[MLIR] Fix duplicated attribute nodes in MLIR bytecode deserialization (#151267)

Fixes #150163 

MLIR bytecode does not preserve alias definitions, so each attribute
encountered during deserialization is treated as a new one. This can
generate duplicate `DISubprogram` nodes during deserialization.

The patch adds a `StringMap` cache that records attributes and fetches
them when encountered again.

Author: hankluo6

mlir/include/mlir/AsmParser/AsmParser.h

@@ -53,7 +53,8 @@ parseAsmSourceFile(const llvm::SourceMgr &sourceMgr, Block *block,
 /// null terminated.
 Attribute parseAttribute(llvm::StringRef attrStr, MLIRContext *context,
                          Type type = {}, size_t *numRead = nullptr,
-                         bool isKnownNullTerminated = false);
+                         bool isKnownNullTerminated = false,
+                         llvm::StringMap<Attribute> *attributesCache = nullptr);
 
 /// This parses a single MLIR type to an MLIR context if it was valid. If not,
 /// an error diagnostic is emitted to the context.

mlir/lib/AsmParser/DialectSymbolParser.cpp

@@ -245,6 +245,15 @@ static Symbol parseExtendedSymbol(Parser &p, AsmParserState *asmState,
       return nullptr;
   }
 
+  if constexpr (std::is_same_v<Symbol, Attribute>) {
+    auto &cache = p.getState().symbols.attributesCache;
+    auto cacheIt = cache.find(symbolData);
+    // Skip cached attribute if it has type.
+    if (cacheIt != cache.end() && !p.getToken().is(Token::colon))
+      return cacheIt->second;
+
+    return cache[symbolData] = createSymbol(dialectName, symbolData, loc);
+  }
   return createSymbol(dialectName, symbolData, loc);
 }
 
@@ -337,6 +346,7 @@ Type Parser::parseExtendedType() {
 template <typename T, typename ParserFn>
 static T parseSymbol(StringRef inputStr, MLIRContext *context,
                      size_t *numReadOut, bool isKnownNullTerminated,
+                     llvm::StringMap<Attribute> *attributesCache,
                      ParserFn &&parserFn) {
   // Set the buffer name to the string being parsed, so that it appears in error
   // diagnostics.
@@ -348,6 +358,9 @@ static T parseSymbol(StringRef inputStr, MLIRContext *context,
   SourceMgr sourceMgr;
   sourceMgr.AddNewSourceBuffer(std::move(memBuffer), SMLoc());
   SymbolState aliasState;
+  if (attributesCache)
+    aliasState.attributesCache = *attributesCache;
+
   ParserConfig config(context);
   ParserState state(sourceMgr, config, aliasState, /*asmState=*/nullptr,
                     /*codeCompleteContext=*/nullptr);
@@ -358,6 +371,11 @@ static T parseSymbol(StringRef inputStr, MLIRContext *context,
   if (!symbol)
     return T();
 
+  if constexpr (std::is_same_v<T, Attribute>) {
+    if (attributesCache)
+      *attributesCache = state.symbols.attributesCache;
+  }
+
   // Provide the number of bytes that were read.
   Token endTok = parser.getToken();
   size_t numRead =
@@ -374,13 +392,15 @@ static T parseSymbol(StringRef inputStr, MLIRContext *context,
 
 Attribute mlir::parseAttribute(StringRef attrStr, MLIRContext *context,
                                Type type, size_t *numRead,
-                               bool isKnownNullTerminated) {
+                               bool isKnownNullTerminated,
+                               llvm::StringMap<Attribute> *attributesCache) {
   return parseSymbol<Attribute>(
-      attrStr, context, numRead, isKnownNullTerminated,
+      attrStr, context, numRead, isKnownNullTerminated, attributesCache,
       [type](Parser &parser) { return parser.parseAttribute(type); });
 }
 Type mlir::parseType(StringRef typeStr, MLIRContext *context, size_t *numRead,
                      bool isKnownNullTerminated) {
   return parseSymbol<Type>(typeStr, context, numRead, isKnownNullTerminated,
+                           /*attributesCache=*/nullptr,
                            [](Parser &parser) { return parser.parseType(); });
 }

mlir/lib/AsmParser/ParserState.h

@@ -40,6 +40,9 @@ struct SymbolState {
 
   /// A map from unique integer identifier to DistinctAttr.
   DenseMap<uint64_t, DistinctAttr> distinctAttributes;
+
+  /// A map from unique string identifier to Attribute.
+  llvm::StringMap<Attribute> attributesCache;
 };
 
 //===----------------------------------------------------------------------===//

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

@@ -895,6 +895,10 @@ class AttrTypeReader {
   SmallVector<AttrEntry> attributes;
   SmallVector<TypeEntry> types;
 
+  /// The map of cached attributes, used to avoid re-parsing the same
+  /// attribute multiple times.
+  llvm::StringMap<Attribute> attributesCache;
+
   /// A location used for error emission.
   Location fileLoc;
 
@@ -1235,7 +1239,7 @@ LogicalResult AttrTypeReader::parseAsmEntry(T &result, EncodingReader &reader,
         ::parseType(asmStr, context, &numRead, /*isKnownNullTerminated=*/true);
   else
     result = ::parseAttribute(asmStr, context, Type(), &numRead,
-                              /*isKnownNullTerminated=*/true);
+                              /*isKnownNullTerminated=*/true, &attributesCache);
   if (!result)
     return failure();
 

mlir/test/IR/recursive-distinct-attr.mlir

@@ -0,0 +1,13 @@
+// RUN: mlir-opt -emit-bytecode %s | mlir-opt --mlir-print-debuginfo | FileCheck %s
+
+// Verify that the distinct attribute which is used transitively
+// through two aliases does not end up duplicated when round-tripped
+// through bytecode.
+
+// CHECK: distinct[0]
+// CHECK-NOT: distinct[1]
+#attr_ugly = #test<attr_ugly begin distinct[0]<> end>
+#attr_ugly1 = #test<attr_ugly begin #attr_ugly end>
+
+module attributes {test.alias = #attr_ugly, test.alias1 = #attr_ugly1} {
+}