xref: /llvm-project/llvm/lib/CodeGen/MacroFusion.cpp (revision 5f746c8e2704d311cc01c8edb5460a0cb19c74e2)
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/Statistic.h"
17 #include "llvm/Support/CommandLine.h"
18 #include "llvm/Target/TargetInstrInfo.h"
19 
20 #define DEBUG_TYPE "misched"
21 
22 STATISTIC(NumFused, "Number of instr pairs fused");
23 
24 using namespace llvm;
25 
26 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
27   cl::desc("Enable scheduling for macro fusion."), cl::init(true));
28 
29 namespace {
30 
31 static void fuseInstructionPair(ScheduleDAGMI &DAG, SUnit &FirstSU,
32                                 SUnit &SecondSU) {
33   // Create a single weak edge between the adjacent instrs. The only effect is
34   // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
35   DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster));
36 
37   // Adjust the latency between the anchor instr and its
38   // predecessors.
39   for (SDep &IDep : SecondSU.Preds)
40     if (IDep.getSUnit() == &FirstSU)
41       IDep.setLatency(0);
42 
43   // Adjust the latency between the dependent instr and its
44   // predecessors.
45   for (SDep &IDep : FirstSU.Succs)
46     if (IDep.getSUnit() == &SecondSU)
47       IDep.setLatency(0);
48 
49   DEBUG(dbgs() << DAG.MF.getName() << "(): Macro fuse ";
50         FirstSU.print(dbgs(), &DAG); dbgs() << " - ";
51         SecondSU.print(dbgs(), &DAG); dbgs() << " /  ";
52         dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - " <<
53                   DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n'; );
54 
55   if (&SecondSU != &DAG.ExitSU)
56     // Make instructions dependent on FirstSU also dependent on SecondSU to
57     // prevent them from being scheduled between FirstSU and and SecondSU.
58     for (const SDep &SI : FirstSU.Succs) {
59       if (SI.getSUnit() == &SecondSU)
60         continue;
61       DEBUG(dbgs() << "  Copy Succ ";
62             SI.getSUnit()->print(dbgs(), &DAG); dbgs() << '\n';);
63       DAG.addEdge(SI.getSUnit(), SDep(&SecondSU, SDep::Artificial));
64     }
65 
66   ++NumFused;
67 }
68 
69 
70 /// \brief Post-process the DAG to create cluster edges between instrs that may
71 /// be fused by the processor into a single operation.
72 class MacroFusion : public ScheduleDAGMutation {
73   ShouldSchedulePredTy shouldScheduleAdjacent;
74   bool FuseBlock;
75   bool scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU);
76 
77 public:
78   MacroFusion(ShouldSchedulePredTy shouldScheduleAdjacent, bool FuseBlock)
79     : shouldScheduleAdjacent(shouldScheduleAdjacent), FuseBlock(FuseBlock) {}
80 
81   void apply(ScheduleDAGInstrs *DAGInstrs) override;
82 };
83 
84 void MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) {
85   ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
86 
87   if (FuseBlock)
88     // For each of the SUnits in the scheduling block, try to fuse the instr in
89     // it with one in its predecessors.
90     for (SUnit &ISU : DAG->SUnits)
91         scheduleAdjacentImpl(*DAG, ISU);
92 
93   if (DAG->ExitSU.getInstr())
94     // Try to fuse the instr in the ExitSU with one in its predecessors.
95     scheduleAdjacentImpl(*DAG, DAG->ExitSU);
96 }
97 
98 /// \brief Implement the fusion of instr pairs in the scheduling DAG,
99 /// anchored at the instr in AnchorSU..
100 bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGMI &DAG, SUnit &AnchorSU) {
101   const MachineInstr &AnchorMI = *AnchorSU.getInstr();
102   const TargetInstrInfo &TII = *DAG.TII;
103   const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
104 
105   // Check if the anchor instr may be fused.
106   if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
107     return false;
108 
109   // Explorer for fusion candidates among the dependencies of the anchor instr.
110   for (SDep &Dep : AnchorSU.Preds) {
111     // Ignore dependencies that don't enforce ordering.
112     if (Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output ||
113         Dep.isWeak())
114       continue;
115 
116     SUnit &DepSU = *Dep.getSUnit();
117     if (DepSU.isBoundaryNode())
118       continue;
119 
120     const MachineInstr *DepMI = DepSU.getInstr();
121     if (!shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
122       continue;
123 
124     fuseInstructionPair(DAG, DepSU, AnchorSU);
125     return true;
126   }
127 
128   return false;
129 }
130 
131 } // end anonymous namespace
132 
133 
134 namespace llvm {
135 
136 std::unique_ptr<ScheduleDAGMutation>
137 createMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent) {
138   if(EnableMacroFusion)
139     return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, true);
140   return nullptr;
141 }
142 
143 std::unique_ptr<ScheduleDAGMutation>
144 createBranchMacroFusionDAGMutation(ShouldSchedulePredTy shouldScheduleAdjacent) {
145   if(EnableMacroFusion)
146     return llvm::make_unique<MacroFusion>(shouldScheduleAdjacent, false);
147   return nullptr;
148 }
149 
150 } // end namespace llvm
151