On Thu, Jan 23, 2020 at 09:14:42AM +0000, Richard Sandiford wrote:
> > The other patch is something suggested by Richard S., avoid using OImode
> > for this and instead use a partial int mode that is smaller. This is still
> > playing with fire because even the partial int mode is larger than
> > MAX_BITSIZE_MODE_ANY_INT, but at least its precision fits into those 3
> > WIDE_INT_MAX_ELTS wide-int.h uses.
>
> I think we should increase MAX_BITSIZE_MODE_ANY_INT to the precision
> of POImode at the same time, and make that precision less than 192
> (191 maybe?). That way everything is "correct" without increasing
> the size of wide_int.
The 192 was chosen just to be nice and round, I guess it could be just 130
or say 160 (128 + 32).
Uros, would you be ok with bumping MAX_BITSIZE_MODE_ANY_INT to 160
and changing the POImode patch to use 160 bits instead of 192
if it passes testing?
I guess it would still affect the sizes of the wide-int.cc arrays:
unsigned HOST_HALF_WIDE_INT
u[4 * MAX_BITSIZE_MODE_ANY_INT / HOST_BITS_PER_HALF_WIDE_INT];
unsigned HOST_HALF_WIDE_INT
v[4 * MAX_BITSIZE_MODE_ANY_INT / HOST_BITS_PER_HALF_WIDE_INT];
/* The '2' in 'R' is because we are internally doing a full
multiply. */
unsigned HOST_HALF_WIDE_INT
r[2 * 4 * MAX_BITSIZE_MODE_ANY_INT / HOST_BITS_PER_HALF_WIDE_INT];
so if we didn't want that, we would need to use 132; but those
are weird anyway, because they round down rather than up.
Jakub
