[clang] [Clang] [NFC] Introduce `DynamicRecursiveASTVisitor` (PR #110040)

Aaron Ballman via cfe-commits cfe-commits at lists.llvm.org
Mon Nov 4 08:46:13 PST 2024


================
@@ -0,0 +1,452 @@
+//=== DynamicRecursiveASTVisitor.cpp - Dynamic AST Visitor Implementation -===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements DynamicRecursiveASTVisitor in terms of the CRTP-based
+// RecursiveASTVisitor.
+//
+//===----------------------------------------------------------------------===//
+#include "clang/AST/DynamicRecursiveASTVisitor.h"
+#include "clang/AST/RecursiveASTVisitor.h"
+
+using namespace clang;
+
+// The implementation of DRAV deserves some explanation:
+//
+// We want to implement DynamicRecursiveASTVisitor without having to inherit or
+// reference RecursiveASTVisitor in any way in the header: if we instantiate
+// RAV in the header, then every user of (or rather every file that uses) DRAV
+// still has to instantiate a RAV, which gets us nowhere. Moreover, even just
+// including RecursiveASTVisitor.h would probably cause some amount of slowdown
+// because we'd have to parse a huge template. For these reasons, the fact that
+// DRAV is implemented using a RAV is solely an implementation detail.
+//
+// As for the implementation itself, DRAV by default acts exactly like a RAV
+// that overrides none of RAV's functions. There are two parts to this:
+//
+//   1. Any function in DRAV has to act like the corresponding function in RAV,
+//      unless overridden by a derived class, of course.
+//
+//   2. Any call to a function by the RAV implementation that DRAV allows to be
+//      overridden must be transformed to a virtual call on the user-provided
+//      DRAV object: if some function in RAV calls e.g. TraverseCallExpr()
+//      during traversal, then the derived class's TraverseCallExpr() must be
+//      called (provided it overrides TraverseCallExpr()).
+//
+// The 'Impl' class is a helper that connects the two implementations; it is
+// a wrapper around a reference to a DRAV that is itself a RecursiveASTVisitor.
+// It overrides every function in RAV *that is virtual in DRAV* to perform a
+// virtual call on its DRAV reference. This accomplishes point 2 above.
+//
+// Point 1 is accomplished by, first, having the base class implementation of
+// each of the virtual functions construct an Impl object (which is actually
+// just a no-op), passing in itself so that any virtual calls use the right
+// vtable. Secondly, it then calls RAV's implementation of that same function
+// *on Impl* (using a qualified call so that we actually call into the RAV
+// implementation instead of Impl's version of that same function); this way,
+// we both execute RAV's implementation for this function only and ensure that
+// calls to subsequent functions call into Impl via CRTP (and Impl then calls
+// back into DRAV and so on).
+//
+// While this ends up constructing a lot of Impl instances (almost one per
+// function call), this doesn't really matter since Impl just holds a single
+// pointer, and everything in this file should get inlined into all the DRAV
+// functions here anyway.
+//
+//===----------------------------------------------------------------------===//
+//
+// The following illustrates how a call to an (overridden) function is actually
+// resolved: given some class 'Derived' that derives from DRAV and overrides
+// TraverseStmt(), if we are traversing some AST, and TraverseStmt() is called
+// by the RAV implementation, the following happens:
+//
+//   1. Impl::TraverseStmt() overrides RAV::TraverseStmt() via CRTP, so the
+//      former is called.
+//
+//   2. Impl::TraverseStmt() performs a virtual call to the visitor (which is
+//      an instance to Derived), so Derived::TraverseStmt() is called.
+//
+//   End result: Derived::TraverseStmt() is executed.
+//
+// Suppose some other function, e.g. TraverseCallExpr(), which is NOT overridden
+// by Derived is called, we get:
+//
+//   1. Impl::TraverseCallExpr() overrides RAV::TraverseCallExpr() via CRTP,
+//      so the former is called.
+//
+//   2. Impl::TraverseCallExpr() performs a virtual call, but since Derived
+//      does not override that function, DRAV::TraverseCallExpr() is called.
+//
+//   3. DRAV::TraverseCallExpr() creates a new instance of Impl, passing in
+//      itself (this doesn't change that the pointer is an instance of Derived);
+//      it then calls RAV::TraverseCallExpr() on the Impl object, which actually
+//      ends up executing RAV's implementation because we used a qualified
+//      function call.
+//
+//   End result: RAV::TraverseCallExpr() is executed,
+namespace {
+struct Impl : RecursiveASTVisitor<Impl> {
+  DynamicRecursiveASTVisitor &Visitor;
+  Impl(DynamicRecursiveASTVisitor &Visitor) : Visitor(Visitor) {}
+
+  bool shouldVisitTemplateInstantiations() const {
+    return Visitor.ShouldVisitTemplateInstantiations;
+  }
+
+  bool shouldWalkTypesOfTypeLocs() const {
+    return Visitor.ShouldWalkTypesOfTypeLocs;
+  }
+
+  bool shouldVisitImplicitCode() const {
+    return Visitor.ShouldVisitImplicitCode;
+  }
+
+  bool shouldVisitLambdaBody() const { return Visitor.ShouldVisitLambdaBody; }
+
+  // Supporting post-order would be very hard because of quirks of the
+  // RAV implementation that only work with CRTP. It also is only used
+  // by less than 5 visitors in the entire code base.
+  bool shouldTraversePostOrder() const { return false; }
+
+  bool TraverseAST(ASTContext &AST) { return Visitor.TraverseAST(AST); }
+  bool TraverseAttr(Attr *At) { return Visitor.TraverseAttr(At); }
+  bool TraverseDecl(Decl *D) { return Visitor.TraverseDecl(D); }
+  bool TraverseType(QualType T) { return Visitor.TraverseType(T); }
+  bool TraverseTypeLoc(TypeLoc TL) { return Visitor.TraverseTypeLoc(TL); }
+  bool TraverseStmt(Stmt *S) { return Visitor.TraverseStmt(S); }
+
+  bool TraverseConstructorInitializer(CXXCtorInitializer *Init) {
+    return Visitor.TraverseConstructorInitializer(Init);
+  }
+
+  bool TraverseTemplateArgument(const TemplateArgument &Arg) {
+    return Visitor.TraverseTemplateArgument(Arg);
+  }
+
+  bool TraverseTemplateArgumentLoc(const TemplateArgumentLoc &ArgLoc) {
+    return Visitor.TraverseTemplateArgumentLoc(ArgLoc);
+  }
----------------
AaronBallman wrote:

I don't think the macro approach is necessary, but it does seem like someday we should be able to specify this in DeclNodes.td/StmtNodes.td so we can tabledrive more of this (and at that point, the macros might pull more weight).

https://github.com/llvm/llvm-project/pull/110040


More information about the cfe-commits mailing list