On Wed, Nov 26, 2014 at 06:39:44PM +0100, Marek Polacek wrote:
> Both C11 and C++14 standards specify that integral promotions are
> performed on both operands of a shift-expression. This we do just
> fine. But then we convert the right operand to integer_type_node.
> Not only is this unnecessary, it can also be harfmul, because for
> e.g.
> void
> foo (unsigned int x)
> {
> if (-1 >> x != -1)
> bar ();
> }
> with (int) x we lose info that x is nonnegative, which means that
> tree_expr_nonnegative_p cannot fold this expr. Another problem
> with the conversion is that we weren't able to detect e.g. shift
> by 0x100000000ULL, since after the conversion this is 0.
Have you checked what the expander does with it? Say trying to
shift something by __int128 count or similar might upset it.
Wonder about if we don't make similar assumptions in the middle-end.
It might make a difference (sometimes positive, sometimes negative)
for vectorization too.
> 2014-11-26 Marek Polacek <pola...@redhat.com>
>
> PR c/63862
> c-family/
> * c-ubsan.c (ubsan_instrument_shift): Change the type of a MINUS_EXPR
> to op1_utype.
This part is ok regardless of the rest.
Jakub
