[CIR][CIRGen] Support array def after decl with unknown bound (#375)

Arrays can be first declared without a known bound, and then defined
with a known bound. For example:

```cpp
extern int data[];

int test() { return data[1]; }

int data[3] {1, 2, 3};
```

Currently `clangir` crashes on generating CIR for this case. This is due
to the type of the `data` definition being different from its
declaration. This patch adds support for such a case.

Author: Lancern

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -471,7 +471,8 @@ mlir::Value CIRGenModule::getGlobalValue(const Decl *D) {
 mlir::cir::GlobalOp CIRGenModule::createGlobalOp(CIRGenModule &CGM,
                                                  mlir::Location loc,
                                                  StringRef name, mlir::Type t,
-                                                 bool isCst) {
+                                                 bool isCst,
+                                                 mlir::Operation *insertPoint) {
   mlir::cir::GlobalOp g;
   auto &builder = CGM.getBuilder();
   {
@@ -486,8 +487,12 @@ mlir::cir::GlobalOp CIRGenModule::createGlobalOp(CIRGenModule &CGM,
       builder.setInsertionPoint(curCGF->CurFn);
 
     g = builder.create<mlir::cir::GlobalOp>(loc, name, t, isCst);
-    if (!curCGF)
-      CGM.getModule().push_back(g);
+    if (!curCGF) {
+      if (insertPoint)
+        CGM.getModule().insert(insertPoint, g);
+      else
+        CGM.getModule().push_back(g);
+    }
 
     // Default to private until we can judge based on the initializer,
     // since MLIR doesn't allow public declarations.
@@ -501,6 +506,35 @@ void CIRGenModule::setCommonAttributes(GlobalDecl GD, mlir::Operation *GV) {
   assert(!UnimplementedFeature::setCommonAttributes());
 }
 
+void CIRGenModule::replaceGlobal(mlir::cir::GlobalOp Old,
+                                 mlir::cir::GlobalOp New) {
+  assert(Old.getSymName() == New.getSymName() && "symbol names must match");
+
+  // If the types does not match, update all references to Old to the new type.
+  auto OldTy = Old.getSymType();
+  auto NewTy = New.getSymType();
+  if (OldTy != NewTy) {
+    auto OldSymUses = Old.getSymbolUses(theModule.getOperation());
+    if (OldSymUses.has_value()) {
+      for (auto Use : *OldSymUses) {
+        auto *UserOp = Use.getUser();
+        assert((isa<mlir::cir::GetGlobalOp>(UserOp) ||
+                isa<mlir::cir::GlobalOp>(UserOp)) &&
+               "GlobalOp symbol user is neither a GetGlobalOp nor a GlobalOp");
+
+        if (auto GGO = dyn_cast<mlir::cir::GetGlobalOp>(Use.getUser())) {
+          auto UseOpResultValue = GGO.getAddr();
+          UseOpResultValue.setType(
+              mlir::cir::PointerType::get(builder.getContext(), NewTy));
+        }
+      }
+    }
+  }
+
+  // Remove old global from the module.
+  Old.erase();
+}
+
 /// If the specified mangled name is not in the module,
 /// create and return an mlir GlobalOp with the specified type (TODO(cir):
 /// address space).
@@ -592,11 +626,14 @@ CIRGenModule::getOrCreateCIRGlobal(StringRef MangledName, mlir::Type Ty,
   // mlir::SymbolTable::Visibility::Public is the default, no need to explicitly
   // mark it as such.
   auto GV = CIRGenModule::createGlobalOp(*this, loc, MangledName, Ty,
-                                         /*isConstant=*/false);
+                                         /*isConstant=*/false,
+                                         /*insertPoint=*/Entry.getOperation());
 
   // If we already created a global with the same mangled name (but different
-  // type) before, take its name and remove it from its parent.
-  assert(!Entry && "not implemented");
+  // type) before, replace it with the new global.
+  if (Entry) {
+    replaceGlobal(Entry, GV);
+  }
 
   // This is the first use or definition of a mangled name.  If there is a
   // deferred decl with this name, remember that we need to emit it at the end

clang/lib/CIR/CodeGen/CIRGenModule.h

@@ -220,7 +220,8 @@ class CIRGenModule : public CIRGenTypeCache {
 
   static mlir::cir::GlobalOp createGlobalOp(CIRGenModule &CGM,
                                             mlir::Location loc, StringRef name,
-                                            mlir::Type t, bool isCst = false);
+                                            mlir::Type t, bool isCst = false,
+                                            mlir::Operation *insertPoint = nullptr);
 
   /// Return the mlir::Value for the address of the given global variable.
   /// If Ty is non-null and if the global doesn't exist, then it will be created
@@ -445,6 +446,14 @@ class CIRGenModule : public CIRGenTypeCache {
   void setGVProperties(mlir::Operation *Op, const NamedDecl *D) const;
   void setGVPropertiesAux(mlir::Operation *Op, const NamedDecl *D) const;
 
+  /// Replace the present global `Old` with the given global `New`. Their symbol
+  /// names must match; their types can be different. Usages of the old global
+  /// will be automatically updated if their types mismatch.
+  ///
+  /// This function will erase the old global. This function will NOT insert the
+  /// new global into the module.
+  void replaceGlobal(mlir::cir::GlobalOp Old, mlir::cir::GlobalOp New);
+
   /// Determine whether the definition must be emitted; if this returns \c
   /// false, the definition can be emitted lazily if it's used.
   bool MustBeEmitted(const clang::ValueDecl *D);

clang/test/CIR/CodeGen/array-unknown-bound.cpp

@@ -0,0 +1,14 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-cir %s -o - | FileCheck %s
+
+extern int table[];
+// CHECK: cir.global external @table = #cir.const_array<[#cir.int<1> : !s32i, #cir.int<2> : !s32i, #cir.int<3> : !s32i]> : !cir.array<!s32i x 3>
+
+int *table_ptr = table;
+// CHECK: cir.global external @table_ptr = #cir.global_view<@table> : !cir.ptr<!s32i>
+
+int test() { return table[1]; }
+//      CHECK: cir.func @_Z4testv() -> !s32i extra( {inline = #cir.inline<no>, optnone = #cir.optnone} ) {
+// CHECK-NEXT:    %0 = cir.alloca !s32i, cir.ptr <!s32i>, ["__retval"] {alignment = 4 : i64}
+// CHECK-NEXT:    %1 = cir.get_global @table : cir.ptr <!cir.array<!s32i x 3>>
+
+int table[3] {1, 2, 3};