Auto indexing on set()

[arunsrinivasan]

set()'s i argument could also be optimised to use auto indexing..

Relevant SO post: http://stackoverflow.com/q/30997215/559784

[jangorecki]

set i argument takes integer rows identifiers already, so there is space to put index optimisation there.
Index can already be utilised by passing dt[, which=TRUE] results to set's i arg.

library(data.table)
dt = data.table(a=1:4, b=5:2)
setindex(dt, b)
dt[b%in%3:4, which=TRUE, verbose=TRUE]
#Optimized subsetting with index 'b'
#on= matches existing index, using index
#Starting bmerge ...done in 0.000s elapsed (0.000s cpu) 
#Reordering 2 rows after bmerge done in ... 0 secs
#[1] 2 3

This of course will include overhead of [, but if set is used to the same subset of rows on multiple columns then the overhead will be implied only once. What we could eventually improve here is to reduce overhead of [ by figuring out that dt[a==b, which=TRUE] call was made, and redirect it early to C. The thing is that there are multiple kind of queries that could be optimised in such a way. Ultimately we need to move [ to C, or make it more modular #852, so different kinds of queries can be more easily/sooner redirected to C.