I'm sorry, this is nonsense. As James Ulrich points out, the convolution in GCC has nothing to do with the implementation language.
Its biggest problem -- still pervasive in the RTL side of things -- was always global, unstated assumptions, often assumptions wired into target machine description files, RTL optimization passes, and reload. Often an RTL optimization pass would assume (or would grow to assume over years, accidentally) some property of md files that was true for all existing md files but not necessarily true, and then reload would come to depend on the form of the RTL emitted by optimization files when when that property was true. Some of these properties are much nastier than CC0 and can't be grepped for -- and fixing them requires understanding a lot of targets, and *testing* them.
This is slowly being sorted out as more machinery migrates into the tree-ssa side of things, and as older and cruftier targets are slowly decommissioned. But it's a slow, slow job, and it would be every bit as slow regardless of the implementation language. (This is one reason why reload has been such a monster to dump and replace: it's where all these unstated assumptions go to roost. Break just one of them and you might find yourself with wrong code on a couple of random targets you'd never heard of, and the poor sod who finds this is going to have the devil of a time tracking it down to your change.)