Turn on argument promotion in gccas. This can give us substantially better
[oota-llvm.git] / tools / gccas / gccas.cpp
index ce22ccdaf3b1e727d681c11c344d22399637f7e8..dd2b307d3b8ffc40490e98f15afe89d9fcbf521d 100644 (file)
+//===-- gccas.cpp - The "optimizing assembler" used by the GCC frontend ---===//
+// 
+//                     The LLVM Compiler Infrastructure
+//
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// 
 //===----------------------------------------------------------------------===//
-// LLVM 'GCCAS' UTILITY 
 //
-//  This utility is designed to be used by the GCC frontend for creating
-// bytecode files from it's intermediate llvm assembly.  The requirements for
-// this utility are thus slightly different than that of the standard as util.
+// This utility is designed to be used by the GCC frontend for creating bytecode
+// files from its intermediate LLVM assembly.  The requirements for this utility
+// are thus slightly different than that of the standard `as' util.
 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Module.h"
 #include "llvm/PassManager.h"
+#include "llvm/Analysis/LoadValueNumbering.h"
+#include "llvm/Analysis/Verifier.h"
 #include "llvm/Assembly/Parser.h"
-#include "llvm/Transforms/CleanupGCCOutput.h"
-#include "llvm/Transforms/LevelChange.h"
-#include "llvm/Transforms/ConstantMerge.h"
-#include "llvm/Transforms/ChangeAllocations.h"
-#include "llvm/Transforms/Scalar/ConstantProp.h"
-#include "llvm/Transforms/Scalar/DCE.h"
-#include "llvm/Transforms/Scalar/GCSE.h"
-#include "llvm/Transforms/Scalar/IndVarSimplify.h"
-#include "llvm/Transforms/Scalar/InstructionCombining.h"
-#include "llvm/Transforms/Scalar/PromoteMemoryToRegister.h"
 #include "llvm/Bytecode/WriteBytecodePass.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Transforms/IPO.h"
+#include "llvm/Transforms/Scalar.h"
 #include "Support/CommandLine.h"
 #include "Support/Signals.h"
 #include <memory>
 #include <fstream>
 
-cl::String InputFilename ("", "Parse <arg> file, compile to bytecode",
-                          cl::Required, "");
-cl::String OutputFilename("o", "Override output filename", cl::NoFlags, "");
-cl::Flag   StopAtLevelRaise("stopraise", "Stop optimization before level raise",
-                            cl::Hidden);
+using namespace llvm;
+
+namespace {
+  cl::opt<std::string>
+  InputFilename(cl::Positional,cl::desc("<input llvm assembly>"),cl::init("-"));
+
+  cl::opt<std::string> 
+  OutputFilename("o", cl::desc("Override output filename"),
+                 cl::value_desc("filename"));
+
+  cl::opt<bool>   
+  Verify("verify", cl::desc("Verify each pass result"));
+
+  cl::opt<bool>
+  DisableInline("disable-inlining", cl::desc("Do not run the inliner pass"));
+
+  cl::opt<bool>
+  DisableOptimizations("disable-opt",
+                       cl::desc("Do not run any optimization passes"));
+}
+
+
+static inline void addPass(PassManager &PM, Pass *P) {
+  // Add the pass to the pass manager...
+  PM.add(P);
+  
+  // If we are verifying all of the intermediate steps, add the verifier...
+  if (Verify) PM.add(createVerifierPass());
+}
+
+
+void AddConfiguredTransformationPasses(PassManager &PM) {
+  PM.add(createVerifierPass());                  // Verify that input is correct
+  addPass(PM, createLowerSetJmpPass());          // Lower llvm.setjmp/.longjmp
+  addPass(PM, createFunctionResolvingPass());    // Resolve (...) functions
+
+  if (DisableOptimizations) return;
+
+  addPass(PM, createRaiseAllocationsPass());     // call %malloc -> malloc inst
+  addPass(PM, createCFGSimplificationPass());    // Clean up disgusting code
+  addPass(PM, createPromoteMemoryToRegister());  // Kill useless allocas
+  addPass(PM, createGlobalConstifierPass());     // Mark read-only globals const
+  addPass(PM, createGlobalDCEPass());            // Remove unused globals
+  addPass(PM, createIPConstantPropagationPass());// IP Constant Propagation
+  addPass(PM, createDeadArgEliminationPass());   // Dead argument elimination
+  addPass(PM, createInstructionCombiningPass()); // Clean up after IPCP & DAE
+  addPass(PM, createCFGSimplificationPass());    // Clean up after IPCP & DAE
+
+  addPass(PM, createPruneEHPass());              // Remove dead EH info
+
+  if (!DisableInline)
+    addPass(PM, createFunctionInliningPass());   // Inline small functions
+  addPass(PM, createArgumentPromotionPass());    // Scalarize uninlined fn args
+
+  addPass(PM, createRaisePointerReferencesPass());// Recover type information
+  addPass(PM, createTailDuplicationPass());      // Simplify cfg by copying code
+  addPass(PM, createCFGSimplificationPass());    // Merge & remove BBs
+  addPass(PM, createScalarReplAggregatesPass()); // Break up aggregate allocas
+  addPass(PM, createInstructionCombiningPass()); // Combine silly seq's
+
+  addPass(PM, createReassociatePass());          // Reassociate expressions
+  addPass(PM, createInstructionCombiningPass()); // Combine silly seq's
+  addPass(PM, createTailCallEliminationPass());  // Eliminate tail calls
+  addPass(PM, createCFGSimplificationPass());    // Merge & remove BBs
+  addPass(PM, createLICMPass());                 // Hoist loop invariants
+  addPass(PM, createLoadValueNumberingPass());   // GVN for load instructions
+  addPass(PM, createGCSEPass());                 // Remove common subexprs
+  addPass(PM, createSCCPPass());                 // Constant prop with SCCP
+
+  // Run instcombine after redundancy elimination to exploit opportunities
+  // opened up by them.
+  addPass(PM, createInstructionCombiningPass());
+  addPass(PM, createIndVarSimplifyPass());       // Canonicalize indvars
+  addPass(PM, createAggressiveDCEPass());        // SSA based 'Aggressive DCE'
+  addPass(PM, createCFGSimplificationPass());    // Merge & remove BBs
+  addPass(PM, createDeadTypeEliminationPass());  // Eliminate dead types
+  addPass(PM, createConstantMergePass());        // Merge dup global constants
+}
+
 
 int main(int argc, char **argv) {
   cl::ParseCommandLineOptions(argc, argv, " llvm .s -> .o assembler for GCC\n");
+  PrintStackTraceOnErrorSignal();
 
   std::auto_ptr<Module> M;
   try {
     // Parse the file now...
     M.reset(ParseAssemblyFile(InputFilename));
   } catch (const ParseException &E) {
-    cerr << E.getMessage() << endl;
+    std::cerr << argv[0] << ": " << E.getMessage() << "\n";
     return 1;
   }
 
   if (M.get() == 0) {
-    cerr << "assembly didn't read correctly.\n";
+    std::cerr << argv[0] << ": assembly didn't read correctly.\n";
     return 1;
   }
-  
+
+  std::ostream *Out = 0;
   if (OutputFilename == "") {   // Didn't specify an output filename?
-    std::string IFN = InputFilename;
-    int Len = IFN.length();
-    if (IFN[Len-2] == '.' && IFN[Len-1] == 's') {   // Source ends in .s?
-      OutputFilename = std::string(IFN.begin(), IFN.end()-2);
+    if (InputFilename == "-") {
+      OutputFilename = "-";
     } else {
-      OutputFilename = IFN;   // Append a .o to it
+      std::string IFN = InputFilename;
+      int Len = IFN.length();
+      if (IFN[Len-2] == '.' && IFN[Len-1] == 's') {   // Source ends in .s?
+        OutputFilename = std::string(IFN.begin(), IFN.end()-2);
+      } else {
+        OutputFilename = IFN;   // Append a .o to it
+      }
+      OutputFilename += ".o";
     }
-    OutputFilename += ".o";
   }
 
-  std::ofstream Out(OutputFilename.c_str(), ios::out);
-  if (!Out.good()) {
-    cerr << "Error opening " << OutputFilename << "!\n";
-    return 1;
+  if (OutputFilename == "-")
+    Out = &std::cout;
+  else {
+    Out = new std::ofstream(OutputFilename.c_str(), std::ios::out);
+
+    // Make sure that the Out file gets unlinked from the disk if we get a
+    // signal
+    RemoveFileOnSignal(OutputFilename);
   }
 
-  // Make sure that the Out file gets unlink'd from the disk if we get a SIGINT
-  RemoveFileOnSignal(OutputFilename);
+  
+  if (!Out->good()) {
+    std::cerr << argv[0] << ": error opening " << OutputFilename << "!\n";
+    return 1;
+  }
 
   // In addition to just parsing the input from GCC, we also want to spiff it up
   // a little bit.  Do this now.
   //
   PassManager Passes;
-  Passes.add(createFunctionResolvingPass());      // Resolve (...) functions
-  Passes.add(createConstantMergePass());          // Merge dup global constants
-  Passes.add(createDeadInstEliminationPass());    // Remove Dead code/vars
-  Passes.add(createRaiseAllocationsPass());       // call %malloc -> malloc inst
-  Passes.add(createCleanupGCCOutputPass());       // Fix gccisms
-  Passes.add(createIndVarSimplifyPass());         // Simplify indvars
-  if (!StopAtLevelRaise) {
-    Passes.add(createRaisePointerReferencesPass()); // Eliminate casts
-    Passes.add(createPromoteMemoryToRegister());    // Promote alloca's to regs
-    Passes.add(createInstructionCombiningPass());   // Combine silly seq's
-    Passes.add(createDeadCodeEliminationPass());    // Remove Dead code/vars
-    Passes.add(createSCCPPass());                   // Constant prop with SCCP
-    Passes.add(createGCSEPass());                   // Remove common subexprs
-  }
-  Passes.add(new WriteBytecodePass(&Out));        // Write bytecode to file...
+
+  // Add an appropriate TargetData instance for this module...
+  Passes.add(new TargetData("gccas", M.get()));
+
+  // Add all of the transformation passes to the pass manager to do the cleanup
+  // and optimization of the GCC output.
+  //
+  AddConfiguredTransformationPasses(Passes);
+
+  // Make sure everything is still good.
+  Passes.add(createVerifierPass());
+
+  // Write bytecode to file...
+  Passes.add(new WriteBytecodePass(Out));
 
   // Run our queue of passes all at once now, efficiently.
-  Passes.run(M.get());
+  Passes.run(*M.get());
+
+  if (Out != &std::cout) delete Out;
   return 0;
 }
-