jess.bench

Also require('@jess/bench'). Two things a benchmark needs that nothing else in Luau gives it, and nothing more: the harness, the batches, and the statistics are a package's business.

A number means something only on the build releases ship. jess is built with L3i's verified toolchain, cross-language link-time optimization included, and a development build (tools/dev/no-lto.sh, or the MSRV check) runs Luau slower than a release does. A harness that can't tell which build it is on will publish a number from the wrong one, so verified says, and jess --version marks an unverified build too.

jess.bench.verified§

Whether this jess was built with the verified toolchain. A harness refuses to produce a number when it isn't; a time from an unverified build measures a slower Luau, and the honest thing to do with it is nothing.

jess.bench.collect§

Runs a full garbage collection, so one batch's garbage isn't collected inside the next one's timing. Call it between batches, not inside one.

--!strict
local bench = require '@jess/bench'

if not bench.verified then
  error('this jess was built without the verified toolchain; its numbers measure a slower Luau', 0)
end

local numbers = table.create(100000, 1)

local function sum(): number
  local total = 0
  for _, number in numbers do
    total += number
  end

  return total
end

bench.collect()
local started = os.clock()
for _ = 1, 50 do
  sum()
end

print(`{(os.clock() - started) * 1000 / 50} ms per sum`)

On a development build, which is what the example was run on:

error: this jess was built without the verified toolchain; its numbers measure a slower Luau

That is the example working. os.clock is a wall clock; a bench checks its answer before and after timing, so a path that stops doing the work fails instead of posting a time, and it reports a median of several batches, never one run. Benchmarks covers the harness jess's own measurements go through.

Related: Performance.