Performance regression for int32 and sfixed32 lists

[themreza]

While benchmarking vtprotobuf in our projects, we noticed a performance regression in case of lists of int32 and sfixed32 numbers.

Marshaling and unmarshaling both seem to be slower with vtprotobuf in case of repeated int32 fields.

Although unmarshaling is faster with vtprotobuf for repeated sfixed32, marshaling is slower.

This repository contains samples of these microbenchmarks: https://github.com/themreza/vtprotobuf-bench/tree/main

What could be causing this? Is there a way to improve the performance?

It would be helpful to have automated benchmarks for different data types comparing vtprotobuf with the built-in proto.Marshal and proto.Unmarshal.