problem with your GIMPLE_FORMAL_TEMP change

[Richard Kenner]

    Fear of some arbitrary unnamed transformation that alters the
    lifetime of the temporary so as to invalidate the assumptions
    of DECL_GIMPLE_FORMAL_TEMP_P when calling back into the 
    gimplifier later.

So how about defering it for functions that have nested procedures
until those are gimplified too?  Or doing the clearing in a new (tiny)
pass after gimplification?