> I note that you have closed this bug by generating them with bc. However, > I'll still reply inline:
Yeah, switching to bc was good for other reasons; it's now testing a lot more precision than it was before. > Might running a different kernel variant evince this? And/or if the > processor has different hardware extensions? (Alas, I stopped > following new CPU extensions somewhere around the introduction of > MMX!) I was reviewing the x86_64 architecture manual to try and figure out if any new hardware extensions could have changed this. I don't think so? Both should be using SSE for the 64-bit arithmetic operations. > So I don't have direct access to the test machines in question. Holger > does, of course. Shall we pursue him for that, or are you happy with > just using bc…? I think it's worth pursuing to see if there is a compiler or hardware bug here; I'll continue using bc for nickle, but we can extract the program which generates that data from the repository and use it for testing along with the shorter possible reproducer I wrote here. Finding another x86 FPU bug would be notable. -- -keith
signature.asc
Description: PGP signature
