On Fri, Jan 29, 2021 at 05:27:13PM +0100, Richard Biener wrote:
> > > + df_process_deferred_rescans ();
> >
> > rather than here?
> > The pass these days does two separate optimizations, and only one of those
> > creates the use uninitialized pseudo and at the end initialize it.
> > The other optimization is replace_constant_pool_with_broadcast, which
> > doesn't do this but still calls validate_change -> df_insn_rescan.
>
> But both are very cheap and IMHO the flow is less obfuscated this way.
The df_set_flags call sure, but it means that all the df_insn_rescan calls
the pass will do will allocate memory to queue the rescans, even when it
doesn't need it.
Anyway, your call on this.
Jakub
