On 05/12/2009 13:53, Patrick LeBoutillier wrote:

Hi,

I sent this question to haskell-cafe, realizing later that perhaps this list is a better place for it.

Patrick

---------- Forwarded message ---------- From: Patrick LeBoutillier Date: Fri, Dec 4, 2009 at 2:09 PM Subject: FFI and ghci To: Haskell Cafe

Hi all,

I have a small FFI-based library that I like to test like this:

$ ghci -I../c -L../c/.libs -lmlp t.hs GHCi, version 6.10.1: http://www.haskell.org/ghc/ :? for help

Please upgrade to 6.10.4, lots of bugs were fixed.

Loading package ghc-prim ... linking ... done. Loading package integer ... linking ... done. Loading package base ... linking ... done. Loading object (dynamic) mlp ... done final link ... done [1 of 9] Compiling MLP ( MLP.hs, interpreted ) [2 of 9] Compiling MLP.Error ( MLP/Error.hs, interpreted ) [3 of 9] Compiling MLP.Context ( MLP/Context.hs, interpreted ) [4 of 9] Compiling MLP.Runtime ( MLP/Runtime.hs, interpreted ) [5 of 9] Compiling MLP.Object ( MLP/Object.hs, interpreted ) [6 of 9] Compiling MLP.Type ( MLP/Type.hs, interpreted ) [7 of 9] Compiling MLP.Value ( MLP/Value.hs, interpreted ) [8 of 9] Compiling MLP.Language ( MLP/Language.hs, interpreted ) [9 of 9] Compiling Main ( t.hs, interpreted ) Ok, modules loaded: MLP, MLP.Object, Main, MLP.Runtime, MLP.Context, MLP.Error, MLP.Value, MLP.Type, MLP.Language. *Main> main

Note: Inside libmlp.so, another shared object is loaded using dlopen().

When running this test program I get some sporadic segmentation faults, which seem to always occur at entrypoint to the shared object that is loaded with dlopen().

But when I compile it with ghc, i.e.

$ ghc --make -I../c -L../c/.libs -lmlp t.hs

I can't reproduce crash.

Can anyone see why this could be happening? It's quite possible that there is really a bug in the C code, but if someone knows about a bug or something in ghci that can cause this behaviour I can stop looking...

This one might apply if you're on x86-64: http://hackage.haskell.org/trac/ghc/ticket/781 other than that, nothing springs to mind. If you can narrow it down to a small(er) test case and submit a bug report, we'll try to take a look. Cheers, Simon