On Tue, Oct 27, 2020 at 05:46:42PM +0000, Richard Sandiford via Gcc wrote:
> >> and proceeds to call
> >>
> >> wide_int new_lb = wi::set_bit (r.lower_bound (0), 127)
> >>
> >> and creates the value:
> >>
> >> p new_lb
> >> {<wide_int_storage> = {val = {-65535, -1, 0}, len = 2, precision = 128},
> >> static is_sign_extended = true}
> >
> > This is non-canonical and so invalid, if the low HWI has the MSB set
> > and the high HWI is -1, it should have been just
> > val = {-65535}, len = 1, precision = 128}
> >
> > I guess the bug is that wi::set_bit_large doesn't call canonize.
>
> Yeah, looks like a micro-optimisation gone wrong.
> LGTM, thanks.
I've now successfully bootstrapped/regtested the patch and committed to
trunk. I'll backport it later.
2020-10-28 Jakub Jelinek <ja...@redhat.com>
* wide-int.cc (wi::set_bit_large): Call canonize unless setting
msb bit and clearing bits above it.
--- gcc/wide-int.cc.jj 2020-10-19 18:42:41.134426398 +0200
+++ gcc/wide-int.cc 2020-10-27 18:33:38.546703763 +0100
@@ -702,8 +702,11 @@ wi::set_bit_large (HOST_WIDE_INT *val, c
/* If the bit we just set is at the msb of the block, make sure
that any higher bits are zeros. */
if (bit + 1 < precision && subbit == HOST_BITS_PER_WIDE_INT - 1)
- val[len++] = 0;
- return len;
+ {
+ val[len++] = 0;
+ return len;
+ }
+ return canonize (val, len, precision);
}
else
{
Jakub
