On Fri, 1 Jun 2012, Eric Botcazou wrote:
> > Ah, I see. So the proper fix would be to use STRIP_NOP()ed op0/op1,
> > something not readily available though.
>
> Why not just add
>
> TYPE_UNSIGNED (TREE_TYPE (op0)) == TYPE_UNSIGNED (TREE_TYPE (arg0))
> && TYPE_UNSIGNED (TREE_TYPE (op1)) == TYPE_UNSIGNED (TREE_TYPE (arg1))
>
> with a small comment?
Well, it would rather be
TYPE_UNSIGNED (type) == TYPE_UNSIGNED (TREE_TYPE (arg0))
&& TYPE_UNSIGNED (type) == TYPE_UNSIGNED (TREE_TYPE (arg1))
but only in the !FLOAT_TYPE_P path. We could even compare
TYPE_OVERFLOW_UNDEFINED I think. Or even just make sure
that when TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (arg0)) also
TYPE_OVERFLOW_UNDEFINED (type), thus
!TYPE_OVERFLOW_UNDEFINED (type)
|| ((TREE_CODE (arg0) != MULT_EXPR
|| TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (arg0)))
&& (TREE_CODE (arg1) != MULT_EXPR
|| TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (arg1))))
That is, the newly created multiplication in type TYPE should
either not have undefined overflow or the inner multiplications
all should already have. Best done with a comment in
fold_plusminus_mult_expr.
Richard.
