MC: Provide MCAssembler with a TargetAsmBackend.
[oota-llvm.git] / lib / CodeGen / LLVMTargetMachine.cpp
index 64e28fb764f593ab82dd4825121a4fd29a01fd51..23ef8ba7ce44fc3a71073192470dddfe6b27de48 100644 (file)
@@ -14,6 +14,7 @@
 #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/MachineFunctionAnalysis.h"
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/MC/MCAsmInfo.h"
+#include "llvm/MC/MCContext.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;
 
@@ -31,27 +37,48 @@ 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> 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> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
     cl::desc("Verify generated machine code"),
     cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
 
-// This is not enabled by default due to 1) high compile time cost, 2) it's not
-// beneficial to all targets. The plan is to let targets decide whether this
-// is enabled.
-static cl::opt<bool>
-EnablePostRAScheduler("post-RA-scheduler",
-                       cl::desc("Enable scheduling after register allocation"),
-                       cl::init(false));
+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
@@ -60,168 +87,140 @@ 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()));
-
-  if (addPreEmitPass(PM, OptLevel) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
+  if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify))
+    return true;
 
-  if (OptLevel != CodeGenOpt::None)
-    PM.add(createCodePlacementOptPass());
+  OwningPtr<MCContext> Context(new MCContext());
+  OwningPtr<MCStreamer> AsmStreamer;
 
+  formatted_raw_ostream *LegacyOutput;
   switch (FileType) {
-  default:
+  default: return true;
+  case CGFT_AssemblyFile: {
+    const MCAsmInfo &MAI = *getMCAsmInfo();
+    MCInstPrinter *InstPrinter =
+      getTarget().createMCInstPrinter(MAI.getAssemblerDialect(), MAI, Out);
+    AsmStreamer.reset(createAsmStreamer(*Context, Out, MAI,
+                                        getTargetData()->isLittleEndian(),
+                                        getVerboseAsm(), InstPrinter,
+                                        /*codeemitter*/0));
+    // Set the AsmPrinter's "O" to the output file.
+    LegacyOutput = &Out;
     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;
-}
-
-bool LLVMTargetMachine::addAssemblyEmitter(PassManagerBase &PM,
-                                           CodeGenOpt::Level OptLevel,
-                                           bool Verbose,
-                                           formatted_raw_ostream &Out) {
+  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(createMachOStreamer(*Context, *TAB, Out, MCE));
+    
+    // Any output to the asmprinter's "O" stream is bad and needs to be fixed,
+    // force it to come out stderr.
+    // FIXME: this is horrible and leaks, eventually remove the raw_ostream from
+    // asmprinter.
+    LegacyOutput = new formatted_raw_ostream(errs());
+    break;
+  }
+  case CGFT_Null:
+    // The Null output is intended for use for performance analysis and testing,
+    // not real users.
+    AsmStreamer.reset(createNullStreamer(*Context));
+    // Any output to the asmprinter's "O" stream is bad and needs to be fixed,
+    // force it to come out stderr.
+    // FIXME: this is horrible and leaks, eventually remove the raw_ostream from
+    // asmprinter.
+    LegacyOutput = new formatted_raw_ostream(errs());
+    break;
+  }
+  
+  // Create the AsmPrinter, which takes ownership of Context and AsmStreamer
+  // if successful.
   FunctionPass *Printer =
-    getTarget().createAsmPrinter(Out, *this, getMCAsmInfo(), Verbose);
-  if (!Printer)
+    getTarget().createAsmPrinter(*LegacyOutput, *this, *Context, *AsmStreamer,
+                                 getMCAsmInfo());
+  if (Printer == 0)
     return true;
-
+  
+  // If successful, createAsmPrinter took ownership of AsmStreamer and Context.
+  Context.take(); AsmStreamer.take();
+  
   PM.add(Printer);
-  return false;
-}
-
-/// 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());
-
+  
+  // Make sure the code model is set.
+  setCodeModelForStatic();
   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.
-///
-bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
-                                                   MachineCodeEmitter &MCE,
-                                                   CodeGenOpt::Level OptLevel) {
-  // Add common CodeGen passes.
-  if (addCommonCodeGenPasses(PM, OptLevel))
-    return true;
-
-  if (addPreEmitPass(PM, OptLevel) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
-
-  addCodeEmitter(PM, OptLevel, MCE);
-  if (PrintEmittedAsm)
-    addAssemblyEmitter(PM, OptLevel, true, ferrs());
-
-  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,
                                                    JITCodeEmitter &JCE,
-                                                   CodeGenOpt::Level OptLevel) {
+                                                   CodeGenOpt::Level OptLevel,
+                                                   bool DisableVerify) {
+  // Make sure the code model is set.
+  setCodeModelForJIT();
+  
   // Add common CodeGen passes.
-  if (addCommonCodeGenPasses(PM, OptLevel))
+  if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify))
     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());
 
   return false; // success!
 }
 
+static void printNoVerify(PassManagerBase &PM,
+                           const char *Banner) {
+  if (PrintMachineCode)
+    PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
+}
+
 static void printAndVerify(PassManagerBase &PM,
+                           const char *Banner,
                            bool allowDoubleDefs = false) {
   if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(errs()));
+    PM.add(createMachineFunctionPrinterPass(dbgs(), Banner));
 
   if (VerifyMachineCode)
     PM.add(createMachineVerifierPass(allowDoubleDefs));
@@ -231,14 +230,26 @@ static void printAndVerify(PassManagerBase &PM,
 /// emitting to assembly files or machine code output.
 ///
 bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
-                                               CodeGenOpt::Level OptLevel) {
+                                               CodeGenOpt::Level OptLevel,
+                                               bool DisableVerify) {
   // 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()));
   }
 
   // Turn exception handling constructs into something the code generators can
@@ -247,8 +258,13 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
   {
   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()));
+    PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None));
     break;
   case ExceptionHandling::Dwarf:
     PM.add(createDwarfEHPass(getTargetLowering(), OptLevel==CodeGenOpt::None));
@@ -263,18 +279,20 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
   // 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()));
 
   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.
 
@@ -291,57 +309,99 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
     return true;
 
   // Print the instruction selected machine code...
-  printAndVerify(PM, /* allowDoubleDefs= */ true);
+  printAndVerify(PM, "After Instruction Selection",
+                 /* allowDoubleDefs= */ true);
+
+  // Optimize PHIs before DCE: removing dead PHI cycles may make more
+  // instructions dead.
+  if (OptLevel != CodeGenOpt::None)
+    PM.add(createOptimizePHIsPass());
+
+  // Delete dead machine instructions regardless of optimization level.
+  PM.add(createDeadMachineInstructionElimPass());
+  printAndVerify(PM, "After codegen DCE pass",
+                 /* allowDoubleDefs= */ true);
 
   if (OptLevel != CodeGenOpt::None) {
-    PM.add(createMachineLICMPass());
-    PM.add(createMachineSinkingPass());
-    printAndVerify(PM, /* allowDoubleDefs= */ true);
+    PM.add(createOptimizeExtsPass());
+    if (!DisableMachineLICM)
+      PM.add(createMachineLICMPass());
+    PM.add(createMachineCSEPass());
+    if (!DisableMachineSink)
+      PM.add(createMachineSinkingPass());
+    printAndVerify(PM, "After Machine LICM, CSE and Sinking passes",
+                   /* allowDoubleDefs= */ true);
+  }
+
+  // Pre-ra tail duplication.
+  if (OptLevel != CodeGenOpt::None && !DisableEarlyTailDup) {
+    PM.add(createTailDuplicatePass(true));
+    printAndVerify(PM, "After Pre-RegAlloc TailDuplicate",
+                   /* allowDoubleDefs= */ true);
   }
 
   // Run pre-ra passes.
   if (addPreRegAlloc(PM, OptLevel))
-    printAndVerify(PM, /* allowDoubleDefs= */ true);
+    printAndVerify(PM, "After PreRegAlloc passes",
+                   /* allowDoubleDefs= */ true);
 
   // Perform register allocation.
   PM.add(createRegisterAllocator());
+  printAndVerify(PM, "After Register Allocation");
 
   // Perform stack slot coloring.
-  if (OptLevel != CodeGenOpt::None)
+  if (OptLevel != CodeGenOpt::None && !DisableSSC) {
     // FIXME: Re-enable coloring with register when it's capable of adding
     // kill markers.
     PM.add(createStackSlotColoringPass(false));
-
-  printAndVerify(PM);           // Print the register-allocated code
+    printAndVerify(PM, "After StackSlotColoring");
+  }
 
   // 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, "After PreSched2 passes");
 
   // Second pass scheduler.
-  if (OptLevel != CodeGenOpt::None && EnablePostRAScheduler) {
-    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;
 }