Question on gimplify.c:create_tmp_var

[Richard Kenner]

    >     Hah!  Every time you drastically change what is valid gimple,
    >     you have to change all of the optimizers.
    > 
    > Certainly true for scalars, but is that true for what's within a
    > chain of references?

    Of course.

Which of the optimizers look inside a chain of _REF nodes?  It would seem that
it's just ones that try to do array/loop optimizations and it's quite
acceptable for them to punt if they don't recognize a particular pattern.
They can be made more intelligent incrementally.  So it wouldn't seem they
all have to "change".