new ICE in 3.3.4 on alpha

[Joe Buck]

See PR 15718.  There is an ICE in g77 in 3.3.4-pre that causes 10 new test
failures (two distinct input programs, failing at multiple optimization
levels) on the Alpha.  Seems it's a subreg problem affecting Fortran
programs that use COMPLEX.  This is a regression from 3.3.3.

Given the high standards for a point release, I would consider this a
showstopper.  We can't tolerate new ICEs in our own regression test suite.