Allocation sinking in git HEAD

  • From: Mike Pall <mike-1207@xxxxxxxxxx>
  • To: luajit@xxxxxxxxxxxxx
  • Date: Tue, 3 Jul 2012 14:21:20 +0200

LuaJIT git HEAD now contains the new allocation sinking and store
sinking optimization.

This optimization is enabled by default. In case you encounter any
problems and want to check whether they are caused by this
optimization, you can turn it off with: -O-sink

The optimization is geared towards the elimination of short-lived
aggregates. It handles plain Lua tables as well as FFI cdata (e.g.
structs, complex or short arrays). It also handles elimination of
immutable FFI types that are implicitly boxed (e.g. 64 bit ints or
pointers) in more contexts (e.g. loop-carried variables).

This optimization adds quite a bit of complexity, so I'd appreciate
it if it would receive wider testing. Feedback welcome!

Here are a few examples that show the improved performance. The
timings in seconds are for Lua 5.1.5 vs. LuaJIT git HEAD on x86
(32 bit). Lower numbers are better:

Typical point class with Lua tables:

  local point
  point = {
    new = function(self, x, y)
      return setmetatable({x=x, y=y}, self)
    __add = function(a, b)
     return point:new(a.x + b.x, a.y + b.y)
  point.__index = point
  local a, b = point:new(1.5, 2.5), point:new(3.25, 4.75)
  for i=1,1e8 do a = (a + b) + b end
  print(a.x, a.y)

140.0  Lua
 26.9  LuaJIT -O-sink
  0.35 LuaJIT -O+sink *** 400x faster than Lua ***

Typical point class with cdata struct:

  local ffi = require("ffi")
  local point
  point = ffi.metatype("struct { double x, y; }", {
    __add = function(a, b)
     return point(a.x + b.x, a.y + b.y)
  local a, b = point(1.5, 2.5), point(3.25, 4.75)
  for i=1,1e8 do a = (a + b) + b end
  print(a.x, a.y)

 10.9  LuaJIT -O-sink
  0.20 LuaJIT -O+sink *** 700x faster than Lua ***

64 bit arithmetic in a loop:

  local x = 0LL
  for i=1,1e9 do x = x + 100 end

 45.8  LuaJIT -O-sink (x86)
 40.9  LuaJIT -O-sink (x64)
  0.84 LuaJIT -O+sink (x86)
  0.48 LuaJIT -O+sink (x64)


Other related posts: