optimized reduce for range

[gustf]

After @josevalim commit yesterday I was playing around with this trying to optimize it a bit more.

What I got seems to be around 30% faster testing with a "noop" reduce function like

def noop(_x, acc) do
  acc
end

Here is also the implementation as a gist that I used for testing.

Example result I got was

iex(30)> :timer.tc(ReduceTest, :reduce_new, [1..10000000, 1, &ReduceTest.noop/2]) 
{331031, 1}
iex(31)> :timer.tc(ReduceTest, :reduce_new, [10000000..1, 1, &ReduceTest.noop/2]) 
{339605, 1}
iex(32)> :timer.tc(ReduceTest, :reduce, [1..10000000, 1, &ReduceTest.noop/2])    
{474536, 1}
iex(33)> :timer.tc(ReduceTest, :reduce, [10000000..1, 1, &ReduceTest.noop/2])    
{481838, 1}

I am not sure about the "loop unrolling" on lines 1784-1786 and 1796-1798, as they are a bit of an obfuscation. But they gave 10-15% of the speedup, so I kept them for now.

[josevalim]

Let's not do the loop unrolling for now. There are many places we could use it, for example, when traversing lists, and we don't in any of those cases.

[josevalim]

Can we please match on the ranges just once and do the first <= last check inside the function clause? I want to avoid multiple root level checks (even though the compiler likely optimizes it away anyway).

[lexmag]

I think this PR is a good opportunity to replace "faceless" x and y with first and last.

[gustf]

Done! :)

[lexmag]

Thank you @gustf!