GCC Status Report (2004-03-09)

[Richard Kenner]

    That would be certainly safest and the UNCHANGING would really mean
    unchanging.  But I hope we could at least extend it to objects which are
    *never* stored into in current function (see if the object is initialized
    in the function currently expanded into RTL, if not, use RTX_UNCHANGING_P
    for it as well, otherwise perhaps use some other bit or alias set).

I think so.  Certainly that would be true for parameters in languages
where they can't be changed.

It seems the trickiest problems come when we try to define what
*writing* to "unchanging" memory means.