear that the starts/ends trick is going to even keep working, unless the splitter is very careful. Another possibility is to axe the initialization function, and try harder to detect the staticclosureinds section at runtime. Doing this properly in a cross-platform way is tricky (I originally tried to do it with linker scripts, but gave up when it (1) didn't work, and (2) wasn't going to work on other platforms.) I'll upload an up-to-date patchset for GHC 7.8 at some point. There's been a recent change to integer-gmp which needs to be updated for HEAP_ALLOCED. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8199#comment:20 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler