On Wednesday, 15 June 2016 at 20:50:25 UTC, deadalnix wrote:
Consider that there are many optimizations that rely on fp not being predictable. For instance, fabs(x) can be transformed into x | 0x80000000 if you don't care about NaN's sign (yes, NaN are signed).
I guess that would be "-fabs(x)"? NaNs are signed, quiet/signalling and may carry an id. Floating point would truly benefit from a more constraint based typing system, but not sure how that would work out with D. (E.g. constraining a type to be non-Nan, non-Infinity, non-Zero etc subtype)
I am thinking that one should keep strict semantics and sloppy semantics separate in the type system. I am not a big fan of letting the compiler switches control it, so maybe float/double should work more like C/C++ or current D.
Then have other types for the strict semantics and allow direct bitmasking operations/simd. The overall solution could be restricted to a useful, but reproducible subset of 754-2008 (avoiding trancendentals). Meaning, you would have to cast to double/float to get non-reproducible behaviour.
Then you can have implicit transition from strict->sloppy, but require explicit casting the other way.
For the strict type the rounding mode probably should be part of the type, enabling interval arithmetics and avoiding the "dynamic typing issues" of run-time rounding mode. Interval arithmetics is something that is hard to get right with the main C/C++ compilers.
