Add #include to get alloca, patch by Kenneth Boyd!
[oota-llvm.git] / lib / CodeGen / LLVMTargetMachine.cpp
index e4ad872094197c479713af72cb71c9869b15f775..cd444552f7ac21aaa9edf5419e97d8c92c68d43a 100644 (file)
 #include "llvm/Assembly/PrintModulePass.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/CodeGen/Passes.h"
-#include "llvm/CodeGen/Collector.h"
+#include "llvm/CodeGen/GCStrategy.h"
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/Target/TargetAsmInfo.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
 using namespace llvm;
 
+namespace llvm {
+  bool EnableFastISel;
+}
+
 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,
@@ -38,12 +43,9 @@ static cl::opt<bool>
 EnableSinking("enable-sinking", cl::init(false), cl::Hidden,
               cl::desc("Perform sinking on machine code"));
 static cl::opt<bool>
-AlignLoops("align-loops", cl::init(true), cl::Hidden,
-           cl::desc("Align loop headers"));
-static cl::opt<bool>
-PerformLICM("machine-licm",
-            cl::init(false), cl::Hidden,
-            cl::desc("Perform loop-invariant code motion on machine code"));
+EnableLICM("machine-licm",
+           cl::init(false), cl::Hidden,
+           cl::desc("Perform loop-invariant code motion on machine code"));
 
 // When this works it will be on by default.
 static cl::opt<bool>
@@ -51,96 +53,35 @@ DisablePostRAScheduler("disable-post-RA-scheduler",
                        cl::desc("Disable scheduling after register allocation"),
                        cl::init(true));
 
+// 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 -fast.
+static cl::opt<bool>
+EnableFastISelOption("fast-isel", cl::Hidden,
+  cl::desc("Enable the experimental \"fast\" instruction selector"));
+static cl::opt<bool>
+DisableFastISelOption("disable-fast-isel", cl::Hidden,
+  cl::desc("Disable the experimental \"fast\" instruction selector"));
+
 FileModel::Model
 LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
-                                       std::ostream &Out,
+                                       raw_ostream &Out,
                                        CodeGenFileType FileType,
                                        bool Fast) {
-  // Standard LLVM-Level Passes.
-  
-  // Run loop strength reduction before anything else.
-  if (!Fast) {
-    PM.add(createLoopStrengthReducePass(getTargetLowering()));
-    if (PrintLSR)
-      PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
-  }
-  
-  PM.add(createGCLoweringPass());
-
-  if (!getTargetAsmInfo()->doesSupportExceptionHandling())
-    PM.add(createLowerInvokePass(getTargetLowering()));
-
-  // Make sure that no unreachable blocks are instruction selected.
-  PM.add(createUnreachableBlockEliminationPass());
-
-  if (!Fast)
-    PM.add(createCodeGenPreparePass(getTargetLowering()));
-
-  if (PrintISelInput)
-    PM.add(new PrintFunctionPass("\n\n*** Final LLVM Code input to ISel ***\n",
-                                 &cerr));
-  
-  // Ask the target for an isel.
-  if (addInstSelector(PM, Fast))
+  // Add common CodeGen passes.
+  if (addCommonCodeGenPasses(PM, Fast))
     return FileModel::Error;
 
-  // Print the instruction selected machine code...
-  if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-
-  if (PerformLICM)
-    PM.add(createMachineLICMPass());
-  
-  if (EnableSinking)
-    PM.add(createMachineSinkingPass());
-
-  // Run pre-ra passes.
-  if (addPreRegAlloc(PM, Fast) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-
-  // Perform register allocation to convert to a concrete x86 representation
-  PM.add(createRegisterAllocator());
-  
-  if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-    
-  PM.add(createLowerSubregsPass());
-  
-  if (PrintMachineCode)  // Print the subreg lowered code
-    PM.add(createMachineFunctionPrinterPass(cerr));
-
-  // Run post-ra passes.
-  if (addPostRegAlloc(PM, Fast) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-
-  // Insert prolog/epilog code.  Eliminate abstract frame index references...
-  PM.add(createPrologEpilogCodeInserter());
-  
-  // Second pass scheduler.
-  if (!Fast && !DisablePostRAScheduler)
-    PM.add(createPostRAScheduler());
-
-  // Branch folding must be run after regalloc and prolog/epilog insertion.
-  if (!Fast)
-    PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
-
-  PM.add(createGCMachineCodeAnalysisPass());
-  if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-  
-  if (PrintGCInfo)
-    PM.add(createCollectorMetadataPrinter(*cerr));
-  
   // Fold redundant debug labels.
   PM.add(createDebugLabelFoldingPass());
-  
-  if (PrintMachineCode)  // Print the register-allocated code
+
+  if (PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
   if (addPreEmitPass(PM, Fast) && PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
-  if (AlignLoops && !Fast && !OptimizeForSize)
+  if (!Fast)
     PM.add(createLoopAlignerPass());
 
   switch (FileType) {
@@ -159,7 +100,7 @@ LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
 
   return FileModel::Error;
 }
+
 /// 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.
@@ -168,8 +109,8 @@ bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
                                                   bool Fast) {
   if (MCE)
     addSimpleCodeEmitter(PM, Fast, PrintEmittedAsm, *MCE);
-    
-  PM.add(createCollectorMetadataDeleter());
+
+  PM.add(createGCInfoDeleter());
 
   // Delete machine code for this function
   PM.add(createMachineCodeDeleter());
@@ -186,20 +127,41 @@ bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
 bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
                                                    MachineCodeEmitter &MCE,
                                                    bool Fast) {
+  // Add common CodeGen passes.
+  if (addCommonCodeGenPasses(PM, Fast))
+    return true;
+
+  if (addPreEmitPass(PM, Fast) && PrintMachineCode)
+    PM.add(createMachineFunctionPrinterPass(cerr));
+
+  addCodeEmitter(PM, Fast, PrintEmittedAsm, MCE);
+
+  PM.add(createGCInfoDeleter());
+
+  // Delete machine code for this function
+  PM.add(createMachineCodeDeleter());
+
+  return false; // success!
+}
+
+/// addCommonCodeGenPasses - Add standard LLVM codegen passes used for
+/// both emitting to assembly files or machine code output.
+///
+bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, bool Fast) {
   // Standard LLVM-Level Passes.
-  
+
   // Run loop strength reduction before anything else.
   if (!Fast) {
     PM.add(createLoopStrengthReducePass(getTargetLowering()));
     if (PrintLSR)
       PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
   }
-  
+
   PM.add(createGCLoweringPass());
-  
+
   if (!getTargetAsmInfo()->doesSupportExceptionHandling())
     PM.add(createLowerInvokePass(getTargetLowering()));
-  
+
   // Make sure that no unreachable blocks are instruction selected.
   PM.add(createUnreachableBlockEliminationPass());
 
@@ -210,6 +172,15 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
     PM.add(new PrintFunctionPass("\n\n*** Final LLVM Code input to ISel ***\n",
                                  &cerr));
 
+  // Standard Lower-Level Passes.
+
+  // Enable FastISel with -fast, but allow that to be overridden.
+  assert((!EnableFastISelOption || !DisableFastISelOption) &&
+         "Both -fast-isel and -disable-fast-isel given!");
+  if (EnableFastISelOption ||
+      (Fast && !DisableFastISelOption))
+    EnableFastISel = true;
+
   // Ask the target for an isel.
   if (addInstSelector(PM, Fast))
     return true;
@@ -218,9 +189,13 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
   if (PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
-  if (PerformLICM)
+  // If we're using Fast-ISel, clean up the mess.
+  if (EnableFastISel)
+    PM.add(createDeadMachineInstructionElimPass());
+
+  if (EnableLICM)
     PM.add(createMachineLICMPass());
-  
+
   if (EnableSinking)
     PM.add(createMachineSinkingPass());
 
@@ -228,29 +203,36 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
   if (addPreRegAlloc(PM, Fast) && PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
-  // Perform register allocation to convert to a concrete x86 representation
+  // Perform register allocation.
   PM.add(createRegisterAllocator());
-  
-  if (PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
-    
-  PM.add(createLowerSubregsPass());
-  
-  if (PrintMachineCode)  // Print the subreg lowered code
+
+  // Perform stack slot coloring.
+  if (!Fast)
+    PM.add(createStackSlotColoringPass());
+
+  if (PrintMachineCode)  // Print the register-allocated code
     PM.add(createMachineFunctionPrinterPass(cerr));
 
   // Run post-ra passes.
   if (addPostRegAlloc(PM, Fast) && PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
+  if (PrintMachineCode)
+    PM.add(createMachineFunctionPrinterPass(cerr));
+
+  PM.add(createLowerSubregsPass());
+
+  if (PrintMachineCode)  // Print the subreg lowered code
+    PM.add(createMachineFunctionPrinterPass(cerr));
+
   // Insert prolog/epilog code.  Eliminate abstract frame index references...
   PM.add(createPrologEpilogCodeInserter());
-  
-  if (PrintMachineCode)  // Print the register-allocated code
+
+  if (PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
-  
+
   // Second pass scheduler.
-  if (!Fast)
+  if (!Fast && !DisablePostRAScheduler)
     PM.add(createPostRAScheduler());
 
   // Branch folding must be run after regalloc and prolog/epilog insertion.
@@ -258,21 +240,12 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
     PM.add(createBranchFoldingPass(getEnableTailMergeDefault()));
 
   PM.add(createGCMachineCodeAnalysisPass());
+
   if (PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
-  
+
   if (PrintGCInfo)
-    PM.add(createCollectorMetadataPrinter(*cerr));
-  
-  if (addPreEmitPass(PM, Fast) && PrintMachineCode)
-    PM.add(createMachineFunctionPrinterPass(cerr));
+    PM.add(createGCInfoPrinter(*cerr));
 
-  addCodeEmitter(PM, Fast, PrintEmittedAsm, MCE);
-  
-  PM.add(createCollectorMetadataDeleter());
-  
-  // Delete machine code for this function
-  PM.add(createMachineCodeDeleter());
-  
-  return false; // success!
+  return false;
 }