Hello everyone.
Below is an introductory high level proposal for Semantic ARC. It attempts to lay out the "vision" without going into too many details. The details will be provided in subsequent mini proposals. The first mini proposal (High Level ARC Memory Operations) has been sent out in a separate thread for discussion on swift-dev.
A version in rendered markdown is provided here:
https://gottesmm.github.io/proposals/semantic-arc-high-level-vision.html
···
----
# Preface
For Swift 4, we plan to introduce "Semantic ARC", a major overhaul of ownership
in SIL. This document outlines the problems solved by "Semantic ARC", how it
works, and an engineering plan for implementing it.
# The Problem and the Goal
As defined today, SIL represents retain/release operations as independent
operations, without surfacing the relationships that would the most aggressive
optimizations and fast static verification. Semantic ARC enhances SIL to
preserve crucial information, allowing us to:
1. Quickly determine the source of any ARC errors from the optimizer and SILGen.
2. Implement a new, simpler ARC optimizer that will be faster and more
maintainable.
3. Eliminate retain/release traffic more aggressively using semantic pairing
information.
4. Lay the groundwork for the representation of other forms of ownership, such
as move semantics, in SIL.
# The Engineering Plan
The engineering plan involves the following steps:
1. Replace Low Level Dataflow ARC Operations with High Level SSA ARC operations.
2. Add ownership conventions to block arguments and terminators.
3. Implement a verifier that ensures all paired ownership operations have
compatible conventions.
4. Implement an algorithm that verifies that all ownership operations are paired
exactly once along paths through the program.
5. Represent Address Only Types using SSA values instead of Memory Locations.
Each of these steps will be described in detail by focused mini-proposals sent to swift-dev.