[CGData][ThinLTO] Global Outlining with Two-CodeGen Rounds

Author: kyulee-com

llvm/include/llvm/CGData/CodeGenData.h

@@ -164,6 +164,22 @@ publishOutlinedHashTree(std::unique_ptr<OutlinedHashTree> HashTree) {
   CodeGenData::getInstance().publishOutlinedHashTree(std::move(HashTree));
 }
 
+/// Initialize the two-codegen rounds.
+void initializeTwoCodegenRounds();
+
+/// Save the current module before the first codegen round.
+void saveModuleForTwoRounds(const Module &TheModule, unsigned Task);
+
+/// Load the current module before the second codegen round.
+std::unique_ptr<Module> loadModuleForTwoRounds(BitcodeModule &OrigModule,
+                                               unsigned Task,
+                                               LLVMContext &Context);
+
+/// Merge the codegen data from the input files in scratch vector in ThinLTO
+/// two-codegen rounds.
+Error mergeCodeGenData(
+    const std::unique_ptr<std::vector<llvm::SmallString<0>>> InputFiles);
+
 void warn(Error E, StringRef Whence = "");
 void warn(Twine Message, std::string Whence = "", std::string Hint = "");
 

llvm/lib/CGData/CodeGenData.cpp

@@ -17,6 +17,7 @@
 #include "llvm/Object/ObjectFile.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/FileSystem.h"
+#include "llvm/Support/Path.h"
 #include "llvm/Support/WithColor.h"
 
 #define DEBUG_TYPE "cg-data"
@@ -30,6 +31,14 @@ cl::opt<bool>
 cl::opt<std::string>
     CodeGenDataUsePath("codegen-data-use-path", cl::init(""), cl::Hidden,
                        cl::desc("File path to where .cgdata file is read"));
+cl::opt<bool> CodeGenDataThinLTOTwoRounds(
+    "codegen-data-thinlto-two-rounds", cl::init(false), cl::Hidden,
+    cl::desc("Enable two-round ThinLTO code generation. The first round "
+             "generates code and emits CodeGen data, while the second round "
+             "uses the emitted data for further optimizations."));
+
+// Path to where the optimized bitcodes are saved and restored for ThinLTO.
+static SmallString<128> CodeGenDataThinLTOTwoRoundsPath;
 
 static std::string getCGDataErrString(cgdata_error Err,
                                       const std::string &ErrMsg = "") {
@@ -139,7 +148,7 @@ CodeGenData &CodeGenData::getInstance() {
   std::call_once(CodeGenData::OnceFlag, []() {
     Instance = std::unique_ptr<CodeGenData>(new CodeGenData());
 
-    if (CodeGenDataGenerate)
+    if (CodeGenDataGenerate || CodeGenDataThinLTOTwoRounds)
       Instance->EmitCGData = true;
     else if (!CodeGenDataUsePath.empty()) {
       // Initialize the global CGData if the input file name is given.
@@ -215,6 +224,76 @@ void warn(Error E, StringRef Whence) {
   }
 }
 
+static std::string getPath(StringRef Dir, unsigned Task) {
+  return (Dir + "/" + llvm::Twine(Task) + ".saved_copy.bc").str();
+}
+
+void initializeTwoCodegenRounds() {
+  assert(CodeGenDataThinLTOTwoRounds);
+  if (auto EC = llvm::sys::fs::createUniqueDirectory(
+          "cgdata", CodeGenDataThinLTOTwoRoundsPath))
+    report_fatal_error(Twine("Failed to create directory: ") + EC.message());
+}
+
+void saveModuleForTwoRounds(const Module &TheModule, unsigned Task) {
+  assert(sys::fs::is_directory(CodeGenDataThinLTOTwoRoundsPath));
+  std::string Path = getPath(CodeGenDataThinLTOTwoRoundsPath, Task);
+  std::error_code EC;
+  raw_fd_ostream OS(Path, EC, sys::fs::OpenFlags::OF_None);
+  if (EC)
+    report_fatal_error(Twine("Failed to open ") + Path +
+                       " to save optimized bitcode: " + EC.message());
+  WriteBitcodeToFile(TheModule, OS, /* ShouldPreserveUseListOrder */ true);
+}
+
+std::unique_ptr<Module> loadModuleForTwoRounds(BitcodeModule &OrigModule,
+                                               unsigned Task,
+                                               LLVMContext &Context) {
+  assert(sys::fs::is_directory(CodeGenDataThinLTOTwoRoundsPath));
+  std::string Path = getPath(CodeGenDataThinLTOTwoRoundsPath, Task);
+  auto FileOrError = MemoryBuffer::getFile(Path);
+  if (auto EC = FileOrError.getError())
+    report_fatal_error(Twine("Failed to open ") + Path +
+                       " to load optimized bitcode: " + EC.message());
+
+  std::unique_ptr<MemoryBuffer> FileBuffer = std::move(*FileOrError);
+  auto RestoredModule = llvm::parseBitcodeFile(*FileBuffer, Context);
+  if (!RestoredModule)
+    report_fatal_error(Twine("Failed to parse optimized bitcode loaded from ") +
+                       Path + "\n");
+
+  // Restore the original module identifier.
+  (*RestoredModule)->setModuleIdentifier(OrigModule.getModuleIdentifier());
+  return std::move(*RestoredModule);
+}
+
+Error mergeCodeGenData(
+    const std::unique_ptr<std::vector<llvm::SmallString<0>>> InputFiles) {
+
+  OutlinedHashTreeRecord GlobalOutlineRecord;
+  for (auto &InputFile : *(InputFiles)) {
+    if (InputFile.empty())
+      continue;
+    StringRef File = StringRef(InputFile.data(), InputFile.size());
+    std::unique_ptr<MemoryBuffer> Buffer = MemoryBuffer::getMemBuffer(
+        File, "in-memory object file", /*RequiresNullTerminator=*/false);
+    Expected<std::unique_ptr<object::ObjectFile>> BinOrErr =
+        object::ObjectFile::createObjectFile(Buffer->getMemBufferRef());
+    if (!BinOrErr)
+      return BinOrErr.takeError();
+
+    std::unique_ptr<object::ObjectFile> &Obj = BinOrErr.get();
+    if (auto E = CodeGenDataReader::mergeFromObjectFile(Obj.get(),
+                                                        GlobalOutlineRecord))
+      return E;
+  }
+
+  if (!GlobalOutlineRecord.empty())
+    cgdata::publishOutlinedHashTree(std::move(GlobalOutlineRecord.HashTree));
+
+  return Error::success();
+}
+
 } // end namespace cgdata
 
 } // end namespace llvm

llvm/lib/LTO/LTO.cpp

@@ -21,6 +21,7 @@
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Bitcode/BitcodeReader.h"
 #include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/CGData/CodeGenData.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/Config/llvm-config.h"
 #include "llvm/IR/AutoUpgrade.h"
@@ -71,6 +72,8 @@ static cl::opt<bool>
     DumpThinCGSCCs("dump-thin-cg-sccs", cl::init(false), cl::Hidden,
                    cl::desc("Dump the SCCs in the ThinLTO index's callgraph"));
 
+extern cl::opt<bool> CodeGenDataThinLTOTwoRounds;
+
 namespace llvm {
 /// Enable global value internalization in LTO.
 cl::opt<bool> EnableLTOInternalization(
@@ -1459,7 +1462,7 @@ class InProcessThinBackend : public ThinBackendProc {
           GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
   }
 
-  Error runThinLTOBackendThread(
+  virtual Error runThinLTOBackendThread(
       AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM,
       ModuleSummaryIndex &CombinedIndex,
       const FunctionImporter::ImportMapTy &ImportList,
@@ -1560,6 +1563,60 @@ class InProcessThinBackend : public ThinBackendProc {
     return BackendThreadPool.getMaxConcurrency();
   }
 };
+
+/// This Backend will run ThinBackend process but throw away all the output from
+/// the codegen. This class facilitates the first codegen round.
+class NoOutputThinBackend : public InProcessThinBackend {
+public:
+  NoOutputThinBackend(
+      const Config &Conf, ModuleSummaryIndex &CombinedIndex,
+      ThreadPoolStrategy ThinLTOParallelism,
+      const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
+      std::unique_ptr<std::vector<llvm::SmallString<0>>> Scratch)
+      : InProcessThinBackend(
+            Conf, CombinedIndex, ThinLTOParallelism, ModuleToDefinedGVSummaries,
+            // Allocate a scratch buffer for each task to write output to.
+            [Allocation = &*Scratch](unsigned Task, const Twine &ModuleName) {
+              return std::make_unique<CachedFileStream>(
+                  std::make_unique<raw_svector_ostream>((*Allocation)[Task]));
+            },
+            FileCache(), nullptr, false, false),
+        Scratch(std::move(Scratch)) {}
+
+  /// Scratch space for writing output during the codegen.
+  std::unique_ptr<std::vector<llvm::SmallString<0>>> Scratch;
+};
+
+/// This Backend performs codegen on bitcode that was previously saved after
+/// going through optimization. This class facilitates the second codegen round.
+class OptimizedBitcodeThinBackend : public InProcessThinBackend {
+public:
+  OptimizedBitcodeThinBackend(
+      const Config &Conf, ModuleSummaryIndex &CombinedIndex,
+      ThreadPoolStrategy ThinLTOParallelism,
+      const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
+      AddStreamFn AddStream)
+      : InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism,
+                             ModuleToDefinedGVSummaries, AddStream, FileCache(),
+                             nullptr, false, false) {}
+
+  virtual Error runThinLTOBackendThread(
+      AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM,
+      ModuleSummaryIndex &CombinedIndex,
+      const FunctionImporter::ImportMapTy &ImportList,
+      const FunctionImporter::ExportSetTy &ExportList,
+      const std::map<GlobalValue::GUID, GlobalValue::LinkageTypes> &ResolvedODR,
+      const GVSummaryMapTy &DefinedGlobals,
+      MapVector<StringRef, BitcodeModule> &ModuleMap) override {
+    LTOLLVMContext BackendContext(Conf);
+    std::unique_ptr<Module> LoadedModule =
+        cgdata::loadModuleForTwoRounds(BM, Task, BackendContext);
+
+    return thinBackend(Conf, Task, AddStream, *LoadedModule, CombinedIndex,
+                       ImportList, DefinedGlobals, &ModuleMap,
+                       /*CodeGenOnly=*/true);
+  }
+};
 } // end anonymous namespace
 
 ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism,
@@ -1880,10 +1937,46 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
     return BackendProcess->wait();
   };
 
-  std::unique_ptr<ThinBackendProc> BackendProc =
-      ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
-                      AddStream, Cache);
-  return RunBackends(BackendProc.get());
+  if (!CodeGenDataThinLTOTwoRounds) {
+    std::unique_ptr<ThinBackendProc> BackendProc =
+        ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
+                        AddStream, Cache);
+    return RunBackends(BackendProc.get());
+  }
+
+  // Perform two rounds of code generation for ThinLTO:
+  // 1. First round: Run optimization and code generation with a scratch output.
+  // 2. Merge codegen data extracted from the scratch output.
+  // 3. Second round: Run code generation again using the merged data.
+  LLVM_DEBUG(dbgs() << "Running ThinLTO two-codegen rounds\n");
+
+  // Initialize a temporary path to store and retrieve optimized IRs for
+  // two-round code generation.
+  cgdata::initializeTwoCodegenRounds();
+
+  // Create a scratch output to hold intermediate results.
+  auto Outputs =
+      std::make_unique<std::vector<llvm::SmallString<0>>>(getMaxTasks());
+  auto FirstRoundLTO = std::make_unique<NoOutputThinBackend>(
+      Conf, ThinLTO.CombinedIndex, llvm::heavyweight_hardware_concurrency(),
+      ModuleToDefinedGVSummaries, std::move(Outputs));
+  // First round: Run optimization and code generation with a scratch output.
+  // Before code generation, serialize modules.
+  if (Error E = RunBackends(FirstRoundLTO.get()))
+    return E;
+
+  // Merge codegen data extracted from the scratch output.
+  if (Error E = cgdata::mergeCodeGenData(std::move(FirstRoundLTO->Scratch)))
+    return E;
+
+  // Second round: Run code generation by reading IRs.
+  std::unique_ptr<ThinBackendProc> SecondRoundLTO =
+      std::make_unique<OptimizedBitcodeThinBackend>(
+          Conf, ThinLTO.CombinedIndex, llvm::heavyweight_hardware_concurrency(),
+          ModuleToDefinedGVSummaries, AddStream);
+  Error E = RunBackends(SecondRoundLTO.get());
+
+  return E;
 }
 
 Expected<std::unique_ptr<ToolOutputFile>> lto::setupLLVMOptimizationRemarks(

llvm/lib/LTO/LTOBackend.cpp

@@ -20,6 +20,7 @@
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Bitcode/BitcodeReader.h"
 #include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/CGData/CodeGenData.h"
 #include "llvm/IR/LLVMRemarkStreamer.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/PassManager.h"
@@ -74,6 +75,8 @@ static cl::opt<bool> ThinLTOAssumeMerged(
     cl::desc("Assume the input has already undergone ThinLTO function "
              "importing and the other pre-optimization pipeline changes."));
 
+extern cl::opt<bool> CodeGenDataThinLTOTwoRounds;
+
 namespace llvm {
 extern cl::opt<bool> NoPGOWarnMismatch;
 }
@@ -599,11 +602,19 @@ Error lto::thinBackend(const Config &Conf, unsigned Task, AddStreamFn AddStream,
   auto OptimizeAndCodegen =
       [&](Module &Mod, TargetMachine *TM,
           std::unique_ptr<ToolOutputFile> DiagnosticOutputFile) {
+        // Perform optimization and code generation for ThinLTO.
         if (!opt(Conf, TM, Task, Mod, /*IsThinLTO=*/true,
                  /*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex,
                  CmdArgs))
           return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
 
+        // Save the current module before the first codegen round.
+        // Note that the second codegen round runs only `codegen()` without
+        // running `opt()`. We're not reaching here as it's bailed out earlier
+        // with CodeGenOnly which has been set in `OptimizedBitcodeThinBackend`.
+        if (CodeGenDataThinLTOTwoRounds)
+          cgdata::saveModuleForTwoRounds(Mod, Task);
+
         codegen(Conf, TM, AddStream, Task, Mod, CombinedIndex);
         return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
       };

llvm/test/ThinLTO/AArch64/cgdata-read-single-outline.ll

@@ -0,0 +1,42 @@
+; This test verifies whether we can outline a singleton instance (i.e., an instance that does not repeat)
+; using codegen data that has been read from a previous codegen run.
+
+; RUN: split-file %s %t
+
+; First, we generate the cgdata file from a local outline instance present in local-two.ll.
+; RUN: llc -mtriple=arm64-apple-darwin -codegen-data-generate=true -filetype=obj %t/local-two.ll -o %t_write
+; RUN: llvm-cgdata --merge %t_write -o %t_cgdata
+; RUN: llvm-cgdata --show %t_cgdata | FileCheck %s --check-prefix=SHOW
+
+; SHOW: Outlined hash tree:
+; SHOW-NEXT:  Total Node Count: 4
+; SHOW-NEXT:  Terminal Node Count: 1
+; SHOW-NEXT:  Depth: 3
+
+; Now, we read the cgdata in the machine outliner, enabling us to optimistically
+; outline a singleton instance in local-one.ll that matches against the cgdata.
+; RUN: llc -mtriple=arm64-apple-darwin -codegen-data-use-path=%t_cgdata -filetype=obj %t/local-one.ll -o %t_read
+; RUN: llvm-objdump -d %t_read | FileCheck %s
+
+; CHECK: _OUTLINED_FUNCTION
+; CHECK-NEXT:  mov
+; CHECK-NEXT:  mov
+; CHECK-NEXT:  b
+
+;--- local-two.ll
+declare i32 @g(i32, i32, i32)
+define i32 @f1() minsize {
+  %1 = call i32 @g(i32 10, i32 1, i32 2);
+  ret i32 %1
+}
+define i32 @f2() minsize {
+  %1 = call i32 @g(i32 20, i32 1, i32 2);
+  ret i32 %1
+}
+
+;--- local-one.ll
+declare i32 @g(i32, i32, i32)
+define i32 @f3() minsize {
+  %1 = call i32 @g(i32 30, i32 1, i32 2);
+ ret i32 %1
+}