There used to not be any way to attach reassociate or contract flags to fadd/fmul from Swift, which meant that we had to use inline C shims to accomplish this. But I added some Builtins to Swift, so now we don't need C to do this any longer.
In cases of underflow, multiplication by reciprocal may not be the same as division. Document this, rather than tightening the behavior, because it's still about as good as multiplication by reciprocal would be for a Real type.
This commit _also_ relaxes the constraint on Augmented algorithms from Real to FloatingPoint. It's worth calling out that fastTwoSum does not work with decimal floating-point types, so the unordered `Augemented.sum(_:_:)` now has a radix check and calls into the ordered twoSum function instead for Decimal. This makes more sense that a `BinaryFloatingPoint` constraint, as otherwise generic code written against Real or FloatingPoint would have to do the check at every call site.
This is potentially useful to identify some bugs, but there are also algorithms (notably Kahan summation) that can be written in terms of fast2sum but would fail this check. Rework documentation to account for the change.
This is potentially useful to identify some bugs, but there are also algorithms (notably Kahan summation) that can be written in terms of fast2sum but would fail this check. Rework documentation to account for the change.
This is formally a source-breaking change. However it is expected to not be source-breaking in practice, because any conforming type either has a real magnitude (in the case of real or complex types), which would necessarily have to support all the AF operations to implement the type, or is perhaps a finite field with a bogus stand-in Magnitude.
It's a useful operation, but it deserves its own API to allow the use of a custom random source.
Also, the existing API blocks some constant-folding operations because the call to the system random source is an opaque optimization barrier, so removing it allows us to produce better codegen for the other rounding modes.
We're also bumping required Swift version to 5.9, and removing the executable test targets for log/log1p, as they're really only useful for development of those algorithms.
IntegerUtilities has RoundingRule that is a superset of FloatingPointRoundingRule; many of these are also useful when working with floating-point. This change provides a `rounding(_:)` function similar to the standard libraries `rounded(_:)` that makes all of them available. It has a different name instead of being a shadow because otherwise existing use sites like `rounded(.down)` would become ambiguous.
These were omitted in the first pass over integer rounding rules, but are generally useful and make good sense to have available. In particular, toNearestOrUp is the natural mode for a lot of fixed-point arithmetic, and frequently has HW support on SIMD units, so it makes good sense to have a name for that. Once you add that, having nearestOrDown also makes sense, and then nearestOrZero should exist by analogy to nearestOrAway. Also beefed up testing and refactored the division rounding code somewhat.
These are used only in TestSupport, but it's still nice to get them right. Signed types had a long-standing bug in how overflow was computed for multiplication, and masking shifts would do the wrong thing when the bitwdith was not a power of two and the shift count was negative. I also added implementations of &*, &+, and &-.
Replaces the rescaling algorithm for Complex division to one inspired by Doug Priest's "Efficient Scaling for Complex Division," with some further tweaks to:
- allow it to work for arbitrary FloatingPoint types, including Float16
- get exactly the same rounding behavior as the un-rescaled path, so that z/w = tz/tw when tz and tw are computed exactly.
- allow future optimizations to hoist a rescaled reciprocal for more speedups.
Unlike Priest, we do not try to avoid spurious overflow in the final computation when the result is very near the overflow boundary but cancellation brings us just inside it. We do not believe that this is a good tradeoff, as complex multiplication overflows in exactly the same way. We will investigate providing opt-in API to avoid this overflow case in a future PR.
Improved how tolerance is applied close to infinity for complex division
(this still could use a more principled approach, but at least we're in
the right ballpark now). Tightened up the tolerance for semi-exhaustive
testing too, since that can be done now.