I noticed that using rmp_serde to deserialize large structs has a considerable effect on the compile time. It also has twice the impact of comparable crates like serde_json when generating deserialization code.
The solution here is to take as much code as possible out of generic methods that are instantiated for each type. This reduces the LLVM IR lines from 3652241 to 3238664 (-11.3%) and the time to compile rmp_serde::from_slice::<MyBigStruct>(&[]) from 52s to 33s (-36%). For reference, doing the same with serde_json takes about 25s to compile.
Here are the LLVM stats of the worst offenders before and after the changes. any_inner is still pretty high up and with some more elbow grease it might be possible to reduce it more, but I don't feel comfortable going any further in this PR.
I noticed that using
rmp_serde
to deserialize large structs has a considerable effect on the compile time. It also has twice the impact of comparable crates likeserde_json
when generating deserialization code.The solution here is to take as much code as possible out of generic methods that are instantiated for each type. This reduces the LLVM IR lines from 3652241 to 3238664 (-11.3%) and the time to compile
rmp_serde::from_slice::<MyBigStruct>(&[])
from 52s to 33s (-36%). For reference, doing the same withserde_json
takes about 25s to compile.Here are the LLVM stats of the worst offenders before and after the changes.
any_inner
is still pretty high up and with some more elbow grease it might be possible to reduce it more, but I don't feel comfortable going any further in this PR.before
after