Re: [PATCH] libgcc: Fix up __divmodbitint4 [PR114755]
On Thu, Apr 18, 2024 at 11:25:43AM +0200, Christophe Lyon wrote: > On Thu, 18 Apr 2024 at 09:37, Jakub Jelinek wrote: > > The following testcase aborts on aarch64-linux but does not on x86_64-linux. > > In both cases there is UB in the __divmodbitint4 implemenetation. > > When the divisor is negative with most significant limb (even when partial) > > all ones, has at least 2 limbs and the second most significant limb has the > > most significant bit clear, when this number is negated, it will have 0 > > in the most significant limb. > > Already in the PR114397 r14-9592 fix I was dealing with such divisors, but > > thought the problem is only if because of that un < vn doesn't imply the > > quotient is 0 and remainder u. > > But as this testcase shows, the problem is with such divisors always. > > What happens is that we use __builtin_clz* on the most significant limb, > > and assume it will not be 0 because that is UB for the builtins. > > Normally the most significant limb of the divisor shouldn't be 0, as > > guaranteed by the bitint_reduce_prec e.g. for the positive numbers, unless > > the divisor is just 0 (but for vn == 1 we have special cases). > > Just curious: could this have been caught by ubsan? (I don't know if > it knows about clz) ubsan does instrument clz, I don't remember right now if even libgcc is built with -fsanitize=undefined during bootstrap-ubsan, if it is, it probably should (but we didn't have this test in the testsuite). Jakub
Re: [PATCH] libgcc: Fix up __divmodbitint4 [PR114755]
On Thu, 18 Apr 2024 at 09:37, Jakub Jelinek wrote:
>
> Hi!
>
> The following testcase aborts on aarch64-linux but does not on x86_64-linux.
> In both cases there is UB in the __divmodbitint4 implemenetation.
> When the divisor is negative with most significant limb (even when partial)
> all ones, has at least 2 limbs and the second most significant limb has the
> most significant bit clear, when this number is negated, it will have 0
> in the most significant limb.
> Already in the PR114397 r14-9592 fix I was dealing with such divisors, but
> thought the problem is only if because of that un < vn doesn't imply the
> quotient is 0 and remainder u.
> But as this testcase shows, the problem is with such divisors always.
> What happens is that we use __builtin_clz* on the most significant limb,
> and assume it will not be 0 because that is UB for the builtins.
> Normally the most significant limb of the divisor shouldn't be 0, as
> guaranteed by the bitint_reduce_prec e.g. for the positive numbers, unless
> the divisor is just 0 (but for vn == 1 we have special cases).
Just curious: could this have been caught by ubsan? (I don't know if
it knows about clz)
Thanks,
Christophe
>
> The following patch moves the handling of this corner case a few lines
> earlier before the un < vn check, because adjusting the vn later is harder.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, plus tested with
> make check-gcc -j32 -k GCC_TEST_RUN_EXPENSIVE=1
> RUNTESTFLAGS="GCC_TEST_RUN_EXPENSIVE=1 dg.exp='*bitint* pr112673.c
> builtin-stdc-bit-*.c pr112566-2.c pr112511.c' dg-torture.exp=*bitint*
> dfp.exp=*bitint*"
> on aarch64-linux, ok for trunk?
>
> 2024-04-18 Jakub Jelinek
>
> PR libgcc/114755
> * libgcc2.c (__divmodbitint4): Perform the decrement on negative
> v with most significant limb all ones and the second least
> significant limb with most significant bit clear always, regardless of
> un < vn.
>
> * gcc.dg/torture/bitint-69.c: New test.
>
> --- libgcc/libgcc2.c.jj 2024-03-21 13:07:43.629886730 +0100
> +++ libgcc/libgcc2.c2024-04-17 19:00:55.453691368 +0200
> @@ -1705,69 +1705,62 @@ __divmodbitint4 (UBILtype *q, SItype qpr
>USItype rn = ((USItype) rprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
>USItype up = auprec % W_TYPE_SIZE;
>USItype vp = avprec % W_TYPE_SIZE;
> + /* If vprec < 0 and 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, because after negation otherwise v2 would have most
> + significant limb clear. */
> + if (vprec < 0
> + && ((v[BITINT_END (0, vn - 1)] | (vp ? ((UWtype) -1 << vp) : 0))
> + == (UWtype) -1)
> + && vn > 1
> + && (Wtype) v[BITINT_END (1, vn - 2)] >= 0)
> +{
> + vp = 0;
> + --vn;
> +#if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
> + ++v;
> +#endif
> +}
>if (__builtin_expect (un < vn, 0))
> {
> - /* 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)
> - {
> - vp = 0;
> - --vn;
> + /* 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__
> - ++v;
> + 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;
> }
> - else
> + if (!rn)
> + return;
> + if (up)
> {
> - /* q is 0 and r is u. */
> - if (q)
> - __builtin_memset (q, 0, qn * sizeof (UWtype));
> - if (r == NULL)
> + if (uprec > 0)
> + *r = *u & (((UWtype) 1 << up) - 1);
> + else
> + *r = *u | ((UWtype) -1 << up);
> + r += BITINT_INC;
> + if (!--rn)
> 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 += BITIN
Re: [PATCH] libgcc: Fix up __divmodbitint4 [PR114755]
On Thu, 18 Apr 2024, Jakub Jelinek wrote:
> Hi!
>
> The following testcase aborts on aarch64-linux but does not on x86_64-linux.
> In both cases there is UB in the __divmodbitint4 implemenetation.
> When the divisor is negative with most significant limb (even when partial)
> all ones, has at least 2 limbs and the second most significant limb has the
> most significant bit clear, when this number is negated, it will have 0
> in the most significant limb.
> Already in the PR114397 r14-9592 fix I was dealing with such divisors, but
> thought the problem is only if because of that un < vn doesn't imply the
> quotient is 0 and remainder u.
> But as this testcase shows, the problem is with such divisors always.
> What happens is that we use __builtin_clz* on the most significant limb,
> and assume it will not be 0 because that is UB for the builtins.
> Normally the most significant limb of the divisor shouldn't be 0, as
> guaranteed by the bitint_reduce_prec e.g. for the positive numbers, unless
> the divisor is just 0 (but for vn == 1 we have special cases).
>
> The following patch moves the handling of this corner case a few lines
> earlier before the un < vn check, because adjusting the vn later is harder.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, plus tested with
> make check-gcc -j32 -k GCC_TEST_RUN_EXPENSIVE=1
> RUNTESTFLAGS="GCC_TEST_RUN_EXPENSIVE=1 dg.exp='*bitint* pr112673.c
> builtin-stdc-bit-*.c pr112566-2.c pr112511.c' dg-torture.exp=*bitint*
> dfp.exp=*bitint*"
> on aarch64-linux, ok for trunk?
OK.
Thanks,
Richard.
> 2024-04-18 Jakub Jelinek
>
> PR libgcc/114755
> * libgcc2.c (__divmodbitint4): Perform the decrement on negative
> v with most significant limb all ones and the second least
> significant limb with most significant bit clear always, regardless of
> un < vn.
>
> * gcc.dg/torture/bitint-69.c: New test.
>
> --- libgcc/libgcc2.c.jj 2024-03-21 13:07:43.629886730 +0100
> +++ libgcc/libgcc2.c 2024-04-17 19:00:55.453691368 +0200
> @@ -1705,69 +1705,62 @@ __divmodbitint4 (UBILtype *q, SItype qpr
>USItype rn = ((USItype) rprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
>USItype up = auprec % W_TYPE_SIZE;
>USItype vp = avprec % W_TYPE_SIZE;
> + /* If vprec < 0 and 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, because after negation otherwise v2 would have most
> + significant limb clear. */
> + if (vprec < 0
> + && ((v[BITINT_END (0, vn - 1)] | (vp ? ((UWtype) -1 << vp) : 0))
> + == (UWtype) -1)
> + && vn > 1
> + && (Wtype) v[BITINT_END (1, vn - 2)] >= 0)
> +{
> + vp = 0;
> + --vn;
> +#if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
> + ++v;
> +#endif
> +}
>if (__builtin_expect (un < vn, 0))
> {
> - /* 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)
> - {
> - vp = 0;
> - --vn;
> + /* 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__
> - ++v;
> + 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;
> }
> - else
> + if (!rn)
> + return;
> + if (up)
> {
> - /* q is 0 and r is u. */
> - if (q)
> - __builtin_memset (q, 0, qn * sizeof (UWtype));
> - if (r == NULL)
> + if (uprec > 0)
> + *r = *u & (((UWtype) 1 << up) - 1);
> + else
> + *r = *u | ((UWtype) -1 << up);
> + r += BITINT_INC;
> + if (!--rn)
> 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
> -
