1 //===- MacroFusion.cpp - Macro Fusion -------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// \file This file contains the implementation of the DAG scheduling mutation
11 /// to pair instructions back to back.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/MacroFusion.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/CodeGen/MachineInstr.h"
19 #include "llvm/CodeGen/MachineScheduler.h"
20 #include "llvm/CodeGen/ScheduleDAG.h"
21 #include "llvm/CodeGen/ScheduleDAGMutation.h"
22 #include "llvm/CodeGen/TargetInstrInfo.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
27 #define DEBUG_TYPE "machine-scheduler"
29 STATISTIC(NumFused, "Number of instr pairs fused");
33 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
34 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
36 static bool isHazard(const SDep &Dep) {
37 return Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output;
40 static bool fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
42 // Check that neither instr is already paired with another along the edge
44 for (SDep &SI : FirstSU.Succs)
48 for (SDep &SI : SecondSU.Preds)
51 // Though the reachability checks above could be made more generic,
52 // perhaps as part of ScheduleDAGMI::addEdge(), since such edges are valid,
53 // the extra computation cost makes it less interesting in general cases.
55 // Create a single weak edge between the adjacent instrs. The only effect is
56 // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
57 if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
60 // Adjust the latency between both instrs.
61 for (SDep &SI : FirstSU.Succs)
62 if (SI.getSUnit() == &SecondSU)
65 for (SDep &SI : SecondSU.Preds)
66 if (SI.getSUnit() == &FirstSU)
69 DEBUG(dbgs() << "Macro fuse: ";
70 FirstSU.print(dbgs(), &DAG); dbgs() << " - ";
71 SecondSU.print(dbgs(), &DAG); dbgs() << " / ";
72 dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - " <<
73 DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n'; );
75 // Make data dependencies from the FirstSU also dependent on the SecondSU to
76 // prevent them from being scheduled between the FirstSU and the SecondSU.
77 if (&SecondSU != &DAG.ExitSU)
78 for (const SDep &SI : FirstSU.Succs) {
79 SUnit *SU = SI.getSUnit();
80 if (SI.isWeak() || isHazard(SI) ||
81 SU == &DAG.ExitSU || SU == &SecondSU || SU->isPred(&SecondSU))
83 DEBUG(dbgs() << " Bind ";
84 SecondSU.print(dbgs(), &DAG); dbgs() << " - ";
85 SU->print(dbgs(), &DAG); dbgs() << '\n';);
86 DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
89 // Make the FirstSU also dependent on the dependencies of the SecondSU to
90 // prevent them from being scheduled between the FirstSU and the SecondSU.
91 if (&FirstSU != &DAG.EntrySU)
92 for (const SDep &SI : SecondSU.Preds) {
93 SUnit *SU = SI.getSUnit();
94 if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
96 DEBUG(dbgs() << " Bind ";
97 SU->print(dbgs(), &DAG); dbgs() << " - ";
98 FirstSU.print(dbgs(), &DAG); dbgs() << '\n';);
99 DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
108 /// \brief Post-process the DAG to create cluster edges between instrs that may
109 /// be fused by the processor into a single operation.
110 class MacroFusion : public ScheduleDAGMutation {
111 ShouldSchedulePredTy shouldScheduleAdjacent;
113 bool scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU);
116 MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
117 : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
119 void apply(ScheduleDAGInstrs *DAGInstrs) override;
122 } // end anonymous namespace
124 void MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) {
125 ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
128 // For each of the SUnits in the scheduling block, try to fuse the instr in
129 // it with one in its predecessors.
130 for (SUnit &ISU : DAG->SUnits)
131 scheduleAdjacentImpl(*DAG, ISU);
133 if (DAG->ExitSU.getInstr())
134 // Try to fuse the instr in the ExitSU with one in its predecessors.
135 scheduleAdjacentImpl(*DAG, DAG->ExitSU);
138 /// \brief Implement the fusion of instr pairs in the scheduling DAG,
139 /// anchored at the instr in AnchorSU..
140 bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
141 const MachineInstr &AnchorMI = *AnchorSU.getInstr();
142 const TargetInstrInfo &TII = *DAG.TII;
143 const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
145 // Check if the anchor instr may be fused.
146 if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
149 // Explorer for fusion candidates among the dependencies of the anchor instr.
150 for (SDep &Dep : AnchorSU.Preds) {
151 // Ignore dependencies other than data or strong ordering.
152 if (Dep.isWeak() || isHazard(Dep))
155 SUnit &DepSU = *Dep.getSUnit();
156 if (DepSU.isBoundaryNode())
159 const MachineInstr *DepMI = DepSU.getInstr();
160 if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
163 if (fuseInstructionPair(DAG, DepSU, AnchorSU))
170 std::unique_ptr<ScheduleDAGMutation>
171 llvm::createMacroFusionDAGMutation(
172 ShouldSchedulePredTy shouldScheduleAdjacent) {
173 if(EnableMacroFusion)
174 return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
178 std::unique_ptr<ScheduleDAGMutation>
179 llvm::createBranchMacroFusionDAGMutation(
180 ShouldSchedulePredTy shouldScheduleAdjacent) {
181 if(EnableMacroFusion)
182 return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, false);