benchmark various implementations of Bench.to_seq discussed in #1006

[gasche]

Results on my machine:

enumeration (89.71 us) is 77.5% faster than
batseq (399.39 us) which is 3.8% faster than
too strict (415.29 us) which is probably (alpha=40.06%) same speed as
simple (416.74 us)
Saving times to times.flat

This suggests that using an enumeration indeed provides a large
performance advantage over the simple approach.

[gasche]

cc @jabot2

[jabot2]

Looks good

[UnixJunkie]

Does this mean that in general BatEnum is faster than BatSeq ?
I was hoping that BatSeq is at least as fast as Enum.

[gasche]

No, enumeration here is an internal datatype that represents a (pervasive) external iterator, and is then used to implement conversion to both Enum and Seq. I am not sure whether Enum is faster or slower than Seq (I think there are cases where the non-allocating nature of the Enum interface makes is faster), but its implementation is certainly more complex and thus harder to trust.