[PATCH] PROMOTE_FOR_CALL_ONLY on sparc

[Richard Kenner]

    If you really want a benchmark to show this, fine...  but I can't say
    I can honestly conjur one up which won't be biased in some way.

No, I don't mean a specific made-up program, I mean running something
like SPEC or even timing the compiler bootstrapping with and without
this defined.  I find that relying on intuition or looking at code
fragments in making decisions like this is very unreliable.

    Particularly avoiding promotion will win when you want to do some
    operation on the variable which will also promote/demote it (SHIFT and
    AND operations in particular on sparc64).  This will be the source of
    the SET, and I suppose you are claiming this is quite uncommon.

Right.  The most common operations are addition and subtraction.