Skip to content

Performance tweaks #142

@3Hren

Description

@3Hren

Currently there are performance issues due to indirection between decoding a marker and a value itself; and non-inlining in the core RMP crate. Need to write benchmarks and fix them. The one approach is to mark some functions as [inline] or [inline(always)]. This works, but I feel it a little bit hacky. The other one - is to match for marker bytes directly and avoid using read_marker, but we lose an abstraction then.

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions