Faster compilation speed

[Richard Kenner]

    Just to make a point: look at copy_rtx_if_shared(), which tries to do 
    this (yeah, I have an older tree, maybe this is fixed these days. I 
    seriously doubt it).

    The code is CRAP. Total and utter sh*t. The damn thing should just
    test a reference count and be done with it. Instead, it has this
    heuristic that knows about some rtx's that might be shared, and knows
    which never can be.  And that _cap_ comes directly from the fact that
    the code uses a lazy GC scheme instead of a more intelligent memory
    manager.

Bad example. That code predates any sort of GC and has to do with 
*correctness* issues involving the semantics of RTL, not anything having
to do with allocation efficiency.