Hi Richard, >On Thu, Feb 21, 2019 at 6:09 PM Wilco Dijkstra <wilco.dijks...@arm.com> wrote: >> >> Hi Richard, >> >> >>Fix an issue with sinl (atanl (sqrtl (LDBL_MAX)) returning 0.0 >> >>instead of 1.0 by using x < sqrtl (LDBL_MAX) in match.pd. >> > >> > Wasn't that a intermediate problem with the mpfr exponent range limiting? >> > Please check whether that's still needed. >> >> I tested it with trunk about an hour ago, and it included Jacub's patch. >> Are there other fixes outstanding which haven't been committed yet? > > Not that I know of. Did we root-cause the bogus folding to 0.0? Because > I don't really understand why using < can "fix" this...
Yes, the underlying issue is that build_sinatan_real returns the first value that does overflow when squared. Maybe that wasn't intended, but using less-than on the first value that does overflow works. With my patch (now committed) the test passes in all rounding modes. Like I mentioned, in the future this check could use a much smaller value based on the size of the mantissa - that's safer since you're not close to infinity. > Latest trunk also still gives an assertion failure in mpc with the > gcc.dg/torture/builtin-math-5.c > which started at the same time as the other mpc/mpfr releated issues: > > build/src/mpc/src/pow.c:631: MPC assertion failed: z_imag || mpfr_number_p > (MPC_RE(u)) > build/src/gcc/gcc/testsuite/gcc.dg/torture/builtin-math-5.c:95:3: internal > compiler error: Aborted > 0x6725ab crash_signal > build/src/gcc/gcc/toplev.c:326 > > Ick. Is there a PR about this? This happens when using an old mpc (0.8.2). It's valid according to the configure check, however it works with the 1.0.3 version that download-prerequisites uses. Maybe we should increase the minimum mpc version in configure? Wilco
