Unnecessary boxing allocation for unsafe_load() on ptr to non-isbits immutable #50243
Description
Currently, unsafe_load(::Ptr{T}) where T is an immutable struct, heap allocates the new T() instance.
However, there's no need to do that, since immutable structs can be heap allocated, even if they are not isbits, as of a few years ago.
Consider this small MRE:
julia> struct X # isbits
x::Int
end
julia> struct Y # immutable, but not isbits
y::Vector{Int}
end
julia> function caller(v::V) where V
r = Ref(v)
GC.@preserve r begin
p = reinterpret(Ptr{V}, pointer_from_objref(r))
callee(p)
end
end
caller (generic function with 1 method)
julia> function callee(p::Ptr)
v = unsafe_load(p)
foo(v)
end
callee (generic function with 1 method)
julia> foo(x::X) = x.x
foo (generic function with 2 methods)
julia> foo(y::Y) = sum(y.y, init=0)
foo (generic function with 2 methods)
julia> const x1 = X(1)
X(1)
julia> const y1 = Y([1,2])
Y([1, 2])
julia> caller(x1)
1
julia> caller(y1)
3
julia> @time caller(x1)
0.000001 seconds
1
julia> @time caller(y1) # One alloc from the Ref(), one alloc from boxing the Y() instance.
0.000002 seconds (2 allocations: 32 bytes)
3I'm not sure why it can only elide the Ref() allocation for ::X, but not for ::Y... Is that a similar optimization opportunity that's currently being missed?
I'm not 100% sure, but i'm pretty sure this can be fixed by just changing this line:
Line 698 in 95749c3
to:
else if (!jl_is_immutable(ety)) {.