Problem with use of TYPE_SIZE in tree optimizers

[Richard Kenner]

I'm running into a problem when SSA_NAME nodes that are actually in use
are being freed.  I traced this down to some improper uses of TYPE_SIZE
in the tree optimizers.

The tree optimizers can only use TYPE_SIZE and TYPE_SIZE_UNIT if they are
known to be INTEGER_CST.  Otherwise, we have "hidden operands", which are
not permitted in GIMPLE.

There are exactly three places that violate this rule.  One is in
tree-mudflap.c, but in a function has no comments in front of it (I thought
these were supposed to be supplied before the merge?), so I don't know what
it's supposed to do.  That was not the one that's causing problems.

The ones that *are* causing problems are in
tree-data-ref.c:estimate_niter_from_size_of_data.

element_size should be array_ref_element_size, but there's no safe way of
referencing the size of the array itself.

I suspect that this routine should simply return if either is not an
INTEGER_CST, but I'm not sure.  Does that seem right?