Kenneth Zadeck <zad...@naturalbridge.com> writes:
> + int blocks_needed = (precision + HOST_BITS_PER_WIDE_INT - 1) /
> HOST_BITS_PER_WIDE_INT;
Watch out for the long lines. There are several others in the patch too.
You'll need to clean up the existing ones before the merge :-)
> + case CONST_WIDE_INT:
> + len = CONST_WIDE_INT_NUNITS (x.first);
> + if (small_prec && blocks_needed == len - 1)
> + {
> + int i;
> + for (i = 0; i < len - 1; i++)
> + scratch[i] = CONST_WIDE_INT_ELT (x.first, i);
> + scratch[len - 1] = sext_hwi (CONST_WIDE_INT_ELT (x.first, i),
> small_prec);
> + }
> + else
> + scratch = &CONST_WIDE_INT_ELT (x.first, 0);
> + break;
> +
> +#if TARGET_SUPPORTS_WIDE_INT == 0
> + case CONST_DOUBLE:
> + len = 2;
> + if (small_prec)
> + {
> + scratch[0] = CONST_DOUBLE_LOW (x.first);
> + scratch[1] = sext_hwi (CONST_DOUBLE_HIGH (x.first), small_prec);
> + }
> + else
> + scratch = &CONST_DOUBLE_LOW (x.first);
> + break;
> +#endif
I'd prefer we kept these as-is if possible, and only canonise CONST_INTs.
That ought to be the only problem case, because of GEN_INT not taking a mode.
Code that creates CONST_DOUBLE or CONST_WIDE_INT should already be going
through immed_wide_int_const and so always needs to specify a mode.
Plus we don't support non-power-of-2 precisions yet anyway.
> @@ -1400,9 +1418,9 @@ wi::fits_uhwi_p (const wide_int_ref &x)
> if (x.precision <= HOST_BITS_PER_WIDE_INT)
> return true;
> if (x.len == 1)
> - return x.sign_mask () == 0;
> + return x.get_val ()[0] >= 0;
The rest of the patch seems to use slow () for this.
> if (x.precision < 2 * HOST_BITS_PER_WIDE_INT)
> - return zext_hwi (x.uhigh (), x.precision % HOST_BITS_PER_WIDE_INT) == 0;
> + return x.uhigh () == 0;
> return x.len == 2 && x.uhigh () == 0;
We don't need the x.precision < 2 * HOST_BITS_PER_WIDE_INT case any more.
x.len == 2 && x.uhigh () == 0 handles anything > HOST_BITS_PER_WIDE_INT.
> @@ -1415,9 +1433,7 @@ wi::neg_p (const wide_int_ref &x, signop
> return false;
> if (x.precision == 0)
> return false;
> - if (x.len * HOST_BITS_PER_WIDE_INT > x.precision)
> - return (x.uhigh () >> (x.precision % HOST_BITS_PER_WIDE_INT - 1)) & 1;
> - return x.sign_mask () < 0;
> + return x.shigh () < 0;
> }
We don't need the x.precision == 0 test now.
Looks good to me otherwise.
Thanks,
Richard
