Re: [tree-ssa] Coalescing/killing redundant labels?

[law at redhat dot com]

In message <1058396499.2837.25.camel@steven.lr-s.tudelft.nl>, Steven Bosscher w
rites:
 >Consider this snippet:
 >
 >int a, b;
 > 
 >void
 >foo (void)
 >{
 >  if (a)
 >    { a = b; goto byebye2; }
 >  else if (b)
 >    { b = a; uselesspos: goto byebye1; }
 > 
 >byebye2:
 >byebye1:
 >  return;
 >}
 >
 >At -O2 we tree-optimize this to:
 > 
 >;; Function foo (foo)
 > 
 >foo ()
 >{
 >  void byebye2 = <<< error >>>;
 >  void byebye1 = <<< error >>>;
 > 
 >  if (a != 0)
 >    {
 >      a = b
 >    }
 >  else
 >    {
 >      if (b != 0)
 >        {
 >          b = a;
 >	  uselesspos:;;
 >          goto byebye1;
 >        }
 >    };
 >  byebye2:;;
 >  byebye1:;;
 >  return;
 >}
 >
 >So the presence of byebye2 apparently prevents the elimination of the
 >`goto' statement.  Why aren't the byebyex labels coasesced?  And what
 >about the unused label `uselesspos'?  It doesn't matter much for the
 >generated code I suppose, but it gives the expanders just a little less
 >cruft to deal with, and also for the future post-ssa optimizers...  So
 >maybe it's worth it anyway.  Has anyone tried it already??
I've got some changes here which will kill unused labels during the
remove_useless_stmts_and_vars pass.

I haven't benchmarked to see if the reduction in trees we have to
expand is worth the bookkeeping work.  It was really quick and
dirty (used/not-used rather than reference counting).

Also note that policy has been that user labels are never removed.
So the label removal code only removed compiler-generated labels.

If you're interested, I can probably provide some diffs...  I probably
won't get back to it for a while since I'm buried in trying to
effectively make cse1's path following code pointless.


Jeff