On Mon, Oct 12, 2020 at 3:02 PM Martin Liška <mli...@suse.cz> wrote:
>
> On 10/12/20 2:44 PM, Martin Liška wrote:
> > On 10/7/20 10:00 AM, Richard Biener wrote:
> >> As said I'd have a BB-local pass over BBs recording the index variable
> >> and the range covered by the BBs gcond, plus recording how many excess
> >> stmts there are for eventual code motion.
> >>
> >> Only after that BB-local pass start to group BBs in a walk from dominated
> >> to
> >> dominating BBs looking for common indexes and building a case vector.
> >>
> >> The main thing is to avoid repeatedly analyzing BBs conditions (so the
> >> first
> >> pass could be also a on-demand precompute thing) and making the
> >> case vector build optimal.
> >
> > I have a patch that does that using the infrastructure from
> > tree-ssa-reassoc.
> > Now I would like to implement the code hoisting. Am I right that we want
> > something like:
> >
> > if (index == C0)
> > goto BB_0;
> > else
> > {
> > BB1_to_hoist_stmts;
> > if (index == C1)
> > goto BB_1;
> > else
> > {
> > BB2_to_hoist_stmts;
> > if (index == C2)
> > goto BB_2;
> > else
> > goto default_BB;
> > }
> > }
> >
> > be converted into:
> >
> > switch(index)
> > {
> > case C0:
> > goto BB_0;
> > case C1:
> > BB1_to_hoist_stmts;
> > goto BB_1;
> > case C2:
> > BB1_to_hoist_stmts;
> > BB2_to_hoist_stmts;
> > goto BB_2;
> > else:
> > BB1_to_hoist_stmts;
> > BB2_to_hoist_stmts;
> > goto default_BB;
> > }
> >
> > ?
>
> Reading again Richi's comment:
>
> > ifcombine simply hoists any stmts without side-effects up the
> > dominator tree and thus only requires BBs without side-effects
> > (IIRC there's a predicate fn for that).
>
> we likely want to hoist the statements "up" to the gswitch BB?
Yes. code sinking later will sink what can be sunk.
> Martin
>
> >
> > Thanks,
> > Martin
>
