gm_microoptimisation_war_crime
An experimental optimisation for Garry's Mod
Wat
This attractively named repository contains a Garry's Mod module that performs a micro optimisation that makes use of LuaJIT's constant folding on all Lua scripts.
The module intercepts the loading of Lua scripts and preprocesses the Lua code to replace all instances of SERVER and CLIENT with true and false depending on the current realm.
The module does not replace these in comments or strings and does not change the size of the file or span of the tokens (by inserting extra whitespace.)
Explanation
In Garry's Mod there are two runtime realms, the server realm and the client realm. Because Lua doesn't have conditional compilation, developers have to use an if statement to determine whether the script is running on the server or the client:
if SERVER then
print("Hello, server!")
end
if CLIENT then
print("Hello, client!")
end
The equivalent C preprocessor code would look like this:
#ifdef SERVER
printf("Hello, server!\n")
#endif
#ifdef CLIENT
printf("Hello, client!\n")
#endif
Results
First I should probably note that this is just a dumb experiment and I expected it to not have much performance impact.
CPU branch predictors are pretty good these days and LuaJIT should hopefully help too to make this not so much of an issue for performance.
Additionally, it's quite rare that a runtime realm check would find itself in a hot loop or hot function, so the impact of overall server/game performance from such an optimisation is probably completely unnoticable.
Bytecode
The generated bytecode is only slightly different after this hack. It would seem that LuaJIT does not do dead code elimination in this case. It does however eliminate the branch and replaces it with a direct jump.
Sample Code
function helloWorld()
if SERVER then
print("Hello SERVER")
end
if CLIENT then
print("Hello CLIENT")
end
end
Unoptimised
0000 FUNCF 3
0001 GGET 0 0 ; "SERVER"
0002 ISF 0
0003 JMP 1 => 0007
0004 GGET 0 1 ; "print"
0005 KSTR 2 2 ; "Hello SERVER"
0006 CALL 0 1 2
0007 GGET 0 3 ; "CLIENT"
0008 ISF 0
0009 JMP 1 => 0013
0010 GGET 0 1 ; "print"
0011 KSTR 2 4 ; "Hello CLIENT"
0012 CALL 0 1 2
0013 RET0 0 1
Optimised
0000 FUNCF 3
0001 GGET 0 0 ; "print"
0002 KSTR 2 1 ; "Hello SERVER"
0003 CALL 0 1 2
0004 JMP 0 => 0005
0005 JMP 0 => 0009
0006 GGET 0 0 ; "print"
0007 KSTR 2 2 ; "Hello CLIENT"
0008 CALL 0 1 2
0009 RET0 0 1
Benchmarks
jit.off() jit.on()
jit.flush()
print(jit.status()) print()
local seed = math.random(1, 2)
do
local function globals()
local low = math.huge
local high = 0
local total = 0
for i = 1, 10 do
local start = SysTime()
for i = 1, 100000 do
local n = 0
if SERVER then
n = n + seed
end
if CLIENT then
n = n + seed
end
end
local delta = SysTime() - start
low = math.min(low, delta)
high = math.max(high, delta)
total = total + delta
end
print("(Global) Unoptimised")
print("Low: " .. low * (10 ^ 6) .. "us")
print("High: " .. high * (10 ^ 6) .. "us")
print("Mean: " .. (total / 10) * (10 ^ 6) .. "us\n")
end
globals()
end
do
local function localised()
local SERVER = SERVER
local CLIENT = CLIENT
local low = math.huge
local high = 0
local total = 0
for i = 1, 10 do
local start = SysTime()
for i = 1, 100000 do
local n = 0
if SERVER then
n = n + seed
end
if CLIENT then
n = n + seed
end
end
local delta = SysTime() - start
low = math.min(low, delta)
high = math.max(high, delta)
total = total + delta
end
print("(Local) Unoptimised")
print("Low: " .. low * (10 ^ 6) .. "us")
print("High: " .. high * (10 ^ 6) .. "us")
print("Mean: " .. (total / 10) * (10 ^ 6) .. "us\n")
end
localised()
end
do
local function localised_no_const()
local SERVER = math.random(1, 2) == 1
local CLIENT = not SERVER
local low = math.huge
local high = 0
local total = 0
for i = 1, 10 do
local start = SysTime()
for i = 1, 100000 do
local n = 0
if SERVER then
n = n + seed
end
if CLIENT then
n = n + seed
end
end
local delta = SysTime() - start
low = math.min(low, delta)
high = math.max(high, delta)
total = total + delta
end
print("(Local NoConst) Unoptimised")
print("Low: " .. low * (10 ^ 6) .. "us")
print("High: " .. high * (10 ^ 6) .. "us")
print("Mean: " .. (total / 10) * (10 ^ 6) .. "us\n")
end
localised_no_const()
end
do
local function optimised()
local low = math.huge
local high = 0
local total = 0
for i = 1, 10 do
local start = SysTime()
for i = 1, 100000 do
local n = 0
if true then
n = n + seed
end
if false then
n = n + seed
end
end
local delta = SysTime() - start
low = math.min(low, delta)
high = math.max(high, delta)
total = total + delta
end
print("Optimised")
print("Low: " .. low * (10 ^ 6) .. "us")
print("High: " .. high * (10 ^ 6) .. "us")
print("Mean: " .. (total / 10) * (10 ^ 6) .. "us")
end
optimised()
end
true SSE2 SSE3 SSE4.1 AMD BMI2 fold cse dce fwd dse narrow loop abc sink fuse
(Global) Unoptimised
Low: 77.999999575695us
High: 115.39999968591us
Mean: 85.85999994466us
(Local) Unoptimised
Low: 34.700000014709us
High: 53.599999773724us
Mean: 39.509999987786us
(Local NoConst) Unoptimised
Low: 34.700000014709us
High: 70.099999902595us
Mean: 39.699999979348us
Optimised
Low: 34.699999559962us
High: 51.699999858101us
Mean: 37.459999975908us
Conclusion
It is clear that most of the overhead from if SERVER and if CLIENT runtime checks come from global lookups.
Unsurprisingly, the branching overhead is extremely small. We can see this because the (Local) Unoptimised average is only ~2.24us difference when compared to the Optimised average.
Because SERVER and CLIENT are just booleans, they will be copied into the localised variables. This could potentially enable additional optimisations since LuaJIT can see that the local value could be constants. I tested this with the Local NoConst benchmark but it seemed to have negligible difference, so in this case I don't think it was relevant.
In conclusion from these observations, this module can optimise runtime checking of realm when looking up from the global table but has negligible impact when those globals have been localised.
Alternative Solution
A module that prepends common globals as locals to the top of scripts could be an alternative, and potentially even better method for this, but is definitely more error prone as Lua has a maximum upvalue amount and adding new locals to the top of a script could cause an error for this reason. However, a more advanced module that implements a real GLua lexer + parser could catch this issue and skip optimising that file.