[MLIR][LLVM] Fix nameless global import to support use before def case (llvm#111797)

This commit fixes a bug in the import of nameless globals. Before this
change, the fake symbol names were only generated during the
transformation of the definition. This caused issues when the symbol was
used before it was defined.

Author: Dinistro

mlir/include/mlir/Target/LLVMIR/ModuleImport.h

@@ -355,6 +355,10 @@ class ModuleImport {
   /// and stores a mapping from the struct to the symbol pointing to the
   /// translated operation.
   void processComdat(const llvm::Comdat *comdat);
+  /// Returns a symbol name for a nameless global. MLIR, in contrast to LLVM,
+  /// always requires a symbol name.
+  FlatSymbolRefAttr
+  getOrCreateNamelessSymbolName(llvm::GlobalVariable *globalVar);
 
   /// Builder pointing at where the next instruction should be generated.
   OpBuilder builder;

mlir/lib/Target/LLVMIR/ModuleImport.cpp

@@ -874,6 +874,24 @@ Attribute ModuleImport::getConstantAsAttr(llvm::Constant *constant) {
   return {};
 }
 
+FlatSymbolRefAttr
+ModuleImport::getOrCreateNamelessSymbolName(llvm::GlobalVariable *globalVar) {
+  assert(globalVar->getName().empty() &&
+         "expected to work with a nameless global");
+  auto [it, success] = namelessGlobals.try_emplace(globalVar);
+  if (!success)
+    return it->second;
+
+  // Make sure the symbol name does not clash with an existing symbol.
+  SmallString<128> globalName = SymbolTable::generateSymbolName<128>(
+      getNamelessGlobalPrefix(),
+      [this](StringRef newName) { return llvmModule->getNamedValue(newName); },
+      namelessGlobalId);
+  auto symbolRef = FlatSymbolRefAttr::get(context, globalName);
+  it->getSecond() = symbolRef;
+  return symbolRef;
+}
+
 LogicalResult ModuleImport::convertGlobal(llvm::GlobalVariable *globalVar) {
   // Insert the global after the last one or at the start of the module.
   OpBuilder::InsertionGuard guard(builder);
@@ -907,17 +925,10 @@ LogicalResult ModuleImport::convertGlobal(llvm::GlobalVariable *globalVar) {
 
   // Workaround to support LLVM's nameless globals. MLIR, in contrast to LLVM,
   // always requires a symbol name.
-  SmallString<128> globalName(globalVar->getName());
-  if (globalName.empty()) {
-    // Make sure the symbol name does not clash with an existing symbol.
-    globalName = SymbolTable::generateSymbolName<128>(
-        getNamelessGlobalPrefix(),
-        [this](StringRef newName) {
-          return llvmModule->getNamedValue(newName);
-        },
-        namelessGlobalId);
-    namelessGlobals[globalVar] = FlatSymbolRefAttr::get(context, globalName);
-  }
+  StringRef globalName = globalVar->getName();
+  if (globalName.empty())
+    globalName = getOrCreateNamelessSymbolName(globalVar).getValue();
+
   GlobalOp globalOp = builder.create<GlobalOp>(
       mlirModule.getLoc(), type, globalVar->isConstant(),
       convertLinkageFromLLVM(globalVar->getLinkage()), StringRef(globalName),
@@ -1100,13 +1111,14 @@ FailureOr<Value> ModuleImport::convertConstant(llvm::Constant *constant) {
   }
 
   // Convert global variable accesses.
-  if (auto *globalVar = dyn_cast<llvm::GlobalObject>(constant)) {
-    Type type = convertType(globalVar->getType());
-    StringRef globalName = globalVar->getName();
+  if (auto *globalObj = dyn_cast<llvm::GlobalObject>(constant)) {
+    Type type = convertType(globalObj->getType());
+    StringRef globalName = globalObj->getName();
     FlatSymbolRefAttr symbolRef;
     // Empty names are only allowed for global variables.
     if (globalName.empty())
-      symbolRef = namelessGlobals[cast<llvm::GlobalVariable>(globalVar)];
+      symbolRef =
+          getOrCreateNamelessSymbolName(cast<llvm::GlobalVariable>(globalObj));
     else
       symbolRef = FlatSymbolRefAttr::get(context, globalName);
     return builder.create<AddressOfOp>(loc, type, symbolRef).getResult();

mlir/test/Target/LLVMIR/Import/global-variables.ll

@@ -330,3 +330,14 @@ declare void @"mlir.llvm.nameless_global_2"()
 !5 = !DIBasicType(name: "char", size: 8, encoding: DW_ATE_signed_char)
 !6 = !{}
 !7 = !{i32 2, !"Debug Info Version", i32 3}
+
+; // -----
+
+; Verify that unnamed globals can also be referenced before they are defined.
+
+; CHECK:  llvm.mlir.global internal constant @reference()
+; CHECK:    llvm.mlir.addressof @mlir.llvm.nameless_global_0 : !llvm.ptr
+@reference = internal constant ptr @0
+
+; CHECK:  llvm.mlir.global private unnamed_addr constant @mlir.llvm.nameless_global_0("0\00")
+@0 = private unnamed_addr constant [2 x i8] c"0\00"