[ThinLTO] Option to invoke ThinLTO backend passes and importing

[tejohnson]

Summary

Adds new option -fthinlto-backend=<file> to invoke the LTO pipeline
along with function importing via clang using the supplied function
summary index file. This supports invoking the parallel ThinLTO
backend processes in a distributed build environment via clang.

Additionally, this causes the module linker to be invoked on the bitcode
file being compiled to perform any necessary promotion and renaming of
locals that are exported via the function summary index file.

Add a couple tests that confirm we get expected errors when we try to
use the new option on a file that isn't bitcode, or specify an invalid
index file. The tests also confirm that we trigger the expected function
import pass.

Depends on D15024

[mehdi_amini]

lib/CodeGen/BackendUtil.cpp

309

It does not seem to be nicely integrated within the context of this function. What about all the options set just a few line below? It is not clear if CodeGenOpts.DisableLLVMOpts is well honored either.

lib/CodeGen/CodeGenAction.cpp

31

Not well sorted :)

190

Is this an unrelated change to the -fthinlto-backend one that could be committed separately?

821

This lambda is the same as the one just above for getFunctionIndexForFile, name it and define it once?

826

So for the renaming we need to duplicate completely the module? It cannot be done by morphing the existing module in place? That seems quite inefficient :(

[tejohnson]

Thanks for the review, comments below.
Teresa

lib/CodeGen/CodeGenAction.cpp

190

It is related in that I am now calling linkerDiagnosticHandler from outside BackendConsumer, which required moving it out of that class, and therefore we need to pass in the DiagnosticsEngine.

821

Ok

826

For now that is how the ModuleLinker works unfortunately. Rafael is working on a change to split this into two parts, and the symbol resolution would be done in place. So this should be resolved eventually.

[tejohnson]

Updated patch coming shortly.

lib/CodeGen/BackendUtil.cpp

309

This is basically the same LTO pipeline we setup in both gold-plugin.cpp and in LTOCodeGenerator::optimize (used by the llvm-lto tool and presumably ld64). I added that info to the comments. It is a special pipeline for this mode. That being said, I have made a couple of changes in my latest version to be uploaded shortly: Skip this special pipeline if DisableLLVMOpts is true; set LoopVectorize to CodeGenOpts.VectorizeLoop as in the below code; set SLPVectorize to CodeGenOpts.VectorizeSLP as in the below code.

One thing I may do first actually is to refactor this code out of gold-plugin and LTOCodeGenerator into a PassManagerBuilder helper function.

[tejohnson]

• Address Mehdi's comments.

[mehdi_amini]

lib/CodeGen/BackendUtil.cpp

291

The !CodeGenOpts.DisableLLVMOpts part is unclear to me. This may lead to unexpected user experience?

309

This duplication of PMBuilder / CodeGenOpts setup is what triggered my initial comment.

323–334

Could you move your code here and reuse the same PBBuilder initialized here?
I expect that you don't need to do (almost) anything else than:

if (!CodeGenOpts.ThinLTOIndexFile.empty()) {
  legacy::PassManager *MPM = getPerModulePasses();
  PMB.populateLTOPassManager(MPM);
  return;
}

398–399

Could we move the inlining setup before your code so that you don't need to it?

398–399

Could this be moved earlier so that it is before your code and you don't need to duplicate it?

lib/CodeGen/CodeGenAction.cpp

791

I think I didn't express what I meant by "name it", it can still be a lambda:

auto DiagHandler = [&] (const DiagnosticInfo &DI) {
   linkerDiagnosticHandler(DI, TheModule.get(),
                        getCompilerInstance().getDiagnostics());
};

(there not real reason to use std::bind() for anything now that we have lambda I think)

819

Side note: I found terrible that we can't do the renaming/linkage upgrade in-place :(

[tejohnson]

New patch coming I think addresses all your comments.
Thanks, Teresa

lib/CodeGen/BackendUtil.cpp

291

I'm not sure what the effect of DisableLLVMOpts should be on importing. With the changes to the latest patch I will upload shortly that moves this down and uses the same PM setup as the below code, we will still perform importing, but the optimization level is turned down to 0. That is probably reasonable, I think.

lib/CodeGen/CodeGenAction.cpp

819

This is an unfortunate effect of the currently module linker code, assuming that I want to use the same renaming/promotion support. I suppose I can add some support for doing this in place on the module, it was convenient to just reuse the support I put in the module linker initially.

[tejohnson]

• Address more comments.

[mehdi_amini]

Looks good to me but I'd like another opinion on the name of "-fthinlto-backend", Duncan?

include/clang/Driver/Types.h

68

I guess you could already commit this separately

lib/CodeGen/CodeGenAction.cpp

819

I know, we can improve it later, I just wanted to share this so we're on the same page on the ultimate goal :)

[mehdi_amini]

Talked with Duncan, we're not convinced about -fthinlto-backend=... for the option name: the word backend does not seems right here. What about -fthinlto-index=...?

lib/CodeGen/CodeGenAction.cpp

822

Could we refactor this in a helper in llvm?

test/Driver/thinlto_backend.c

6

Why not pipe to FileCheck here?

[tejohnson]

Talked with Duncan, we're not convinced about -fthinlto-backend=... for the option name: the word backend does not seems right here. What about -fthinlto-index=...?

I wanted to express that this was invoking a different pipeline, and distinguish from a normal compilation that takes a source file. But I'm ok with your suggestion if you prefer that.

lib/CodeGen/CodeGenAction.cpp

822

I can create a new Linker interface that takes the index path, builds the index, does the link and returns the ErrorOr index.

test/Driver/thinlto_backend.c

6

Will do. I think I cloned this from another test.

[mehdi_amini]

lib/CodeGen/CodeGenAction.cpp

822

I was thinking about a free function that would be something like: std::unique_ptr<Module> llvm::renameModuleForThinLTO(std::unique_ptr<Module> &M, const FunctionInfoIndex &Index);

[tejohnson]

lib/CodeGen/CodeGenAction.cpp

822

We also need to get the Index back so that it can be passed down to EmitBackendOutput below (to get it on the pass manager builder). So what I was thinking was create the Combined Module unique_ptr here and pass it in (as we do currently for LinkModules), and return the std::unique_ptr<FunctionInfoIndex> .

[mehdi_amini]

lib/CodeGen/CodeGenAction.cpp

822

You pass a reference to the index, you already have it.

[tejohnson]

lib/CodeGen/CodeGenAction.cpp

822

Oh I see, you are just talking about the Combined module creation and LinkModules call.

I was thinking you meant refactor out the getFunctionIndexForFile + LinkModules. Just doing the module creation and LinkModules is easier, will do that.

[tejohnson]

• Use new renameModuleForThinLTO helper function from llvm.
• Fix tests
  
  I have not change the option name yet. Wanted to be sure we have a name that reflects the fact that this is a special compile that changes the compilation pipeline and is only used for the last part of a ThinLTO build in a distributed build situation. Let me know if you feel -fthinlto-index is sufficient for this, or if there is a better alternative to my current -fthinlto-backend.

[tejohnson]

• Move ownership of FunctionInfoIndex onto EmitAssemblyHelper

[tejohnson]

• Rename -fthinlto-backend to -fthinlto-index as per review and IRC.

[mehdi_amini]

LGTM, thanks!

include/clang/CodeGen/BackendUtil.h

 //===--- BackendUtil.h - LLVM Backend Utilities -----------------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_CLANG_CODEGEN_BACKENDUTIL_H
 #define LLVM_CLANG_CODEGEN_BACKENDUTIL_H
 
 #include "clang/Basic/LLVM.h"
+#include "llvm/IR/FunctionInfo.h"
+#include <memory>
 
 namespace llvm {
   class Module;
 }
 
 namespace clang {
   class DiagnosticsEngine;
   class CodeGenOptions;
   class TargetOptions;
   class LangOptions;
 
   enum BackendAction {
     Backend_EmitAssembly,  ///< Emit native assembly files
     Backend_EmitBC,        ///< Emit LLVM bitcode files
     Backend_EmitLL,        ///< Emit human-readable LLVM assembly
     Backend_EmitNothing,   ///< Don't emit anything (benchmarking mode)
     Backend_EmitMCNull,    ///< Run CodeGen, but don't emit anything
     Backend_EmitObj        ///< Emit native object files
   };
 
-  void EmitBackendOutput(DiagnosticsEngine &Diags, const CodeGenOptions &CGOpts,
-                         const TargetOptions &TOpts, const LangOptions &LOpts,
-                         StringRef TDesc, llvm::Module *M, BackendAction Action,
-                         raw_pwrite_stream *OS);
+  void
+  EmitBackendOutput(DiagnosticsEngine &Diags, const CodeGenOptions &CGOpts,
+                    raw_pwrite_stream *OS,
+                    std::unique_ptr<llvm::FunctionInfoIndex> Index = nullptr);
 }
 
 #endif

include/clang/Driver/Options.td

 def flax_vector_conversions : Flag<["-"], "flax-vector-conversions">, Group<f_Group>;
 def flimited_precision_EQ : Joined<["-"], "flimited-precision=">, Group<f_Group>;
 def flto_EQ : Joined<["-"], "flto=">, Flags<[CC1Option]>, Group<f_Group>,
   HelpText<"Set LTO mode to either 'full' or 'thin'">;
 def flto : Flag<["-"], "flto">, Flags<[CC1Option]>, Group<f_Group>,
   HelpText<"Enable LTO in 'full' mode">;
 def fno_lto : Flag<["-"], "fno-lto">, Group<f_Group>,
   HelpText<"Disable LTO mode (default)">;
+def fthinlto_backend_EQ : Joined<["-"], "fthinlto-backend=">,
+  Flags<[CC1Option]>, Group<f_Group>,
+  HelpText<"Perform ThinLTO backend using provided function summary index">;
 def fmacro_backtrace_limit_EQ : Joined<["-"], "fmacro-backtrace-limit=">,
                                 Group<f_Group>, Flags<[DriverOption, CoreOption]>;
 def fmerge_all_constants : Flag<["-"], "fmerge-all-constants">, Group<f_Group>;
 def fmessage_length_EQ : Joined<["-"], "fmessage-length=">, Group<f_Group>;
 def fms_extensions : Flag<["-"], "fms-extensions">, Group<f_Group>, Flags<[CC1Option, CoreOption]>,
   HelpText<"Accept some non-standard constructs supported by the Microsoft compiler">;
 def fms_compatibility : Flag<["-"], "fms-compatibility">, Group<f_Group>, Flags<[CC1Option, CoreOption]>,
   HelpText<"Enable full Microsoft Visual C++ compatibility">;

include/clang/Driver/Types.h

#undef TYPE
   bool canLipoType(ID Id);
 
   /// isAcceptedByClang - Can clang handle this input type.
   bool isAcceptedByClang(ID Id);
 
   /// isCXX - Is this a "C++" input (C++ and Obj-C++ sources and headers).
   bool isCXX(ID Id);
 
+  /// Is this LLVM IR.
+  bool isLLVMIR(ID Id);
+
   /// isCuda - Is this a CUDA input.
   bool isCuda(ID Id);
 
   /// isObjC - Is this an "ObjC" input (Obj-C and Obj-C++ sources and headers).
   bool isObjC(ID Id);
 
   /// lookupTypeForExtension - Lookup the type to use for the file
   /// extension \p Ext.

include/clang/Frontend/CodeGenOptions.h

public:
   std::string InstrProfileOutput;
 
   /// Name of the profile file to use with -fprofile-sample-use.
   std::string SampleProfileFile;
 
   /// Name of the profile file to use as input for -fprofile-instr-use
   std::string InstrProfileInput;
 
+  /// Name of the function summary index file to use for ThinLTO function
+  /// importing.
+  std::string ThinLTOIndexFile;
+
   /// The EABI version to use
   std::string EABIVersion;
 
   /// A list of file names passed with -fcuda-include-gpubinary options to
   /// forward to CUDA runtime back-end for incorporating them into host-side
   /// object file.
   std::vector<std::string> CudaGpuBinaryFileNames;
 

lib/CodeGen/BackendUtil.cpp

 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Bitcode/BitcodeWriterPass.h"
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/CodeGen/SchedulerRegistry.h"
 #include "llvm/IR/DataLayout.h"
+#include "llvm/IR/FunctionInfo.h"
 #include "llvm/IR/IRPrintingPasses.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/Verifier.h"
 #include "llvm/MC/SubtargetFeature.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/PrettyStackTrace.h"
 #include "llvm/Support/TargetRegistry.h"

 namespace {
 
 class EmitAssemblyHelper {
   DiagnosticsEngine &Diags;
   const CodeGenOptions &CodeGenOpts;
   const clang::TargetOptions &TargetOpts;
   const LangOptions &LangOpts;
   Module *TheModule;
+  std::unique_ptr<FunctionInfoIndex> FunctionIndex;
 
   Timer CodeGenerationTime;
 
   mutable legacy::PassManager *CodeGenPasses;
   mutable legacy::PassManager *PerModulePasses;
   mutable legacy::FunctionPassManager *PerFunctionPasses;
 
 private:
private:
   TargetMachine *CreateTargetMachine(bool MustCreateTM);
 
   /// Add passes necessary to emit assembly or LLVM IR.
   ///
   /// \return True on success.
   bool AddEmitPasses(BackendAction Action, raw_pwrite_stream &OS);
 
 public:
-  EmitAssemblyHelper(DiagnosticsEngine &_Diags,
-                     const CodeGenOptions &CGOpts,
+  EmitAssemblyHelper(DiagnosticsEngine &_Diags, const CodeGenOptions &CGOpts,
                      const clang::TargetOptions &TOpts,
-                     const LangOptions &LOpts,
-                     Module *M)
+                     const LangOptions &LOpts, Module *M,
+                     std::unique_ptr<FunctionInfoIndex> Index)
       : Diags(_Diags), CodeGenOpts(CGOpts), TargetOpts(TOpts), LangOpts(LOpts),
-      TheModule(M), CodeGenerationTime("Code Generation Time"),
-      CodeGenPasses(nullptr), PerModulePasses(nullptr),
-      PerFunctionPasses(nullptr) {}
+        TheModule(M), FunctionIndex(std::move(Index)),
+        CodeGenerationTime("Code Generation Time"), CodeGenPasses(nullptr),
+        PerModulePasses(nullptr), PerFunctionPasses(nullptr) {}
 
   ~EmitAssemblyHelper() {
     delete CodeGenPasses;
     delete PerModulePasses;
     delete PerFunctionPasses;
     if (CodeGenOpts.DisableFree)
       BuryPointer(std::move(TM));
   }
void EmitAssemblyHelper::CreatePasses() {
   // Handle disabling of LLVM optimization, where we want to preserve the
   // internal module before any optimization.
   if (CodeGenOpts.DisableLLVMOpts) {
     OptLevel = 0;
     Inlining = CodeGenOpts.NoInlining;
   }
 
   PassManagerBuilderWrapper PMBuilder(CodeGenOpts, LangOpts);
+
+  // Figure out TargetLibraryInfo.
+  Triple TargetTriple(TheModule->getTargetTriple());
+  PMBuilder.LibraryInfo = createTLII(TargetTriple, CodeGenOpts);
+
+  switch (Inlining) {
+  case CodeGenOptions::NoInlining:
+    break;
+  case CodeGenOptions::NormalInlining: {
+    PMBuilder.Inliner =
+        createFunctionInliningPass(OptLevel, CodeGenOpts.OptimizeSize);
+    break;
+  }
+  case CodeGenOptions::OnlyAlwaysInlining:
+    // Respect always_inline.
+    if (OptLevel == 0)
+      // Do not insert lifetime intrinsics at -O0.
+      PMBuilder.Inliner = createAlwaysInlinerPass(false);
+    else
+      PMBuilder.Inliner = createAlwaysInlinerPass();
+    break;
+  }
+
   PMBuilder.OptLevel = OptLevel;
   PMBuilder.SizeLevel = CodeGenOpts.OptimizeSize;
   PMBuilder.BBVectorize = CodeGenOpts.VectorizeBB;
   PMBuilder.SLPVectorize = CodeGenOpts.VectorizeSLP;
   PMBuilder.LoopVectorize = CodeGenOpts.VectorizeLoop;
 
   PMBuilder.DisableUnitAtATime = !CodeGenOpts.UnitAtATime;
   PMBuilder.DisableUnrollLoops = !CodeGenOpts.UnrollLoops;
   PMBuilder.MergeFunctions = CodeGenOpts.MergeFunctions;
   PMBuilder.PrepareForLTO = CodeGenOpts.PrepareForLTO;
   PMBuilder.RerollLoops = CodeGenOpts.RerollLoops;
 
+  legacy::PassManager *MPM = getPerModulePasses();
+
+  // If we are performing a ThinLTO backend compile, invoke the LTO
+  // pipeline and pass down the in-memory function index.
+  if (!CodeGenOpts.ThinLTOIndexFile.empty()) {
+    assert(FunctionIndex && "Expected non-empty function index");
+    PMBuilder.FunctionIndex = std::move(FunctionIndex);
+    PMBuilder.populateLTOPassManager(*MPM);
+    return;
+  }
+
   PMBuilder.addExtension(PassManagerBuilder::EP_EarlyAsPossible,
                          addAddDiscriminatorsPass);
 
   // In ObjC ARC mode, add the main ARC optimization passes.
   if (LangOpts.ObjCAutoRefCount) {
     PMBuilder.addExtension(PassManagerBuilder::EP_EarlyAsPossible,
                            addObjCARCExpandPass);
     PMBuilder.addExtension(PassManagerBuilder::EP_ModuleOptimizerEarly,
void EmitAssemblyHelper::CreatePasses() {
   }
 
   if (LangOpts.Sanitize.has(SanitizerKind::DataFlow)) {
     PMBuilder.addExtension(PassManagerBuilder::EP_OptimizerLast,
                            addDataFlowSanitizerPass);
     PMBuilder.addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
                            addDataFlowSanitizerPass);
   }
 
-  // Figure out TargetLibraryInfo.
-  Triple TargetTriple(TheModule->getTargetTriple());
-  PMBuilder.LibraryInfo = createTLII(TargetTriple, CodeGenOpts);
-
-  switch (Inlining) {
-  case CodeGenOptions::NoInlining: break;
-  case CodeGenOptions::NormalInlining: {
-    PMBuilder.Inliner =
-        createFunctionInliningPass(OptLevel, CodeGenOpts.OptimizeSize);
-    break;
-  }
-  case CodeGenOptions::OnlyAlwaysInlining:
-    // Respect always_inline.
-    if (OptLevel == 0)
-      // Do not insert lifetime intrinsics at -O0.
-      PMBuilder.Inliner = createAlwaysInlinerPass(false);
-    else
-      PMBuilder.Inliner = createAlwaysInlinerPass();
-    break;
-  }
-
   // Set up the per-function pass manager.
   legacy::FunctionPassManager *FPM = getPerFunctionPasses();
   if (CodeGenOpts.VerifyModule)
     FPM->add(createVerifierPass());
   PMBuilder.populateFunctionPassManager(*FPM);
 
   // Set up the per-module pass manager.
-  legacy::PassManager *MPM = getPerModulePasses();
   if (!CodeGenOpts.RewriteMapFiles.empty())
     addSymbolRewriterPass(CodeGenOpts, MPM);
 
   if (!CodeGenOpts.DisableGCov &&
       (CodeGenOpts.EmitGcovArcs || CodeGenOpts.EmitGcovNotes)) {
     // Not using 'GCOVOptions::getDefault' allows us to avoid exiting if
     // LLVM's -default-gcov-version flag is set to something invalid.
     GCOVOptions Options;
void EmitAssemblyHelper::EmitAssembly(BackendAction Action,
   }
 }
 
 void clang::EmitBackendOutput(DiagnosticsEngine &Diags,
                               const CodeGenOptions &CGOpts,
                               const clang::TargetOptions &TOpts,
                               const LangOptions &LOpts, StringRef TDesc,
                               Module *M, BackendAction Action,
-                              raw_pwrite_stream *OS) {
-  EmitAssemblyHelper AsmHelper(Diags, CGOpts, TOpts, LOpts, M);
+                              raw_pwrite_stream *OS,
+                              std::unique_ptr<FunctionInfoIndex> Index) {
+  EmitAssemblyHelper AsmHelper(Diags, CGOpts, TOpts, LOpts, M,
+                               std::move(Index));
 
   AsmHelper.EmitAssembly(Action, OS);
 
   // If an optional clang TargetInfo description string was passed in, use it to
   // verify the LLVM TargetMachine's DataLayout.
   if (AsmHelper.TM && !TDesc.empty()) {
     std::string DLDesc = M->getDataLayout().getStringRepresentation();
     if (DLDesc != TDesc) {
       unsigned DiagID = Diags.getCustomDiagID(
           DiagnosticsEngine::Error, "backend data layout '%0' does not match "
                                     "expected target description '%1'");
       Diags.Report(DiagID) << DLDesc << TDesc;
     }
   }
 }

lib/CodeGen/CodeGenAction.cpp

 #include "clang/Frontend/CompilerInstance.h"
 #include "clang/Frontend/FrontendDiagnostic.h"
 #include "clang/Lex/Preprocessor.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/Bitcode/ReaderWriter.h"
 #include "llvm/IR/DebugInfo.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/DiagnosticPrinter.h"
+#include "llvm/IR/FunctionInfo.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IRReader/IRReader.h"
 #include "llvm/Linker/Linker.h"
+#include "llvm/Object/FunctionIndexObjectFile.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/SourceMgr.h"
 #include "llvm/Support/Timer.h"
 #include <memory>
 using namespace clang;
 using namespace llvm;
 
 namespace clang {
+/// Diagnostic handler used by invocations of Linker::LinkModules
+static void linkerDiagnosticHandler(const DiagnosticInfo &DI,
+                                    const llvm::Module *LinkModule,
+                                    DiagnosticsEngine &Diags) {
+  if (DI.getSeverity() != DS_Error)
+    return;
+
+  std::string MsgStorage;
+  {
+    raw_string_ostream Stream(MsgStorage);
+    DiagnosticPrinterRawOStream DP(Stream);
+    DI.print(DP);
+  }
+
+  Diags.Report(diag::err_fe_cannot_link_module)
+      << LinkModule->getModuleIdentifier() << MsgStorage;
+}
+
   class BackendConsumer : public ASTConsumer {
     virtual void anchor();
     DiagnosticsEngine &Diags;
     BackendAction Action;
     const CodeGenOptions &CodeGenOpts;
     const TargetOptions &TargetOpts;
     const LangOptions &LangOpts;
     raw_pwrite_stream *AsmOutStream;
void HandleTranslationUnit(ASTContext &C) override {
              "Unexpected module change during IR generation");
 
       // Link LinkModule into this module if present, preserving its validity.
       for (auto &I : LinkModules) {
         unsigned LinkFlags = I.first;
         llvm::Module *LinkModule = I.second.get();
         if (Linker::LinkModules(M, LinkModule,
                                 [=](const DiagnosticInfo &DI) {
-                                  linkerDiagnosticHandler(DI, LinkModule);
+                                  linkerDiagnosticHandler(DI, LinkModule, Diags);
                                 },
                                 LinkFlags))
           return;
       }
 
       // Install an inline asm handler so that diagnostics get printed through
       // our diagnostics hooks.
       LLVMContext &Ctx = TheModule->getContext();
public:
     }
 
     static void InlineAsmDiagHandler(const llvm::SMDiagnostic &SM,void *Context,
                                      unsigned LocCookie) {
       SourceLocation Loc = SourceLocation::getFromRawEncoding(LocCookie);
       ((BackendConsumer*)Context)->InlineAsmDiagHandler2(SM, Loc);
     }
 
-    void linkerDiagnosticHandler(const llvm::DiagnosticInfo &DI,
-                                 const llvm::Module *LinkModule);
-
     static void DiagnosticHandler(const llvm::DiagnosticInfo &DI,
                                   void *Context) {
       ((BackendConsumer *)Context)->DiagnosticHandlerImpl(DI);
     }
 
     void InlineAsmDiagHandler2(const llvm::SMDiagnostic &,
                                SourceLocation LocCookie);
 
EmitOptimizationMessage(
         D, diag::remark_fe_backend_optimization_remark_analysis_aliasing);
 }
 
 void BackendConsumer::OptimizationFailureHandler(
     const llvm::DiagnosticInfoOptimizationFailure &D) {
   EmitOptimizationMessage(D, diag::warn_fe_backend_optimization_failure);
 }
 
-void BackendConsumer::linkerDiagnosticHandler(const DiagnosticInfo &DI,
-                                              const llvm::Module *LinkModule) {
-  if (DI.getSeverity() != DS_Error)
-    return;
-
-  std::string MsgStorage;
-  {
-    raw_string_ostream Stream(MsgStorage);
-    DiagnosticPrinterRawOStream DP(Stream);
-    DI.print(DP);
-  }
-
-  Diags.Report(diag::err_fe_cannot_link_module)
-      << LinkModule->getModuleIdentifier() << MsgStorage;
-}
-
 /// \brief This function is invoked when the backend needs
 /// to report something to the user.
 void BackendConsumer::DiagnosticHandlerImpl(const DiagnosticInfo &DI) {
   unsigned DiagID = diag::err_fe_inline_asm;
   llvm::DiagnosticSeverity Severity = DI.getSeverity();
   // Get the diagnostic ID based.
   switch (DI.getKind()) {
   case llvm::DK_InlineAsm:
if (getCurrentFileKind() == IK_LLVM_IR) {
     const TargetOptions &TargetOpts = CI.getTargetOpts();
     if (TheModule->getTargetTriple() != TargetOpts.Triple) {
       CI.getDiagnostics().Report(SourceLocation(),
                                  diag::warn_fe_override_module)
           << TargetOpts.Triple;
       TheModule->setTargetTriple(TargetOpts.Triple);
     }
 
+    auto DiagHandler = [&](const DiagnosticInfo &DI) {
+      linkerDiagnosticHandler(DI, TheModule.get(),
+                              getCompilerInstance().getDiagnostics());
+    };
+
+    // If we are performing ThinLTO backend compilation (indicated by
+    // a non-empty index file option), then we need to invoke the module
+    // linker for the current module that we are importing into. This
+    // is because the module linker will need to promote to global scope
+    // and rename any local values that are potentially exported to other
+    // modules. Do this early so that the rest of the compilation sees the
+    // promoted symbols.
+    std::unique_ptr<FunctionInfoIndex> Index;
+    if (!CI.getCodeGenOpts().ThinLTOIndexFile.empty()) {
+      ErrorOr<std::unique_ptr<FunctionInfoIndex>> IndexOrErr =
+          llvm::getFunctionIndexForFile(CI.getCodeGenOpts().ThinLTOIndexFile,
+                                        DiagHandler, TheModule.get());
+      if (std::error_code EC = IndexOrErr.getError()) {
+        std::string Error = EC.message();
+        errs() << "Error loading index file '"
+               << CI.getCodeGenOpts().ThinLTOIndexFile << "': " << Error
+               << "\n";
+        return;
+      }
+      Index = std::move(IndexOrErr.get());
+      assert(Index);
+      std::unique_ptr<llvm::Module> Combined(new llvm::Module(
+          TheModule->getModuleIdentifier(), TheModule->getContext()));
+      if (Linker::LinkModules(Combined.get(), TheModule.get(), DiagHandler,
+                              llvm::Linker::Flags::None, Index.get()))
+        return;
+
+      TheModule = std::move(Combined);
+    }
+
     LLVMContext &Ctx = TheModule->getContext();
     Ctx.setInlineAsmDiagnosticHandler(BitcodeInlineAsmDiagHandler);
     EmitBackendOutput(CI.getDiagnostics(), CI.getCodeGenOpts(), TargetOpts,
                       CI.getLangOpts(), CI.getTarget().getDataLayoutString(),
-                      TheModule.get(), BA, OS);
+                      TheModule.get(), BA, OS, std::move(Index));
     return;
   }
 
   // Otherwise follow the normal AST path.
   this->ASTFrontendAction::ExecuteAction();
 }
 
 //

lib/Driver/Tools.cpp

if (isa<AnalyzeJobAction>(JA)) {
     // the use-list order, since serialization to bitcode is part of the flow.
     if (JA.getType() == types::TY_LLVM_BC)
       CmdArgs.push_back("-emit-llvm-uselists");
 
     if (D.isUsingLTO())
       Args.AddLastArg(CmdArgs, options::OPT_flto, options::OPT_flto_EQ);
   }
 
+  if (const Arg *A = Args.getLastArg(options::OPT_fthinlto_backend_EQ)) {
+    if (!types::isLLVMIR(Input.getType()))
+      D.Diag(diag::err_drv_argument_only_allowed_with) << A->getAsString(Args)
+                                                       << "-x ir";
+    Args.AddLastArg(CmdArgs, options::OPT_fthinlto_backend_EQ);
+  }
+
   // We normally speed up the clang process a bit by skipping destructors at
   // exit, but when we're generating diagnostics we can rely on some of the
   // cleanup.
   if (!C.isForDiagnostics())
     CmdArgs.push_back("-disable-free");
 
 // Disable the verification pass in -asserts builds.
 #ifdef NDEBUG

lib/Driver/Types.cpp

bool types::isCXX(ID Id) {
   case TY_ObjCXX: case TY_PP_ObjCXX: case TY_PP_ObjCXX_Alias:
   case TY_CXXHeader: case TY_PP_CXXHeader:
   case TY_ObjCXXHeader: case TY_PP_ObjCXXHeader:
   case TY_CUDA: case TY_PP_CUDA: case TY_CUDA_DEVICE:
     return true;
   }
 }
 
+bool types::isLLVMIR(ID Id) {
+  switch (Id) {
+  default:
+    return false;
+
+  case TY_LLVM_IR:
+  case TY_LLVM_BC:
+  case TY_LTO_IR:
+  case TY_LTO_BC:
+    return true;
+  }
+}
+
 bool types::isCuda(ID Id) {
   switch (Id) {
   default:
     return false;
 
   case TY_CUDA:
   case TY_PP_CUDA:
   case TY_CUDA_DEVICE:

lib/Frontend/CompilerInvocation.cpp

static bool ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args, InputKind IK,
   Opts.UniqueSectionNames = Args.hasFlag(OPT_funique_section_names,
                                          OPT_fno_unique_section_names, true);
 
   Opts.MergeFunctions = Args.hasArg(OPT_fmerge_functions);
 
   Opts.PrepareForLTO = Args.hasArg(OPT_flto, OPT_flto_EQ);
   const Arg *A = Args.getLastArg(OPT_flto, OPT_flto_EQ);
   Opts.EmitFunctionSummary = A && A->containsValue("thin");
+  if (Arg *A = Args.getLastArg(OPT_fthinlto_backend_EQ)) {
+    if (IK != IK_LLVM_IR)
+      Diags.Report(diag::err_drv_argument_only_allowed_with)
+          << A->getAsString(Args) << "-x ir";
+    Opts.ThinLTOIndexFile = Args.getLastArgValue(OPT_fthinlto_backend_EQ);
+  }
 
   Opts.MSVolatile = Args.hasArg(OPT_fms_volatile);
 
   Opts.VectorizeBB = Args.hasArg(OPT_vectorize_slp_aggressive);
   Opts.VectorizeLoop = Args.hasArg(OPT_vectorize_loops);
   Opts.VectorizeSLP = Args.hasArg(OPT_vectorize_slp);
 
   Opts.MainFileName = Args.getLastArgValue(OPT_main_file_name);

test/CodeGen/thinlto_backend.c

+// RUN: %clang -target x86_64-unknown-linux -O2 %s -flto=thin -c -o %t.o
+// RUN: %clang -target x86_64-unknown-linux -O2 -flto=thin -fuse-ld=gold -o %t %t.o
+
+// Ensure clang -cc1 give expected error for incorrect input type
+// RUN: not %clang_cc1 -target x86_64-unknown-linux -O2 -o %t1.o %s -c -fthinlto-backend=%t.thinlto.bc 2> %t
+// RUN: FileCheck -check-prefix=CHECK-WARNING < %t %s
+// CHECK-WARNING: error: invalid argument '-fthinlto-backend={{.*}}' only allowed with '-x ir'
+
+// Ensure we get expected error for missing index file
+// RUN: %clang -target x86_64-unknown-linux -O2 -o %t1.o -x ir %t.o -c -fthinlto-backend=bad.thinlto.bc 2> %t
+// RUN: FileCheck -check-prefix=CHECK-ERROR < %t %s
+// CHECK-ERROR: Error loading index file 'bad.thinlto.bc': No such file or directory
+
+// Ensure Function Importing pass added
+// RUN: %clang -target x86_64-unknown-linux -O2 -o %t1.o -x ir %t.o -c -fthinlto-backend=%t.thinlto.bc -mllvm -debug-pass=Structure 2> %t
+// RUN: FileCheck -check-prefix=CHECK-PASS < %t %s
+// CHECK-PASS: Function Importing

test/Driver/thinlto_backend.c

+// RUN: %clang -target x86_64-unknown-linux -O2 %s -flto=thin -c -o %t.o
+// RUN: %clang -target x86_64-unknown-linux -O2 -flto=thin -fuse-ld=gold -o %t %t.o
+
+// -fthinlto_backend should be passed to cc1
+// RUN: %clang -target x86_64-unknown-linux -O2 -o %t1.o -x ir %t.o -c -fthinlto-backend=%t.thinlto.bc -### 2> %t
+// RUN: FileCheck -check-prefix=CHECK-THINLTOBE-ACTION < %t %s
+// CHECK-THINLTOBE-ACTION: -fthinlto-backend=
+
+// Ensure clang driver gives the expected error for incorrect input type
+// RUN: not %clang -target x86_64-unknown-linux -O2 -o %t1.o %s -c -fthinlto-backend=%t.thinlto.bc 2> %t
+// RUN: FileCheck -check-prefix=CHECK-WARNING < %t %s
+// CHECK-WARNING: error: invalid argument '-fthinlto-backend={{.*}}' only allowed with '-x ir'