On Tue, 2024-07-16 at 15:38 -0700, Yonghong Song wrote: [...] > diff --git a/tools/testing/selftests/bpf/prog_tests/reg_bounds.c b/tools/testing/selftests/bpf/prog_tests/reg_bounds.c > index eb74363f9f70..c88602908cfe 100644 > --- a/tools/testing/selftests/bpf/prog_tests/reg_bounds.c > +++ b/tools/testing/selftests/bpf/prog_tests/reg_bounds.c > @@ -441,6 +441,22 @@ static struct range range_refine(enum num_t x_t, struct range x, enum num_t y_t, > if (t_is_32(y_t) && !t_is_32(x_t)) { > struct range x_swap; > > + /* If we know that > + * - *x* is in the range of signed 32bit value > + * - *y_cast* range is 32-bit sign non-negative, and > + * then *x* range can be narrowed to the interaction of > + * *x* and *y_cast*. Otherwise, if the new range for *x* > + * allows upper 32-bit 0xffffffff then the eventual new > + * range for *x* will be out of signed 32-bit range > + * which violates the origin *x* range. > + */ > + if (x_t == S64 && y_t == S32 && > + !(y_cast.a & 0xffffffff80000000ULL) && !(y_cast.b & 0xffffffff80000000) && > + (long long)x.a >= S32_MIN && (long long)x.b <= S32_MAX) { > + return range(S64, max_t(S64, y_cast.a, x.a), > + min_t(S64, y_cast.b, x.b)); > + } > + > /* some combinations of upper 32 bits and sign bit can lead to > * invalid ranges, in such cases it's easier to detect them > * after cast/swap than try to enumerate all the conditions > @@ -2108,6 +2124,9 @@ static struct subtest_case crafted_cases[] = { > {S32, U32, {(u32)S32_MIN, 0}, {0, 0}}, > {S32, U32, {(u32)S32_MIN, 0}, {(u32)S32_MIN, (u32)S32_MIN}}, > {S32, U32, {(u32)S32_MIN, S32_MAX}, {S32_MAX, S32_MAX}}, > + {S64, U32, {0x0, 0x1f}, {0xffffffff80000000ULL, 0x000000007fffffffULL}}, > + {S64, U32, {0x0, 0x1f}, {0xffffffffffff8000ULL, 0x0000000000007fffULL}}, > + {S64, U32, {0x0, 0x1f}, {0xffffffffffffff80ULL, 0x000000000000007fULL}}, > }; > > /* Go over crafted hard-coded cases. This is fast, so we do it as part of > > The logic is very similar to kernel implementation but has a difference in generating > the final range. In reg_bounds implementation, the range is narrowed by intersecting > y_cast and x range which are necessary. > > In kernel implementation, there is no interection since we only have one register > and two register has been compared before. > > Eduard, could you take a look at the above code? I think this change is correct. The return clause could be simplified a bit: return range_improve(x_t, x, y_cast); [...]