reference, declarationdefinition
definition → references, declarations, derived classes, virtual overrides
reference to multiple definitions → definitions
unreferenced
    1
    2
    3
    4
    5
    6
    7
    8
    9
   10
   11
   12
   13
   14
   15
   16
   17
   18
   19
   20
   21
   22
   23
   24
   25
   26
   27
   28
   29
   30
   31
   32
   33
   34
   35
   36
   37
   38
   39
   40
   41
   42
   43
   44
   45
   46
   47
   48
   49
   50
   51
   52
   53
   54
   55
   56
   57
   58
   59
   60
   61
   62
   63
   64
   65
   66
   67
   68
   69
   70
   71
   72
   73
   74
   75
   76
   77
   78
   79
   80
   81
   82
   83
   84
   85
   86
   87
   88
   89
   90
   91
   92
   93
   94
   95
   96
   97
   98
   99
  100
  101
  102
  103
  104
  105
  106
  107
  108
  109
  110
  111
  112
  113
  114
  115
  116
  117
  118
  119
  120
  121
  122
  123
  124
  125
  126
  127
  128
  129
  130
  131
  132
  133
  134
  135
  136
  137
  138
  139
  140
  141
  142
  143
  144
  145
  146
  147
  148
  149
  150
  151
  152
  153
  154
  155
  156
  157
  158
  159
  160
  161
  162
  163
  164
  165
  166
  167
  168
  169
  170
  171
  172
//===--- AArch64StorePairSuppress.cpp --- Suppress store pair formation ---===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This pass identifies floating point stores that should not be combined into
// store pairs. Later we may do the same for floating point loads.
// ===---------------------------------------------------------------------===//

#include "AArch64InstrInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineTraceMetrics.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetSchedule.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"

using namespace llvm;

#define DEBUG_TYPE "aarch64-stp-suppress"

#define STPSUPPRESS_PASS_NAME "AArch64 Store Pair Suppression"

namespace {
class AArch64StorePairSuppress : public MachineFunctionPass {
  const AArch64InstrInfo *TII;
  const TargetRegisterInfo *TRI;
  const MachineRegisterInfo *MRI;
  TargetSchedModel SchedModel;
  MachineTraceMetrics *Traces;
  MachineTraceMetrics::Ensemble *MinInstr;

public:
  static char ID;
  AArch64StorePairSuppress() : MachineFunctionPass(ID) {
    initializeAArch64StorePairSuppressPass(*PassRegistry::getPassRegistry());
  }

  StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; }

  bool runOnMachineFunction(MachineFunction &F) override;

private:
  bool shouldAddSTPToBlock(const MachineBasicBlock *BB);

  bool isNarrowFPStore(const MachineInstr &MI);

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<MachineTraceMetrics>();
    AU.addPreserved<MachineTraceMetrics>();
    MachineFunctionPass::getAnalysisUsage(AU);
  }
};
char AArch64StorePairSuppress::ID = 0;
} // anonymous

INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress",
                STPSUPPRESS_PASS_NAME, false, false)

FunctionPass *llvm::createAArch64StorePairSuppressPass() {
  return new AArch64StorePairSuppress();
}

/// Return true if an STP can be added to this block without increasing the
/// critical resource height. STP is good to form in Ld/St limited blocks and
/// bad to form in float-point limited blocks. This is true independent of the
/// critical path. If the critical path is longer than the resource height, the
/// extra vector ops can limit physreg renaming. Otherwise, it could simply
/// oversaturate the vector units.
bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) {
  if (!MinInstr)
    MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);

  MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(BB);
  unsigned ResLength = BBTrace.getResourceLength();

  // Get the machine model's scheduling class for STPQi.
  // Bypass TargetSchedule's SchedClass resolution since we only have an opcode.
  unsigned SCIdx = TII->get(AArch64::STPDi).getSchedClass();
  const MCSchedClassDesc *SCDesc =
      SchedModel.getMCSchedModel()->getSchedClassDesc(SCIdx);

  // If a subtarget does not define resources for STPQi, bail here.
  if (SCDesc->isValid() && !SCDesc->isVariant()) {
    unsigned ResLenWithSTP = BBTrace.getResourceLength(None, SCDesc);
    if (ResLenWithSTP > ResLength) {
      LLVM_DEBUG(dbgs() << "  Suppress STP in BB: " << BB->getNumber()
                        << " resources " << ResLength << " -> " << ResLenWithSTP
                        << "\n");
      return false;
    }
  }
  return true;
}

/// Return true if this is a floating-point store smaller than the V reg. On
/// cyclone, these require a vector shuffle before storing a pair.
/// Ideally we would call getMatchingPairOpcode() and have the machine model
/// tell us if it's profitable with no cpu knowledge here.
///
/// FIXME: We plan to develop a decent Target abstraction for simple loads and
/// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer.
bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) {
  switch (MI.getOpcode()) {
  default:
    return false;
  case AArch64::STRSui:
  case AArch64::STRDui:
  case AArch64::STURSi:
  case AArch64::STURDi:
    return true;
  }
}

bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) {
  if (skipFunction(MF.getFunction()))
    return false;

  const TargetSubtargetInfo &ST = MF.getSubtarget();
  TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo());
  TRI = ST.getRegisterInfo();
  MRI = &MF.getRegInfo();
  SchedModel.init(&ST);
  Traces = &getAnalysis<MachineTraceMetrics>();
  MinInstr = nullptr;

  LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n');

  if (!SchedModel.hasInstrSchedModel()) {
    LLVM_DEBUG(dbgs() << "  Skipping pass: no machine model present.\n");
    return false;
  }

  // Check for a sequence of stores to the same base address. We don't need to
  // precisely determine whether a store pair can be formed. But we do want to
  // filter out most situations where we can't form store pairs to avoid
  // computing trace metrics in those cases.
  for (auto &MBB : MF) {
    bool SuppressSTP = false;
    unsigned PrevBaseReg = 0;
    for (auto &MI : MBB) {
      if (!isNarrowFPStore(MI))
        continue;
      const MachineOperand *BaseOp;
      int64_t Offset;
      if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, TRI) &&
          BaseOp->isReg()) {
        Register BaseReg = BaseOp->getReg();
        if (PrevBaseReg == BaseReg) {
          // If this block can take STPs, skip ahead to the next block.
          if (!SuppressSTP && shouldAddSTPToBlock(MI.getParent()))
            break;
          // Otherwise, continue unpairing the stores in this block.
          LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n");
          SuppressSTP = true;
          TII->suppressLdStPair(MI);
        }
        PrevBaseReg = BaseReg;
      } else
        PrevBaseReg = 0;
    }
  }
  // This pass just sets some internal MachineMemOperand flags. It can't really
  // invalidate anything.
  return false;
}