Incremental generators

Incremental generators

The classic ISourceGenerator re-runs everything on every change — fine for tiny tools, brutal for IDEs. IIncrementalGenerator instead builds a pipeline of transformations the compiler tracks by value equality: a keystroke re-runs only the steps whose inputs actually changed.

var provider = ctx.SyntaxProvider.ForAttributeWithMetadataName(...)
    .Select(static (name, _) => name.ToUpperInvariant())   // pure transform
    .Collect();                                            // gather into ImmutableArray

ctx.RegisterSourceOutput(provider, static (spc, names) => { /* emit once */ });

The performance contract, from the Roslyn team's own guidance:

1. Transformations must be pure and cheap — they may run many times; their results are cached by equality.
2. ForAttributeWithMetadataName first. It is a heavily-optimized provider for the dominant pattern (find types with attribute X); hand- rolled CreateSyntaxProvider predicates run more often and do more work.
3. Collect() deliberately — collecting defeats incrementalism for the steps after it; do per-item work before collecting, then emit once.
4. Lambda hygiene: all pipeline lambdas must be static (no captured variables) — capturing silently breaks the caching contract.

When generators go wrong in production, it is almost always a pipeline that recomputes a whole-project model per keystroke — the fix is always "push work later, make steps cheap, cache by value".