precompile fails to cache code for calls that are not fully specialized (non-concrete signature)

[KristofferC]

In #46690, I want to deserialize an object on a precompile worker and it is quite important that this is fully precompiled since a lot of precompile workers are spawned and each of these have a separate process and thus they would all need to compile the same code over and over if this isn't the case.

For arbitrary data types, deserialize has a function that takes ::DataType and will not get specialized:

julia/stdlib/Serialization/src/Serialization.jl

Line 1468 in aae8c48

function deserialize(s::AbstractSerializer, t::DataType)

As an example workload we can look at the following code

using Serialization

struct MyStruct
    x::String
end

s = MyStruct("foo")

io = IOBuffer()
serialize(io, s)
seekstart(io)
@time deserialize(io);

Running that with --trace-compile=stderr the following precompile statements are emitted from the deserialize call:

precompile(Tuple{typeof(Serialization.deserialize), Serialization.Serializer{Base.GenericIOBuffer{Array{UInt8, 1}}}, DataType})

However, restarting Julia and executing these before the deserialize call we can see that the precompile call for the DataType fails:

julia> precompile(Tuple{typeof(Serialization.deserialize), Serialization.Serializer{Base.IOStream}, DataType})
false

and the compilation time is still there.

However, if we specialize the function on the data type (as done in 6e0b2da#diff-0fbb91f060d958d93fcf0a101ade4d87b365f3664a409276c591d7faf77acb67) the precompilation statements we get is:

precompile(Tuple{typeof(Serialization.deserialize), Serialization.Serializer{Base.GenericIOBuffer{Array{UInt8, 1}}}, Type{Main.MyStruct}})

the compilation time is then removed:

julia> @time deserialize(io);
  0.000014 seconds (14 allocations: 1.148 KiB)

It would be good if the code that is compiled for signatures that are non-concrete could be cached. They clearly are when you run the code since you don't have to compile it for every call.

[timholy]

What happens if you explicitly add @nospecialize(t::DataType)?

[KristofferC]

That doesn't seem to change anything (running with --trace-compile=stderr):

julia> precompile(Tuple{typeof(Serialization.deserialize), Base.GenericIOBuffer{Array{UInt8, 1}}})
precompile(Tuple{typeof(Serialization.deserialize), Base.GenericIOBuffer{Array{UInt8, 1}}})
true

julia> precompile(Tuple{typeof(Serialization.deserialize), Serialization.Serializer{Base.GenericIOBuffer{Array{UInt8, 1}}}, DataType})
false

julia> @time deserialize(io);
precompile(Tuple{typeof(Serialization.deserialize), Serialization.Serializer{Base.GenericIOBuffer{Array{UInt8, 1}}}, DataType})
  0.013337 seconds (15.33 k allocations: 1011.251 KiB, 98.18% compilation time)

You can see, the precompile statement for the DataType one is printed again, showing how it failed to cache the one from the explicit precompile call.