PATCH to make return in memory explicit

[Jason Merrill]

Hmm, on second thought, maybe forcing CALL_EXPR_HAS_RETURN_SLOT) isn't such
a good idea; it might result in taking the address of a variable which
could otherwise be scalarized.  Now I'm thinking that we only want to make
passing by invisible reference explicit when it's necessary for language
semantics, not when it's a detail of the target ABI.

I'm reverting (most of) this patch.

Jason