lib/CodeGen/MacroFusion.cpp

   1 //===- MacroFusion.cpp - Macro Fusion -------------------------------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 /// \file This file contains the implementation of the DAG scheduling mutation
  10 /// to pair instructions back to back.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/CodeGen/MacroFusion.h"
  15 #include "llvm/ADT/STLExtras.h"
  16 #include "llvm/ADT/Statistic.h"
  17 #include "llvm/CodeGen/MachineInstr.h"
  18 #include "llvm/CodeGen/MachineScheduler.h"
  19 #include "llvm/CodeGen/ScheduleDAG.h"
  20 #include "llvm/CodeGen/ScheduleDAGMutation.h"
  21 #include "llvm/CodeGen/TargetInstrInfo.h"
  22 #include "llvm/Support/CommandLine.h"
  23 #include "llvm/Support/Debug.h"
  24 #include "llvm/Support/raw_ostream.h"
  25
  26 #define DEBUG_TYPE "machine-scheduler"
  27
  28 STATISTIC(NumFused, "Number of instr pairs fused");
  29
  30 using namespace llvm;
  31
  32 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
  33   cl::desc("Enable scheduling for macro fusion."), cl::init(true));
  34
  35 static bool isHazard(const SDep &Dep) {
  36   return Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output;
  37 }
  38
  39 static bool fuseInstructionPair(ScheduleDAGInstrs &DAG, SUnit &FirstSU,
  40                                 SUnit &SecondSU) {
  41   // Check that neither instr is already paired with another along the edge
  42   // between them.
  43   for (SDep &SI : FirstSU.Succs)
  44     if (SI.isCluster())
  45       return false;
  46
  47   for (SDep &SI : SecondSU.Preds)
  48     if (SI.isCluster())
  49       return false;
  50   // Though the reachability checks above could be made more generic,
  51   // perhaps as part of ScheduleDAGInstrs::addEdge(), since such edges are valid,
  52   // the extra computation cost makes it less interesting in general cases.
  53
  54   // Create a single weak edge between the adjacent instrs. The only effect is
  55   // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
  56   if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
  57     return false;
  58
  59   // Adjust the latency between both instrs.
  60   for (SDep &SI : FirstSU.Succs)
  61     if (SI.getSUnit() == &SecondSU)
  62       SI.setLatency(0);
  63
  64   for (SDep &SI : SecondSU.Preds)
  65     if (SI.getSUnit() == &FirstSU)
  66       SI.setLatency(0);
  67
  68   LLVM_DEBUG(
  69       dbgs() << "Macro fuse: "; DAG.dumpNodeName(FirstSU); dbgs() << " - ";
  70       DAG.dumpNodeName(SecondSU); dbgs() << " /  ";
  71       dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - "
  72              << DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n';);
  73
  74   // Make data dependencies from the FirstSU also dependent on the SecondSU to
  75   // prevent them from being scheduled between the FirstSU and the SecondSU.
  76   if (&SecondSU != &DAG.ExitSU)
  77     for (const SDep &SI : FirstSU.Succs) {
  78       SUnit *SU = SI.getSUnit();
  79       if (SI.isWeak() || isHazard(SI) ||
  80           SU == &DAG.ExitSU || SU == &SecondSU || SU->isPred(&SecondSU))
  81         continue;
  82       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(SecondSU);
  83                  dbgs() << " - "; DAG.dumpNodeName(*SU); dbgs() << '\n';);
  84       DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
  85     }
  86
  87   // Make the FirstSU also dependent on the dependencies of the SecondSU to
  88   // prevent them from being scheduled between the FirstSU and the SecondSU.
  89   if (&FirstSU != &DAG.EntrySU) {
  90     for (const SDep &SI : SecondSU.Preds) {
  91       SUnit *SU = SI.getSUnit();
  92       if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
  93         continue;
  94       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(*SU); dbgs() << " - ";
  95                  DAG.dumpNodeName(FirstSU); dbgs() << '\n';);
  96       DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
  97     }
  98     // ExitSU comes last by design, which acts like an implicit dependency
  99     // between ExitSU and any bottom root in the graph. We should transfer
 100     // this to FirstSU as well.
 101     if (&SecondSU == &DAG.ExitSU) {
 102       for (SUnit &SU : DAG.SUnits) {
 103         if (SU.Succs.empty())
 104           DAG.addEdge(&FirstSU, SDep(&SU, SDep::Artificial));
 105       }
 106     }
 107   }
 108
 109   ++NumFused;
 110   return true;
 111 }
 112
 113 namespace {
 114
 115 /// Post-process the DAG to create cluster edges between instrs that may
 116 /// be fused by the processor into a single operation.
 117 class MacroFusion : public ScheduleDAGMutation {
 118   ShouldSchedulePredTy shouldScheduleAdjacent;
 119   bool FuseBlock;
 120   bool scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU);
 121
 122 public:
 123   MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
 124     : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
 125
 126   void apply(ScheduleDAGInstrs *DAGInstrs) override;
 127 };
 128
 129 } // end anonymous namespace
 130
 131 void MacroFusion::apply(ScheduleDAGInstrs *DAG) {
 132   if (FuseBlock)
 133     // For each of the SUnits in the scheduling block, try to fuse the instr in
 134     // it with one in its predecessors.
 135     for (SUnit &ISU : DAG->SUnits)
 136         scheduleAdjacentImpl(*DAG, ISU);
 137
 138   if (DAG->ExitSU.getInstr())
 139     // Try to fuse the instr in the ExitSU with one in its predecessors.
 140     scheduleAdjacentImpl(*DAG, DAG->ExitSU);
 141 }
 142
 143 /// Implement the fusion of instr pairs in the scheduling DAG,
 144 /// anchored at the instr in AnchorSU..
 145 bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU) {
 146   const MachineInstr &AnchorMI = *AnchorSU.getInstr();
 147   const TargetInstrInfo &TII = *DAG.TII;
 148   const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
 149
 150   // Check if the anchor instr may be fused.
 151   if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
 152     return false;
 153
 154   // Explorer for fusion candidates among the dependencies of the anchor instr.
 155   for (SDep &Dep : AnchorSU.Preds) {
 156     // Ignore dependencies other than data or strong ordering.
 157     if (Dep.isWeak() || isHazard(Dep))
 158       continue;
 159
 160     SUnit &DepSU = *Dep.getSUnit();
 161     if (DepSU.isBoundaryNode())
 162       continue;
 163
 164     const MachineInstr *DepMI = DepSU.getInstr();
 165     if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
 166       continue;
 167
 168     if (fuseInstructionPair(DAG, DepSU, AnchorSU))
 169       return true;
 170   }
 171
 172   return false;
 173 }
 174
 175 std::unique_ptr<ScheduleDAGMutation>
 176 llvm::createMacroFusionDAGMutation(
 177      ShouldSchedulePredTy shouldScheduleAdjacent) {
 178   if(EnableMacroFusion)
 179     return std::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
 180   return nullptr;
 181 }
 182
 183 std::unique_ptr<ScheduleDAGMutation>
 184 llvm::createBranchMacroFusionDAGMutation(
 185      ShouldSchedulePredTy shouldScheduleAdjacent) {
 186   if(EnableMacroFusion)
 187     return std::make_unique<MacroFusion>(shouldScheduleAdjacent, false);
 188   return nullptr;
 189 }