Finally found the ghci bug on Snow Leopard
After much digging through otool output and tracing Haskell code with dtrace, log messages, and gdb, I finally found the cause of the SIGBUS errors with ghci on Snow Leopard. At the end, only a tiny change was required[1]: the homegrown dynamic linker of ghci had simply ignored a relocation type that ld on Snow Leopard chose to use.
What this really shows again is that re-implementing parts of basic infrastructure in an ad hoc manner always comes back to bite you. Such infrastructure is usually full of low-level, platform-dependent details, which you are bound to get wrong once in a while and then it can be rather involved to find the problem. In this case, the relocation of the memory access of a bit of C code in GMP, reading a global variable, went wrong. As a result, the wrong memory location was accessed, which happened to contain a NULL pointer. Upon dereferencing the NULL, the program crashed. All of this was further obscured by the fact that the homegrown linker approach of ghci implies that we have two copies of the basic libraries in memory and in use at the same time. One copy has been statically linked with the ghc executable and the second copy is dynamically loaded by ghci. The relocation in the statically linked copy was perfectly fine, but the relocation in the dynamically loaded one was broken. [1] http://www.haskell.org/pipermail/cvs-ghc/2009-October/050863.html