Login

Floating Point

Floating Point

Miscellaneous Thoughts

Many compilers such as GCC have flags which specify the degree of strictness floating point code requires viz. optimization. Unlike true real-valued arithmetic, most float operations are not actually commutative. For Anselm, it would be nice to have something more fine-grained. As such we should expose operators which tell the compiler where we care about allowing relaxed behavior.

One idea, if we were to make operators containing . signal non-associativity, this might make sense as a suffix for "strict" FP operators like e.g. +..

I'm not sure what the syntax should be, but rounding modes should also be exposed in some manner. Probably, uglier syntax can be allowed for in such circumstances. Alternatively, implicits might be of use here.