[clang-repl] Simplify the value printing logic to enable out-of-process. (#107737)

This patch improves the design of the IncrementalParser and Interpreter
classes. Now the incremental parser is only responsible for building the
partial translation unit declaration and the AST, while the Interpreter
fills in the lower level llvm::Module and other JIT-related
infrastructure. Finally the Interpreter class now orchestrates the AST
and the LLVM IR with the IncrementalParser and IncrementalExecutor
classes.

The design improvement allows us to rework some of the logic that
extracts an interpreter value into the clang::Value object. The new
implementation simplifies use-cases which are used for out-of-process
execution by allowing interpreter to be inherited or customized with an
clang::ASTConsumer.

This change will enable completing the pretty printing work which is in
#84769

Author: vgvassilev

clang/include/clang/Frontend/MultiplexConsumer.h

@@ -53,6 +53,7 @@ class MultiplexConsumer : public SemaConsumer {
 public:
   // Takes ownership of the pointers in C.
   MultiplexConsumer(std::vector<std::unique_ptr<ASTConsumer>> C);
+  MultiplexConsumer(std::unique_ptr<ASTConsumer> C);
   ~MultiplexConsumer() override;
 
   // ASTConsumer
@@ -80,7 +81,7 @@ class MultiplexConsumer : public SemaConsumer {
   void InitializeSema(Sema &S) override;
   void ForgetSema() override;
 
-private:
+protected:
   std::vector<std::unique_ptr<ASTConsumer>> Consumers; // Owns these.
   std::unique_ptr<MultiplexASTMutationListener> MutationListener;
   std::unique_ptr<MultiplexASTDeserializationListener> DeserializationListener;

clang/include/clang/Interpreter/Interpreter.h

@@ -14,11 +14,9 @@
 #ifndef LLVM_CLANG_INTERPRETER_INTERPRETER_H
 #define LLVM_CLANG_INTERPRETER_INTERPRETER_H
 
-#include "clang/AST/Decl.h"
 #include "clang/AST/GlobalDecl.h"
 #include "clang/Interpreter/PartialTranslationUnit.h"
 #include "clang/Interpreter/Value.h"
-#include "clang/Sema/Ownership.h"
 
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ExecutionEngine/JITSymbol.h"
@@ -38,6 +36,9 @@ class ThreadSafeContext;
 namespace clang {
 
 class CompilerInstance;
+class CodeGenerator;
+class CXXRecordDecl;
+class Decl;
 class IncrementalExecutor;
 class IncrementalParser;
 
@@ -77,42 +78,45 @@ class IncrementalCompilerBuilder {
   llvm::StringRef CudaSDKPath;
 };
 
-/// Generate glue code between the Interpreter's built-in runtime and user code.
-class RuntimeInterfaceBuilder {
-public:
-  virtual ~RuntimeInterfaceBuilder() = default;
-
-  using TransformExprFunction = ExprResult(RuntimeInterfaceBuilder *Builder,
-                                           Expr *, ArrayRef<Expr *>);
-  virtual TransformExprFunction *getPrintValueTransformer() = 0;
-};
+class IncrementalAction;
+class InProcessPrintingASTConsumer;
 
 /// Provides top-level interfaces for incremental compilation and execution.
 class Interpreter {
+  friend class Value;
+  friend InProcessPrintingASTConsumer;
+
   std::unique_ptr<llvm::orc::ThreadSafeContext> TSCtx;
+  /// Long-lived, incremental parsing action.
+  std::unique_ptr<IncrementalAction> Act;
   std::unique_ptr<IncrementalParser> IncrParser;
   std::unique_ptr<IncrementalExecutor> IncrExecutor;
-  std::unique_ptr<RuntimeInterfaceBuilder> RuntimeIB;
 
   // An optional parser for CUDA offloading
   std::unique_ptr<IncrementalParser> DeviceParser;
 
+  /// List containing information about each incrementally parsed piece of code.
+  std::list<PartialTranslationUnit> PTUs;
+
   unsigned InitPTUSize = 0;
 
   // This member holds the last result of the value printing. It's a class
   // member because we might want to access it after more inputs. If no value
   // printing happens, it's in an invalid state.
   Value LastValue;
 
-  // Add a call to an Expr to report its result. We query the function from
-  // RuntimeInterfaceBuilder once and store it as a function pointer to avoid
-  // frequent virtual function calls.
-  RuntimeInterfaceBuilder::TransformExprFunction *AddPrintValueCall = nullptr;
+  /// When CodeGen is created the first llvm::Module gets cached in many places
+  /// and we must keep it alive.
+  std::unique_ptr<llvm::Module> CachedInCodeGenModule;
+
+  /// Compiler instance performing the incremental compilation.
+  std::unique_ptr<CompilerInstance> CI;
 
 protected:
   // Derived classes can use an extended interface of the Interpreter.
-  Interpreter(std::unique_ptr<CompilerInstance> CI, llvm::Error &Err,
-              std::unique_ptr<llvm::orc::LLJITBuilder> JITBuilder = nullptr);
+  Interpreter(std::unique_ptr<CompilerInstance> Instance, llvm::Error &Err,
+              std::unique_ptr<llvm::orc::LLJITBuilder> JITBuilder = nullptr,
+              std::unique_ptr<clang::ASTConsumer> Consumer = nullptr);
 
   // Create the internal IncrementalExecutor, or re-create it after calling
   // ResetExecutor().
@@ -122,15 +126,8 @@ class Interpreter {
   // JIT engine. In particular, it doesn't run cleanup or destructors.
   void ResetExecutor();
 
-  // Lazily construct the RuntimeInterfaceBuilder. The provided instance will be
-  // used for the entire lifetime of the interpreter. The default implementation
-  // targets the in-process __clang_Interpreter runtime. Override this to use a
-  // custom runtime.
-  virtual std::unique_ptr<RuntimeInterfaceBuilder> FindRuntimeInterface();
-
 public:
   virtual ~Interpreter();
-
   static llvm::Expected<std::unique_ptr<Interpreter>>
   create(std::unique_ptr<CompilerInstance> CI);
   static llvm::Expected<std::unique_ptr<Interpreter>>
@@ -145,7 +142,6 @@ class Interpreter {
   llvm::Expected<PartialTranslationUnit &> Parse(llvm::StringRef Code);
   llvm::Error Execute(PartialTranslationUnit &T);
   llvm::Error ParseAndExecute(llvm::StringRef Code, Value *V = nullptr);
-  llvm::Expected<llvm::orc::ExecutorAddr> CompileDtorCall(CXXRecordDecl *CXXRD);
 
   /// Undo N previous incremental inputs.
   llvm::Error Undo(unsigned N = 1);
@@ -167,8 +163,6 @@ class Interpreter {
   llvm::Expected<llvm::orc::ExecutorAddr>
   getSymbolAddressFromLinkerName(llvm::StringRef LinkerName) const;
 
-  enum InterfaceKind { NoAlloc, WithAlloc, CopyArray, NewTag };
-
   const llvm::SmallVectorImpl<Expr *> &getValuePrintingInfo() const {
     return ValuePrintingInfo;
   }
@@ -178,7 +172,15 @@ class Interpreter {
 private:
   size_t getEffectivePTUSize() const;
   void markUserCodeStart();
+  llvm::Expected<Expr *> ExtractValueFromExpr(Expr *E);
+  llvm::Expected<llvm::orc::ExecutorAddr> CompileDtorCall(CXXRecordDecl *CXXRD);
+
+  CodeGenerator *getCodeGen() const;
+  std::unique_ptr<llvm::Module> GenModule();
+  PartialTranslationUnit &RegisterPTU(TranslationUnitDecl *TU);
 
+  // A cache for the compiled destructors used to for de-allocation of managed
+  // clang::Values.
   llvm::DenseMap<CXXRecordDecl *, llvm::orc::ExecutorAddr> Dtors;
 
   llvm::SmallVector<Expr *, 4> ValuePrintingInfo;

clang/lib/Frontend/MultiplexConsumer.cpp

@@ -298,6 +298,13 @@ MultiplexConsumer::MultiplexConsumer(
   }
 }
 
+MultiplexConsumer::MultiplexConsumer(std::unique_ptr<ASTConsumer> C)
+    : MultiplexConsumer([](std::unique_ptr<ASTConsumer> Consumer) {
+        std::vector<std::unique_ptr<ASTConsumer>> Consumers;
+        Consumers.push_back(std::move(Consumer));
+        return Consumers;
+      }(std::move(C))) {}
+
 MultiplexConsumer::~MultiplexConsumer() {}
 
 void MultiplexConsumer::Initialize(ASTContext &Context) {

clang/lib/Interpreter/CMakeLists.txt

@@ -22,6 +22,7 @@ add_clang_library(clangInterpreter
   IncrementalExecutor.cpp
   IncrementalParser.cpp
   Interpreter.cpp
+  InterpreterValuePrinter.cpp
   InterpreterUtils.cpp
   Value.cpp
   ${WASM_SRC}

clang/lib/Interpreter/DeviceOffload.cpp

@@ -15,6 +15,7 @@
 #include "clang/Basic/TargetOptions.h"
 #include "clang/CodeGen/ModuleBuilder.h"
 #include "clang/Frontend/CompilerInstance.h"
+#include "clang/Interpreter/PartialTranslationUnit.h"
 
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/Module.h"
@@ -24,15 +25,17 @@
 namespace clang {
 
 IncrementalCUDADeviceParser::IncrementalCUDADeviceParser(
-    Interpreter &Interp, std::unique_ptr<CompilerInstance> Instance,
-    IncrementalParser &HostParser, llvm::LLVMContext &LLVMCtx,
+    std::unique_ptr<CompilerInstance> DeviceInstance,
+    CompilerInstance &HostInstance,
     llvm::IntrusiveRefCntPtr<llvm::vfs::InMemoryFileSystem> FS,
-    llvm::Error &Err)
-    : IncrementalParser(Interp, std::move(Instance), LLVMCtx, Err),
-      HostParser(HostParser), VFS(FS) {
+    llvm::Error &Err, const std::list<PartialTranslationUnit> &PTUs)
+    : IncrementalParser(*DeviceInstance, Err), PTUs(PTUs), VFS(FS),
+      CodeGenOpts(HostInstance.getCodeGenOpts()),
+      TargetOpts(HostInstance.getTargetOpts()) {
   if (Err)
     return;
-  StringRef Arch = CI->getTargetOpts().CPU;
+  DeviceCI = std::move(DeviceInstance);
+  StringRef Arch = TargetOpts.CPU;
   if (!Arch.starts_with("sm_") || Arch.substr(3).getAsInteger(10, SMVersion)) {
     Err = llvm::joinErrors(std::move(Err), llvm::make_error<llvm::StringError>(
                                                "Invalid CUDA architecture",
@@ -41,7 +44,7 @@ IncrementalCUDADeviceParser::IncrementalCUDADeviceParser(
   }
 }
 
-llvm::Expected<PartialTranslationUnit &>
+llvm::Expected<TranslationUnitDecl *>
 IncrementalCUDADeviceParser::Parse(llvm::StringRef Input) {
   auto PTU = IncrementalParser::Parse(Input);
   if (!PTU)
@@ -62,7 +65,7 @@ IncrementalCUDADeviceParser::Parse(llvm::StringRef Input) {
                    llvm::StringRef(FatbinContent.data(), FatbinContent.size()),
                    "", false));
 
-  HostParser.getCI()->getCodeGenOpts().CudaGpuBinaryFileName = FatbinFileName;
+  CodeGenOpts.CudaGpuBinaryFileName = FatbinFileName;
 
   FatbinContent.clear();
 
@@ -80,7 +83,7 @@ llvm::Expected<llvm::StringRef> IncrementalCUDADeviceParser::GeneratePTX() {
                                                std::error_code());
   llvm::TargetOptions TO = llvm::TargetOptions();
   llvm::TargetMachine *TargetMachine = Target->createTargetMachine(
-      PTU.TheModule->getTargetTriple(), getCI()->getTargetOpts().CPU, "", TO,
+      PTU.TheModule->getTargetTriple(), TargetOpts.CPU, "", TO,
       llvm::Reloc::Model::PIC_);
   PTU.TheModule->setDataLayout(TargetMachine->createDataLayout());
 

clang/lib/Interpreter/DeviceOffload.h

@@ -18,19 +18,24 @@
 #include "llvm/Support/VirtualFileSystem.h"
 
 namespace clang {
+struct PartialTranslationUnit;
+class CompilerInstance;
+class CodeGenOptions;
+class TargetOptions;
 
 class IncrementalCUDADeviceParser : public IncrementalParser {
+  const std::list<PartialTranslationUnit> &PTUs;
+
 public:
   IncrementalCUDADeviceParser(
-      Interpreter &Interp, std::unique_ptr<CompilerInstance> Instance,
-      IncrementalParser &HostParser, llvm::LLVMContext &LLVMCtx,
+      std::unique_ptr<CompilerInstance> DeviceInstance,
+      CompilerInstance &HostInstance,
       llvm::IntrusiveRefCntPtr<llvm::vfs::InMemoryFileSystem> VFS,
-      llvm::Error &Err);
+      llvm::Error &Err, const std::list<PartialTranslationUnit> &PTUs);
 
-  llvm::Expected<PartialTranslationUnit &>
-  Parse(llvm::StringRef Input) override;
+  llvm::Expected<TranslationUnitDecl *> Parse(llvm::StringRef Input) override;
 
-  // Generate PTX for the last PTU
+  // Generate PTX for the last PTU.
   llvm::Expected<llvm::StringRef> GeneratePTX();
 
   // Generate fatbinary contents in memory
@@ -39,11 +44,13 @@ class IncrementalCUDADeviceParser : public IncrementalParser {
   ~IncrementalCUDADeviceParser();
 
 protected:
-  IncrementalParser &HostParser;
+  std::unique_ptr<CompilerInstance> DeviceCI;
   int SMVersion;
   llvm::SmallString<1024> PTXCode;
   llvm::SmallVector<char, 1024> FatbinContent;
   llvm::IntrusiveRefCntPtr<llvm::vfs::InMemoryFileSystem> VFS;
+  CodeGenOptions &CodeGenOpts; // Intentionally a reference.
+  const TargetOptions &TargetOpts;
 };
 
 } // namespace clang

clang/lib/Interpreter/IncrementalExecutor.cpp

@@ -118,4 +118,4 @@ IncrementalExecutor::getSymbolAddress(llvm::StringRef Name,
   return SymOrErr->getAddress();
 }
 
-} // end namespace clang
+} // namespace clang