type gimplification, continued

[Richard Kenner]

    Given that we're not explicitly handling DECL_SIZEs for any other 
    variable, doing so here looks like a hack.  

No, not at all!  Read what I wrote: the walk_tree below is most definitely
handling the DECL_SIZE for *every* other variable except for PARM_DECLs.

    So I think I prefer the unsharing in gimplify_one_sizepos.

That's the hack because it unshares too much and is too late: you want the
GENERIC trees unshared and that's what the above function is supposed to to.