[PATCH] tree-optimization/101280 - revise interchange fix for PR101173

[Richard Biener]

On Thu, 1 Jul 2021, Michael Matz wrote:

> Hello,
> 
> On Thu, 1 Jul 2021, Richard Biener wrote:
> 
> > diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
> > index 43045c5455e..43ef112a2d0 100644
> > --- a/gcc/gimple-loop-interchange.cc
> > +++ b/gcc/gimple-loop-interchange.cc
> > @@ -1043,8 +1043,11 @@ tree_loop_interchange::valid_data_dependences (unsigned i_idx, unsigned o_idx,
> >  	    continue;
> >  
> >  	  /* Be conservative, skip case if either direction at i_idx/o_idx
> > -	     levels is not '=' (for the inner loop) or '<'.  */
> > -	  if (dist_vect[i_idx] < 0 || dist_vect[o_idx] <= 0)
> > +	     levels is not '=' or '<'.  */
> > +	  if (dist_vect[i_idx] < 0
> > +	      || (DDR_REVERSED_P (ddr) && dist_vect[i_idx] > 0)
> > +	      || dist_vect[o_idx] < 0
> > +	      || (DDR_REVERSED_P (ddr) && dist_vect[o_idx] > 0))
> 
> Hmm, if DDR_REVERSED_P matters here, then it should matter for all arms.  
> IOW: < 0 should be tested only when !DDR_REVERSED_P, not always:
> 
>        if ((!DDR_REVERSED_P (ddr) && dist_vect[i_idx] < 0)
>            || (DDR_REVERSED_P (ddr) && dist_vect[i_idx] > 0)
>            || (!DDR_REVERSED_P (ddr) && dist_vect[o_idx] < 0)
>            || (DDR_REVERSED_P (ddr) && dist_vect[o_idx] > 0))
> 
> (what you have effectively written is a condition that allows only 0 when 
> DDR_REVERSED_P)

True - been fixing bugs too fast I guess.  Maybe the 3rd time's the charm 
then ;)

Richard.