RefactoringGenerator emits the new method (promoted signature, parameter
names = source symbol names) plus the rewritten call site for every
em 02/03 bucket shape (return slot declared inside/outside, tuple
deconstruction, ref write-back, return-call when the selection ends a
non-void method) and refuses loudly where v1 cannot be sound (composite
trailing return, return not ending the method).
The transformed tree must compile with zero diagnostics (the yak's
round-trip check, enforced in both the CLI and the tests).
Findings that shaped the implementation, all enforced by tests:
- NormalizeWhitespace on the WHOLE root corrupts doc-comment trivia so
Roslyn's XML-doc writer fails with CS1569 ('count (-3) must be
non-negative') — and it would reformat the entire file, killing the
diff. Only generated nodes are normalized; original trivia is kept and
indentation/end-of-line is grafted onto the inserted statements/method.
- ParseTypeName("void") yields a node rejected as a method return type
(CS1547) — a void signature must be PredefinedType(Token(VoidKeyword)).
- NormalizeWhitespace drops the space after the contextual keyword 'var'
before '(' — the parsed 'var (x, y) = ...' deconstruction skeleton is
used verbatim; builder-made statements are normalized.
- The new method is normalized inside a throwaway class wrapper: the
normalizer computes indentation from nesting depth, and a bare method
has none.
v1 decision: print a unified-diff PREVIEW to stdout (hand-rolled LCS
unified diff), never an in-place rewrite. Also fixes the hunk header
off-by-one from the phantom trailing empty line of Split('\n').
399 lines
20 KiB
C#
399 lines
20 KiB
C#
using Microsoft.CodeAnalysis;
|
|
using Microsoft.CodeAnalysis.CSharp;
|
|
using Microsoft.CodeAnalysis.CSharp.Syntax;
|
|
|
|
namespace ExtractMethod.Tooling;
|
|
|
|
/// <summary>
|
|
/// The generated extract-method refactoring (yak em 04 codegen): the NEW
|
|
/// method (the em 03 promoted signature; parameter names come straight from
|
|
/// the source symbols), the CALL STATEMENTS that replace the selection in the
|
|
/// enclosing method, and the fully transformed syntax tree.
|
|
/// </summary>
|
|
/// <param name="NewMethod">The extracted method, ready to be read. Built
|
|
/// with SyntaxFactory and NORMALIZED (whitespace) before insertion, so its
|
|
/// text is the readable shape that lands in the file — not the trivia-less
|
|
/// builder output.</param>
|
|
/// <param name="CallStatements">The statement(s) that replace the selected
|
|
/// run at the call site (one statement, or a capture + destructure pair for
|
|
/// a mixed tuple return), normalized like <see cref="NewMethod"/>.</param>
|
|
/// <param name="TransformedTree">The whole input tree with the selection
|
|
/// moved out and the new method added next to the enclosing one — the
|
|
/// round-trip check compiles exactly this tree.</param>
|
|
public sealed record Refactoring(
|
|
MethodDeclarationSyntax NewMethod,
|
|
IReadOnlyList<StatementSyntax> CallStatements,
|
|
SyntaxTree TransformedTree)
|
|
{
|
|
public string TransformedText => TransformedTree.ToString();
|
|
}
|
|
|
|
/// <summary>
|
|
/// Emits the refactoring behind the em 03 suggestion.
|
|
///
|
|
/// WHY no identifier rewriting at all: the suggested parameter names ARE the
|
|
/// source variable names (em 02/03 promotion), so the moved statements bind
|
|
/// unchanged — every local/parameter read inside the selection is a parameter
|
|
/// of the new method, every variable declared inside moves with the body, and
|
|
/// the remaining reads are type members visible from the same class
|
|
/// (the new method is inserted into the SAME containing type). This is the
|
|
/// property the resolver (one method body, whole statements) and the
|
|
/// classifier (params bucket = exactly the local/parameter reads) jointly
|
|
/// guarantee; a generic SyntaxRewriter renaming pass would be dead weight in
|
|
/// v1. What IS generated structurally:
|
|
/// - the new method declaration (signature line, static-ness mirrored from
|
|
/// the enclosing method, body = selected statements + appended
|
|
/// <c>return ...</c> for return slots the selection did not already end
|
|
/// with a return of),
|
|
/// - the call statement(s) replacing the selection,
|
|
/// - the insertion of the new method right after the enclosing one.
|
|
///
|
|
/// CALL-SITE RULES (per promoted return slots, v1):
|
|
/// - void → a plain call expression statement. A ref parameter already flows
|
|
/// the value back (e.g. Accumulate: <c>Extract(ref acc, n);</c>).
|
|
/// - one slot DECLARED inside the selection → redeclared at the call site
|
|
/// from the result: <c>int total = Extract(limit);</c>.
|
|
/// - one slot declared OUTSIDE → plain assignment: <c>acc = Extract(...);</c>.
|
|
/// - several slots ALL declared inside → tuple deconstruction:
|
|
/// <c>var (x, y) = Extract(n);</c>.
|
|
/// - several slots, any declared outside → capture + ItemN assignments
|
|
/// (the result is evaluated ONCE, then assigned member-wise).
|
|
/// - the enclosing method's return type is non-void and the selection ends
|
|
/// the method → the call site must be <c>return Extract(...);</c> and the
|
|
/// suggested return type must equal the enclosing one (otherwise: refuse).
|
|
///
|
|
/// REFUSALS (loud, never a silent wrong refactoring):
|
|
/// - the selection contains a return that does NOT end the enclosing method
|
|
/// (v1: select to the end of the method);
|
|
/// - the suggested return type does not match the enclosing method's return
|
|
/// type — including the em 03 "composite trailing return" case, where the
|
|
/// signature is void although the enclosing method returns a value (the
|
|
/// report's note already says: extract a local first, then re-run).
|
|
///
|
|
/// BUILDER STYLE: single-slot declarations and assignments are hand-built
|
|
/// with SyntaxFactory; the deconstruction call <c>var (x, y) = ...;</c> and
|
|
/// the tuple return <c>return (x, y);</c> are PARSED from a text skeleton and
|
|
/// the call slot swapped in via ReplaceNode — the parser knows the
|
|
/// designation/tuple-element shapes a v1 builder would get wrong. (Parse
|
|
/// diagnostics would surface as red nodes and fail the round-trip check
|
|
/// loudly, not silently.)
|
|
///
|
|
/// KNOWN v1 APPROXIMATIONS: type display strings are re-parsed for the
|
|
/// generated declarations (a tuple return with a comma-bearing generic type
|
|
/// element is not supported); the capture name <c>result</c> could in theory
|
|
/// collide with an enclosing local; the name check "suggested return type ==
|
|
/// enclosing return type" is a string comparison, not symbol equality.
|
|
/// </summary>
|
|
public static class RefactoringGenerator
|
|
{
|
|
public static Refactoring Generate(SemanticModel model, SelectionReport selection, SignatureSuggestion signature)
|
|
{
|
|
var method = selection.Method!;
|
|
var body = method.Body!;
|
|
|
|
var declaredInside = SignatureBuilder.DeclaredInsideNames(model, selection);
|
|
var slots = signature.ReturnSlots.ToList();
|
|
var slotNames = slots.Select(s => s.Name).ToList();
|
|
|
|
var containsReturn = selection.Statements.OfType<ReturnStatementSyntax>().Any();
|
|
var selectionEndsBody = ReferenceEquals(selection.Statements[^1], body.Statements[^1]);
|
|
var enclosingNonVoid = method.ReturnType is not PredefinedTypeSyntax voidType
|
|
|| !voidType.Keyword.IsKind(SyntaxKind.VoidKeyword);
|
|
var enclosingReturnType = method.ReturnType.ToString();
|
|
|
|
// ---- soundness refusals (see class comment) ----
|
|
if (containsReturn && enclosingNonVoid)
|
|
{
|
|
if (!selectionEndsBody)
|
|
{
|
|
throw new InvalidOperationException(
|
|
"the selection contains a return that does not end the enclosing method — select to the end of the method (v1 codegen)");
|
|
}
|
|
|
|
if (signature.ReturnType != enclosingReturnType)
|
|
{
|
|
throw new InvalidOperationException(
|
|
signature.ReturnType == "void"
|
|
? "the selection ends with the method's return, but its value is not nameable in v1 — " +
|
|
"extract it into a local first (see the report note), then re-run"
|
|
: $"the suggested return type {signature.ReturnType} does not match the enclosing method's " +
|
|
$"return type {enclosingReturnType} — refusing to generate");
|
|
}
|
|
}
|
|
|
|
// ---- the call expression (parameter names = source names, ref-ness kept) ----
|
|
var arguments = signature.Params
|
|
.Select(p =>
|
|
{
|
|
var argument = SyntaxFactory.Argument(SyntaxFactory.IdentifierName(p.Name));
|
|
return p.ByRef
|
|
? argument.WithRefOrOutKeyword(SyntaxFactory.Token(SyntaxKind.RefKeyword))
|
|
: argument;
|
|
})
|
|
.ToList();
|
|
var invocation = SyntaxFactory.InvocationExpression(
|
|
SyntaxFactory.IdentifierName(signature.MethodName))
|
|
.WithArgumentList(SyntaxFactory.ArgumentList(SyntaxFactory.SeparatedList(arguments)))
|
|
// Normalized once, here: the PARSED deconstruction skeleton (see
|
|
// BuildCallStatements) must not be re-normalized (the normalizer
|
|
// drops the space after the contextual keyword `var` before `(`),
|
|
// so it receives the invocation already properly spaced.
|
|
.NormalizeWhitespace();
|
|
|
|
// Normalizing only the GENERATED nodes keeps the original trivia —
|
|
// and therefore the doc comments — untouched (a whole-root
|
|
// NormalizeWhitespace corrupts doc-comment trivia in a way Roslyn's
|
|
// XML-doc writer rejects with CS1569, and it would also reformat the
|
|
// entire file, making the printed diff useless). The call statements
|
|
// and the new method get their indentation/end-of-line grafted from
|
|
// the selection they replace, so the edit slots into the original
|
|
// formatting.
|
|
var callStatements = BuildCallStatements(slots, slotNames, invocation, declaredInside,
|
|
callMustReturn: containsReturn && enclosingNonVoid).ToList();
|
|
var indent = SyntaxFactory.TriviaList(
|
|
selection.Statements[0].GetLeadingTrivia().Where(t => t.IsKind(SyntaxKind.WhitespaceTrivia)));
|
|
for (var index = 0; index < callStatements.Count; index++)
|
|
{
|
|
var call = callStatements[index].WithLeadingTrivia(indent);
|
|
if (index == callStatements.Count - 1)
|
|
{
|
|
// The replaced statement's trailing end-of-line kept the
|
|
// following token (the next statement or the closing brace)
|
|
// on its own line — carry it over.
|
|
call = call.WithTrailingTrivia(selection.Statements[^1].GetTrailingTrivia());
|
|
}
|
|
|
|
callStatements[index] = call;
|
|
}
|
|
|
|
// Normalize the new method INSIDE a throwaway class wrapper: the
|
|
// normalizer computes indentation from nesting depth, and a bare
|
|
// method node has none — its body would come out flush-left. At class
|
|
// depth the signature lands one level in and the body two. The grafted
|
|
// leading/trailing trivia below replaces whatever the wrapper put
|
|
// around the member.
|
|
var newMethod = (MethodDeclarationSyntax)SyntaxFactory.ClassDeclaration("__normalizer__")
|
|
.WithMembers(SyntaxFactory.SingletonList<MemberDeclarationSyntax>(
|
|
BuildNewMethod(method, selection, signature, slots, slotNames)))
|
|
.NormalizeWhitespace()
|
|
.Members[0];
|
|
var methodIndent = method.GetLeadingTrivia()
|
|
.LastOrDefault(t => t.IsKind(SyntaxKind.WhitespaceTrivia));
|
|
newMethod = newMethod
|
|
.WithLeadingTrivia(SyntaxFactory.TriviaList(SyntaxFactory.EndOfLine("\n"), methodIndent))
|
|
.WithTrailingTrivia(SyntaxFactory.EndOfLine("\n"));
|
|
|
|
// ---- apply to the tree: swap the selection for the call, add the method ----
|
|
var allStatements = body.Statements;
|
|
var firstIndex = allStatements.IndexOf(selection.Statements[0]); // reference identity: same tree
|
|
var newStatements = new List<StatementSyntax>();
|
|
newStatements.AddRange(allStatements.Take(firstIndex));
|
|
newStatements.AddRange(callStatements);
|
|
newStatements.AddRange(allStatements.Skip(firstIndex + selection.Statements.Count));
|
|
var rewrittenMethod = method.WithBody(body.WithStatements(SyntaxFactory.List(newStatements)));
|
|
|
|
// The new method lands right after the enclosing one, in the SAME type
|
|
// (that is what keeps the extract-first own-class members in scope).
|
|
var containingType = method.AncestorsAndSelf().OfType<TypeDeclarationSyntax>().First();
|
|
var memberIndex = containingType.Members.IndexOf(method);
|
|
var newType = containingType.WithMembers(
|
|
containingType.Members.Replace(method, rewrittenMethod).Insert(memberIndex + 1, newMethod));
|
|
|
|
var tree = model.SyntaxTree;
|
|
var visited = new SingleNodeReplacer(containingType, newType).Visit(tree.GetRoot());
|
|
if (visited is null)
|
|
{
|
|
throw new InvalidOperationException("the type replacement rewrote the root away (internal error)");
|
|
}
|
|
|
|
return new Refactoring(newMethod, callStatements, tree.WithRootAndOptions(visited, tree.Options));
|
|
}
|
|
|
|
/// <summary>
|
|
/// The one place a SyntaxRewriter enters this tool (yak em 04's stated
|
|
/// curriculum): swapping a single node by REFERENCE identity. The typed
|
|
/// With*-slices (WithBody, WithMembers) cover the local changes; only the
|
|
/// "put the rewritten type back into the root" step is a whole-tree walk.
|
|
/// </summary>
|
|
private sealed class SingleNodeReplacer : CSharpSyntaxRewriter
|
|
{
|
|
private readonly SyntaxNode _oldNode;
|
|
private readonly SyntaxNode _newNode;
|
|
|
|
public SingleNodeReplacer(SyntaxNode oldNode, SyntaxNode newNode)
|
|
{
|
|
_oldNode = oldNode;
|
|
_newNode = newNode;
|
|
}
|
|
|
|
public override SyntaxNode? Visit(SyntaxNode? node)
|
|
=> ReferenceEquals(node, _oldNode) ? _newNode : base.Visit(node);
|
|
}
|
|
|
|
/// <summary>The call statement(s) replacing the selection (see class comment).</summary>
|
|
private static IReadOnlyList<StatementSyntax> BuildCallStatements(
|
|
IReadOnlyList<ReturnSuggestion> slots,
|
|
IReadOnlyList<string> slotNames,
|
|
InvocationExpressionSyntax invocation,
|
|
ISet<string> declaredInside,
|
|
bool callMustReturn)
|
|
{
|
|
if (callMustReturn)
|
|
{
|
|
// The selection ends the enclosing (non-void) method: the call is
|
|
// the return value, the slots live in the new method.
|
|
return [Norm(SyntaxFactory.ReturnStatement(invocation))];
|
|
}
|
|
|
|
if (slots.Count == 0)
|
|
{
|
|
return [Norm(SyntaxFactory.ExpressionStatement(invocation))];
|
|
}
|
|
|
|
if (slots.Count == 1)
|
|
{
|
|
var slot = slots[0];
|
|
if (declaredInside.Contains(slot.Name))
|
|
{
|
|
// Declared inside the selection: redeclare it at the call site
|
|
// from the result (the declaration statement moved with the body).
|
|
var declarator = SyntaxFactory.VariableDeclarator(slot.Name)
|
|
.WithInitializer(SyntaxFactory.EqualsValueClause(invocation));
|
|
var declaration = SyntaxFactory.VariableDeclaration(
|
|
ParseType(slot.Type), SyntaxFactory.SingletonSeparatedList(declarator));
|
|
return [Norm(SyntaxFactory.LocalDeclarationStatement(declaration))];
|
|
}
|
|
|
|
// Declared outside: the caller already owns it (a parameter, or a
|
|
// local declared before the selection) — plain assignment.
|
|
return [Norm(SyntaxFactory.ExpressionStatement(SyntaxFactory.AssignmentExpression(
|
|
SyntaxKind.SimpleAssignmentExpression,
|
|
SyntaxFactory.IdentifierName(slot.Name),
|
|
invocation)))];
|
|
}
|
|
|
|
// Several return slots.
|
|
if (slots.All(s => declaredInside.Contains(s.Name)))
|
|
{
|
|
// All declared inside: tuple deconstruction at the call site.
|
|
// Parsed from a skeleton (see class comment — builder style), the
|
|
// call swapped in at the placeholder identifier. NOT normalized:
|
|
// the skeleton text is already well-spaced, and NormalizeWhitespace
|
|
// would drop the space after the contextual keyword `var` (it is
|
|
// an identifier token to the normalizer, so `var (` becomes
|
|
// `var(`).
|
|
return [ParseStatementWithCall($"var ({string.Join(", ", slotNames)}) = __call__;", invocation)];
|
|
}
|
|
|
|
// Mixed: evaluate the call ONCE into a capture, then assign each slot
|
|
// from the matching ItemN. (Capture name collision with an enclosing
|
|
// local is a documented v1 limitation.)
|
|
var statements = new List<StatementSyntax>
|
|
{
|
|
Norm(SyntaxFactory.LocalDeclarationStatement(SyntaxFactory.VariableDeclaration(
|
|
SyntaxFactory.PredefinedType(SyntaxFactory.Token(SyntaxKind.VarKeyword)),
|
|
SyntaxFactory.SingletonSeparatedList(SyntaxFactory.VariableDeclarator("result")
|
|
.WithInitializer(SyntaxFactory.EqualsValueClause(invocation)))))),
|
|
};
|
|
for (var index = 0; index < slots.Count; index++)
|
|
{
|
|
statements.Add(Norm(SyntaxFactory.ExpressionStatement(SyntaxFactory.AssignmentExpression(
|
|
SyntaxKind.SimpleAssignmentExpression,
|
|
SyntaxFactory.IdentifierName(slots[index].Name),
|
|
SyntaxFactory.MemberAccessExpression(SyntaxKind.SimpleMemberAccessExpression,
|
|
SyntaxFactory.IdentifierName("result"),
|
|
SyntaxFactory.IdentifierName($"Item{index + 1}"))))));
|
|
}
|
|
|
|
return statements;
|
|
}
|
|
|
|
/// <summary>
|
|
/// Normalizes a builder-made statement: SyntaxFactory tokens carry no
|
|
/// trivia, so without this <c>return</c> and <c>Extract</c> would render
|
|
/// glued together. Only used for BUILDER statements — the parsed skeleton
|
|
/// statements are already well-spaced (see the deconstruction comment).
|
|
/// </summary>
|
|
private static StatementSyntax Norm(StatementSyntax statement) => statement.NormalizeWhitespace();
|
|
|
|
/// <summary>
|
|
/// Parses <paramref name="statementText"/> (containing the placeholder
|
|
/// identifier <c>__call__</c>) and swaps the placeholder for the built
|
|
/// <paramref name="invocation"/> — the parser supplies the syntax shapes
|
|
/// (tuple designations, tuple elements) the v1 builder avoids.
|
|
/// </summary>
|
|
private static StatementSyntax ParseStatementWithCall(string statementText, InvocationExpressionSyntax invocation)
|
|
{
|
|
var parsed = SyntaxFactory.ParseStatement(statementText);
|
|
var placeholder = parsed.DescendantNodes()
|
|
.OfType<IdentifierNameSyntax>()
|
|
.First(n => n.Identifier.ValueText == "__call__");
|
|
return parsed.ReplaceNode(placeholder, invocation);
|
|
}
|
|
|
|
/// <summary>
|
|
/// The extracted method: the promoted signature, static-ness mirrored from
|
|
/// the enclosing method, body = the selected statements plus an appended
|
|
/// <c>return</c> for the return slots (skipped when the selection already
|
|
/// ends in a return — a second one would be unreachable). The appended
|
|
/// tuple return is parsed from its text skeleton (see class comment).
|
|
/// </summary>
|
|
private static MethodDeclarationSyntax BuildNewMethod(
|
|
MethodDeclarationSyntax enclosingMethod,
|
|
SelectionReport selection,
|
|
SignatureSuggestion signature,
|
|
IReadOnlyList<ReturnSuggestion> slots,
|
|
IReadOnlyList<string> slotNames)
|
|
{
|
|
// A static enclosing method can only read statics and parameters, so
|
|
// the extracted body needs no instance — mirror the static modifier.
|
|
var modifiers = enclosingMethod.Modifiers.Any(m => m.IsKind(SyntaxKind.StaticKeyword))
|
|
? new[] { SyntaxFactory.Token(SyntaxKind.StaticKeyword) }
|
|
: Array.Empty<SyntaxToken>();
|
|
|
|
var parameters = signature.Params
|
|
.Select(p =>
|
|
{
|
|
var parameter = SyntaxFactory.Parameter(SyntaxFactory.Identifier(p.Name))
|
|
.WithType(ParseType(p.Type));
|
|
return p.ByRef
|
|
? parameter.WithModifiers(SyntaxFactory.TokenList(
|
|
SyntaxFactory.Token(SyntaxKind.RefKeyword)))
|
|
: parameter;
|
|
})
|
|
.ToList();
|
|
|
|
var bodyStatements = selection.Statements.ToList();
|
|
if (slots.Count > 0 && selection.Statements[^1] is not ReturnStatementSyntax)
|
|
{
|
|
// The slot order here is the same name-ordered sequence the tuple
|
|
// ReturnType was built from, so the emitted value matches the
|
|
// declared element types.
|
|
bodyStatements.Add(slots.Count == 1
|
|
? SyntaxFactory.ReturnStatement(SyntaxFactory.IdentifierName(slotNames[0]))
|
|
: SyntaxFactory.ParseStatement($"return ({string.Join(", ", slotNames)});"));
|
|
}
|
|
|
|
return SyntaxFactory.MethodDeclaration(ParseType(signature.ReturnType), SyntaxFactory.Identifier(signature.MethodName))
|
|
.WithModifiers(SyntaxFactory.TokenList(modifiers))
|
|
.WithParameterList(SyntaxFactory.ParameterList(SyntaxFactory.SeparatedList(parameters)))
|
|
.WithBody(SyntaxFactory.Block(SyntaxFactory.List(bodyStatements)));
|
|
}
|
|
|
|
/// <summary>
|
|
/// Re-parses a type display string (em 02/03 are plain strings) in type
|
|
/// context — <c>ParseTypeName</c> handles plain names, generics and tuple
|
|
/// display strings <c>(T1, T2)</c> alike.
|
|
///
|
|
/// EXCEPT <c>void</c>: parsing "void" in TYPE context yields a node the
|
|
/// compiler rejects as a method return type with CS1547 (keyword 'void'
|
|
/// cannot be used in this context) — the parser that reads real method
|
|
/// declarations produces <c>PredefinedType(Token(VoidKeyword))</c>, so
|
|
/// that is what a void signature must be built from.
|
|
/// Single choke point for the string-to-syntax hop; its correctness is
|
|
/// enforced by the round-trip check and the tests, not by construction.
|
|
/// </summary>
|
|
private static TypeSyntax ParseType(string text) => text == "void"
|
|
? SyntaxFactory.PredefinedType(SyntaxFactory.Token(SyntaxKind.VoidKeyword))
|
|
: SyntaxFactory.ParseTypeName(text);
|
|
} |