[ABI] Introduce indirect symbolic references to context descriptors.

Extending the mangling of symbolic references to also include indirect symbolic references. This allows mangled names to refer to context descriptors (both type and protocol) not in the current source file. For now, only permit indirect symbolic references within the current module, because remote mirrors (among other things) is unable to handle relocations. Co-authored-by: Joe Groff <[email protected]>
daeken · Oct 23, 2018 · 5b41ac1 · 5b41ac1
1 parent 8d3da66
commit 5b41ac1
Show file tree

Hide file tree

Showing 23 changed files with 610 additions and 233 deletions.
diff --git a/docs/ABI/Mangling.rst b/docs/ABI/Mangling.rst
@@ -46,6 +46,48 @@ mangled name will start with the module name (after the ``_S``).
 In the following, productions which are only _part_ of an operator, are
 named with uppercase letters.
 
+Symbolic references
+~~~~~~~~~~~~~~~~~~~
+
+The Swift compiler emits mangled names into binary images to encode
+references to types for runtime instantiation and reflection. In a binary,
+these mangled names may embed pointers to runtime data
+structures in order to more efficiently represent locally-defined types.
+We call these pointers **symbolic references**.
+These references will be introduced by a control character in the range
+`\x01` ... `\x1F`, which indicates the kind of symbolic reference, followed by
+some number of arbitrary bytes *which may include null bytes*. Code that
+processes mangled names out of Swift binaries needs to be aware of symbolic
+references in order to properly terminate strings; a null terminator may be
+part of a symbolic reference.
+
+::
+
+  symbolic-reference ::= [\x01-\x17] .{4} // Relative symbolic reference
+   #if sizeof(void*) == 8
+     symbolic-reference ::= [\x18-\x1F] .{8} // Absolute symbolic reference
+   #elif sizeof(void*) == 4
+     symbolic-reference ::= [\x18-\x1F] .{4} // Absolute symbolic reference
+   #endif
+
+Symbolic references are only valid in compiler-emitted metadata structures
+and must only appear in read-only parts of a binary image. APIs and tools
+that interpret Swift mangled names from potentially uncontrolled inputs must
+refuse to interpret symbolic references.
+
+The following symbolic reference kinds are currently implemented:
+
+::
+
+   {any-generic-type, protocol} ::= '\x01' .{4}              // Reference points directly to context descriptor
+   {any-generic-type, protocol} ::= '\x02' .{4}              // Reference points indirectly to context descriptor
+   // The grammatical role of the symbolic reference is determined by the
+   // kind of context descriptor referenced
+
+   protocol-conformance-ref ::= '\x03' .{4}  // Reference points directly to protocol conformance descriptor (NOT IMPLEMENTED)
+   protocol-conformance-ref ::= '\x04' .{4}  // Reference points indirectly to protocol conformance descriptor (NOT IMPLEMENTED)
+
+
 Globals
 ~~~~~~~
 
@@ -553,18 +595,36 @@ Generics
 
 ::
 
-  protocol-conformance ::= type protocol module generic-signature?
+  protocol-conformance-context ::= protocol module generic-signature?
+
+  protocol-conformance ::= type protocol-conformance-context
 
 ``<protocol-conformance>`` refers to a type's conformance to a protocol. The
 named module is the one containing the extension or type declaration that
 declared the conformance.
 
 ::
 
+  protocol-conformance ::= type protocol
+
+If ``type`` is a generic parameter or associated type of one, then no module
+is mangled, because the conformance must be resolved from the generic
+environment.
+
   protocol-conformance ::= context identifier protocol identifier generic-signature?  // Property behavior conformance
 
 Property behaviors are implemented using private protocol conformances.
 
+::
+
+  concrete-protocol-conformance ::= type protocol-conformance-ref
+  protocol-conformance-ref ::= protocol module?
+
+A compact representation used to represent mangled protocol conformance witness
+arguments at runtime. The ``module`` is only specified for conformances that
+are "retroactive", meaning that the context in which the conformance is defined
+is in neither the protocol or type module.
+
 ::
 
   generic-signature ::= requirement* 'l'     // one generic parameter

diff --git a/include/swift/AST/ASTMangler.h b/include/swift/AST/ASTMangler.h
@@ -46,12 +46,22 @@ class ASTMangler : public Mangler {
   /// If disabled, it is an error to try to mangle such an entity.
   bool AllowNamelessEntities = false;
 
-  /// If nonnull, provides a callback to encode symbolic references to
-  /// type contexts.
-  std::function<bool (const DeclContext *Context)>
-    CanSymbolicReference;
-
-  std::vector<std::pair<const DeclContext *, unsigned>> SymbolicReferences;
+  /// If enabled, some entities will be emitted as symbolic reference
+  /// placeholders. The offsets of these references will be stored in the
+  /// `SymbolicReferences` vector, and it is up to the consumer of the mangling
+  /// to fill these in.
+  bool AllowSymbolicReferences = false;
+
+public:
+  using SymbolicReferent = llvm::PointerUnion<const NominalTypeDecl *,
+                                              const ProtocolConformance *>;
+protected:
+
+  /// If set, the mangler calls this function to determine whether to symbolic
+  /// reference a given entity. Defaults to always returning true.
+  std::function<bool (SymbolicReferent)> CanSymbolicReference;
+
+  std::vector<std::pair<SymbolicReferent, unsigned>> SymbolicReferences;
 
 public:
   enum class SymbolKind {
@@ -292,7 +302,7 @@ class ASTMangler : public Mangler {
 
   void appendOpParamForLayoutConstraint(LayoutConstraint Layout);
 
-  void appendSymbolicReference(const DeclContext *context);
+  void appendSymbolicReference(SymbolicReferent referent);
 
   std::string mangleTypeWithoutPrefix(Type type) {
     appendType(type);

diff --git a/include/swift/Demangling/Demangle.h b/include/swift/Demangling/Demangle.h
@@ -33,6 +33,8 @@ namespace llvm {
 namespace swift {
 namespace Demangle {
 
+enum class SymbolicReferenceKind : uint8_t;
+
 struct DemangleOptions {
   bool SynthesizeSugarOnTypes = false;
   bool DisplayDebuggerGeneratedModule = true;
@@ -473,7 +475,9 @@ void mangleIdentifier(const char *data, size_t length,
 /// This should always round-trip perfectly with demangleSymbolAsNode.
 std::string mangleNode(const NodePointer &root);
 
-using SymbolicResolver = llvm::function_ref<Demangle::NodePointer (const void *)>;
+using SymbolicResolver =
+  llvm::function_ref<Demangle::NodePointer (SymbolicReferenceKind,
+                                            const void *)>;
 
 /// \brief Remangle a demangled parse tree, using a callback to resolve
 /// symbolic references.
@@ -537,6 +541,8 @@ class DemanglerPrinter {
     return std::move(*this << std::forward<T>(x));
   }
 
+  DemanglerPrinter &writeHex(unsigned long long n) &;
+
   std::string &&str() && { return std::move(Stream); }
 
   llvm::StringRef getStringRef() const { return Stream; }

diff --git a/include/swift/Demangling/DemangleNodes.def b/include/swift/Demangling/DemangleNodes.def
@@ -143,6 +143,7 @@ NODE(PrefixOperator)
 NODE(PrivateDeclName)
 NODE(PropertyDescriptor)
 CONTEXT_NODE(Protocol)
+CONTEXT_NODE(ProtocolSymbolicReference)
 NODE(ProtocolConformance)
 NODE(ProtocolDescriptor)
 NODE(ProtocolConformanceDescriptor)
@@ -172,13 +173,13 @@ NODE(SpecializationIsFragile)
 CONTEXT_NODE(Static)
 CONTEXT_NODE(Structure)
 CONTEXT_NODE(Subscript)
-CONTEXT_NODE(SymbolicReference)
 NODE(Suffix)
 NODE(ThinFunctionType)
 NODE(Tuple)
 NODE(TupleElement)
 NODE(TupleElementName)
 NODE(Type)
+CONTEXT_NODE(TypeSymbolicReference)
 CONTEXT_NODE(TypeAlias)
 NODE(TypeList)
 NODE(TypeMangling)
@@ -192,7 +193,6 @@ NODE(TypeMetadataLazyCache)
 NODE(UncurriedFunctionType)
 #define REF_STORAGE(Name, ...) NODE(Name)
 #include "swift/AST/ReferenceStorage.def"
-CONTEXT_NODE(UnresolvedSymbolicReference)
 CONTEXT_NODE(UnsafeAddressor)
 CONTEXT_NODE(UnsafeMutableAddressor)
 NODE(ValueWitness)

diff --git a/include/swift/Demangling/Demangler.h b/include/swift/Demangling/Demangler.h
@@ -280,6 +280,17 @@ class CharVector : public Vector<char> {
   }
 };
 
+/// Kinds of symbolic reference supported.
+enum class SymbolicReferenceKind : uint8_t {
+  /// A symbolic reference to a context descriptor, representing the
+  /// (unapplied generic) context.
+  Context,  
+};
+
+using SymbolicReferenceResolver_t = NodePointer (SymbolicReferenceKind,
+                                                 Directness,
+                                                 int32_t, const void *);
+
 /// The demangler.
 ///
 /// It de-mangles a string and it also owns the returned node-tree. This means
@@ -301,7 +312,7 @@ class Demangler : public NodeFactory {
   StringRef Words[MaxNumWords];
   int NumWords = 0;
 
-  std::function<NodePointer (int32_t, const void *)> SymbolicReferenceResolver;
+  std::function<SymbolicReferenceResolver_t> SymbolicReferenceResolver;
 
   bool nextIf(StringRef str) {
     if (!Text.substr(Pos).startswith(str)) return false;
@@ -472,7 +483,8 @@ class Demangler : public NodeFactory {
 
   NodePointer demangleObjCTypeName();
   NodePointer demangleTypeMangling();
-  NodePointer demangleSymbolicReference(const void *at);
+  NodePointer demangleSymbolicReference(unsigned char rawKind,
+                                        const void *at);
 
   void dump();
 
@@ -483,7 +495,7 @@ class Demangler : public NodeFactory {
 
   /// Install a resolver for symbolic references in a mangled string.
   void setSymbolicReferenceResolver(
-                  std::function<NodePointer (int32_t, const void*)> resolver) {
+                          std::function<SymbolicReferenceResolver_t> resolver) {
     SymbolicReferenceResolver = resolver;
   }
 

diff --git a/include/swift/Demangling/TypeDecoder.h b/include/swift/Demangling/TypeDecoder.h
@@ -114,7 +114,7 @@ class TypeDecoder {
     case NodeKind::Enum:
     case NodeKind::Structure:
     case NodeKind::TypeAlias: // This can show up for imported Clang decls.
-    case NodeKind::SymbolicReference:
+    case NodeKind::TypeSymbolicReference:
     {
       BuiltNominalTypeDecl typeDecl = BuiltNominalTypeDecl();
       BuiltType parent = BuiltType();
@@ -228,7 +228,8 @@ class TypeDecoder {
                                                    IsClassBound);
     }
 
-    case NodeKind::Protocol: {
+    case NodeKind::Protocol:
+    case NodeKind::ProtocolSymbolicReference: {
       if (auto Proto = decodeMangledProtocolType(Node)) {
         return Builder.createProtocolCompositionType(Proto, BuiltType(),
                                                      /*IsClassBound=*/false);
@@ -473,14 +474,14 @@ class TypeDecoder {
   }
 
 private:
-  bool decodeMangledNominalType(const Demangle::NodePointer &node,
+  bool decodeMangledNominalType(Demangle::NodePointer node,
                                 BuiltNominalTypeDecl &typeDecl,
                                 BuiltType &parent) {
     if (node->getKind() == NodeKind::Type)
       return decodeMangledNominalType(node->getChild(0), typeDecl, parent);
 
     Demangle::NodePointer nominalNode;
-    if (node->getKind() == NodeKind::SymbolicReference) {
+    if (node->getKind() == NodeKind::TypeSymbolicReference) {
       // A symbolic reference can be directly resolved to a nominal type.
       nominalNode = node;
     } else {
@@ -519,19 +520,19 @@ class TypeDecoder {
     return true;
   }
 
-  BuiltProtocolDecl decodeMangledProtocolType(
-                                            const Demangle::NodePointer &node) {
+  BuiltProtocolDecl decodeMangledProtocolType(Demangle::NodePointer node) {
     if (node->getKind() == NodeKind::Type)
       return decodeMangledProtocolType(node->getChild(0));
 
-    if (node->getNumChildren() < 2 || node->getKind() != NodeKind::Protocol)
+    if ((node->getNumChildren() < 2 || node->getKind() != NodeKind::Protocol)
+        && node->getKind() != NodeKind::ProtocolSymbolicReference)
       return BuiltProtocolDecl();
 
     return Builder.createProtocolDecl(node);
   }
 
   bool decodeMangledFunctionInputType(
-      const Demangle::NodePointer &node,
+      Demangle::NodePointer node,
       std::vector<FunctionParam<BuiltType>> &params,
       bool &hasParamFlags) {
     // Look through a couple of sugar nodes.
@@ -542,7 +543,7 @@ class TypeDecoder {
     }
 
     auto decodeParamTypeAndFlags =
-        [&](const Demangle::NodePointer &typeNode,
+        [&](Demangle::NodePointer typeNode,
             FunctionParam<BuiltType> &param) -> bool {
       Demangle::NodePointer node = typeNode;
 

diff --git a/include/swift/Reflection/TypeRefBuilder.h b/include/swift/Reflection/TypeRefBuilder.h
@@ -354,15 +354,30 @@ class TypeRefBuilder {
     // demangling out of the referenced context descriptors in the target
     // process.
     Dem.setSymbolicReferenceResolver(
-      [this, &reader](int32_t offset, const void *base) -> Demangle::NodePointer {
-        // Resolve the reference to a remote address.
-        auto remoteAddress = getRemoteAddrOfTypeRefPointer(base);
-        if (remoteAddress == 0)
+    [this, &reader](SymbolicReferenceKind kind,
+                    Directness directness,
+                    int32_t offset, const void *base) -> Demangle::NodePointer {
+      // Resolve the reference to a remote address.
+      auto remoteAddress = getRemoteAddrOfTypeRefPointer(base);
+      if (remoteAddress == 0)
+        return nullptr;
+
+      auto address = remoteAddress + offset;
+      if (directness == Directness::Indirect) {
+        if (auto indirectAddress = reader.readPointerValue(address)) {
+          address = *indirectAddress;
+        } else {
           return nullptr;
-
-        return reader.readDemanglingForContextDescriptor(remoteAddress + offset,
-                                                         Dem);
-      });
+        }
+      }
+
+      switch (kind) {
+      case Demangle::SymbolicReferenceKind::Context:
+        return reader.readDemanglingForContextDescriptor(address, Dem);
+      }
+
+      return nullptr;
+    });
   }
 
   TypeConverter &getTypeConverter() { return TC; }