Analyzer: Introduce an analyzer to recommend the more efficient string splitting features in .NET 8

[geeknoid]

Introduce a performance-centric analyzer that detects uses of String.Split and recommends the new more efficient string splitting functionality from MemoryExtensions. I'm not sure a fixer can be provided here, it might be too complicated for that. But having an analyzer to point out an obvious place that can get a perf boost is still useful.

Category : Performance
Severity : Info

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Introduce a performance-centric analyzer that detects uses of String.Split and recommends the new more efficient string splitting functionality from MemoryExtensions. I'm not sure a fixer can be provided here, it might be too complicated for that. But having an analyzer to point out an obvious place that can get a perf boost is still useful.

Author:	geeknoid
Assignees:	-
Labels:	area-System.Runtime, untriaged, code-analyzer
Milestone:	-

[buyaa-n]

[Triage note]: We shouldn't flag for all string.Split cases, need to look for patterns where the new API could be used. Based on the pattern a fixer will be offered. When #934 implemented we could replace more string.Split cases

Category : Performance
Severity : Info

Question:

• What pattern the analyzer should look for?

• We might also want to flag String.Trim, String.TrimEnt, String.TrimStart and suggest new Trim, TrimEnd, TrimStart from MemoryExtensions. Maybe with different ID as these could have code-fixers.

[bartonjs]

Video

Out of a concern of false positives, it seems like the initial version of this analyzer should be limited to those callsites where a count was passed into string.Split, as that pattern has a perfect analogue in the span-based split.

Using uncounted string.Split in a foreach would be better handled by an iterator-based split (which we don't have yet). Other uses of uncounted split might be analyzable, but they're harder to describe, so they should probably be a different diagnostic ID (once the pattern can be analyzed) just to keep the docs/explanation sane.

The suggestion regarding Trim/TrimEnd/TrimStart may have merit, but seems like it might need some more thought (as it depends heavily on what happens after the call to Trim)... and should be addressed in a separate issue, with more details/examples.

Category: Performance
Severity: Info