As pointed Michael Weber in a sparc-unknown-linux porting,the key point is

"gcc on sparc-linux (mips-irix) doesn't mark the beginning of .data(like "data_start" on i386-linux) . Using etext is definitely wrong here, since .rodata comes afterwards and therefore etext doesn't cover the infotables for static-closures"

You may be interested to know that recent CVS versions of GHC don't have the requirement that the section boundaries need to be known. You'll undoubtedly find it easier to port newer versions (although in order to bootstrap you might need to start from an older version...). Cheers, Simon