Performance regression in packDecimal/numDecimalDigits

[wrengr]

The current version of packDecimal shows a 3x slowdown compared to the version in the benchmark. This is due to the n0 > fromIntegral (maxBound :: Word64) guard in numDecimalDigits. This must be corrected. a 1.2x slowdown may be acceptable, but 3x is not.

[wrengr]

On further investigation, the slowdown is almost surely due to taking the slow-path when we shouldn't, rather than due to the guard itself introducing slowness; that is, the guard is returning true when it shouldn't. The fromIntegral function has modular-arithmetic semantics for all the built-in Int/Word types, and this is causing things to go haywire.

For example, when a ~ Int64 we have that fromIntegral(maxBound::Word64) == -1, which in turn means n0 > -1 will always return true (since we've already filtered out the negative numbers), thus we'll always take the slow-path— even though in truth every single non-negative 'Int64' can be perfectly well embedded into Word64, so there's no actual risk of the fast-path breaking.

The easiest fix is to change the guard to toInteger n0 > toInteger (maxBound::Word64). If the guard itself seems to be introducing the slowness, then we can do some bit-bashing to speed that up.