LLVM 19.0.0git
MacroFusion.cpp
Go to the documentation of this file.
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
15#include "llvm/ADT/Statistic.h"
22#include "llvm/Support/Debug.h"
24
25#define DEBUG_TYPE "machine-scheduler"
26
27STATISTIC(NumFused, "Number of instr pairs fused");
28
29using namespace llvm;
30
32 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
33
34static bool isHazard(const SDep &Dep) {
35 return Dep.getKind() == SDep::Anti || Dep.getKind() == SDep::Output;
36}
37
38static SUnit *getPredClusterSU(const SUnit &SU) {
39 for (const SDep &SI : SU.Preds)
40 if (SI.isCluster())
41 return SI.getSUnit();
42
43 return nullptr;
44}
45
46bool llvm::hasLessThanNumFused(const SUnit &SU, unsigned FuseLimit) {
47 unsigned Num = 1;
48 const SUnit *CurrentSU = &SU;
49 while ((CurrentSU = getPredClusterSU(*CurrentSU)) && Num < FuseLimit) Num ++;
50 return Num < FuseLimit;
51}
52
54 SUnit &SecondSU) {
55 // Check that neither instr is already paired with another along the edge
56 // between them.
57 for (SDep &SI : FirstSU.Succs)
58 if (SI.isCluster())
59 return false;
60
61 for (SDep &SI : SecondSU.Preds)
62 if (SI.isCluster())
63 return false;
64 // Though the reachability checks above could be made more generic,
65 // perhaps as part of ScheduleDAGInstrs::addEdge(), since such edges are valid,
66 // the extra computation cost makes it less interesting in general cases.
67
68 // Create a single weak edge between the adjacent instrs. The only effect is
69 // to cause bottom-up scheduling to heavily prioritize the clustered instrs.
70 if (!DAG.addEdge(&SecondSU, SDep(&FirstSU, SDep::Cluster)))
71 return false;
72
73 // TODO - If we want to chain more than two instructions, we need to create
74 // artifical edges to make dependencies from the FirstSU also dependent
75 // on other chained instructions, and other chained instructions also
76 // dependent on the dependencies of the SecondSU, to prevent them from being
77 // scheduled into these chained instructions.
78 assert(hasLessThanNumFused(FirstSU, 2) &&
79 "Currently we only support chaining together two instructions");
80
81 // Adjust the latency between both instrs.
82 for (SDep &SI : FirstSU.Succs)
83 if (SI.getSUnit() == &SecondSU)
84 SI.setLatency(0);
85
86 for (SDep &SI : SecondSU.Preds)
87 if (SI.getSUnit() == &FirstSU)
88 SI.setLatency(0);
89
91 dbgs() << "Macro fuse: "; DAG.dumpNodeName(FirstSU); dbgs() << " - ";
92 DAG.dumpNodeName(SecondSU); dbgs() << " / ";
93 dbgs() << DAG.TII->getName(FirstSU.getInstr()->getOpcode()) << " - "
94 << DAG.TII->getName(SecondSU.getInstr()->getOpcode()) << '\n';);
95
96 // Make data dependencies from the FirstSU also dependent on the SecondSU to
97 // prevent them from being scheduled between the FirstSU and the SecondSU.
98 if (&SecondSU != &DAG.ExitSU)
99 for (const SDep &SI : FirstSU.Succs) {
100 SUnit *SU = SI.getSUnit();
101 if (SI.isWeak() || isHazard(SI) ||
102 SU == &DAG.ExitSU || SU == &SecondSU || SU->isPred(&SecondSU))
103 continue;
104 LLVM_DEBUG(dbgs() << " Bind "; DAG.dumpNodeName(SecondSU);
105 dbgs() << " - "; DAG.dumpNodeName(*SU); dbgs() << '\n';);
106 DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
107 }
108
109 // Make the FirstSU also dependent on the dependencies of the SecondSU to
110 // prevent them from being scheduled between the FirstSU and the SecondSU.
111 if (&FirstSU != &DAG.EntrySU) {
112 for (const SDep &SI : SecondSU.Preds) {
113 SUnit *SU = SI.getSUnit();
114 if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
115 continue;
116 LLVM_DEBUG(dbgs() << " Bind "; DAG.dumpNodeName(*SU); dbgs() << " - ";
117 DAG.dumpNodeName(FirstSU); dbgs() << '\n';);
118 DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
119 }
120 // ExitSU comes last by design, which acts like an implicit dependency
121 // between ExitSU and any bottom root in the graph. We should transfer
122 // this to FirstSU as well.
123 if (&SecondSU == &DAG.ExitSU) {
124 for (SUnit &SU : DAG.SUnits) {
125 if (SU.Succs.empty())
126 DAG.addEdge(&FirstSU, SDep(&SU, SDep::Artificial));
127 }
128 }
129 }
130
131 ++NumFused;
132 return true;
133}
134
135namespace {
136
137/// Post-process the DAG to create cluster edges between instrs that may
138/// be fused by the processor into a single operation.
139class MacroFusion : public ScheduleDAGMutation {
140 std::vector<MacroFusionPredTy> Predicates;
141 bool FuseBlock;
142 bool scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU);
143
144public:
145 MacroFusion(ArrayRef<MacroFusionPredTy> Predicates, bool FuseBlock)
146 : Predicates(Predicates.begin(), Predicates.end()), FuseBlock(FuseBlock) {
147 }
148
149 void apply(ScheduleDAGInstrs *DAGInstrs) override;
150
152 const TargetSubtargetInfo &STI,
153 const MachineInstr *FirstMI,
154 const MachineInstr &SecondMI);
155};
156
157} // end anonymous namespace
158
159bool MacroFusion::shouldScheduleAdjacent(const TargetInstrInfo &TII,
160 const TargetSubtargetInfo &STI,
161 const MachineInstr *FirstMI,
162 const MachineInstr &SecondMI) {
163 return llvm::any_of(Predicates, [&](MacroFusionPredTy Predicate) {
164 return Predicate(TII, STI, FirstMI, SecondMI);
165 });
166}
167
168void MacroFusion::apply(ScheduleDAGInstrs *DAG) {
169 if (FuseBlock)
170 // For each of the SUnits in the scheduling block, try to fuse the instr in
171 // it with one in its predecessors.
172 for (SUnit &ISU : DAG->SUnits)
173 scheduleAdjacentImpl(*DAG, ISU);
174
175 if (DAG->ExitSU.getInstr())
176 // Try to fuse the instr in the ExitSU with one in its predecessors.
177 scheduleAdjacentImpl(*DAG, DAG->ExitSU);
178}
179
180/// Implement the fusion of instr pairs in the scheduling DAG,
181/// anchored at the instr in AnchorSU..
182bool MacroFusion::scheduleAdjacentImpl(ScheduleDAGInstrs &DAG, SUnit &AnchorSU) {
183 const MachineInstr &AnchorMI = *AnchorSU.getInstr();
184 const TargetInstrInfo &TII = *DAG.TII;
185 const TargetSubtargetInfo &ST = DAG.MF.getSubtarget();
186
187 // Check if the anchor instr may be fused.
188 if (!shouldScheduleAdjacent(TII, ST, nullptr, AnchorMI))
189 return false;
190
191 // Explorer for fusion candidates among the dependencies of the anchor instr.
192 for (SDep &Dep : AnchorSU.Preds) {
193 // Ignore dependencies other than data or strong ordering.
194 if (Dep.isWeak() || isHazard(Dep))
195 continue;
196
197 SUnit &DepSU = *Dep.getSUnit();
198 if (DepSU.isBoundaryNode())
199 continue;
200
201 // Only chain two instructions together at most.
202 const MachineInstr *DepMI = DepSU.getInstr();
203 if (!hasLessThanNumFused(DepSU, 2) ||
204 !shouldScheduleAdjacent(TII, ST, DepMI, AnchorMI))
205 continue;
206
207 if (fuseInstructionPair(DAG, DepSU, AnchorSU))
208 return true;
209 }
210
211 return false;
212}
213
214std::unique_ptr<ScheduleDAGMutation>
216 bool BranchOnly) {
218 return std::make_unique<MacroFusion>(Predicates, !BranchOnly);
219 return nullptr;
220}
static bool shouldScheduleAdjacent(const TargetInstrInfo &TII, const TargetSubtargetInfo &TSI, const MachineInstr *FirstMI, const MachineInstr &SecondMI)
Check if the instr pair, FirstMI and SecondMI, should be fused together.
#define LLVM_DEBUG(X)
Definition: Debug.h:101
const HexagonInstrInfo * TII
static cl::opt< bool > EnableMacroFusion("misched-fusion", cl::Hidden, cl::desc("Enable scheduling for macro fusion."), cl::init(true))
static SUnit * getPredClusterSU(const SUnit &SU)
Definition: MacroFusion.cpp:38
static bool isHazard(const SDep &Dep)
Definition: MacroFusion.cpp:34
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the 'Statistic' class, which is designed to be an easy way to expose various metric...
#define STATISTIC(VARNAME, DESC)
Definition: Statistic.h:167
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
StringRef getName(unsigned Opcode) const
Returns the name for the instructions with the given opcode.
Definition: MCInstrInfo.h:70
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
Representation of each machine instruction.
Definition: MachineInstr.h:69
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:544
Scheduling dependency.
Definition: ScheduleDAG.h:49
SUnit * getSUnit() const
Definition: ScheduleDAG.h:480
Kind getKind() const
Returns an enum value representing the kind of the dependence.
Definition: ScheduleDAG.h:486
@ Output
A register output-dependence (aka WAW).
Definition: ScheduleDAG.h:55
@ Anti
A register anti-dependence (aka WAR).
Definition: ScheduleDAG.h:54
bool isWeak() const
Tests if this a weak dependence.
Definition: ScheduleDAG.h:194
@ Cluster
Weak DAG edge linking a chain of clustered instrs.
Definition: ScheduleDAG.h:74
@ Artificial
Arbitrary strong DAG edge (no real dependence).
Definition: ScheduleDAG.h:72
Scheduling unit. This is a node in the scheduling DAG.
Definition: ScheduleDAG.h:242
bool isSucc(const SUnit *N) const
Tests if node N is a successor of this node.
Definition: ScheduleDAG.h:439
bool isPred(const SUnit *N) const
Tests if node N is a predecessor of this node.
Definition: ScheduleDAG.h:431
bool isBoundaryNode() const
Boundary nodes are placeholders for the boundary of the scheduling region.
Definition: ScheduleDAG.h:344
SmallVector< SDep, 4 > Succs
All sunit successors.
Definition: ScheduleDAG.h:257
SmallVector< SDep, 4 > Preds
All sunit predecessors.
Definition: ScheduleDAG.h:256
MachineInstr * getInstr() const
Returns the representative MachineInstr for this SUnit.
Definition: ScheduleDAG.h:373
A ScheduleDAG for scheduling lists of MachineInstr.
bool addEdge(SUnit *SuccSU, const SDep &PredDep)
Add a DAG edge to the given SU with the given predecessor dependence data.
Mutate the DAG as a postpass after normal DAG building.
virtual void apply(ScheduleDAGInstrs *DAG)=0
const TargetInstrInfo * TII
Target instruction information.
Definition: ScheduleDAG.h:557
std::vector< SUnit > SUnits
The scheduling units.
Definition: ScheduleDAG.h:561
SUnit EntrySU
Special node for the region entry.
Definition: ScheduleDAG.h:562
MachineFunction & MF
Machine function.
Definition: ScheduleDAG.h:559
void dumpNodeName(const SUnit &SU) const
SUnit ExitSU
Special node for the region exit.
Definition: ScheduleDAG.h:563
TargetInstrInfo - Interface to description of machine instruction set.
TargetSubtargetInfo - Generic base class for all target subtargets.
Predicate
Predicate - These are "(BI << 5) | BO" for various predicates.
Definition: PPCPredicates.h:26
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:450
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
std::unique_ptr< ScheduleDAGMutation > createMacroFusionDAGMutation(ArrayRef< MacroFusionPredTy > Predicates, bool BranchOnly=false)
Create a DAG scheduling mutation to pair instructions back to back for instructions that benefit acco...
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1738
bool fuseInstructionPair(ScheduleDAGInstrs &DAG, SUnit &FirstSU, SUnit &SecondSU)
Create an artificial edge between FirstSU and SecondSU.
Definition: MacroFusion.cpp:53
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
bool(*)(const TargetInstrInfo &TII, const TargetSubtargetInfo &STI, const MachineInstr *FirstMI, const MachineInstr &SecondMI) MacroFusionPredTy
Check if the instr pair, FirstMI and SecondMI, should be fused together.
Definition: MacroFusion.h:35
static bool shouldScheduleAdjacent(const TargetInstrInfo &TII, const TargetSubtargetInfo &TSI, const MachineInstr *FirstMI, const MachineInstr &SecondMI)
Check if the instr pair, FirstMI and SecondMI, should be fused together.
bool hasLessThanNumFused(const SUnit &SU, unsigned FuseLimit)
Checks if the number of cluster edges between SU and its predecessors is less than FuseLimit.
Definition: MacroFusion.cpp:46