[PATCH]: Structure aliasing, part 1

[Diego Novillo]

Daniel Berlin wrote:

> I moved it when i moved get_subvars_ and friends (they were originally
> all in tree-ssa-alias.c).
> I'm happy to move it back to tree-dfa.c if you like :)
> 
Yeah, tree-flow-inline.h is only supposed to be there for one-liners.


> I concluded at this point that it wasn't worth it to run it after SRA,
> at least at the moment, and it was on my todo list to reevaluate after
> TCB merges structure ccp/copy prop.
> 
OK, thanks.  That's good enough for me.

> Which seemed more confusing to follow than making it explicit :)
> 
Hmm, yeah.  Good point.

>   /* Memory tags are by definition addressable.  This also prevents
>      is_gimple_ref frome confusing memory tags with optimizable
>      variables.  */
>   TREE_ADDRESSABLE (tag) = 1;
> 
Ah, right.  A comment above it should help future confusion.  Perhaps we 
want an alias for TREE_ADDRESSABLE, like TREE_NOT_GIMPLE_REG, but that's 
overkill.  The comment should be fine.

> Though if you really feel strongly about this, i'll move it.  I just
> don't currently see advantages (Not that i don't believe that with
> structure CCP/copy prop, we might not have some, but i'd rather move it
> after that point, when we have more benchmarks)
> 
Agreed.  Let's run it early now.  We can always change our minds in the 
future.


> The problem is that you need a {VAR, FIELD_UID} mapping, because fields
> are specific to types, not vars.
> 
Ah, OK.


Diego.