Re: fold: &a[i1] + c * i2 -> &a[i1 + i2] messing up types

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> > > > > If you prefer removing the type argument that would be fine, too.
> > > >
> > > > this might be a good idea, yes.
> > >
> > > Actually, in
> > >
> > > 	  step = array_ref_element_size (ref);
> > >
> > > 	  if (TREE_CODE (step) != INTEGER_CST)
> > > 	    continue;
> > >
> > > 	  itype = TREE_TYPE (step);
> > >
> > > 	  /* If the type sizes do not match, we might run into problems
> > > 	     when one of them would overflow.  */
> > > 	  if (TYPE_PRECISION (itype) != TYPE_PRECISION (type))
> > > 	    continue;
> > >
> > > 	  if (!operand_equal_p (step, fold_convert (itype, s), 0))
> > > 	    continue;
> > >
> > > do you really mean to compare TYPE_PRECISION (itype) and TYPE_PRECISION
> > > (type)?  I think type should refer to TREE_TYPE (s) here, not to
> > > TREE_TYPE (addr), no?
> >
> > they have the same size.
> 
> Consider (C++)
> 
> int a[1024];
> 
> int *foo(unsigned short i, long long int j)
> {
>         return &a[i]+j;
> }
> 
> transformed to
> 
> int* foo(short unsigned int, long long int) (i, j)
> {
>   int * D.1567;
>   int D.1568;
>   unsigned int D.1569;
>   unsigned int D.1570;
> 
>   D.1568 = (int) i;
>   D.1569 = (unsigned int) j;
>   D.1570 = D.1568 + D.1569;
>   D.1567 = &a[D.1570];
>   return D.1567;
> }
> 
> is the cast of j to (unsigned int) working by pure luck - what about
> j == -1?

then j will be casted to unsigned int.  That's perfectly ok.

> It seems at least the frontend is promoting the unsigned short to int
> for the array index.
> 
> Note that the type returned by array_ref_element_size() seems to
> be unsigned int always.

no, it should be (equivalent of) size_t.

Zdenek