[Bug tree-optimization/64454] optimize (x%5)%5

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64454

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Marc Glisse from comment #6)
> (In reply to Jakub Jelinek from comment #5)
> > The reason for tree_int_cst_sgn (vr->min) >= 0 was that I don't want to let
> > 0 through and for negative values, handling those would require computing
> > absolute value, but as match.pd already folds x % -5 already into x % 5,
> > there is no need to bother with it, so I'm just trying to play safe.
> 
> I don't think we are talking about the same thing. Restricting to positive
> op1 is good. What I find a little strange is:
> 
> +      if (TYPE_UNSIGNED (TREE_TYPE (op0))
> +	  || tree_int_cst_sgn (vr->min) >= 0
> +	  || tree_int_cst_lt (fold_unary (NEGATE_EXPR, TREE_TYPE (op1), op1),
> +			      vr->min))
> 
> where condition 2: min>=0 is more restrictive than condition 3: min>-op1
> (since op1 is known to be positive) so we could skip condition 2.

Ah, this, I just didn't want to call fold_unary to create GC garbage when I can
cheaply see that it is ok.  If I were to remove something, the TYPE_UNSIGNED
(TREE_TYPE (op0)) check could go (as tree_int_csg_sgn (vr->min) will be then >=
0 always).