This commit adds the following to the RealFunctions protocol:
static func _relaxedAdd(_:Self, _:Self) -> Self
static func _relaxedMul(_:Self, _:Self) -> Self
These are equivalent to + and *, but have "relaxed semantics"; specifically, they license the compiler to reassociate them and to form FMA nodes, which are both significant optimizations that can easily make many common loops 8-10x faster. These transformation perturb results slightly, so they should not be enabled without care, but the results with the relaxed operations are--for most purposes--"just as good as" (and often better than) what strict operations produce. The main thing to beware of is that they are no longer portable; different compiler versions and different targets and optimization flags will result in different results.
They are underscored because they are not stable API. In particular:
- `RealFunctions` is not really the right protocol for these (and neither is `Real`). I need to do some thinking about where to attach them.
- Even if it were the right protocol, these are more like implementation hooks than the API I really want people to use (TBD).
- I like "relaxed" more than other commonly used idioms ("fast"), but I'm not sure it's the name I ultimately want.
A new compiler check reports a missing import for a conformance
implicitly imported:
```
Sources/_TestSupport/Error.swift:21:13: error: cannot use conformance of
'Float' to 'Real' here; 'RealModule' was not imported by this file
public func componentwiseError(_ tst: Complex<Float>, _ ref:
Complex<Double>) -> Double {
^
```
Address this issue in the project by adding the missing import.
The conformance was breaking CI testing, because the CI bots use a macOS version that predates availability of the _Differentiation module, and the compiler-generated differentation operations have neither availability attributes nor checks, so any reference to those operations simply crashes in dyld because the library is not present. We can add availability, but that's not really accurate either, because these operations are not actually binary-stable, so while we wouldn't crash in dyld, we also might not work correctly in the future. We could mark them unavailable on Apple platforms, but that's also not quite right (and further means that they're not going to get a lot of attention). The best solution seems to be to remove the conformance (which was never source-stable).
The standard-library documentation for `trailingZeroBitCount` states:
> If the value is zero, then `trailingZeroBitCount` is equal to `bitWidth`.
Thus, if `T` is not a fixed-width integer type, and `x == 0`, and `y` has more trailing zeros than the representation of `x`, then the gcd calculation will be incorrect.
In particular, the left-shift by `min(xtz, ytz) == xtz` will not “undo” the earlier right-shift by `ytz`, so the final result will be incorrectly right-shifted by `ytz - xtz`.
To fix this I have added an early exit when `x == 0`. I am not sure if this is the optimal solution, and I’m open to alternatives.
***
This change ensures that `x != 0` at the top of the loop, so I have also converted it from a `while` to a `repeat-while` loop, in order to eliminate a redundant extra comparison on the first iteration.
I don’t know if this actually affects performance, and I’m happy to change it back if that’s preferred.