On Thu, 21 Mar 2024, Jakub Jelinek wrote:
> Hi!
>
> The Knuth's division algorithm relies on the number of dividend limbs
> to be greater ore equal to number of divisor limbs, which is why
> I've added a special case for un < vn at the start of __divmodbitint4.
> Unfortunately, my assumption that it then implies abs(v) > abs(u) and
> so quotient must be 0 and remainder same as dividend is incorrect.
> This is because this check is done before negation of the operands.
> While bitint_reduce_prec reduces precision from clearly useless limbs,
> the problematic case is when the dividend is unsigned or non-negative
> and divisor is negative. We can have limbs (from MS to LS):
> dividend: 0 M ?...
> divisor: -1 -N ?...
> where M has most significant bit set and M >= N (if M == N then it
> also the following limbs matter) and the most significant limbs can
> be even partial. In this case, the quotient should be -1 rather than
> 0. bitint_reduce_prec will reduce the precision of the dividend so
> that M is the most significant limb, but can't reduce precision of the
> divisor to more than having the -1 as most significant limb, because
> -N doesn't have the most significant bit set.
>
> The following patch fixes it by detecting this problematic case in the
> un < vn handling, and instead of assuming q is 0 and r is u will
> decrease vn by 1 because it knows the later code will negate the divisor
> and it can be then expressed after negation in one fewer limbs.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Richard.
> 2024-03-21 Jakub Jelinek <ja...@redhat.com>
>
> PR libgcc/114397
> * libgcc2.c (__divmodbitint4): Don't assume un < vn always means
> abs(v) > abs(u), check for a special case of un + 1 == vn where
> u is non-negative and v negative and after v's negation vn could
> be reduced by 1.
>
> * gcc.dg/torture/bitint-65.c: New test.
>
> --- libgcc/libgcc2.c.jj 2024-03-15 19:04:27.000000000 +0100
> +++ libgcc/libgcc2.c 2024-03-20 18:23:51.956879521 +0100
> @@ -1707,44 +1707,67 @@ __divmodbitint4 (UBILtype *q, SItype qpr
> USItype vp = avprec % W_TYPE_SIZE;
> if (__builtin_expect (un < vn, 0))
> {
> - /* If abs(v) > abs(u), then q is 0 and r is u. */
> - if (q)
> - __builtin_memset (q, 0, qn * sizeof (UWtype));
> - if (r == NULL)
> - return;
> -#if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
> - r += rn - 1;
> - u += un - 1;
> -#endif
> - if (up)
> - --un;
> - if (rn < un)
> - un = rn;
> - for (rn -= un; un; --un)
> + /* If abs(v) > abs(u), then q is 0 and r is u.
> + Unfortunately un < vn doesn't always mean abs(v) > abs(u).
> + If uprec > 0 and vprec < 0 and vn == un + 1, if the
> + top limb of v is all ones and the second most significant
> + limb has most significant bit clear, then just decrease
> + vn/avprec/vp and continue, after negation both numbers
> + will have the same number of limbs. */
> + if (un + 1 == vn
> + && uprec >= 0
> + && vprec < 0
> + && ((v[BITINT_END (0, vn - 1)] | (vp ? ((UWtype) -1 << vp) : 0))
> + == (UWtype) -1)
> + && (Wtype) v[BITINT_END (1, vn - 2)] >= 0)
> {
> - *r = *u;
> - r += BITINT_INC;
> - u += BITINT_INC;
> + vp = 0;
> + --vn;
> +#if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
> + ++v;
> +#endif
> }
> - if (!rn)
> - return;
> - if (up)
> + else
> {
> - if (uprec > 0)
> - *r = *u & (((UWtype) 1 << up) - 1);
> - else
> - *r = *u | ((UWtype) -1 << up);
> - r += BITINT_INC;
> - if (!--rn)
> + /* q is 0 and r is u. */
> + if (q)
> + __builtin_memset (q, 0, qn * sizeof (UWtype));
> + if (r == NULL)
> return;
> +#if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
> + r += rn - 1;
> + u += un - 1;
> +#endif
> + if (up)
> + --un;
> + if (rn < un)
> + un = rn;
> + for (rn -= un; un; --un)
> + {
> + *r = *u;
> + r += BITINT_INC;
> + u += BITINT_INC;
> + }
> + if (!rn)
> + return;
> + if (up)
> + {
> + if (uprec > 0)
> + *r = *u & (((UWtype) 1 << up) - 1);
> + else
> + *r = *u | ((UWtype) -1 << up);
> + r += BITINT_INC;
> + if (!--rn)
> + return;
> + }
> + UWtype c = uprec < 0 ? (UWtype) -1 : (UWtype) 0;
> + for (; rn; --rn)
> + {
> + *r = c;
> + r += BITINT_INC;
> + }
> + return;
> }
> - UWtype c = uprec < 0 ? (UWtype) -1 : (UWtype) 0;
> - for (; rn; --rn)
> - {
> - *r = c;
> - r += BITINT_INC;
> - }
> - return;
> }
> USItype qn2 = un - vn + 1;
> if (qn >= qn2)
> --- gcc/testsuite/gcc.dg/torture/bitint-65.c.jj 2024-03-20
> 18:41:38.026311007 +0100
> +++ gcc/testsuite/gcc.dg/torture/bitint-65.c 2024-03-20 18:40:18.604397871
> +0100
> @@ -0,0 +1,44 @@
> +/* PR libgcc/114397 */
> +/* { dg-do run { target bitint } } */
> +/* { dg-options "-std=c23" } */
> +/* { dg-skip-if "" { ! run_expensive_tests } { "*" } { "-O0" "-O2" } } */
> +/* { dg-skip-if "" { ! run_expensive_tests } { "-flto" } { "" } } */
> +
> +#if __BITINT_MAXWIDTH__ >= 129
> +int
> +foo (unsigned _BitInt (128) a, _BitInt (129) b)
> +{
> + return a / b;
> +}
> +#endif
> +
> +#if __BITINT_MAXWIDTH__ >= 192
> +int
> +bar (unsigned _BitInt (128) a, _BitInt (192) b)
> +{
> + return a / b;
> +}
> +#endif
> +
> +int
> +main ()
> +{
> +#if __BITINT_MAXWIDTH__ >= 129
> + if (foo (336225022742818342628768636932743029911uwb,
> + -336225022742818342628768636932743029911wb) != -1
> + || foo (336225022742818342628768636932743029912uwb,
> + -336225022742818342628768636932743029911wb) != -1
> + || foo (336225022742818342628768636932743029911uwb,
> + -336225022742818342628768636932743029912wb) != 0)
> + __builtin_abort ();
> +#endif
> +#if __BITINT_MAXWIDTH__ >= 192
> + if (bar (336225022742818342628768636932743029911uwb,
> + -336225022742818342628768636932743029911wb) != -1
> + || bar (336225022742818342628768636932743029912uwb,
> + -336225022742818342628768636932743029911wb) != -1
> + || bar (336225022742818342628768636932743029911uwb,
> + -336225022742818342628768636932743029912wb) != 0)
> + __builtin_abort ();
> +#endif
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
