[MLIR][LLVMIR] Add support for empty global ctor/dtor lists (#128969)

LLVM IR emitted in from C++ may contain `@llvm.global_ctors = appending
global [0 x { i32, ptr, ptr }] zeroinitializer`. Before this PR, if we
try to roundtrip code like this from the importer, we'll end up with
nothing in place.

Note that `llvm::appendToGlobalCtors` ignores empty lists and this PR
uses the same approach as `llvm-as`, which doesn't use the utilities
from `llvm/lib/Transforms/Utils/ModuleUtils.cpp` in order to build this
- it calls into creating a global variable from scratch.

Author: bcardosolopes

mlir/lib/Target/LLVMIR/ModuleImport.cpp

@@ -1071,11 +1071,21 @@ LogicalResult
 ModuleImport::convertGlobalCtorsAndDtors(llvm::GlobalVariable *globalVar) {
   if (!globalVar->hasInitializer() || !globalVar->hasAppendingLinkage())
     return failure();
-  auto *initializer =
-      dyn_cast<llvm::ConstantArray>(globalVar->getInitializer());
-  if (!initializer)
+  llvm::Constant *initializer = globalVar->getInitializer();
+
+  bool knownInit = isa<llvm::ConstantArray>(initializer) ||
+                   isa<llvm::ConstantAggregateZero>(initializer);
+  if (!knownInit)
     return failure();
 
+  // ConstantAggregateZero does not engage with the operand initialization
+  // in the loop that follows - there should be no operands. This implies
+  // empty ctor/dtor lists.
+  if (auto *caz = dyn_cast<llvm::ConstantAggregateZero>(initializer)) {
+    if (caz->getElementCount().getFixedValue() != 0)
+      return failure();
+  }
+
   SmallVector<Attribute> funcs;
   SmallVector<int32_t> priorities;
   for (llvm::Value *operand : initializer->operands()) {

mlir/lib/Target/LLVMIR/ModuleTranslation.cpp

@@ -1258,16 +1258,35 @@ LogicalResult ModuleTranslation::convertGlobalsAndAliases() {
     auto dtorOp = dyn_cast<GlobalDtorsOp>(op);
     if (!ctorOp && !dtorOp)
       continue;
-    auto range = ctorOp ? llvm::zip(ctorOp.getCtors(), ctorOp.getPriorities())
-                        : llvm::zip(dtorOp.getDtors(), dtorOp.getPriorities());
-    auto appendGlobalFn =
-        ctorOp ? llvm::appendToGlobalCtors : llvm::appendToGlobalDtors;
-    for (auto symbolAndPriority : range) {
-      llvm::Function *f = lookupFunction(
-          cast<FlatSymbolRefAttr>(std::get<0>(symbolAndPriority)).getValue());
-      appendGlobalFn(*llvmModule, f,
-                     cast<IntegerAttr>(std::get<1>(symbolAndPriority)).getInt(),
-                     /*Data=*/nullptr);
+
+    // The empty / zero initialized version of llvm.global_(c|d)tors cannot be
+    // handled by appendGlobalFn logic below, which just ignores empty (c|d)tor
+    // lists. Make sure it gets emitted.
+    if ((ctorOp && ctorOp.getCtors().empty()) ||
+        (dtorOp && dtorOp.getDtors().empty())) {
+      llvm::IRBuilder<llvm::TargetFolder> builder(
+          llvmModule->getContext(),
+          llvm::TargetFolder(llvmModule->getDataLayout()));
+      llvm::Type *eltTy = llvm::StructType::get(
+          builder.getInt32Ty(), builder.getPtrTy(), builder.getPtrTy());
+      llvm::ArrayType *at = llvm::ArrayType::get(eltTy, 0);
+      llvm::Constant *zeroInit = llvm::Constant::getNullValue(at);
+      (void)new llvm::GlobalVariable(
+          *llvmModule, zeroInit->getType(), false,
+          llvm::GlobalValue::AppendingLinkage, zeroInit,
+          ctorOp ? "llvm.global_ctors" : "llvm.global_dtors");
+    } else {
+      auto range = ctorOp
+                       ? llvm::zip(ctorOp.getCtors(), ctorOp.getPriorities())
+                       : llvm::zip(dtorOp.getDtors(), dtorOp.getPriorities());
+      auto appendGlobalFn =
+          ctorOp ? llvm::appendToGlobalCtors : llvm::appendToGlobalDtors;
+      for (const auto &[sym, prio] : range) {
+        llvm::Function *f =
+            lookupFunction(cast<FlatSymbolRefAttr>(sym).getValue());
+        appendGlobalFn(*llvmModule, f, cast<IntegerAttr>(prio).getInt(),
+                       /*Data=*/nullptr);
+      }
     }
   }
 

mlir/test/Dialect/LLVMIR/global.mlir

@@ -233,6 +233,14 @@ llvm.mlir.global_ctors { ctors = [@ctor], priorities = [0 : i32]}
 
 // -----
 
+// CHECK: llvm.mlir.global_ctors {ctors = [], priorities = []}
+llvm.mlir.global_ctors {ctors = [], priorities = []}
+
+// CHECK: llvm.mlir.global_dtors {dtors = [], priorities = []}
+llvm.mlir.global_dtors {dtors = [], priorities = []}
+
+// -----
+
 llvm.func @dtor() {
   llvm.return
 }

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

@@ -256,6 +256,14 @@ define void @bar() {
 
 ; // -----
 
+; CHECK: llvm.mlir.global_ctors {ctors = [], priorities = []}
+@llvm.global_ctors = appending global [0 x { i32, ptr, ptr }] zeroinitializer
+
+; CHECK: llvm.mlir.global_dtors {dtors = [], priorities = []}
+@llvm.global_dtors = appending global [0 x { i32, ptr, ptr }] zeroinitializer
+
+; // -----
+
 ; Visibility attribute.
 
 ; CHECK: llvm.mlir.global external hidden constant @hidden("string")

mlir/test/Target/LLVMIR/llvmir.mlir

@@ -1859,6 +1859,14 @@ llvm.func @foo() {
 
 // -----
 
+// CHECK: @llvm.global_ctors = appending global [0 x { i32, ptr, ptr }] zeroinitializer
+llvm.mlir.global_ctors {ctors = [], priorities = []}
+
+// CHECK: @llvm.global_dtors = appending global [0 x { i32, ptr, ptr }] zeroinitializer
+llvm.mlir.global_dtors {dtors = [], priorities = []}
+
+// -----
+
 // CHECK: @llvm.global_dtors = appending global [1 x { i32, ptr, ptr }] [{ i32, ptr, ptr } { i32 0, ptr @foo, ptr null }]
 llvm.mlir.global_dtors { dtors = [@foo], priorities = [0 : i32]}