path: root/lib/CodeGen/ExecutionDepsFix.cpp

//===- ExecutionDepsFix.cpp - Fix execution dependecy issues ----*- C++ -*-===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains the execution dependency fix pass.
//
// Some X86 SSE instructions like mov, and, or, xor are available in different
// variants for different operand types. These variant instructions are
// equivalent, but on Nehalem and newer cpus there is extra latency
// transferring data between integer and floating point domains.  ARM cores
// have similar issues when they are configured with both VFP and NEON
// pipelines.
//
// This pass changes the variant instructions to minimize domain crossings.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "execution-fix"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/Support/Allocator.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;

/// A DomainValue is a bit like LiveIntervals' ValNo, but it also keeps track
/// of execution domains.
///
/// An open DomainValue represents a set of instructions that can still switch
/// execution domain. Multiple registers may refer to the same open
/// DomainValue - they will eventually be collapsed to the same execution
/// domain.
///
/// A collapsed DomainValue represents a single register that has been forced
/// into one of more execution domains. There is a separate collapsed
/// DomainValue for each register, but it may contain multiple execution
/// domains. A register value is initially created in a single execution
/// domain, but if we were forced to pay the penalty of a domain crossing, we
/// keep track of the fact that the register is now available in multiple
/// domains.
namespace {
struct DomainValue {
  // Basic reference counting.
  unsigned Refs;

  // Bitmask of available domains. For an open DomainValue, it is the still
  // possible domains for collapsing. For a collapsed DomainValue it is the
  // domains where the register is available for free.
  unsigned AvailableDomains;

  // Pointer to the next DomainValue in a chain.  When two DomainValues are
  // merged, Victim.Next is set to point to Victor, so old DomainValue
  // references can be updated by following the chain.
  DomainValue *Next;

  // Twiddleable instructions using or defining these registers.
  SmallVector<MachineInstr*, 8> Instrs;

  // A collapsed DomainValue has no instructions to twiddle - it simply keeps
  // track of the domains where the registers are already available.
  bool isCollapsed() const { return Instrs.empty(); }

  // Is domain available?
  bool hasDomain(unsigned domain) const {
    return AvailableDomains & (1u << domain);
  }

  // Mark domain as available.
  void addDomain(unsigned domain) {
    AvailableDomains |= 1u << domain;
  }

  // Restrict to a single domain available.
  void setSingleDomain(unsigned domain) {
    AvailableDomains = 1u << domain;
  }

  // Return bitmask of domains that are available and in mask.
  unsigned getCommonDomains(unsigned mask) const {
    return AvailableDomains & mask;
  }

  // First domain available.
  unsigned getFirstDomain() const {
    return CountTrailingZeros_32(AvailableDomains);
  }

  DomainValue() : Refs(0) { clear(); }

  // Clear this DomainValue and point to next which has all its data.
  void clear() {
    AvailableDomains = 0;
    Next = 0;
    Instrs.clear();
  }
};
}

namespace {
/// LiveReg - Information about a live register.
struct LiveReg {
  /// Value currently in this register, or NULL when no value is being tracked.
  /// This counts as a DomainValue reference.
  DomainValue *Value;

  /// Instruction that defined this register, relative to the beginning of the
  /// current basic block.  When a LiveReg is used to represent a live-out
  /// register, this value is relative to the end of the basic block, so it
  /// will be a negative number.
  int Def;
};
} // anonynous namespace

namespace {
class ExeDepsFix : public MachineFunctionPass {
  static char ID;
  SpecificBumpPtrAllocator<DomainValue> Allocator;
  SmallVector<DomainValue*,16> Avail;

  const TargetRegisterClass *const RC;
  MachineFunction *MF;
  const TargetInstrInfo *TII;
  const TargetRegisterInfo *TRI;
  std::vector<int> AliasMap;
  const unsigned NumRegs;
  LiveReg *LiveRegs;
  typedef DenseMap<MachineBasicBlock*, LiveReg*> LiveOutMap;
  LiveOutMap LiveOuts;

  /// Current instruction number.
  /// The first instruction in each basic block is 0.
  int CurInstr;

  /// True when the current block has a predecessor that hasn't been visited
  /// yet.
  bool SeenUnknownBackEdge;

public:
  ExeDepsFix(const TargetRegisterClass *rc)
    : MachineFunctionPass(ID), RC(rc), NumRegs(RC->getNumRegs()) {}

  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
    AU.setPreservesAll();
    MachineFunctionPass::getAnalysisUsage(AU);
  }

  virtual bool runOnMachineFunction(MachineFunction &MF);

  virtual const char *getPassName() const {
    return "Execution dependency fix";
  }

private:
  // Register mapping.
  int regIndex(unsigned Reg);

  // DomainValue allocation.
  DomainValue *alloc(int domain = -1);
  DomainValue *retain(DomainValue *DV) {
    if (DV) ++DV->Refs;
    return DV;
  }
  void release(DomainValue*);
  DomainValue *resolve(DomainValue*&);

  // LiveRegs manipulations.
  void setLiveReg(int rx, DomainValue *DV);
  void kill(int rx);
  void force(int rx, unsigned domain);
  void collapse(DomainValue *dv, unsigned domain);
  bool merge(DomainValue *A, DomainValue *B);

  void enterBasicBlock(MachineBasicBlock*);
  void leaveBasicBlock(MachineBasicBlock*);
  void visitInstr(MachineInstr*);
  void processDefs(MachineInstr*, bool Kill);
  void visitSoftInstr(MachineInstr*, unsigned mask);
  void visitHardInstr(MachineInstr*, unsigned domain);
};
}

char ExeDepsFix::ID = 0;

/// Translate TRI register number to an index into our smaller tables of
/// interesting registers. Return -1 for boring registers.
int ExeDepsFix::regIndex(unsigned Reg) {
  assert(Reg < AliasMap.size() && "Invalid register");
  return AliasMap[Reg];
}

DomainValue *ExeDepsFix::alloc(int domain) {
  DomainValue *dv = Avail.empty() ?
                      new(Allocator.Allocate()) DomainValue :
                      Avail.pop_back_val();
  if (domain >= 0)
    dv->addDomain(domain);
  assert(dv->Refs == 0 && "Reference count wasn't cleared");
  assert(!dv->Next && "Chained DomainValue shouldn't have been recycled");
  return dv;
}

/// release - Release a reference to DV.  When the last reference is released,
/// collapse if needed.
void ExeDepsFix::release(DomainValue *DV) {
  while (DV) {
    assert(DV->Refs && "Bad DomainValue");
    if (--DV->Refs)
      return;

    // There are no more DV references. Collapse any contained instructions.
    if (DV->AvailableDomains && !DV->isCollapsed())
      collapse(DV, DV->getFirstDomain());

    DomainValue *Next = DV->Next;
    DV->clear();
    Avail.push_back(DV);
    // Also release the next DomainValue in the chain.
    DV = Next;
  }
}

/// resolve - Follow the chain of dead DomainValues until a live DomainValue is
/// reached.  Update the referenced pointer when necessary.
DomainValue *ExeDepsFix::resolve(DomainValue *&DVRef)