On Thu, Dec 29, 2011 at 04:24:31PM +0000, Richard Sandiford wrote:
> AIUI, the outcome of PR38964 was that we can't use TBAA for testing an
> anti_dependence between a load X and store Y because Y might be defining
> a new object in the same space as the object that was being read by X.
> But it looks like we still use component-based disambiguation
> (nonoverlapping_component_refs_p) in this case. Is it true that
> that's also a problem? E.g. for:
>
> struct s { int f; float g; };
> struct t { int header; struct s s; };
>
> float foo (struct t *newt, struct s *olds, int x, int y)
> {
> float ret = olds[x * y].g;
> newt->header = 0;
> newt->s.f = 1;
> newt->s.g = 1.0;
> return ret;
> }
>
> we can (and on ARM Cortex A8, do) move the store to newt->s.f above
> the load from olds[...].g. If we view the assignment to newt
> as defining a new object in the same space as the now-defunct olds,
> and if x * y happens to be zero, then the accesses might well be
> to the same address.
You would need a placement new in between the read from olds[x * y].g
and newt->* stores, without that it is definitely valid to move it
ahead of the store.
IMHO we should keep placement new in the IL in some form, and decide how
conservative we need to be based on whether the current function
or current loop etc. contains (or could contain) a placement new or not.
E.g. if a loop doesn't contain placement new anywhere, we could be less
conservative in optimizing that loop, do more vectorization on it etc.
Jakub
