Re: [PATCH] Provide special function to find ranges used for aliasing

[Richard Guenther <rguenther at suse dot de>]

On Wed, 14 Dec 2005, Daniel Berlin wrote:

> On Mon, 2005-12-12 at 16:00 +0100, Richard Guenther wrote:
> > This replaces the use of get_inner_reference in the aliasing machinery
> > whenever it tries to find out what parts of a structure is 
> > (possibly) affected by an access expression.  It provides a cleaner
> > interface for this use and at the same time improves the precision of
> > the information generated (this is needed for array aliasing, too).
> > 
> 
> 
> > Bootstrapped and regtested on x86_64-unknown-linux-gnu on mainline and
> > IAB.  Ada won't bootstrap for me with a gcc 3.3 based installation
> > (insert random rant here), so it got no testing.
> > 
> > Ok for mainline?
> > 
> 
> 
> > *************** find_used_portions (tree *tp, int *walk_
> > *** 2712,2748 ****
> >   
> >   	    if (bitpos <= up->minused)
> >   	      up->minused = bitpos;
> > ! 	    if ((bitpos + bitsize >= up->maxused))
> > ! 	      up->maxused = bitpos + bitsize;	    
> >   
> > ! 	    up->explicit_uses = true;
> >   	    up_insert (uid, up);
> >   
> >   	    *walk_subtrees = 0;
> >   	    return NULL_TREE;
> >   	  }
> > - 	else if (DECL_P (ref))
> > - 	  {
> > - 	    if (DECL_SIZE (ref)
> > - 		&& var_can_have_subvars (ref)
> > - 		&& TREE_CODE (DECL_SIZE (ref)) == INTEGER_CST)
> > - 	      {
> > - 		used_part_t up;
> > - 		size_t uid = DECL_UID (ref);
> > - 
> > - 		up = get_or_create_used_part_for (uid);
> > - 
> > - 		up->minused = 0;
> > - 		up->maxused = TREE_INT_CST_LOW (DECL_SIZE (ref));
> > - 
> > - 		up->implicit_uses = true;
> > - 
> > - 		up_insert (uid, up);
> > - 
> > - 		*walk_subtrees = 0;
> > - 		return NULL_TREE;
> > - 	      }
> > - 	  }
> >         }
> >         break;
> >         /* This is here to make sure we mark the entire base variable as used
> 
> 
> What is with this part?
> It's not clear why you are removing it 

I have convinced myself that we can merge the two cases.  Consider
the original code re-written as:

        if (DECL_P (ref))
          {
            size_t uid = DECL_UID (ref);
            used_part_t up;
            if ((offset == NULL && bitsize != -1)
                || (DECL_SIZE (ref)
                    && var_can_have_subvars (ref)
                    && TREE_CODE (DECL_SIZE (ref)) == INTEGER_CST))
              {
                up = get_or_create_used_part_for (uid);
              }
            if (offset == NULL && bitsize != -1)
              {
                /* We have a use that we know precisely.  Not recording
		   anything if !var_can_have_subvars is ok.  Note that
		   bitmaxsize is only -1 if the DECL does have variable
		   size, else it is at most DECL_SIZE (ref).  We didn't
		   record used portions for variable sized types before
		   (see DEC_SIZE == INTEGER_CST check below).  */
                if (bitpos <= up->minused)
                  up->minused = bitpos;
                if ((bitpos + bitsize >= up->maxused))
                  up->maxused = bitpos + bitsize;
                up->explicit_uses = true;
              }
            else if (DECL_SIZE (ref)
                && var_can_have_subvars (ref)
                && TREE_CODE (DECL_SIZE (ref)) == INTEGER_CST)
              {
                /* We have a use but don't know exactly.  In this
		   case bitsize != bitmaxsize.  We still know
		   partly, so we just adjust the ranges like in
		   the case we know.  */
                up->minused = 0;
                up->maxused = TREE_INT_CST_LOW (DECL_SIZE (ref));
                up->implicit_uses = true;
              }
            if ((offset == NULL && bitsize != -1)
                || (DECL_SIZE (ref)
                    && var_can_have_subvars (ref)
                    && TREE_CODE (DECL_SIZE (ref)) == INTEGER_CST))
              {
                up_insert (uid, up);

                *walk_subtrees = 0;
                return NULL_TREE;
              }
          }

and look at the added comments to the part where the two cases differ.

So I think it is safe to merge the two cases.

Richard.