Testcase for PR2264.
[oota-llvm.git] / lib / CodeGen / LLVMTargetMachine.cpp
index 641c04653a07a2f68b24cf99ad2f201cb5729836..4004cf1e1800a3e5b0f8eca818262089f6a64bf2 100644 (file)
@@ -2,8 +2,8 @@
 //
 //                     The LLVM Compiler Infrastructure
 //
-// This file was developed by Chris Lattner and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 #include "llvm/Assembly/PrintModulePass.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/Collector.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Target/TargetAsmInfo.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Support/CommandLine.h"
 using namespace llvm;
 
-static cl::opt<bool> PrintLSR("print-lsr-output");
+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"));
+
+// Hidden options to help debugging
+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"));
+
+// When this works it will be on by default.
+static cl::opt<bool>
+DisablePostRAScheduler("disable-post-RA-scheduler",
+                       cl::desc("Disable scheduling after register allocation"),
+                       cl::init(true));
 
 FileModel::Model
-LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
+LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM,
                                        std::ostream &Out,
                                        CodeGenFileType FileType,
                                        bool Fast) {
@@ -35,19 +62,24 @@ LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
   if (!Fast) {
     PM.add(createLoopStrengthReducePass(getTargetLowering()));
     if (PrintLSR)
-      PM.add(new PrintFunctionPass("\n\n*** Code after LSR *** \n", &cerr));
+      PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
   }
   
-  // FIXME: Implement efficient support for garbage collection intrinsics.
-  PM.add(createLowerGCPass());
-  
-  // FIXME: Implement the invoke/unwind instructions!
-  if (!ExceptionHandling)
+  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))
     return FileModel::Error;
@@ -55,12 +87,27 @@ LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
   // 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)
@@ -69,10 +116,21 @@ LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
   // 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());
-    
+    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());
   
@@ -82,6 +140,9 @@ LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
   if (addPreEmitPass(PM, Fast) && PrintMachineCode)
     PM.add(createMachineFunctionPrinterPass(cerr));
 
+  if (AlignLoops && !OptimizeForSize)
+    PM.add(createLoopAlignerPass());
+
   switch (FileType) {
   default:
     break;
@@ -102,11 +163,13 @@ LLVMTargetMachine::addPassesToEmitFile(FunctionPassManager &PM,
 /// 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(FunctionPassManager &PM,
+bool LLVMTargetMachine::addPassesToEmitFileFinish(PassManagerBase &PM,
                                                   MachineCodeEmitter *MCE,
                                                   bool Fast) {
   if (MCE)
-    addSimpleCodeEmitter(PM, Fast, *MCE);
+    addSimpleCodeEmitter(PM, Fast, PrintEmittedAsm, *MCE);
+    
+  PM.add(createCollectorMetadataDeleter());
 
   // Delete machine code for this function
   PM.add(createMachineCodeDeleter());
@@ -120,23 +183,33 @@ bool LLVMTargetMachine::addPassesToEmitFileFinish(FunctionPassManager &PM,
 /// of functions.  This method should returns true if machine code emission is
 /// not supported.
 ///
-bool LLVMTargetMachine::addPassesToEmitMachineCode(FunctionPassManager &PM,
+bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM,
                                                    MachineCodeEmitter &MCE,
                                                    bool Fast) {
   // Standard LLVM-Level Passes.
   
   // Run loop strength reduction before anything else.
-  if (!Fast) PM.add(createLoopStrengthReducePass(getTargetLowering()));
+  if (!Fast) {
+    PM.add(createLoopStrengthReducePass(getTargetLowering()));
+    if (PrintLSR)
+      PM.add(new PrintFunctionPass("\n\n*** Code after LSR ***\n", &cerr));
+  }
   
-  // FIXME: Implement efficient support for garbage collection intrinsics.
-  PM.add(createLowerGCPass());
+  PM.add(createGCLoweringPass());
   
-  // FIXME: Implement the invoke/unwind instructions!
-  PM.add(createLowerInvokePass(getTargetLowering()));
+  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))
     return true;
@@ -144,12 +217,27 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(FunctionPassManager &PM,
   // 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)
@@ -161,14 +249,27 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(FunctionPassManager &PM,
   if (PrintMachineCode)  // Print the register-allocated code
     PM.add(createMachineFunctionPrinterPass(cerr));
   
+  // Second pass scheduler.
+  if (!Fast)
+    PM.add(createPostRAScheduler());
+
   // Branch folding must be run after regalloc and prolog/epilog insertion.
   if (!Fast)
-    PM.add(createBranchFoldingPass());
+    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));
 
-  addCodeEmitter(PM, Fast, MCE);
+  addCodeEmitter(PM, Fast, PrintEmittedAsm, MCE);
+  
+  PM.add(createCollectorMetadataDeleter());
   
   // Delete machine code for this function
   PM.add(createMachineCodeDeleter());