1 //===- MacroFusion.cpp - Macro Fusion -------------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file This file contains the implementation of the DAG scheduling mutation
11 /// to pair instructions back to back.
12 //
13 //===----------------------------------------------------------------------===//
14 
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"
26 
27 #define DEBUG_TYPE "machine-scheduler"
28 
29 STATISTIC(NumFused, "Number of instr pairs fused");
30 
31 using namespace llvm;
32 
33 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
34   cl::desc("Enable scheduling for macro fusion."), cl::init(true));
35 
36 static bool isHazard(const SDep &Dep) {
37   return Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output;
38 }
39 
40 static bool fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
41                                 SUnit &SecondSU) {
42   // Check that neither instr is already paired with another along the edge
43   // between them.
44   for (SDep &SI : FirstSU.Succs)
45     if (SI.isCluster())
46       return false;
47 
48   for (SDep &SI : SecondSU.Preds)
49     if (SI.isCluster())
50       return false;
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.
54 
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)))
58     return false;
59 
60   // Adjust the latency between both instrs.
61   for (SDep &SI : FirstSU.Succs)
62     if (SI.getSUnit() == &SecondSU)
63       SI.setLatency(0);
64 
65   for (SDep &SI : SecondSU.Preds)
66     if (SI.getSUnit() == &FirstSU)
67       SI.setLatency(0);
68 
69   LLVM_DEBUG(
70       dbgs() << "Macro fuse: "; DAG.dumpNodeName(FirstSU); dbgs() << " - ";
71       DAG.dumpNodeName(SecondSU); dbgs() << " /  ";
72       dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - "
73              << DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n';);
74 
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))
82         continue;
83       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(SecondSU);
84                  dbgs() << " - "; DAG.dumpNodeName(*SU); dbgs() << '\n';);
85       DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
86     }
87 
88   // Make the FirstSU also dependent on the dependencies of the SecondSU to
89   // prevent them from being scheduled between the FirstSU and the SecondSU.
90   if (&FirstSU != &DAG.EntrySU) {
91     for (const SDep &SI : SecondSU.Preds) {
92       SUnit *SU = SI.getSUnit();
93       if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
94         continue;
95       LLVM_DEBUG(dbgs() << "  Bind "; DAG.dumpNodeName(*SU); dbgs() << " - ";
96                  DAG.dumpNodeName(FirstSU); dbgs() << '\n';);
97       DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
98     }
99     // ExitSU comes last by design, which acts like an implicit dependency
100     // between ExitSU and any bottom root in the graph. We should transfer
101     // this to FirstSU as well.
102     if (&SecondSU == &DAG.ExitSU) {
103       for (SUnit &SU : DAG.SUnits) {
104         if (SU.Succs.empty())
105           DAG.addEdge(&FirstSU, SDep(&SU, SDep::Artificial));
106       }
107     }
108   }
109 
110   ++NumFused;
111   return true;
112 }
113 
114 namespace {
115 
116 /// Post-process the DAG to create cluster edges between instrs that may
117 /// be fused by the processor into a single operation.
118 class MacroFusion : public ScheduleDAGMutation {
119   ShouldSchedulePredTy shouldScheduleAdjacent;
120   bool FuseBlock;
121   bool scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU);
122 
123 public:
124   MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
125     : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
126 
127   void apply(ScheduleDAGInstrs *DAGInstrs) override;
128 };
129 
130 } // end anonymous namespace
131 
132 void MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) {
133   ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
134 
135   if (FuseBlock)
136     // For each of the SUnits in the scheduling block, try to fuse the instr in
137     // it with one in its predecessors.
138     for (SUnit &ISU : DAG->SUnits)
139         scheduleAdjacentImpl(*DAG, ISU);
140 
141   if (DAG->ExitSU.getInstr())
142     // Try to fuse the instr in the ExitSU with one in its predecessors.
143     scheduleAdjacentImpl(*DAG, DAG->ExitSU);
144 }
145 
146 /// Implement the fusion of instr pairs in the scheduling DAG,
147 /// anchored at the instr in AnchorSU..
148 bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
149   const MachineInstr &AnchorMI = *AnchorSU.getInstr();
150   const TargetInstrInfo &TII = *DAG.TII;
151   const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
152 
153   // Check if the anchor instr may be fused.
154   if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
155     return false;
156 
157   // Explorer for fusion candidates among the dependencies of the anchor instr.
158   for (SDep &Dep : AnchorSU.Preds) {
159     // Ignore dependencies other than data or strong ordering.
160     if (Dep.isWeak() || isHazard(Dep))
161       continue;
162 
163     SUnit &DepSU = *Dep.getSUnit();
164     if (DepSU.isBoundaryNode())
165       continue;
166 
167     const MachineInstr *DepMI = DepSU.getInstr();
168     if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
169       continue;
170 
171     if (fuseInstructionPair(DAG, DepSU, AnchorSU))
172       return true;
173   }
174 
175   return false;
176 }
177 
178 std::unique_ptr<ScheduleDAGMutation>
179 llvm::createMacroFusionDAGMutation(
180      ShouldSchedulePredTy shouldScheduleAdjacent) {
181   if(EnableMacroFusion)
182     return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
183   return nullptr;
184 }
185 
186 std::unique_ptr<ScheduleDAGMutation>
187 llvm::createBranchMacroFusionDAGMutation(
188      ShouldSchedulePredTy shouldScheduleAdjacent) {
189   if(EnableMacroFusion)
190     return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, false);
191   return nullptr;
192 }
193