5.1 KiB
No support for user inlining (yet)
Summary
Luau's bytecode compiler is smart about inlining functions. We would rather focus on improving this automation than risk the chance of its misuse.
Motivation
context: function inlining replaces function calls with the function's definition to reduce overhead costs.
Whether the goal is to improve instruction locality or perform context specific optimizations, function inlining can be a great resource for developers. At -02
or higher, the compiler is able to automatically decide if functions ought to be inlined. In the currrent state of the language, exposing this to users directly could do more harm than good. Not only do these use cases only cater to a tiny subset of the language's users, it exposes our bytecode compiler to FILL when the larger subset who may not have a working understanding of the compiler adopts it. For Luau, this would happen at around the 5000th recursive call.
local function sum(lst: {number})
return sumAcc(0, 1, lst)
end
@inline
local function sumAcc(acc: number, i: number, lst: {number})
return if i > #lst then acc else sumAcc(acc + lst[i], i + 1, lst)
end
local s = sum({1, 2, 3})
Consider this "inlined" recursive call, it attempts to sum up numbers in a list but it never terminates as sum
is recursively called on lst
which does not shrink. If Luau attempted to inline such a call, it would unroll this and one or two things could happen. If we did have an unroll limit then we would attempt our unrolling up until that point and end up making the program larger than it needs to be or thrashing our stack. The stack will grow beyond the system's stack limit and could potentially eat up resources meant for other processes.
More specifically, due to limited number of cases where the compiler supports inlining, exposing this feature would not be entirely useful. This section further explores common cases were @inline
would not apply.
- Exported a
table
-- A.luau
local mod = {}
@inline
function mod.foo(t)
print(t.foo)
end
return mod -- no table.freeze
-- B.luau
local A = require("./A")
local mod = {}
function mod.foo()
A.foo(mod)
end
return table.freeze(mod)
Consider this example which exports a mutable table mod
. Module B
imports this table and calls it within another function, but because the compiler decides what implementation of this method should be called at runtime through the metatables, it makes the VM unable to guarantee that A.foo()
has not been mutated somewhere between the boundaries of both modules making mod.foo
not suitable for an inline. Even with a frozen table being exported by module B
, its function foo could be mutated before its return leading us back to the same issue.
- "OOP" code
local Account = {}
Account.__index = Account
function Account.new(name)
return setmetatable({ name = name, balance = 0 }, Account)
end
@inline
function Account:deposit(bal)
self.balance += bal
end
Even OOP style code, this mutable resolution problem persists. In this code sample, we define a base Account
class with a deposit method. However, if an instance of this class were made, say DavesAccount
, it could go on to overload the deposit method making this another case that is not suitable for inlining.
- Metamethods
local Vec3 = {}
Vec3.__index = Vec3
function Vec3.new(x, y, z)
return setmetatable({ x = x, y = y, z = z }, Vec3)
end
@inline
function Vec3:__add(rhs)
return Vec3.new(self.x + rhs.x, self.y + rhs.y, self.z + rhs.z)
end
In this example, we declare Vec3
with an inlined __add
(+
) operator overload. To make this inlining work, we would need to do one of two things, (1) confirm the rhs
and lhs
share the same shape (i.e, they both havex
, y
and z
) or (2) ensure the rhs
has a similar operator overload. Due to dynamic dispatch, we cannot make either guarantee making this case unsuitable for overloads.
Design
None (yet).
Drawbacks
None (yet).
Alternatives
Automatic function inlining in Luau is very efficient. We suggest opening github issues with code samples of legitimate cases that we can use to further improve Luau's automatic (function inlining and unrolling) optimizations.