1 | //===- MCSchedule.cpp - Scheduling ------------------------------*- C++ -*-===// |
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 | // This file defines the default scheduling model. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/MC/MCSchedule.h" |
14 | #include "llvm/MC/MCInst.h" |
15 | #include "llvm/MC/MCInstrDesc.h" |
16 | #include "llvm/MC/MCInstrInfo.h" |
17 | #include "llvm/MC/MCSubtargetInfo.h" |
18 | #include <optional> |
19 | #include <type_traits> |
20 | |
21 | using namespace llvm; |
22 | |
23 | static_assert(std::is_trivial_v<MCSchedModel>, |
24 | "MCSchedModel is required to be a trivial type" ); |
25 | const MCSchedModel MCSchedModel::Default = {.IssueWidth: DefaultIssueWidth, |
26 | .MicroOpBufferSize: DefaultMicroOpBufferSize, |
27 | .LoopMicroOpBufferSize: DefaultLoopMicroOpBufferSize, |
28 | .LoadLatency: DefaultLoadLatency, |
29 | .HighLatency: DefaultHighLatency, |
30 | .MispredictPenalty: DefaultMispredictPenalty, |
31 | .PostRAScheduler: false, |
32 | .CompleteModel: true, |
33 | /*EnableIntervals=*/false, |
34 | .ProcID: 0, |
35 | .ProcResourceTable: nullptr, |
36 | .SchedClassTable: nullptr, |
37 | .NumProcResourceKinds: 0, |
38 | .NumSchedClasses: 0, |
39 | .InstrItineraries: nullptr, |
40 | .ExtraProcessorInfo: nullptr}; |
41 | |
42 | int MCSchedModel::computeInstrLatency(const MCSubtargetInfo &STI, |
43 | const MCSchedClassDesc &SCDesc) { |
44 | int Latency = 0; |
45 | for (unsigned DefIdx = 0, DefEnd = SCDesc.NumWriteLatencyEntries; |
46 | DefIdx != DefEnd; ++DefIdx) { |
47 | // Lookup the definition's write latency in SubtargetInfo. |
48 | const MCWriteLatencyEntry *WLEntry = |
49 | STI.getWriteLatencyEntry(SC: &SCDesc, DefIdx); |
50 | // Early exit if we found an invalid latency. |
51 | if (WLEntry->Cycles < 0) |
52 | return WLEntry->Cycles; |
53 | Latency = std::max(a: Latency, b: static_cast<int>(WLEntry->Cycles)); |
54 | } |
55 | return Latency; |
56 | } |
57 | |
58 | int MCSchedModel::computeInstrLatency(const MCSubtargetInfo &STI, |
59 | unsigned SchedClass) const { |
60 | const MCSchedClassDesc &SCDesc = *getSchedClassDesc(SchedClassIdx: SchedClass); |
61 | if (!SCDesc.isValid()) |
62 | return 0; |
63 | if (!SCDesc.isVariant()) |
64 | return MCSchedModel::computeInstrLatency(STI, SCDesc); |
65 | |
66 | llvm_unreachable("unsupported variant scheduling class" ); |
67 | } |
68 | |
69 | int MCSchedModel::computeInstrLatency(const MCSubtargetInfo &STI, |
70 | const MCInstrInfo &MCII, |
71 | const MCInst &Inst) const { |
72 | unsigned SchedClass = MCII.get(Opcode: Inst.getOpcode()).getSchedClass(); |
73 | const MCSchedClassDesc *SCDesc = getSchedClassDesc(SchedClassIdx: SchedClass); |
74 | if (!SCDesc->isValid()) |
75 | return 0; |
76 | |
77 | unsigned CPUID = getProcessorID(); |
78 | while (SCDesc->isVariant()) { |
79 | SchedClass = STI.resolveVariantSchedClass(SchedClass, MI: &Inst, MCII: &MCII, CPUID); |
80 | SCDesc = getSchedClassDesc(SchedClassIdx: SchedClass); |
81 | } |
82 | |
83 | if (SchedClass) |
84 | return MCSchedModel::computeInstrLatency(STI, SCDesc: *SCDesc); |
85 | |
86 | llvm_unreachable("unsupported variant scheduling class" ); |
87 | } |
88 | |
89 | double |
90 | MCSchedModel::getReciprocalThroughput(const MCSubtargetInfo &STI, |
91 | const MCSchedClassDesc &SCDesc) { |
92 | std::optional<double> Throughput; |
93 | const MCSchedModel &SM = STI.getSchedModel(); |
94 | const MCWriteProcResEntry *I = STI.getWriteProcResBegin(SC: &SCDesc); |
95 | const MCWriteProcResEntry *E = STI.getWriteProcResEnd(SC: &SCDesc); |
96 | for (; I != E; ++I) { |
97 | if (!I->ReleaseAtCycle) |
98 | continue; |
99 | unsigned NumUnits = SM.getProcResource(ProcResourceIdx: I->ProcResourceIdx)->NumUnits; |
100 | double Temp = NumUnits * 1.0 / I->ReleaseAtCycle; |
101 | Throughput = Throughput ? std::min(a: *Throughput, b: Temp) : Temp; |
102 | } |
103 | if (Throughput) |
104 | return 1.0 / *Throughput; |
105 | |
106 | // If no throughput value was calculated, assume that we can execute at the |
107 | // maximum issue width scaled by number of micro-ops for the schedule class. |
108 | return ((double)SCDesc.NumMicroOps) / SM.IssueWidth; |
109 | } |
110 | |
111 | double |
112 | MCSchedModel::getReciprocalThroughput(const MCSubtargetInfo &STI, |
113 | const MCInstrInfo &MCII, |
114 | const MCInst &Inst) const { |
115 | unsigned SchedClass = MCII.get(Opcode: Inst.getOpcode()).getSchedClass(); |
116 | const MCSchedClassDesc *SCDesc = getSchedClassDesc(SchedClassIdx: SchedClass); |
117 | |
118 | // If there's no valid class, assume that the instruction executes/completes |
119 | // at the maximum issue width. |
120 | if (!SCDesc->isValid()) |
121 | return 1.0 / IssueWidth; |
122 | |
123 | unsigned CPUID = getProcessorID(); |
124 | while (SCDesc->isVariant()) { |
125 | SchedClass = STI.resolveVariantSchedClass(SchedClass, MI: &Inst, MCII: &MCII, CPUID); |
126 | SCDesc = getSchedClassDesc(SchedClassIdx: SchedClass); |
127 | } |
128 | |
129 | if (SchedClass) |
130 | return MCSchedModel::getReciprocalThroughput(STI, SCDesc: *SCDesc); |
131 | |
132 | llvm_unreachable("unsupported variant scheduling class" ); |
133 | } |
134 | |
135 | double |
136 | MCSchedModel::getReciprocalThroughput(unsigned SchedClass, |
137 | const InstrItineraryData &IID) { |
138 | std::optional<double> Throughput; |
139 | const InstrStage *I = IID.beginStage(ItinClassIndx: SchedClass); |
140 | const InstrStage *E = IID.endStage(ItinClassIndx: SchedClass); |
141 | for (; I != E; ++I) { |
142 | if (!I->getCycles()) |
143 | continue; |
144 | double Temp = llvm::popcount(Value: I->getUnits()) * 1.0 / I->getCycles(); |
145 | Throughput = Throughput ? std::min(a: *Throughput, b: Temp) : Temp; |
146 | } |
147 | if (Throughput) |
148 | return 1.0 / *Throughput; |
149 | |
150 | // If there are no execution resources specified for this class, then assume |
151 | // that it can execute at the maximum default issue width. |
152 | return 1.0 / DefaultIssueWidth; |
153 | } |
154 | |
155 | unsigned |
156 | MCSchedModel::getForwardingDelayCycles(ArrayRef<MCReadAdvanceEntry> Entries, |
157 | unsigned WriteResourceID) { |
158 | if (Entries.empty()) |
159 | return 0; |
160 | |
161 | int DelayCycles = 0; |
162 | for (const MCReadAdvanceEntry &E : Entries) { |
163 | if (E.WriteResourceID != WriteResourceID) |
164 | continue; |
165 | DelayCycles = std::min(a: DelayCycles, b: E.Cycles); |
166 | } |
167 | |
168 | return std::abs(x: DelayCycles); |
169 | } |
170 | |