FloatMathPlugin is based on Sun microsystems fdlibm. This library does not compile correctly on modern C compiler because it uses pointer aliasing to access bits of a double as two int32. Nowadays, pointer aliasing is undefined behavior. The same library was used in java and compiled with special flags, but this is fragile. It is easy to replace this UB, but that mean forking this now unmaintained code, unless we find another already existing fork. No time to search now.
Le lun. 3 déc. 2018 à 22:31, Eliot Miranda notifications@github.com a écrit :
Hi Ben,
ideally we'd have bit-identical FP on all platforms. Currently this is only achievable using Andreas Raab's FloatMath plugin, used by Croquet and Terf, and not yet tested in 64-bits. What that has to say about whether we support the default or 53-bit on x64 I don't know. i guess we should for now simply support the default m ode. In any case it's unlikely to make much difference because the computational model yields results being truncated to 53 bits after each operation (i.e. other than in the graphics subsystem, floating point operations are performed on double-precision floats so there's no scope for extended calculations to use 64-bit precision internally. So feel free to include the above as commentary but I think we want to apply your pull request asap and so we should leave the precision argument for a later day when we have more data (such as a 64-bit Terf to test).
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/OpenSmalltalk/opensmalltalk-vm/pull/313#issuecomment-443877985, or mute the thread https://github.com/notifications/unsubscribe-auth/AhLyW19NRT5NIHT-ZwfbaMs_Mn160JzRks5u1ZhKgaJpZM4Y84NR .