array_population_benchmark.lua

local function time(times, fn)
	local before = os.clock()
	for i=1,times do
		fn()
	end
	local after  = os.clock()
	return after - before
end

local function bench(name, times, fn)
	for i=1,times/100 do
		fn()
	end
	local n = 10
	local total, min, max = 0, math.huge, -math.huge
	for i=1,n do
		local res = time(times, fn)
		total = total + res
		min = math.min(min, res)
		max = math.max(max, res)
	end
	print(string.format("%-20s avg: %.6f min: %.6f max: %.6f)", name..":", total/n, min, max))
end

-----

local iterations = 1e4
local length = 100

bench("table.insert", iterations, function()
	local tab = {}
	for i=1, length do
		table.insert(tab, n)
	end
end)

bench("# operator", iterations, function()
	local tab = {}
	for i=1, length do
		tab[#tab+1] = i
	end
end)

bench("loop index", iterations, function()
	local tab = {}
	for i=1, length do
		tab[i] = i
	end
end)

print("----------")

bench("table.insert with n", iterations, function()
	local tab = { n = length }
	for i=1, length do
		table.insert(tab, n)
	end
end)

bench("# operator with n", iterations, function()
	local tab = { n = length }
	for i=1, length do
		tab[#tab+1] = i
	end
end)

bench("loop index with n", iterations, function()
	local tab = { n = length }
	for i=1, length do
		tab[i] = i
	end
end)

results.md

Using the current version of PUC Lua:

table.insert:        avg: 0.063740 min: 0.062325 max: 0.066512)
# operator:          avg: 0.026996 min: 0.026450 max: 0.027567)
loop index:          avg: 0.016690 min: 0.016462 max: 0.017085)
----------
table.insert with n: avg: 0.066462 min: 0.065171 max: 0.067876)
# operator with n:   avg: 0.030463 min: 0.029716 max: 0.031586)
loop index with n:   avg: 0.019961 min: 0.019849 max: 0.020091)

As expected, using the loop index is considerably faster.

Using LuaJIT 2.1:

table.insert:        avg: 0.004689 min: 0.004570 max: 0.004863)
# operator:          avg: 0.021941 min: 0.021793 max: 0.022196)
loop index:          avg: 0.009231 min: 0.008349 max: 0.012566)
----------
table.insert with n: avg: 0.008029 min: 0.007963 max: 0.008175)
# operator with n:   avg: 0.028864 min: 0.028739 max: 0.029108)
loop index with n:   avg: 0.010937 min: 0.010106 max: 0.011259)

With -joff (JIT compiler disabled)

table.insert:        avg: 0.027672 min: 0.027415 max: 0.029117)
# operator:          avg: 0.030651 min: 0.030455 max: 0.031265)
loop index:          avg: 0.012623 min: 0.012569 max: 0.012662)
----------
table.insert with n: avg: 0.030695 min: 0.030344 max: 0.032600)
# operator with n:   avg: 0.035854 min: 0.035749 max: 0.036166)
loop index with n:   avg: 0.017069 min: 0.016996 max: 0.017261)

Here we can see that the JIT compiler seems to have a much easier time optimizing the table.insert call, regardless of whether there are non-array indices in the table. The length operator is by far the slowest, unlike in the PUC Lua benchmark.