Roslyn's two views of code

Roslyn's two views of code

Roslyn is the compiler-as-a-service: the same engine that builds your code is a library you reference. It exposes two cooperating layers:

• Syntax layer (SyntaxTree, CSharpSyntaxWalker): the exact text as a tree — every brace, every keyword, including trivia (comments, whitespace). Purely lexical: an unresolvable typo still parses.
• Semantic layer (SemanticModel, ISymbol): what the identifiers MEAN — resolved types, method overloads, accessibility, nullable flow. Needs a Compilation with references.

var tree = CSharpSyntaxTree.ParseText("int x = 1;");
var comp = CSharpCompilation.Create("scratch", new[] { tree },
    new[] { MetadataReference.CreateFromFile(typeof(object).Assembly.Location) });
var model = comp.GetSemanticModel(tree);

Two design decisions make the IDE world possible: syntax trees are immutable (share and cache freely) and repirseparable (a tree remembers how to re-create its exact text, so round-tripping is lossless). Roslyn's red/green tree makes incremental edits cheap — you mostly notice this as "analyzers don't reanalyze the whole file per keystroke".

When to use which: linting shape (naming, structure, forbidden patterns) is syntax; correctness rules (null-ability, type mismatches, symbol misuse) are semantic. Analyzers that need semantic info get it lazily — the registration API takes a predicate deciding whether a node is worth resolving at all.