This was a bug in GHC's native code generator for SPARC, which sometimes misaligned the stack pointer. This has already been fixed in CVS (HEAD and branch), but it didn't make it into 5.04, see: http://cvs.haskell.org/cgi-bin/cvsweb.cgi/fptools/ghc/compiler/nativeGen/MachCode.lhs.diff?r1=1.100&r2=1.101 Cheers, S.