Fix processing of ADDR_EXPR in get_expr_operands

[Diego Novillo]

On Fri, 2004-07-30 at 09:27, Richard Kenner wrote:
>     No.  It is not a bug.  I don't know how else to tell you the same
>     thing.  Address constness is checked with TREE_INVARIANT.  The address
>     of a variable may be invariant in one invocation to a function, but have
>     different values from call to call.
> 
> The two mean two different thing: "invariant" means it won't change during
> the execution of the function while "constant" means it won't change *at all*.
>
I feel like I'm talking to myself here.  Yes, that's exactly what it
means and that's exactly what I've been saying all this time.

> I can easily see cases where the first is what's wanted and where the second
> is what's wanted.  It may well be that in that particular case, "invariant"
> is the wanted property, but I don't accept that that's *always* the case.
> 
For address constants?  Yes, that's always the case.  CONSTANT is a
superset of INVARIANT.  In an intra-procedural setting, we care for
invariant addresses, we don't care if they are constant or not.

>     Yes.  Overflow.
> 
> Why do we care about overflow?  If an addressing computation overflows, it's
> still a constant!
>
No.  That's for real constants.  I was describing what
is_gimple_min_invariant does.  Check the code out.


Diego.