Automerge: [CIR] Add support for normal cleanups (#149948)

This change adds basic handling for normal cleanups. This is a very
minimal implemention. In particular, it uses a naive substitute for the
rich cleanup and EH stack handling that is present in classic codegen
and the CIR incubator. This is intended as a temporary implementation to
allow incremental progress. It is not expected to scale well enough to
be used in a production environment. It will be replaced with the full
EHScopeStack handling when such an implementation is needed.

Author: andykaylor

clang/include/clang/CIR/MissingFeatures.h

@@ -196,6 +196,8 @@ struct MissingFeatures {
   static bool cxxRecordStaticMembers() { return false; }
   static bool dataLayoutTypeAllocSize() { return false; }
   static bool deferredCXXGlobalInit() { return false; }
+  static bool ehCleanupFlags() { return false; }
+  static bool ehstackBranches() { return false; }
   static bool emitCheckedInBoundsGEP() { return false; }
   static bool emitCondLikelihoodViaExpectIntrinsic() { return false; }
   static bool emitLifetimeMarkers() { return false; }

clang/lib/CIR/CodeGen/CIRGenClass.cpp

@@ -480,6 +480,19 @@ void CIRGenFunction::emitImplicitAssignmentOperatorBody(FunctionArgList &args) {
                        s->getStmtClassName());
 }
 
+void CIRGenFunction::destroyCXXObject(CIRGenFunction &cgf, Address addr,
+                                      QualType type) {
+  const RecordType *rtype = type->castAs<RecordType>();
+  const CXXRecordDecl *record = cast<CXXRecordDecl>(rtype->getDecl());
+  const CXXDestructorDecl *dtor = record->getDestructor();
+  // TODO(cir): Unlike traditional codegen, CIRGen should actually emit trivial
+  // dtors which shall be removed on later CIR passes. However, only remove this
+  // assertion after we have a test case to exercise this path.
+  assert(!dtor->isTrivial());
+  cgf.emitCXXDestructorCall(dtor, Dtor_Complete, /*forVirtualBase*/ false,
+                            /*delegating=*/false, addr, type);
+}
+
 void CIRGenFunction::emitDelegatingCXXConstructorCall(
     const CXXConstructorDecl *ctor, const FunctionArgList &args) {
   assert(ctor->isDelegatingConstructor());

clang/lib/CIR/CodeGen/CIRGenCleanup.cpp

@@ -0,0 +1,69 @@
+//===--- CIRGenCleanup.cpp - Bookkeeping and code emission for cleanups ---===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains code dealing with the IR generation for cleanups
+// and related information.
+//
+// A "cleanup" is a piece of code which needs to be executed whenever
+// control transfers out of a particular scope.  This can be
+// conditionalized to occur only on exceptional control flow, only on
+// normal control flow, or both.
+//
+//===----------------------------------------------------------------------===//
+
+#include "CIRGenFunction.h"
+
+#include "clang/CIR/MissingFeatures.h"
+
+using namespace clang;
+using namespace clang::CIRGen;
+
+//===----------------------------------------------------------------------===//
+// CIRGenFunction cleanup related
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// EHScopeStack
+//===----------------------------------------------------------------------===//
+
+void EHScopeStack::Cleanup::anchor() {}
+
+static mlir::Block *getCurCleanupBlock(CIRGenFunction &cgf) {
+  mlir::OpBuilder::InsertionGuard guard(cgf.getBuilder());
+  mlir::Block *cleanup =
+      cgf.curLexScope->getOrCreateCleanupBlock(cgf.getBuilder());
+  return cleanup;
+}
+
+/// Pops a cleanup block. If the block includes a normal cleanup, the
+/// current insertion point is threaded through the cleanup, as are
+/// any branch fixups on the cleanup.
+void CIRGenFunction::popCleanupBlock() {
+  assert(!ehStack.cleanupStack.empty() && "cleanup stack is empty!");
+  mlir::OpBuilder::InsertionGuard guard(builder);
+  std::unique_ptr<EHScopeStack::Cleanup> cleanup =
+      ehStack.cleanupStack.pop_back_val();
+
+  assert(!cir::MissingFeatures::ehCleanupFlags());
+  mlir::Block *cleanupEntry = getCurCleanupBlock(*this);
+  builder.setInsertionPointToEnd(cleanupEntry);
+  cleanup->emit(*this);
+}
+
+/// Pops cleanup blocks until the given savepoint is reached.
+void CIRGenFunction::popCleanupBlocks(size_t oldCleanupStackDepth) {
+  assert(!cir::MissingFeatures::ehstackBranches());
+
+  assert(ehStack.getStackDepth() >= oldCleanupStackDepth);
+
+  // Pop cleanup blocks until we reach the base stack depth for the
+  // current scope.
+  while (ehStack.getStackDepth() > oldCleanupStackDepth) {
+    popCleanupBlock();
+  }
+}

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -183,8 +183,8 @@ void CIRGenFunction::emitAutoVarCleanups(
   const VarDecl &d = *emission.Variable;
 
   // Check the type for a cleanup.
-  if (d.needsDestruction(getContext()))
-    cgm.errorNYI(d.getSourceRange(), "emitAutoVarCleanups: type cleanup");
+  if (QualType::DestructionKind dtorKind = d.needsDestruction(getContext()))
+    emitAutoVarTypeCleanup(emission, dtorKind);
 
   assert(!cir::MissingFeatures::opAllocaPreciseLifetime());
 
@@ -648,3 +648,96 @@ void CIRGenFunction::emitNullabilityCheck(LValue lhs, mlir::Value rhs,
 
   assert(!cir::MissingFeatures::sanitizers());
 }
+
+/// Immediately perform the destruction of the given object.
+///
+/// \param addr - the address of the object; a type*
+/// \param type - the type of the object; if an array type, all
+///   objects are destroyed in reverse order
+/// \param destroyer - the function to call to destroy individual
+///   elements
+void CIRGenFunction::emitDestroy(Address addr, QualType type,
+                                 Destroyer *destroyer) {
+  if (getContext().getAsArrayType(type))
+    cgm.errorNYI("emitDestroy: array type");
+
+  return destroyer(*this, addr, type);
+}
+
+CIRGenFunction::Destroyer *
+CIRGenFunction::getDestroyer(QualType::DestructionKind kind) {
+  switch (kind) {
+  case QualType::DK_none:
+    llvm_unreachable("no destroyer for trivial dtor");
+  case QualType::DK_cxx_destructor:
+    return destroyCXXObject;
+  case QualType::DK_objc_strong_lifetime:
+  case QualType::DK_objc_weak_lifetime:
+  case QualType::DK_nontrivial_c_struct:
+    cgm.errorNYI("getDestroyer: other destruction kind");
+    return nullptr;
+  }
+  llvm_unreachable("Unknown DestructionKind");
+}
+
+namespace {
+struct DestroyObject final : EHScopeStack::Cleanup {
+  DestroyObject(Address addr, QualType type,
+                CIRGenFunction::Destroyer *destroyer)
+      : addr(addr), type(type), destroyer(destroyer) {}
+
+  Address addr;
+  QualType type;
+  CIRGenFunction::Destroyer *destroyer;
+
+  void emit(CIRGenFunction &cgf) override {
+    cgf.emitDestroy(addr, type, destroyer);
+  }
+};
+} // namespace
+
+/// Enter a destroy cleanup for the given local variable.
+void CIRGenFunction::emitAutoVarTypeCleanup(
+    const CIRGenFunction::AutoVarEmission &emission,
+    QualType::DestructionKind dtorKind) {
+  assert(dtorKind != QualType::DK_none);
+
+  // Note that for __block variables, we want to destroy the
+  // original stack object, not the possibly forwarded object.
+  Address addr = emission.getObjectAddress(*this);
+
+  const VarDecl *var = emission.Variable;
+  QualType type = var->getType();
+
+  CleanupKind cleanupKind = NormalAndEHCleanup;
+  CIRGenFunction::Destroyer *destroyer = nullptr;
+
+  switch (dtorKind) {
+  case QualType::DK_none:
+    llvm_unreachable("no cleanup for trivially-destructible variable");
+
+  case QualType::DK_cxx_destructor:
+    // If there's an NRVO flag on the emission, we need a different
+    // cleanup.
+    if (emission.NRVOFlag) {
+      cgm.errorNYI(var->getSourceRange(), "emitAutoVarTypeCleanup: NRVO");
+      return;
+    }
+    // Otherwise, this is handled below.
+    break;
+
+  case QualType::DK_objc_strong_lifetime:
+  case QualType::DK_objc_weak_lifetime:
+  case QualType::DK_nontrivial_c_struct:
+    cgm.errorNYI(var->getSourceRange(),
+                 "emitAutoVarTypeCleanup: other dtor kind");
+    return;
+  }
+
+  // If we haven't chosen a more specific destroyer, use the default.
+  if (!destroyer)
+    destroyer = getDestroyer(dtorKind);
+
+  assert(!cir::MissingFeatures::ehCleanupFlags());
+  ehStack.pushCleanup<DestroyObject>(cleanupKind, addr, type, destroyer);
+}

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -26,7 +26,11 @@ namespace clang::CIRGen {
 
 CIRGenFunction::CIRGenFunction(CIRGenModule &cgm, CIRGenBuilderTy &builder,
                                bool suppressNewContext)
-    : CIRGenTypeCache(cgm), cgm{cgm}, builder(builder) {}
+    : CIRGenTypeCache(cgm), cgm{cgm}, builder(builder) {
+  ehStack.setCGF(this);
+  currentCleanupStackDepth = 0;
+  assert(ehStack.getStackDepth() == 0);
+}
 
 CIRGenFunction::~CIRGenFunction() {}
 
@@ -227,6 +231,14 @@ void CIRGenFunction::LexicalScope::cleanup() {
   CIRGenBuilderTy &builder = cgf.builder;
   LexicalScope *localScope = cgf.curLexScope;
 
+  auto applyCleanup = [&]() {
+    if (performCleanup) {
+      // ApplyDebugLocation
+      assert(!cir::MissingFeatures::generateDebugInfo());
+      forceCleanup();
+    }
+  };
+
   if (returnBlock != nullptr) {
     // Write out the return block, which loads the value from `__retval` and
     // issues the `cir.return`.
@@ -235,46 +247,93 @@ void CIRGenFunction::LexicalScope::cleanup() {
     (void)emitReturn(*returnLoc);
   }
 
-  mlir::Block *curBlock = builder.getBlock();
-  if (isGlobalInit() && !curBlock)
-    return;
-  if (curBlock->mightHaveTerminator() && curBlock->getTerminator())
-    return;
-
-  // Get rid of any empty block at the end of the scope.
-  bool entryBlock = builder.getInsertionBlock()->isEntryBlock();
-  if (!entryBlock && curBlock->empty()) {
-    curBlock->erase();
-    if (returnBlock != nullptr && returnBlock->getUses().empty())
-      returnBlock->erase();
-    return;
-  }
-
-  // Reached the end of the scope.
-  {
+  auto insertCleanupAndLeave = [&](mlir::Block *insPt) {
     mlir::OpBuilder::InsertionGuard guard(builder);
-    builder.setInsertionPointToEnd(curBlock);
+    builder.setInsertionPointToEnd(insPt);
+
+    // If we still don't have a cleanup block, it means that `applyCleanup`
+    // below might be able to get us one.
+    mlir::Block *cleanupBlock = localScope->getCleanupBlock(builder);
+
+    // Leverage and defers to RunCleanupsScope's dtor and scope handling.
+    applyCleanup();
+
+    // If we now have one after `applyCleanup`, hook it up properly.
+    if (!cleanupBlock && localScope->getCleanupBlock(builder)) {
+      cleanupBlock = localScope->getCleanupBlock(builder);
+      builder.create<cir::BrOp>(insPt->back().getLoc(), cleanupBlock);
+      if (!cleanupBlock->mightHaveTerminator()) {
+        mlir::OpBuilder::InsertionGuard guard(builder);
+        builder.setInsertionPointToEnd(cleanupBlock);
+        builder.create<cir::YieldOp>(localScope->endLoc);
+      }
+    }
 
     if (localScope->depth == 0) {
       // Reached the end of the function.
       if (returnBlock != nullptr) {
-        if (returnBlock->getUses().empty())
+        if (returnBlock->getUses().empty()) {
           returnBlock->erase();
-        else {
+        } else {
+          // Thread return block via cleanup block.
+          if (cleanupBlock) {
+            for (mlir::BlockOperand &blockUse : returnBlock->getUses()) {
+              cir::BrOp brOp = mlir::cast<cir::BrOp>(blockUse.getOwner());
+              brOp.setSuccessor(cleanupBlock);
+            }
+          }
+
           builder.create<cir::BrOp>(*returnLoc, returnBlock);
           return;
         }
       }
       emitImplicitReturn();
       return;
     }
-    // Reached the end of a non-function scope.  Some scopes, such as those
-    // used with the ?: operator, can return a value.
-    if (!localScope->isTernary() && !curBlock->mightHaveTerminator()) {
+
+    // End of any local scope != function
+    // Ternary ops have to deal with matching arms for yielding types
+    // and do return a value, it must do its own cir.yield insertion.
+    if (!localScope->isTernary() && !insPt->mightHaveTerminator()) {
       !retVal ? builder.create<cir::YieldOp>(localScope->endLoc)
               : builder.create<cir::YieldOp>(localScope->endLoc, retVal);
     }
+  };
+
+  // If a cleanup block has been created at some point, branch to it
+  // and set the insertion point to continue at the cleanup block.
+  // Terminators are then inserted either in the cleanup block or
+  // inline in this current block.
+  mlir::Block *cleanupBlock = localScope->getCleanupBlock(builder);
+  if (cleanupBlock)
+    insertCleanupAndLeave(cleanupBlock);
+
+  // Now deal with any pending block wrap up like implicit end of
+  // scope.
+
+  mlir::Block *curBlock = builder.getBlock();
+  if (isGlobalInit() && !curBlock)
+    return;
+  if (curBlock->mightHaveTerminator() && curBlock->getTerminator())
+    return;
+
+  // Get rid of any empty block at the end of the scope.
+  bool entryBlock = builder.getInsertionBlock()->isEntryBlock();
+  if (!entryBlock && curBlock->empty()) {
+    curBlock->erase();
+    if (returnBlock != nullptr && returnBlock->getUses().empty())
+      returnBlock->erase();
+    return;
   }
+
+  // If there's a cleanup block, branch to it, nothing else to do.
+  if (cleanupBlock) {
+    builder.create<cir::BrOp>(curBlock->back().getLoc(), cleanupBlock);
+    return;
+  }
+
+  // No pre-existent cleanup block, emit cleanup code and yield/return.
+  insertCleanupAndLeave(curBlock);
 }
 
 cir::ReturnOp CIRGenFunction::LexicalScope::emitReturn(mlir::Location loc) {
@@ -408,7 +467,19 @@ void CIRGenFunction::startFunction(GlobalDecl gd, QualType returnType,
   }
 }
 
-void CIRGenFunction::finishFunction(SourceLocation endLoc) {}
+void CIRGenFunction::finishFunction(SourceLocation endLoc) {
+  // Pop any cleanups that might have been associated with the
+  // parameters.  Do this in whatever block we're currently in; it's
+  // important to do this before we enter the return block or return
+  // edges will be *really* confused.
+  // TODO(cir): Use prologueCleanupDepth here.
+  bool hasCleanups = ehStack.getStackDepth() != currentCleanupStackDepth;
+  if (hasCleanups) {
+    assert(!cir::MissingFeatures::generateDebugInfo());
+    // FIXME(cir): should we clearInsertionPoint? breaks many testcases
+    popCleanupBlocks(currentCleanupStackDepth);
+  }
+}
 
 mlir::LogicalResult CIRGenFunction::emitFunctionBody(const clang::Stmt *body) {
   auto result = mlir::LogicalResult::success();