llvm/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 SUnit *getPredClusterSU(const SUnit &SU) {
  40   for (const SDep &SI : SU.Preds)
  41     if (SI.isCluster())
  42       return SI.getSUnit();
  43
  44   return nullptr;
  45 }
  46
  47 static bool hasLessThanNumFused(const SUnit &SU, unsigned FuseLimit) {
  48   unsigned Num = 1;
  49   const SUnit *CurrentSU = &SU;
  50   while ((CurrentSU = getPredClusterSU(*CurrentSU)) && Num < FuseLimit) Num ++;
  51   return Num < FuseLimit;
  52 }
  53
  54 static bool fuseInstructionPair(ScheduleDAGInstrs &DAG, SUnit &FirstSU,
  55                                 SUnit &SecondSU) {
  56   // Check that neither instr is already paired with another along the edge
  57   // between them.
  58   for (SDep &SI : FirstSU.Succs)
  59     if (SI.isCluster())
  60       return false;
  61
  62   for (SDep &SI : SecondSU.Preds)
  63     if (SI.isCluster())
  64       return false;
  65   // Though the reachability checks above could be made more generic,
  66   // perhaps as part of ScheduleDAGInstrs::addEdge(), since such edges are valid,
  67   // the extra computation cost makes it less interesting in general cases.
  68
  69   // Create a single weak edge between the adjacent instrs. The only effect is
  70   // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
  71   if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
  72     return false;
  73
  74   // TODO - If we want to chain more than two instructions, we need to create
  75   // artifical edges to make dependencies from the FirstSU also dependent
  76   // on other chained instructions, and other chained instructions also
  77   // dependent on the dependencies of the SecondSU, to prevent them from being
  78   // scheduled into these chained instructions.
  79   assert(hasLessThanNumFused(FirstSU, 2) &&
  80          "Currently we only support chaining together two instructions");
  81
  82   // Adjust the latency between both instrs.
  83   for (SDep &SI : FirstSU.Succs)
  84     if (SI.getSUnit() == &SecondSU)
  85       SI.setLatency(0);
  86
  87   for (SDep &SI : SecondSU.Preds)
  88     if (SI.getSUnit() == &FirstSU)
  89       SI.setLatency(0);
  90
  91   LLVM_DEBUG(
  92       dbgs() << "Macro fuse: "; DAG.dumpNodeName(FirstSU); dbgs() << " - ";
  93       DAG.dumpNodeName(SecondSU); dbgs() << " /  ";
  94       dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - "
  95              << DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n';);
  96
  97   // Make data dependencies from the FirstSU also dependent on the SecondSU to
  98   // prevent them from being scheduled between the FirstSU and the SecondSU.
  99   if (&SecondSU != &DAG.ExitSU)
 100     for (const SDep &SI : FirstSU.Succs) {
 101       SUnit *SU = SI.getSUnit();
 102       if (SI.isWeak() || isHazard(SI) ||
 103           SU == &DAG.ExitSU || SU == &SecondSU || SU->isPred(&SecondSU))
 104         continue;
 105       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(SecondSU);
 106                  dbgs() << " - "; DAG.dumpNodeName(*SU); dbgs() << '\n';);
 107       DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
 108     }
 109
 110   // Make the FirstSU also dependent on the dependencies of the SecondSU to
 111   // prevent them from being scheduled between the FirstSU and the SecondSU.
 112   if (&FirstSU != &DAG.EntrySU) {
 113     for (const SDep &SI : SecondSU.Preds) {
 114       SUnit *SU = SI.getSUnit();
 115       if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
 116         continue;
 117       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(*SU); dbgs() << " - ";
 118                  DAG.dumpNodeName(FirstSU); dbgs() << '\n';);
 119       DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
 120     }
 121     // ExitSU comes last by design, which acts like an implicit dependency
 122     // between ExitSU and any bottom root in the graph. We should transfer
 123     // this to FirstSU as well.
 124     if (&SecondSU == &DAG.ExitSU) {
 125       for (SUnit &SU : DAG.SUnits) {
 126         if (SU.Succs.empty())
 127           DAG.addEdge(&FirstSU, SDep(&SU, SDep::Artificial));
 128       }
 129     }
 130   }
 131
 132   ++NumFused;
 133   return true;
 134 }
 135
 136 namespace {
 137
 138 /// Post-process the DAG to create cluster edges between instrs that may
 139 /// be fused by the processor into a single operation.
 140 class MacroFusion : public ScheduleDAGMutation {
 141   ShouldSchedulePredTy shouldScheduleAdjacent;
 142   bool FuseBlock;
 143   bool scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU);
 144
 145 public:
 146   MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
 147     : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
 148
 149   void apply(ScheduleDAGInstrs *DAGInstrs) override;
 150 };
 151
 152 } // end anonymous namespace
 153
 154 void MacroFusion::apply(ScheduleDAGInstrs *DAG) {
 155   if (FuseBlock)
 156     // For each of the SUnits in the scheduling block, try to fuse the instr in
 157     // it with one in its predecessors.
 158     for (SUnit &ISU : DAG->SUnits)
 159         scheduleAdjacentImpl(*DAG, ISU);
 160
 161   if (DAG->ExitSU.getInstr())
 162     // Try to fuse the instr in the ExitSU with one in its predecessors.
 163     scheduleAdjacentImpl(*DAG, DAG->ExitSU);
 164 }
 165
 166 /// Implement the fusion of instr pairs in the scheduling DAG,
 167 /// anchored at the instr in AnchorSU..
 168 bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU) {
 169   const MachineInstr &AnchorMI = *AnchorSU.getInstr();
 170   const TargetInstrInfo &TII = *DAG.TII;
 171   const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
 172
 173   // Check if the anchor instr may be fused.
 174   if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
 175     return false;
 176
 177   // Explorer for fusion candidates among the dependencies of the anchor instr.
 178   for (SDep &Dep : AnchorSU.Preds) {
 179     // Ignore dependencies other than data or strong ordering.
 180     if (Dep.isWeak() || isHazard(Dep))
 181       continue;
 182
 183     SUnit &DepSU = *Dep.getSUnit();
 184     if (DepSU.isBoundaryNode())
 185       continue;
 186
 187     // Only chain two instructions together at most.
 188     const MachineInstr *DepMI = DepSU.getInstr();
 189     if (!hasLessThanNumFused(DepSU, 2) ||
 190         !shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
 191       continue;
 192
 193     if (fuseInstructionPair(DAG, DepSU, AnchorSU))
 194       return true;
 195   }
 196
 197   return false;
 198 }
 199
 200 std::unique_ptr<ScheduleDAGMutation>
 201 llvm::createMacroFusionDAGMutation(
 202      ShouldSchedulePredTy shouldScheduleAdjacent) {
 203   if(EnableMacroFusion)
 204     return std::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
 205   return nullptr;
 206 }
 207
 208 std::unique_ptr<ScheduleDAGMutation>
 209 llvm::createBranchMacroFusionDAGMutation(
 210      ShouldSchedulePredTy shouldScheduleAdjacent) {
 211   if(EnableMacroFusion)
 212     return std::make_unique<MacroFusion>(shouldScheduleAdjacent, false);
 213   return nullptr;
 214 }