Stop using the dom frontier in DwarfEHPrepare by not promoting alloca's
[oota-llvm.git] / lib / CodeGen / LLVMTargetMachine.cpp
index 4e713a6ed3165ac1265cf92fa481515d913280a1..36038027b259436fe5e9ef2225abfba30df5998e 100644 (file)
 
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/PassManager.h"
-#include "llvm/Pass.h"
+#include "llvm/Analysis/Verifier.h"
 #include "llvm/Assembly/PrintModulePass.h"
 #include "llvm/CodeGen/AsmPrinter.h"
-#include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/GCStrategy.h"
 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "llvm/CodeGen/GCStrategy.h"
+#include "llvm/CodeGen/Passes.h"
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCStreamer.h"
+#include "llvm/Target/TargetData.h"
 #include "llvm/Target/TargetRegistry.h"
 #include "llvm/Transforms/Scalar.h"
+#include "llvm/ADT/OwningPtr.h"
 #include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
 #include "llvm/Support/FormattedStream.h"
 using namespace llvm;
 
@@ -31,20 +36,58 @@ namespace llvm {
   bool EnableFastISel;
 }
 
+static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden,
+    cl::desc("Disable Post Regalloc"));
+static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
+    cl::desc("Disable branch folding"));
+static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
+    cl::desc("Disable tail duplication"));
+static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
+    cl::desc("Disable pre-register allocation tail duplication"));
+static cl::opt<bool> DisableCodePlace("disable-code-place", cl::Hidden,
+    cl::desc("Disable code placement"));
+static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
+    cl::desc("Disable Stack Slot Coloring"));
+static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
+    cl::desc("Disable Machine LICM"));
+static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
+    cl::Hidden,
+    cl::desc("Disable Machine LICM"));
+static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
+    cl::desc("Disable Machine Sinking"));
+static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
+    cl::desc("Disable Loop Strength Reduction Pass"));
+static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
+    cl::desc("Disable Codegen Prepare"));
 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
     cl::desc("Print LLVM IR produced by the loop-reduce pass"));
 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
     cl::desc("Print LLVM IR input to isel pass"));
-static cl::opt<bool> PrintEmittedAsm("print-emitted-asm", cl::Hidden,
-    cl::desc("Dump emitter generated instructions as assembly"));
 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
     cl::desc("Dump garbage collector data"));
-static cl::opt<bool> HoistConstants("hoist-constants", cl::Hidden,
-    cl::desc("Hoist constants out of loops"));
+static cl::opt<bool> ShowMCEncoding("show-mc-encoding", cl::Hidden,
+    cl::desc("Show encoding in .s output"));
+static cl::opt<bool> ShowMCInst("show-mc-inst", cl::Hidden,
+    cl::desc("Show instruction structure in .s output"));
+static cl::opt<bool> EnableMCLogging("enable-mc-api-logging", cl::Hidden,
+    cl::desc("Enable MC API logging"));
 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
     cl::desc("Verify generated machine code"),
     cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
 
+static cl::opt<cl::boolOrDefault>
+AsmVerbose("asm-verbose", cl::desc("Add comments to directives."),
+           cl::init(cl::BOU_UNSET));
+
+static bool getVerboseAsm() {
+  switch (AsmVerbose) {
+  default:
+  case cl::BOU_UNSET: return TargetMachine::getAsmVerbosityDefault();
+  case cl::BOU_TRUE:  return true;
+  case cl::BOU_FALSE: return false;
+  }
+}
+
 // Enable or disable FastISel. Both options are needed, because
 // FastISel is enabled by default with -fast, and we wish to be
 // able to enable or disable fast-isel independently from -O0.
@@ -52,224 +95,235 @@ static cl::opt<cl::boolOrDefault>
 EnableFastISelOption("fast-isel", cl::Hidden,
   cl::desc("Enable the \"fast\" instruction selector"));
 
+// Enable or disable an experimental optimization to split GEPs
+// and run a special GVN pass which does not examine loads, in
+// an effort to factor out redundancy implicit in complex GEPs.
+static cl::opt<bool> EnableSplitGEPGVN("split-gep-gvn", cl::Hidden,
+    cl::desc("Split GEPs and run no-load GVN"));
 
 LLVMTargetMachine::LLVMTargetMachine(const Target &T,
-                                     const std::string &TargetTriple)
-  : TargetMachine(T) {
+                                     const std::string &Triple)
+  : TargetMachine(T), TargetTriple(Triple) {
   AsmInfo = T.createAsmInfo(TargetTriple);
 }
 
+// Set the default code model for the JIT for a generic target.
+// FIXME: Is small right here? or .is64Bit() ? Large : Small?
+void LLVMTargetMachine::setCodeModelForJIT() {
+  setCodeModel(CodeModel::Small);
+}
 
+// Set the default code model for static compilation for a generic target.
+void LLVMTargetMachine::setCodeModelForStatic() {
+  setCodeModel(CodeModel::Small);
+}
 
-FileModel::Model
-LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
-                                       formatted_raw_ostream &Out,
-                                       CodeGenFileType FileType,
-                                       CodeGenOpt::Level OptLevel) {
+bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
+                                            formatted_raw_ostream &Out,
+                                            CodeGenFileType FileType,
+                                            CodeGenOpt::Level OptLevel,
+                                            bool DisableVerify) {
   // Add common CodeGen passes.
-  if (addCommonCodeGenPasses(PM, OptLevel))
-    return FileModel::Error;
-
-  // Fold redundant debug labels.
-  PM.add(createDebugLabelFoldingPass());
-
-  if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
+  MCContext *Context = 0;
+  if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify, Context))
+    return true;
+  assert(Context != 0 && "Failed to get MCContext");
 
-  if (addPreEmitPass(PM, OptLevel) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
-
-  if (OptLevel != CodeGenOpt::None)
-    PM.add(createCodePlacementOptPass());
+  const MCAsmInfo &MAI = *getMCAsmInfo();
+  OwningPtr<MCStreamer> AsmStreamer;
 
   switch (FileType) {
-  default:
+  default: return true;
+  case CGFT_AssemblyFile: {
+    MCInstPrinter *InstPrinter =
+      getTarget().createMCInstPrinter(MAI.getAssemblerDialect(), MAI);
+
+    // Create a code emitter if asked to show the encoding.
+    MCCodeEmitter *MCE = 0;
+    if (ShowMCEncoding)
+      MCE = getTarget().createCodeEmitter(*this, *Context);
+
+    AsmStreamer.reset(createAsmStreamer(*Context, Out,
+                                        getTargetData()->isLittleEndian(),
+                                        getVerboseAsm(), InstPrinter,
+                                        MCE, ShowMCInst));
+    break;
+  }
+  case CGFT_ObjectFile: {
+    // Create the code emitter for the target if it exists.  If not, .o file
+    // emission fails.
+    MCCodeEmitter *MCE = getTarget().createCodeEmitter(*this, *Context);
+    TargetAsmBackend *TAB = getTarget().createAsmBackend(TargetTriple);
+    if (MCE == 0 || TAB == 0)
+      return true;
+
+    AsmStreamer.reset(getTarget().createObjectStreamer(TargetTriple, *Context,
+                                                       *TAB, Out, MCE,
+                                                       hasMCRelaxAll()));
+    break;
+  }
+  case CGFT_Null:
+    // The Null output is intended for use for performance analysis and testing,
+    // not real users.
+    AsmStreamer.reset(createNullStreamer(*Context));
     break;
-  case TargetMachine::AssemblyFile:
-    if (addAssemblyEmitter(PM, OptLevel, getAsmVerbosityDefault(), Out))
-      return FileModel::Error;
-    return FileModel::AsmFile;
-  case TargetMachine::ObjectFile:
-    if (getMachOWriterInfo())
-      return FileModel::MachOFile;
-    else if (getELFWriterInfo())
-      return FileModel::ElfFile;
   }
 
-  return FileModel::Error;
-}
+  if (EnableMCLogging)
+    AsmStreamer.reset(createLoggingStreamer(AsmStreamer.take(), errs()));
 
-bool LLVMTargetMachine::addAssemblyEmitter(PassManagerBase &PM,
-                                           CodeGenOpt::Level OptLevel,
-                                           bool Verbose,
-                                           formatted_raw_ostream &Out) {
-  FunctionPass *Printer =
-    getTarget().createAsmPrinter(Out, *this, getMCAsmInfo(), Verbose);
-  if (!Printer)
+  // Create the AsmPrinter, which takes ownership of AsmStreamer if successful.
+  FunctionPass *Printer = getTarget().createAsmPrinter(*this, *AsmStreamer);
+  if (Printer == 0)
     return true;
 
-  PM.add(Printer);
-  return false;
-}
+  // If successful, createAsmPrinter took ownership of AsmStreamer.
+  AsmStreamer.take();
 
-/// addPassesToEmitFileFinish - If the passes to emit the specified file had to
-/// be split up (e.g., to add an object writer pass), this method can be used to
-/// finish up adding passes to emit the file, if necessary.
-bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
-                                                  MachineCodeEmitter *MCE,
-                                                  CodeGenOpt::Level OptLevel) {
-  if (MCE)
-    addSimpleCodeEmitter(PM, OptLevel, *MCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
-
-  PM.add(createGCInfoDeleter());
-
-  return false; // success!
-}
-
-/// addPassesToEmitFileFinish - If the passes to emit the specified file had to
-/// be split up (e.g., to add an object writer pass), this method can be used to
-/// finish up adding passes to emit the file, if necessary.
-bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
-                                                  JITCodeEmitter *JCE,
-                                                  CodeGenOpt::Level OptLevel) {
-  if (JCE)
-    addSimpleCodeEmitter(PM, OptLevel, *JCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
-
-  PM.add(createGCInfoDeleter());
-
-  return false; // success!
-}
-
-/// addPassesToEmitFileFinish - If the passes to emit the specified file had to
-/// be split up (e.g., to add an object writer pass), this method can be used to
-/// finish up adding passes to emit the file, if necessary.
-bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
-                                                  ObjectCodeEmitter *OCE,
-                                                  CodeGenOpt::Level OptLevel) {
-  if (OCE)
-    addSimpleCodeEmitter(PM, OptLevel, *OCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
+  PM.add(Printer);
 
+  // Make sure the code model is set.
+  setCodeModelForStatic();
   PM.add(createGCInfoDeleter());
-
-  return false; // success!
+  return false;
 }
 
 /// addPassesToEmitMachineCode - Add passes to the specified pass manager to
-/// get machine code emitted.  This uses a MachineCodeEmitter object to handle
+/// get machine code emitted.  This uses a JITCodeEmitter object to handle
 /// actually outputting the machine code and resolving things like the address
 /// of functions.  This method should returns true if machine code emission is
 /// not supported.
 ///
 bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
-                                                   MachineCodeEmitter &MCE,
-                                                   CodeGenOpt::Level OptLevel) {
+                                                   JITCodeEmitter &JCE,
+                                                   CodeGenOpt::Level OptLevel,
+                                                   bool DisableVerify) {
+  // Make sure the code model is set.
+  setCodeModelForJIT();
+
   // Add common CodeGen passes.
-  if (addCommonCodeGenPasses(PM, OptLevel))
+  MCContext *Ctx = 0;
+  if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify, Ctx))
     return true;
 
-  if (addPreEmitPass(PM, OptLevel) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
-
-  addCodeEmitter(PM, OptLevel, MCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
-
+  addCodeEmitter(PM, OptLevel, JCE);
   PM.add(createGCInfoDeleter());
 
   return false; // success!
 }
 
-/// addPassesToEmitMachineCode - Add passes to the specified pass manager to
-/// get machine code emitted.  This uses a MachineCodeEmitter object to handle
-/// actually outputting the machine code and resolving things like the address
-/// of functions.  This method should returns true if machine code emission is
-/// not supported.
+/// addPassesToEmitMC - Add passes to the specified pass manager to get
+/// machine code emitted with the MCJIT. This method returns true if machine
+/// code is not supported. It fills the MCContext Ctx pointer which can be
+/// used to build custom MCStreamer.
 ///
-bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
-                                                   JITCodeEmitter &JCE,
-                                                   CodeGenOpt::Level OptLevel) {
+bool LLVMTargetMachine::addPassesToEmitMC(PassManagerBase &PM,
+                                          MCContext *&Ctx,
+                                          CodeGenOpt::Level OptLevel,
+                                          bool DisableVerify) {
   // Add common CodeGen passes.
-  if (addCommonCodeGenPasses(PM, OptLevel))
+  if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify, Ctx))
     return true;
-
-  if (addPreEmitPass(PM, OptLevel) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
-
-  addCodeEmitter(PM, OptLevel, JCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
-
-  PM.add(createGCInfoDeleter());
+  // Make sure the code model is set.
+  setCodeModelForJIT();
 
   return false; // success!
 }
 
+static void printNoVerify(PassManagerBase &PM, const char *Banner) {
+  if (PrintMachineCode)
+    PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
+}
+
 static void printAndVerify(PassManagerBase &PM,
-                           bool allowDoubleDefs = false) {
+                           const char *Banner) {
   if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
+    PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
 
   if (VerifyMachineCode)
-    PM.add(createMachineVerifierPass(allowDoubleDefs));
+    PM.add(createMachineVerifierPass());
 }
 
 /// addCommonCodeGenPasses - Add standard LLVM codegen passes used for both
 /// emitting to assembly files or machine code output.
 ///
 bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
-                                               CodeGenOpt::Level OptLevel) {
+                                               CodeGenOpt::Level OptLevel,
+                                               bool DisableVerify,
+                                               MCContext *&OutContext) {
   // Standard LLVM-Level Passes.
 
+  // Before running any passes, run the verifier to determine if the input
+  // coming from the front-end and/or optimizer is valid.
+  if (!DisableVerify)
+    PM.add(createVerifierPass());
+
+  // Optionally, tun split-GEPs and no-load GVN.
+  if (EnableSplitGEPGVN) {
+    PM.add(createGEPSplitterPass());
+    PM.add(createGVNPass(/*NoLoads=*/true));
+  }
+
   // Run loop strength reduction before anything else.
-  if (OptLevel != CodeGenOpt::None) {
+  if (OptLevel != CodeGenOpt::None && !DisableLSR) {
     PM.add(createLoopStrengthReducePass(getTargetLowering()));
     if (PrintLSR)
-      PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &errs()));
+      PM.add(createPrintFunctionPass("\n\n*** Code after LSR ***\n", &dbgs()));
   }
 
+  PM.add(createGCLoweringPass());
+
+  // Make sure that no unreachable blocks are instruction selected.
+  PM.add(createUnreachableBlockEliminationPass());
+
   // Turn exception handling constructs into something the code generators can
   // handle.
-  switch (getMCAsmInfo()->getExceptionHandlingType())
-  {
+  switch (getMCAsmInfo()->getExceptionHandlingType()) {
   case ExceptionHandling::SjLj:
     // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
-    PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None));
+    // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
+    // catch info can get misplaced when a selector ends up more than one block
+    // removed from the parent invoke(s). This could happen when a landing
+    // pad is shared by multiple invokes and is also a target of a normal
+    // edge from elsewhere.
     PM.add(createSjLjEHPass(getTargetLowering()));
-    break;
+    // FALLTHROUGH
   case ExceptionHandling::Dwarf:
-    PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None));
+    PM.add(createDwarfEHPass(this));
     break;
   case ExceptionHandling::None:
     PM.add(createLowerInvokePass(getTargetLowering()));
+
+    // The lower invoke pass may create unreachable code. Remove it.
+    PM.add(createUnreachableBlockEliminationPass());
     break;
   }
 
-  PM.add(createGCLoweringPass());
-
-  // Make sure that no unreachable blocks are instruction selected.
-  PM.add(createUnreachableBlockEliminationPass());
-
-  if (OptLevel != CodeGenOpt::None) {
-    if (HoistConstants)
-      PM.add(createCodeGenLICMPass());
+  if (OptLevel != CodeGenOpt::None && !DisableCGP)
     PM.add(createCodeGenPreparePass(getTargetLowering()));
-  }
 
   PM.add(createStackProtectorPass(getTargetLowering()));
 
+  addPreISel(PM, OptLevel);
+
   if (PrintISelInput)
     PM.add(createPrintFunctionPass("\n\n"
                                    "*** Final LLVM Code input to ISel ***\n",
-                                   &errs()));
+                                   &dbgs()));
+
+  // All passes which modify the LLVM IR are now complete; run the verifier
+  // to ensure that the IR is valid.
+  if (!DisableVerify)
+    PM.add(createVerifierPass());
 
   // Standard Lower-Level Passes.
 
+  // Install a MachineModuleInfo class, which is an immutable pass that holds
+  // all the per-module stuff we're generating, including MCContext.
+  MachineModuleInfo *MMI = new MachineModuleInfo(*getMCAsmInfo());
+  PM.add(MMI);
+  OutContext = &MMI->getContext(); // Return the MCContext specifically by-ref.
+
   // Set up a MachineFunction for the rest of CodeGen to work on.
   PM.add(new MachineFunctionAnalysis(*this, OptLevel));
 
@@ -283,61 +337,107 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
     return true;
 
   // Print the instruction selected machine code...
-  printAndVerify(PM, /* allowDoubleDefs= */ true);
+  printAndVerify(PM, "After Instruction Selection");
+
+  // Optimize PHIs before DCE: removing dead PHI cycles may make more
+  // instructions dead.
+  if (OptLevel != CodeGenOpt::None)
+    PM.add(createOptimizePHIsPass());
+
+  // If the target requests it, assign local variables to stack slots relative
+  // to one another and simplify frame index references where possible.
+  PM.add(createLocalStackSlotAllocationPass());
 
   if (OptLevel != CodeGenOpt::None) {
-    PM.add(createMachineLICMPass());
-    PM.add(createMachineSinkingPass());
-    printAndVerify(PM, /* allowDoubleDefs= */ true);
+    // With optimization, dead code should already be eliminated. However
+    // there is one known exception: lowered code for arguments that are only
+    // used by tail calls, where the tail calls reuse the incoming stack
+    // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
+    PM.add(createDeadMachineInstructionElimPass());
+    printAndVerify(PM, "After codegen DCE pass");
+
+    PM.add(createPeepholeOptimizerPass());
+    if (!DisableMachineLICM)
+      PM.add(createMachineLICMPass());
+    PM.add(createMachineCSEPass());
+    if (!DisableMachineSink)
+      PM.add(createMachineSinkingPass());
+    printAndVerify(PM, "After Machine LICM, CSE and Sinking passes");
+  }
+
+  // Pre-ra tail duplication.
+  if (OptLevel != CodeGenOpt::None && !DisableEarlyTailDup) {
+    PM.add(createTailDuplicatePass(true));
+    printAndVerify(PM, "After Pre-RegAlloc TailDuplicate");
   }
 
   // Run pre-ra passes.
   if (addPreRegAlloc(PM, OptLevel))
-    printAndVerify(PM, /* allowDoubleDefs= */ true);
+    printAndVerify(PM, "After PreRegAlloc passes");
 
   // Perform register allocation.
-  PM.add(createRegisterAllocator());
+  PM.add(createRegisterAllocator(OptLevel));
+  printAndVerify(PM, "After Register Allocation");
 
-  // Perform stack slot coloring.
-  if (OptLevel != CodeGenOpt::None)
+  // Perform stack slot coloring and post-ra machine LICM.
+  if (OptLevel != CodeGenOpt::None) {
     // FIXME: Re-enable coloring with register when it's capable of adding
     // kill markers.
-    PM.add(createStackSlotColoringPass(false));
+    if (!DisableSSC)
+      PM.add(createStackSlotColoringPass(false));
+
+    // Run post-ra machine LICM to hoist reloads / remats.
+    if (!DisablePostRAMachineLICM)
+      PM.add(createMachineLICMPass(false));
 
-  printAndVerify(PM);           // Print the register-allocated code
+    printAndVerify(PM, "After StackSlotColoring and postra Machine LICM");
+  }
 
   // Run post-ra passes.
   if (addPostRegAlloc(PM, OptLevel))
-    printAndVerify(PM);
+    printAndVerify(PM, "After PostRegAlloc passes");
 
   PM.add(createLowerSubregsPass());
-  printAndVerify(PM);
+  printAndVerify(PM, "After LowerSubregs");
 
   // Insert prolog/epilog code.  Eliminate abstract frame index references...
   PM.add(createPrologEpilogCodeInserter());
-  printAndVerify(PM);
+  printAndVerify(PM, "After PrologEpilogCodeInserter");
 
   // Run pre-sched2 passes.
   if (addPreSched2(PM, OptLevel))
-    printAndVerify(PM);
+    printAndVerify(PM, "After PreSched2 passes");
 
   // Second pass scheduler.
-  if (OptLevel != CodeGenOpt::None) {
-    PM.add(createPostRAScheduler());
-    printAndVerify(PM);
+  if (OptLevel != CodeGenOpt::None && !DisablePostRA) {
+    PM.add(createPostRAScheduler(OptLevel));
+    printAndVerify(PM, "After PostRAScheduler");
   }
 
   // Branch folding must be run after regalloc and prolog/epilog insertion.
-  if (OptLevel != CodeGenOpt::None) {
+  if (OptLevel != CodeGenOpt::None && !DisableBranchFold) {
     PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
-    printAndVerify(PM);
+    printNoVerify(PM, "After BranchFolding");
+  }
+
+  // Tail duplication.
+  if (OptLevel != CodeGenOpt::None && !DisableTailDuplicate) {
+    PM.add(createTailDuplicatePass(false));
+    printNoVerify(PM, "After TailDuplicate");
   }
 
   PM.add(createGCMachineCodeAnalysisPass());
-  printAndVerify(PM);
 
   if (PrintGCInfo)
-    PM.add(createGCInfoPrinter(errs()));
+    PM.add(createGCInfoPrinter(dbgs()));
+
+  if (OptLevel != CodeGenOpt::None && !DisableCodePlace) {
+    PM.add(createCodePlacementOptPass());
+    printNoVerify(PM, "After CodePlacementOpt");
+  }
+
+  if (addPreEmitPass(PM, OptLevel))
+    printNoVerify(PM, "After PreEmit passes");
 
   return false;
 }