Re: [PATCH v2] Add if-chain to switch conversion pass.

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Nov 9, 2020 at 1:26 PM Martin Liška <mliska@suse.cz> wrote:
>
> On 11/6/20 1:31 PM, Richard Biener wrote:
> > On Fri, Oct 16, 2020 at 4:04 PM Martin Liška <mliska@suse.cz> wrote:
> >>
> >> Hello.
> >>
> >> There's another version of the patch that should be based on what
> >> I discussed with Richi and Jakub:
> >>
> >> - the first patch introduces a new option -fbit-tests that analogue to -fjump-tables
> >>     and will control the new if-to-switch conversion pass
> >>
> >> - the second patch adds the pass
> >> - I share code with tree-ssa-reassoc.c (range_entry and init_range_entry)
> >> - a local discovery phase is run first
> >> - later than these local BBs are chained into a candidate list for the conversion
> >>
> >> I'm also sending transformed chains for 'make all-host' (620 transformations).
> >> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
> >
> > -static bool
> > +bool
> >   no_side_effect_bb (basic_block bb)
> >   {
> >
> > exporting this with this name is dangerous I think because the function
> > seems to allow side-effects in the last stmt - not sure exactly what
> > it tries to allow - there's no comment to that :/
>
> All right, will fix that.
>
> >
> > +  free (rpo);
> > +  free_dominance_info (CDI_DOMINATORS);
> > +
> > +  if (!all_candidates.is_empty ())
> > +    mark_virtual_operands_for_renaming (fun);
> >
> > please avoid freeing dominance info when there was no change done
> > (move it to the !all_candidates.is_empty () block).
> >
> > +  basic_block bb;
> > +  FOR_EACH_BB_FN (bb, fun)
> > +    find_conditions (bb, &conditions_in_bbs);
> > +
> >
> > if we didn't find any conditions (or found just one?) we can elide the
> > rest of the function, no?
>
> Sure.
>
> >
> > +         if_chain *chain = new if_chain ();
> > +         chain->m_entries.safe_push (info);
> > +         /* Try to find a chain starting in this BB.  */
> > +         while (true)
> > +           {
> > +             if (!single_pred_p (gimple_bb (info->m_cond)))
> > +               break;
> > +             edge e = single_pred_edge (gimple_bb (info->m_cond));
> > +             condition_info *info2 = conditions_in_bbs.get (e->src);
> > +             if (!info2 || info->m_ranges[0].exp != info2->m_ranges[0].exp)
> > +               break;
> > +
> > +             chain->m_entries.safe_push (info2);
> > +             bitmap_set_bit (seen_bbs, e->src->index);
> > +             info = info2;
> > +           }
> >
> > so while we now record conditions per BB the above doesn't really
> > allow matching a binary tree.
>
> Yes. The pass currently only supports conditions of the following form:
> 1) index in {min, max}
> 2) index out of {min, max}
>
> which means one edge in form 1). I don't see how can be useful handling
> of a situation where both edges contain a such-chain? Can you please
> come up with a test-case that can be interesting?

Sth like

   if (val < 10)
    {
       if (val == 1)
        ...
       else if (val == 2)
         ...
      else
        default;
   }
 else
   {
        if (val == 99)
           ...
        else if (val == 42)
          ...
        else
          gcc_unreachable ();
   }

but the most trivial thing would be to feed the pass the
balanced-tree generated by switch expansion where I
would expect us to be able to detect it as the original switch again.

> > What I was thinking of is to record
> > if_chain * per BB as well and look at successors, thus (pseudo-code)
> >
> >     if (block ends in cond)
> >       if (if_chain on true edge && if_chain on false edge)
> >        try merge
> >      else if (if_chain on true edge && this-cond tests same var)
> >        try merge
> >      else if (if_chan on false edge && ...)
> >        try merge
> >      record if_chain for block
> >
> > where merging would eventually detach the if_chains from the successors.
> > For now we'd just handle the true (and maybe false) edge combos to handle
> > linear chains.  Walking reverse RPO (I'm not 100% sure reverse RPO is what
> > we want here, but guess it will work fine for now) will gather chains
> > accordingly.
> > When merging from a successor to a BB fails we push the successor chain
> > to the candidate list.
> >
> > +/* Algorithm of the pass runs in the following steps:
> > +   a) We walk basic blocks in DOMINATOR order so that we first reach
> > +      a first condition of a future switch.
> > +   b) We follow false edges of a if-else-chain and we record chain
> > +      of GIMPLE conditions.  These blocks are only used for comparison
> > +      of a common SSA_NAME and we do not allow any side effect.
> > +   c) We remove all basic blocks (except first) of such chain and
> > +      GIMPLE switch replaces the condition in the first basic block.
> > +   d) We move all GIMPLE statements in the removed blocks into the
> > +      first one.  */
> >
> > the overall comment is now a bit out-of-date?
> >
> > Please remove the PHI mapping as I outlined in earlier review.
> >
> > The 0001-Add-fbit-tests-option.patch is OK for trunk.
>
> Installed to master.
>
> Martin
>
> >
> > Thanks,
> > Richard.
> >
> >
> >> Thoughts?
> >> Thanks,
> >> Martin
>