faster bmerge numeric and roll

[mattdowle]

Thanks to a report comparing a single column case to base::findInterval, I had a look.
Some low hanging fruit was that INHERITS() (an R API call looking up class attribute and looping through the class vector) was being calling deeply in the recursive call to support integer64. Also the XIND branch, although a trinary, is called in many places and the branch could be raised. dtwiddle() has long been sand in the cogs given that the default rounding is 0 anyway, and there was a todo to get NA for double out of the way up front. Tackling those yielded a 2x speedup. There's too much recursion in the case of a large table rolling to small (many rows rolling to the same row) and I'm hoping for another 2x speedup on that.
Related #2490

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.85%. Comparing base (deb9acc) to head (8e8cbbe).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #5187   +/-   ##
=======================================
  Coverage   98.85%   98.85%           
=======================================
  Files          87       87           
  Lines       17140    17176   +36     
=======================================
+ Hits        16944    16980   +36     
  Misses        196      196           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MichaelChirico]

LMK if you'd like an extra pair of eyes on this / when it's ready

[mattdowle]

Status so far in the branch on this particular large-to-small benchmark. Low hanging fruit without tackling the too-much-recursion yet. The code rework allows the too-much-recursion to be done next.

setDTthreads(1)
n <- 1e7 
x.df <- data.frame(date = as.POSIXct("2000-1-1") + 1:n, price = rnorm(n))
x.dt <- data.table(x.df, key = "date")

n2 <- 100
y.df <- data.frame(date = sort(sample(x.df$date, n2)), adjustment = runif(n2))
y.dt <- data.table(y.df, key = "date")

falign <- function(x, y) {
  i <- findInterval(as.numeric(x$date), as.numeric(y$date))
  i[which(i == 0)] <- NA
  x$price.adj <- x[, "price"] + y[i, "adjustment"]
  x
}

falign.dt <- function(x, y) {
  i <- findInterval(as.numeric(x$date), as.numeric(y$date))
  i[which(i == 0)] <- NA
  x[, price.adj := x[, "price"] + y[i, "adjustment"]]
  x
}

microbenchmark(
  falign(x.df, y.df),
  falign.dt(x.dt, y.dt),
  x.dt[, price.adj := y.dt[x.dt, price+adjustment, roll=TRUE]],
  times=10, unit="s"
))

# v1.14.0
Unit: seconds
                                                              expr       min        lq      mean    median        uq       max neval
                                                falign(x.df, y.df) 0.1488798 0.1542454 0.1617588 0.1593861 0.1620447 0.1992783    10
                                             falign.dt(x.dt, y.dt) 0.2815458 0.2901276 0.2958419 0.2960433 0.3009906 0.3185344    10
 x.dt[, `:=`(price.adj, y.dt[x.dt, price + adjustment, roll = T])] 0.8177566 0.8206021 0.8365346 0.8349606 0.8461528 0.8593044    10

# this branch as of now (3rd line down from 0.83 to 0.57 but still a lot slower than `findInterval`) : 
                                                              expr       min        lq      mean    median        uq       max neval
                                                falign(x.df, y.df) 0.1460495 0.1486709 0.1554336 0.1537538 0.1616508 0.1732326    10
                                             falign.dt(x.dt, y.dt) 0.2766880 0.3006418 0.3010593 0.3025785 0.3060092 0.3161539    10
 x.dt[, `:=`(price.adj, y.dt[x.dt, price + adjustment, roll = T])] 0.5418650 0.5635949 0.5735462 0.5717389 0.5876390 0.5971523    10

[MichaelChirico]

Anyone want to take over this PR? @jangorecki and @tlapak are the only two I see with commits to this file in recent memory.

[jangorecki]

This PR will have conflicts with mergelist, which has been already merged and conflicts resolved to recent master

[tdhock]

@Anirban166 @DorisAmoakohene please investigate speedups with atime

[MichaelChirico]

Marking as draft over label:WIP

[github-actions]

No obvious timing issues in HEAD=bmerge_numeric

Generated via commit 8e8cbbe

Download link for the artifact containing the test results: ↓ atime-results.zip

Task	Duration
R setup and installing dependencies	6 minutes and 28 seconds
Installing different package versions	11 minutes and 40 seconds
Running and plotting the test cases	5 minutes and 18 seconds