[CIR] Add support for constructor aliases

[andykaylor]

This change adds support for handling the -mconstructor-aliases option in CIR. Aliases are not yet correctly lowered to LLVM IR. That will be implemented in a future change.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clangir

Author: Andy Kaylor (andykaylor)

Changes

This change adds support for handling the -mconstructor-aliases option in CIR. Aliases are not yet correctly lowered to LLVM IR. That will be implemented in a future change.

---

Full diff: https://github.com/llvm/llvm-project/pull/145792.diff

7 Files Affected:

• (modified) clang/include/clang/CIR/Dialect/IR/CIROps.td (+2-1)
• (modified) clang/include/clang/CIR/MissingFeatures.h (-1)
• (modified) clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp (+87-2)
• (modified) clang/lib/CIR/CodeGen/CIRGenModule.cpp (+104)
• (modified) clang/lib/CIR/CodeGen/CIRGenModule.h (+17)
• (modified) clang/lib/CIR/Dialect/IR/CIRDialect.cpp (+15-5)
• (added) clang/test/CIR/CodeGen/ctor-alias.cpp (+75)

diff --git a/clang/include/clang/CIR/Dialect/IR/CIROps.td b/clang/include/clang/CIR/Dialect/IR/CIROps.td
index ef77c46b011f7..7e45fa464f9d4 100644
--- a/clang/include/clang/CIR/Dialect/IR/CIROps.td
+++ b/clang/include/clang/CIR/Dialect/IR/CIROps.td
@@ -1772,7 +1772,8 @@ def FuncOp : CIR_Op<"func", [
                        OptionalAttr<StrAttr>:$sym_visibility,
                        UnitAttr:$comdat,
                        OptionalAttr<DictArrayAttr>:$arg_attrs,
-                       OptionalAttr<DictArrayAttr>:$res_attrs);
+                       OptionalAttr<DictArrayAttr>:$res_attrs,
+                       OptionalAttr<FlatSymbolRefAttr>:$aliasee);
 
   let regions = (region AnyRegion:$body);
 
diff --git a/clang/include/clang/CIR/MissingFeatures.h b/clang/include/clang/CIR/MissingFeatures.h
index 9e8944d1114b8..7009d6d7d702a 100644
--- a/clang/include/clang/CIR/MissingFeatures.h
+++ b/clang/include/clang/CIR/MissingFeatures.h
@@ -81,7 +81,6 @@ struct MissingFeatures {
   static bool opFuncMultipleReturnVals() { return false; }
   static bool opFuncAttributesForDefinition() { return false; }
   static bool opFuncMaybeHandleStaticInExternC() { return false; }
-  static bool opFuncGlobalAliases() { return false; }
   static bool setLLVMFunctionFEnvAttributes() { return false; }
   static bool setFunctionAttributes() { return false; }
 
diff --git a/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp b/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
index cd9096a0188a7..1044dfe507471 100644
--- a/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
@@ -79,17 +79,102 @@ void CIRGenItaniumCXXABI::emitInstanceFunctionProlog(SourceLocation loc,
   }
 }
 
+// Find out how to cirgen the complete destructor and constructor
+namespace {
+enum class StructorCIRGen { Emit, RAUW, Alias, COMDAT };
+}
+
+static StructorCIRGen getCIRGenToUse(CIRGenModule &cgm,
+                                     const CXXMethodDecl *md) {
+  if (!cgm.getCodeGenOpts().CXXCtorDtorAliases)
+    return StructorCIRGen::Emit;
+
+  // The complete and base structors are not equivalent if there are any virtual
+  // bases, so emit separate functions.
+  if (md->getParent()->getNumVBases()) {
+    // The return value is correct here, but other support for this is NYI.
+    cgm.errorNYI(md->getSourceRange(), "getCIRGenToUse: virtual bases");
+    return StructorCIRGen::Emit;
+  }
+
+  GlobalDecl aliasDecl;
+  if (const auto *dd = dyn_cast<CXXDestructorDecl>(md)) {
+    // The assignment is correct here, but other support for this is NYI.
+    cgm.errorNYI(md->getSourceRange(), "getCIRGenToUse: dtor");
+    aliasDecl = GlobalDecl(dd, Dtor_Complete);
+  } else {
+    const auto *cd = cast<CXXConstructorDecl>(md);
+    aliasDecl = GlobalDecl(cd, Ctor_Complete);
+  }
+
+  cir::GlobalLinkageKind linkage = cgm.getFunctionLinkage(aliasDecl);
+
+  if (cir::isDiscardableIfUnused(linkage))
+    return StructorCIRGen::RAUW;
+
+  // FIXME: Should we allow available_externally aliases?
+  if (!cir::isValidLinkage(linkage))
+    return StructorCIRGen::RAUW;
+
+  if (cir::isWeakForLinker(linkage)) {
+    // Only ELF and wasm support COMDATs with arbitrary names (C5/D5).
+    if (cgm.getTarget().getTriple().isOSBinFormatELF() ||
+        cgm.getTarget().getTriple().isOSBinFormatWasm())
+      return StructorCIRGen::COMDAT;
+    return StructorCIRGen::Emit;
+  }
+
+  return StructorCIRGen::Alias;
+}
+
+static void emitConstructorDestructorAlias(CIRGenModule &cgm,
+                                           GlobalDecl aliasDecl,
+                                           GlobalDecl targetDecl) {
+  cir::GlobalLinkageKind linkage = cgm.getFunctionLinkage(aliasDecl);
+
+  // Does this function alias already exists?
+  StringRef mangledName = cgm.getMangledName(aliasDecl);
+  auto globalValue = dyn_cast_or_null<cir::CIRGlobalValueInterface>(
+      cgm.getGlobalValue(mangledName));
+  if (globalValue && !globalValue.isDeclaration())
+    return;
+
+  auto entry = cast_or_null<cir::FuncOp>(cgm.getGlobalValue(mangledName));
+
+  // Retrieve aliasee info.
+  auto aliasee = cast<cir::FuncOp>(cgm.getAddrOfGlobal(targetDecl));
+
+  // Populate actual alias.
+  cgm.emitAliasForGlobal(mangledName, entry, aliasDecl, aliasee, linkage);
+}
+
 void CIRGenItaniumCXXABI::emitCXXStructor(GlobalDecl gd) {
   auto *md = cast<CXXMethodDecl>(gd.getDecl());
   auto *cd = dyn_cast<CXXConstructorDecl>(md);
 
+  StructorCIRGen cirGenType = getCIRGenToUse(cgm, md);
+
   if (!cd) {
     cgm.errorNYI(md->getSourceRange(), "CXCABI emit destructor");
     return;
   }
 
-  if (cgm.getCodeGenOpts().CXXCtorDtorAliases)
-    cgm.errorNYI(md->getSourceRange(), "Ctor/Dtor aliases");
+  if (gd.getCtorType() == Ctor_Complete) {
+    GlobalDecl baseDecl = gd.getWithCtorType(Ctor_Base);
+
+    if (cirGenType == StructorCIRGen::Alias ||
+        cirGenType == StructorCIRGen::COMDAT) {
+      emitConstructorDestructorAlias(cgm, gd, baseDecl);
+      return;
+    }
+
+    if (cirGenType == StructorCIRGen::RAUW) {
+      StringRef mangledName = cgm.getMangledName(gd);
+      mlir::Operation *aliasee = cgm.getAddrOfGlobal(baseDecl);
+      cgm.addReplacement(mangledName, aliasee);
+      return;
+    }
+  }
 
   auto fn = cgm.codegenCXXStructor(gd);
 
diff --git a/clang/lib/CIR/CodeGen/CIRGenModule.cpp b/clang/lib/CIR/CodeGen/CIRGenModule.cpp
index f24bee44f26a7..1271f290d6193 100644
--- a/clang/lib/CIR/CodeGen/CIRGenModule.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenModule.cpp
@@ -888,6 +888,69 @@ void CIRGenModule::updateCompletedType(const TagDecl *td) {
   genTypes.updateCompletedType(td);
 }
 
+void CIRGenModule::addReplacement(StringRef name, mlir::Operation *op) {
+  replacements[name] = op;
+}
+
+void CIRGenModule::replacePointerTypeArgs(cir::FuncOp oldF, cir::FuncOp newF) {
+  std::optional<mlir::SymbolTable::UseRange> optionalUseRange =
+      oldF.getSymbolUses(theModule);
+  if (!optionalUseRange)
+    return;
+
+  for (const mlir::SymbolTable::SymbolUse &u : *optionalUseRange) {
+    // CallTryOp only shows up after FlattenCFG.
+    auto call = mlir::dyn_cast<cir::CallOp>(u.getUser());
+    if (!call)
+      continue;
+
+    mlir::OperandRange argOps = call.getArgs();
+    mlir::ArrayRef<mlir::Type> funcArgTypes =
+        newF.getFunctionType().getInputs();
+    // In the case of variadic functions, the call may have more arguments that
+    // the function type, so we can't use llvm::enumerate here.
+    for (unsigned i = 0; i < funcArgTypes.size(); i++) {
+      if (argOps[i].getType() == funcArgTypes[i])
+        continue;
+
+      // The purpose of this entire function is to insert bitcasts in the case
+      // where these types don't match, but I haven't seen a case where that
+      // happens.
+      errorNYI(call.getLoc(), "replace call with mismatched types");
+    }
+  }
+}
+
+void CIRGenModule::applyReplacements() {
+  for (auto &i : replacements) {
+    StringRef mangledName = i.first();
+    mlir::Operation *replacement = i.second;
+    mlir::Operation *entry = getGlobalValue(mangledName);
+    if (!entry)
+      continue;
+    assert(isa<cir::FuncOp>(entry) && "expected function");
+    auto oldF = cast<cir::FuncOp>(entry);
+    auto newF = dyn_cast<cir::FuncOp>(replacement);
+    if (!newF) {
+      // In classic codegen, this can be a global alias, a bitcast, or a GEP.
+      errorNYI(replacement->getLoc(), "replacement is not a function");
+      continue;
+    }
+
+    // LLVM has opaque pointer but CIR not. So we may have to handle these
+    // different pointer types when performing replacement.
+    replacePointerTypeArgs(oldF, newF);
+
+    // Replace old with new, but keep the old order.
+    if (oldF.replaceAllSymbolUses(newF.getSymNameAttr(), theModule).failed())
+      llvm_unreachable("internal error, cannot RAUW symbol");
+    if (newF) {
+      newF->moveBefore(oldF);
+      oldF->erase();
+    }
+  }
+}
+
 // TODO(CIR): this could be a common method between LLVM codegen.
 static bool isVarDeclStrongDefinition(const ASTContext &astContext,
                                       CIRGenModule &cgm, const VarDecl *vd,
@@ -1797,11 +1860,52 @@ CIRGenModule::getGlobalVisibilityAttrFromDecl(const Decl *decl) {
 
 void CIRGenModule::release() {
   emitDeferred();
+  applyReplacements();
 
   // There's a lot of code that is not implemented yet.
   assert(!cir::MissingFeatures::cgmRelease());
 }
 
+void CIRGenModule::emitAliasForGlobal(StringRef mangledName,
+                                      mlir::Operation *op, GlobalDecl aliasGD,
+                                      cir::FuncOp aliasee,
+                                      cir::GlobalLinkageKind linkage) {
+
+  auto *aliasFD = dyn_cast<FunctionDecl>(aliasGD.getDecl());
+  assert(aliasFD && "expected FunctionDecl");
+
+  // The aliasee function type is different from the alias one, this difference
+  // is specific to CIR because in LLVM the ptr types are already erased at this
+  // point.
+  const CIRGenFunctionInfo &fnInfo =
+      getTypes().arrangeCXXStructorDeclaration(aliasGD);
+  cir::FuncType fnType = getTypes().getFunctionType(fnInfo);
+
+  cir::FuncOp alias =
+      createCIRFunction(getLoc(aliasGD.getDecl()->getSourceRange()),
+                        mangledName, fnType, aliasFD);
+  alias.setAliasee(aliasee.getName());
+  alias.setLinkage(linkage);
+  // Declarations cannot have public MLIR visibility, just mark them private
+  // but this really should have no meaning since CIR should not be using
+  // this information to derive linkage information.
+  mlir::SymbolTable::setSymbolVisibility(
+      alias, mlir::SymbolTable::Visibility::Private);
+
+  // Alias constructors and destructors are always unnamed_addr.
+  assert(!cir::MissingFeatures::opGlobalUnnamedAddr());
+
+  // Switch any previous uses to the alias.
+  if (op) {
+    errorNYI(aliasFD->getSourceRange(), "emitAliasForGlobal: previous uses");
+  } else {
+    // Name already set by createCIRFunction
+  }
+
+  // Finally, set up the alias with its proper name and attributes.
+  setCommonAttributes(aliasGD, alias);
+}
+
 mlir::Type CIRGenModule::convertType(QualType type) {
   return genTypes.convertType(type);
 }
diff --git a/clang/lib/CIR/CodeGen/CIRGenModule.h b/clang/lib/CIR/CodeGen/CIRGenModule.h
index 9f6a57c31d291..5a1eb9dea35c6 100644
--- a/clang/lib/CIR/CodeGen/CIRGenModule.h
+++ b/clang/lib/CIR/CodeGen/CIRGenModule.h
@@ -256,6 +256,10 @@ class CIRGenModule : public CIRGenTypeCache {
   /// declarations are emitted lazily.
   void emitGlobal(clang::GlobalDecl gd);
 
+  void emitAliasForGlobal(llvm::StringRef mangledName, mlir::Operation *op,
+                          GlobalDecl aliasGD, cir::FuncOp aliasee,
+                          cir::GlobalLinkageKind linkage);
+
   mlir::Type convertType(clang::QualType type);
 
   /// Set the visibility for the given global.
@@ -358,6 +362,8 @@ class CIRGenModule : public CIRGenTypeCache {
   cir::GlobalLinkageKind getCIRLinkageVarDefinition(const VarDecl *vd,
                                                     bool isConstant);
 
+  void addReplacement(llvm::StringRef name, mlir::Operation *op);
+
   /// Helpers to emit "not yet implemented" error diagnostics
   DiagnosticBuilder errorNYI(SourceLocation, llvm::StringRef);
 
@@ -397,6 +403,17 @@ class CIRGenModule : public CIRGenTypeCache {
   llvm::MapVector<clang::GlobalDecl, llvm::StringRef> mangledDeclNames;
   llvm::StringMap<clang::GlobalDecl, llvm::BumpPtrAllocator> manglings;
 
+  // FIXME: should we use llvm::TrackingVH<mlir::Operation> here?
+  typedef llvm::StringMap<mlir::Operation *> ReplacementsTy;
+  ReplacementsTy replacements;
+  /// Call replaceAllUsesWith on all pairs in replacements.
+  void applyReplacements();
+
+  /// A helper function to replace all uses of OldF to NewF that replace
+  /// the type of pointer arguments. This is not needed to tradtional
+  /// pipeline since LLVM has opaque pointers but CIR not.
+  void replacePointerTypeArgs(cir::FuncOp oldF, cir::FuncOp newF);
+
   void setNonAliasAttributes(GlobalDecl gd, mlir::Operation *op);
 };
 } // namespace CIRGen
diff --git a/clang/lib/CIR/Dialect/IR/CIRDialect.cpp b/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
index 17157561357f9..d14fb6af5a988 100644
--- a/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
+++ b/clang/lib/CIR/Dialect/IR/CIRDialect.cpp
@@ -1419,13 +1419,17 @@ ParseResult cir::FuncOp::parse(OpAsmParser &parser, OperationState &state) {
 }
 
 // This function corresponds to `llvm::GlobalValue::isDeclaration` and should
-// have a similar implementation. We don't currently support aliases, ifuncs,
-// or materializable functions, but those should be handled here as they are
-// implemented.
+// have a similar implementation. We don't currently ifuncs or materializable
+// functions, but those should be handled here as they are implemented.
 bool cir::FuncOp::isDeclaration() {
-  assert(!cir::MissingFeatures::opFuncGlobalAliases());
   assert(!cir::MissingFeatures::supportIFuncAttr());
-  return getFunctionBody().empty();
+
+  std::optional<StringRef> aliasee = getAliasee();
+  if (!aliasee)
+    return getFunctionBody().empty();
+
+  // Aliases are always definitions.
+  return false;
 }
 
 mlir::Region *cir::FuncOp::getCallableRegion() {
@@ -1460,6 +1464,12 @@ void cir::FuncOp::print(OpAsmPrinter &p) {
   function_interface_impl::printFunctionSignature(
       p, *this, fnType.getInputs(), fnType.isVarArg(), fnType.getReturnTypes());
 
+  if (std::optional<StringRef> aliaseeName = getAliasee()) {
+    p << " alias(";
+    p.printSymbolName(*aliaseeName);
+    p << ")";
+  }
+
   // Print the body if this is not an external function.
   Region &body = getOperation()->getRegion(0);
   if (!body.empty()) {
diff --git a/clang/test/CIR/CodeGen/ctor-alias.cpp b/clang/test/CIR/CodeGen/ctor-alias.cpp
new file mode 100644
index 0000000000000..a20e206c6d714
--- /dev/null
+++ b/clang/test/CIR/CodeGen/ctor-alias.cpp
@@ -0,0 +1,75 @@
+// RUN: %clang_cc1 -std=c++17 -triple x86_64-unknown-linux-gnu -mconstructor-aliases -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --input-file=%t.cir %s
+// RUN: %clang_cc1 -std=c++17 -triple x86_64-unknown-linux-gnu -mconstructor-aliases -fclangir -emit-llvm %s -o %t-cir.ll
+// RUN: FileCheck --check-prefix=LLVM --input-file=%t-cir.ll %s
+// RUN: %clang_cc1 -std=c++17 -triple x86_64-unknown-linux-gnu -mconstructor-aliases -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=OGCG --input-file=%t.ll %s
+
+struct B {
+  B();
+};
+B::B() {
+}
+
+// OGCG: @_ZN1BC1Ev = unnamed_addr alias void (ptr), ptr @_ZN1BC2Ev
+
+// CHECK: cir.func{{.*}} @_ZN1BC2Ev(%arg0: !cir.ptr<!rec_B>
+// CHECK:   %[[THIS_ADDR:.*]] = cir.alloca !cir.ptr<!rec_B>, !cir.ptr<!cir.ptr<!rec_B>>, ["this", init]
+// CHECK:   cir.store %arg0, %[[THIS_ADDR]]
+// CHECK:   %[[THIS:.*]] = cir.load %[[THIS_ADDR]] : !cir.ptr<!cir.ptr<!rec_B>>, !cir.ptr<!rec_B>
+
+// CHECK: cir.func{{.*}} private dso_local @_ZN1BC1Ev(!cir.ptr<!rec_B>) alias(@_ZN1BC2Ev)
+
+// LLVM: define{{.*}} @_ZN1BC2Ev(ptr %[[THIS_ARG:.*]])
+// LLVM:   %[[THIS_ADDR:.*]] = alloca ptr
+// LLVM:   store ptr %[[THIS_ARG]], ptr %[[THIS_ADDR]]
+// LLVM:   %[[THIS:.*]] = load ptr, ptr %[[THIS_ADDR]]
+
+// This should be an alias, like the similar OGCG alias above, but that's not
+// implemented yet.
+// LLVM: declare dso_local void @_ZN1BC1Ev(ptr)
+
+// OGCG: define{{.*}} @_ZN1BC2Ev(ptr{{.*}} %[[THIS_ARG:.*]])
+// OGCG:   %[[THIS_ADDR:.*]] = alloca ptr
+// OGCG:   store ptr %[[THIS_ARG]], ptr %[[THIS_ADDR]]
+// OGCG:   %[[THIS:.*]] = load ptr, ptr %[[THIS_ADDR]]
+
+// The constructor in this cases is handled by RAUW rather than aliasing.
+struct Struk {
+  Struk() {}
+};
+
+void baz() {
+  Struk s;
+}
+
+// CHECK:   cir.func{{.*}} @_ZN5StrukC2Ev(%arg0: !cir.ptr<!rec_Struk>
+// CHECK:     %[[THIS_ADDR:.*]] = cir.alloca !cir.ptr<!rec_Struk>, !cir.ptr<!cir.ptr<!rec_Struk>>, ["this", init]
+// CHECK:     cir.store %arg0, %[[THIS_ADDR]] : !cir.ptr<!rec_Struk>, !cir.ptr<!cir.ptr<!rec_Struk>>
+// CHECK:     %[[THIS:.*]] = cir.load %[[THIS_ADDR]] : !cir.ptr<!cir.ptr<!rec_Struk>>, !cir.ptr<!rec_Struk>
+// CHECK:     cir.return
+
+// CHECK-NOT:   cir.func{{.*}} @_ZN5StrukC1Ev
+
+// CHECK:   cir.func{{.*}} @_Z3bazv()
+// CHECK:     %[[S_ADDR:.*]] = cir.alloca !rec_Struk, !cir.ptr<!rec_Struk>, ["s", init]
+// CHECK:     cir.call @_ZN5StrukC2Ev(%[[S_ADDR]]) : (!cir.ptr<!rec_Struk>) -> ()
+
+// LLVM: define linkonce_odr void @_ZN5StrukC2Ev(ptr %[[THIS_ARG]])
+// LLVM:   %[[THIS_ADDR:.*]] = alloca ptr
+// LLVM:   store ptr %[[THIS_ARG]], ptr %[[THIS_ADDR]]
+// LLVM:   %[[THIS:.*]] = load ptr, ptr %[[THIS_ADDR]]
+
+// LLVM: define{{.*}} void @_Z3bazv()
+// LLVM:   %[[S_ADDR:.*]] = alloca %struct.Struk
+// LLVM:   call void @_ZN5StrukC2Ev(ptr{{.*}} %[[S_ADDR]])
+
+// This function gets emitted before the constructor in OGCG.
+// OGCG: define{{.*}} void @_Z3bazv()
+// OGCG:   %[[S_ADDR:.*]] = alloca %struct.Struk
+// OGCG:   call void @_ZN5StrukC2Ev(ptr{{.*}} %[[S_ADDR]])
+
+// OGCG: define linkonce_odr void @_ZN5StrukC2Ev(ptr{{.*}} %[[THIS_ARG]])
+// OGCG:   %[[THIS_ADDR:.*]] = alloca ptr
+// OGCG:   store ptr %[[THIS_ARG]], ptr %[[THIS_ADDR]]
+// OGCG:   %[[THIS:.*]] = load ptr, ptr %[[THIS_ADDR]]

[erichkeane]

Small suggestions, looks good anyway.

[erichkeane]

Destructor doesn't require special handling here? Seems a little odd to me...

[andykaylor]

That's an artifact of destructors not being implemented yet. See the errorNYI above.

[erichkeane]

You could use a zip to get the range of each... Something like:

for (auto [arg, funcArgTy] = llvm::zip(call.getArgs(), funcArgTypes)) {...}

if you cared to. llvm::zip will give you a range of the size of the shortest thing.

[andykaylor]

In the full implementation from the incubator, the index is needed to replace the call arg with a bitcast if one is needed. However, I've convinced myself that we never hit the bitcast case, so I guess as long as this is just a check of that assumption, I can use llvm::zip.

[bcardosolopes]

LGTM one minor comment

[bcardosolopes]

Can you add a cir to cir test to exercise the parser?