path: root/lib/Target/R600/AMDILPeepholeOptimizer.cpp

//===-- AMDILPeepholeOptimizer.cpp - AMDGPU Peephole optimizations ---------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
/// \file
//==-----------------------------------------------------------------------===//

#define DEBUG_TYPE "PeepholeOpt"
#ifdef DEBUG
#define DEBUGME (DebugFlag && isCurrentDebugType(DEBUG_TYPE))
#else
#define DEBUGME 0
#endif

#include "AMDILDevices.h"
#include "AMDGPUInstrInfo.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/Twine.h"
#include "llvm/IR/Constants.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionAnalysis.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"

#include <sstream>

#if 0
STATISTIC(PointerAssignments, "Number of dynamic pointer "
    "assigments discovered");
STATISTIC(PointerSubtract, "Number of pointer subtractions discovered");
#endif

using namespace llvm;
// The Peephole optimization pass is used to do simple last minute optimizations
// that are required for correct code or to remove redundant functions
namespace {

class OpaqueType;

class LLVM_LIBRARY_VISIBILITY AMDGPUPeepholeOpt : public FunctionPass {
public:
  TargetMachine &TM;
  static char ID;
  AMDGPUPeepholeOpt(TargetMachine &tm);
  ~AMDGPUPeepholeOpt();
  const char *getPassName() const;
  bool runOnFunction(Function &F);
  bool doInitialization(Module &M);
  bool doFinalization(Module &M);
  void getAnalysisUsage(AnalysisUsage &AU) const;
protected:
private:
  // Function to initiate all of the instruction level optimizations.
  bool instLevelOptimizations(BasicBlock::iterator *inst);
  // Quick check to see if we need to dump all of the pointers into the
  // arena. If this is correct, then we set all pointers to exist in arena. This
  // is a workaround for aliasing of pointers in a struct/union.
  bool dumpAllIntoArena(Function &F);
  // Because I don't want to invalidate any pointers while in the
  // safeNestedForEachFunction. I push atomic conversions to a vector and handle
  // it later. This function does the conversions if required.
  void doAtomicConversionIfNeeded(Function &F);
  // Because __amdil_is_constant cannot be properly evaluated if
  // optimizations are disabled, the call's are placed in a vector
  // and evaluated after the __amdil_image* functions are evaluated
  // which should allow the __amdil_is_constant function to be
  // evaluated correctly.
  void doIsConstCallConversionIfNeeded();
  bool mChanged;
  bool mDebug;
  bool mConvertAtomics;
  CodeGenOpt::Level optLevel;
  // Run a series of tests to see if we can optimize a CALL instruction.
  bool optimizeCallInst(BasicBlock::iterator *bbb);
  // A peephole optimization to optimize bit extract sequences.
  bool optimizeBitExtract(Instruction *inst);
  // A peephole optimization to optimize bit insert sequences.
  bool optimizeBitInsert(Instruction *inst);
  bool setupBitInsert(Instruction *base, 
                      Instruction *&src, 
                      Constant *&mask, 
                      Constant *&shift);
  // Expand the bit field insert instruction on versions of OpenCL that
  // don't support it.
  bool expandBFI(CallInst *CI);
  // Expand the bit field mask instruction on version of OpenCL that 
  // don't support it.
  bool expandBFM(CallInst *CI);
  // On 7XX and 8XX operations, we do not have 24 bit signed operations. So in
  // this case we need to expand them. These functions check for 24bit functions
  // and then expand.
  bool isSigned24BitOps(CallInst *CI);
  void expandSigned24BitOps(CallInst *CI);
  // One optimization that can occur is that if the required workgroup size is
  // specified then the result of get_local_size is known at compile time and
  // can be returned accordingly.
  bool isRWGLocalOpt(CallInst *CI);
  // On northern island cards, the division is slightly less accurate than on
  // previous generations, so we need to utilize a more accurate division. So we
  // can translate the accurate divide to a normal divide on all other cards.
  bool