1 | //===- llvm/Target/TargetSchedule.cpp - Sched Machine Model ---------------===// |
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 implements a wrapper around MCSchedModel that allows the interface |
10 | // to benefit from information currently only available in TargetInstrInfo. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/CodeGen/TargetSchedule.h" |
15 | #include "llvm/CodeGen/MachineFunction.h" |
16 | #include "llvm/CodeGen/MachineInstr.h" |
17 | #include "llvm/CodeGen/MachineOperand.h" |
18 | #include "llvm/CodeGen/TargetInstrInfo.h" |
19 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
20 | #include "llvm/MC/MCInstrDesc.h" |
21 | #include "llvm/MC/MCInstrItineraries.h" |
22 | #include "llvm/MC/MCSchedule.h" |
23 | #include "llvm/Support/CommandLine.h" |
24 | #include "llvm/Support/ErrorHandling.h" |
25 | #include "llvm/Support/raw_ostream.h" |
26 | #include <algorithm> |
27 | #include <cassert> |
28 | #include <numeric> |
29 | |
30 | using namespace llvm; |
31 | |
32 | static cl::opt<bool> ForceEnableIntervals( |
33 | "sched-model-force-enable-intervals" , cl::Hidden, cl::init(Val: false), |
34 | cl::desc("Force the use of resource intervals in the schedule model" )); |
35 | |
36 | bool TargetSchedModel::hasInstrSchedModel() const { |
37 | return EnableSchedModel && SchedModel.hasInstrSchedModel(); |
38 | } |
39 | |
40 | bool TargetSchedModel::hasInstrItineraries() const { |
41 | return EnableSchedItins && !InstrItins.isEmpty(); |
42 | } |
43 | |
44 | void TargetSchedModel::init(const TargetSubtargetInfo *TSInfo, |
45 | bool EnableSModel, bool EnableSItins) { |
46 | STI = TSInfo; |
47 | SchedModel = TSInfo->getSchedModel(); |
48 | TII = TSInfo->getInstrInfo(); |
49 | STI->initInstrItins(InstrItins); |
50 | |
51 | EnableSchedModel = EnableSModel; |
52 | EnableSchedItins = EnableSItins; |
53 | |
54 | unsigned NumRes = SchedModel.getNumProcResourceKinds(); |
55 | ResourceFactors.resize(N: NumRes); |
56 | ResourceLCM = SchedModel.IssueWidth; |
57 | for (unsigned Idx = 0; Idx < NumRes; ++Idx) { |
58 | unsigned NumUnits = SchedModel.getProcResource(ProcResourceIdx: Idx)->NumUnits; |
59 | if (NumUnits > 0) |
60 | ResourceLCM = std::lcm(m: ResourceLCM, n: NumUnits); |
61 | } |
62 | MicroOpFactor = ResourceLCM / SchedModel.IssueWidth; |
63 | for (unsigned Idx = 0; Idx < NumRes; ++Idx) { |
64 | unsigned NumUnits = SchedModel.getProcResource(ProcResourceIdx: Idx)->NumUnits; |
65 | ResourceFactors[Idx] = NumUnits ? (ResourceLCM / NumUnits) : 0; |
66 | } |
67 | } |
68 | |
69 | /// Returns true only if instruction is specified as single issue. |
70 | bool TargetSchedModel::mustBeginGroup(const MachineInstr *MI, |
71 | const MCSchedClassDesc *SC) const { |
72 | if (hasInstrSchedModel()) { |
73 | if (!SC) |
74 | SC = resolveSchedClass(MI); |
75 | if (SC->isValid()) |
76 | return SC->BeginGroup; |
77 | } |
78 | return false; |
79 | } |
80 | |
81 | bool TargetSchedModel::mustEndGroup(const MachineInstr *MI, |
82 | const MCSchedClassDesc *SC) const { |
83 | if (hasInstrSchedModel()) { |
84 | if (!SC) |
85 | SC = resolveSchedClass(MI); |
86 | if (SC->isValid()) |
87 | return SC->EndGroup; |
88 | } |
89 | return false; |
90 | } |
91 | |
92 | unsigned TargetSchedModel::getNumMicroOps(const MachineInstr *MI, |
93 | const MCSchedClassDesc *SC) const { |
94 | if (hasInstrItineraries()) { |
95 | int UOps = InstrItins.getNumMicroOps(ItinClassIndx: MI->getDesc().getSchedClass()); |
96 | return (UOps >= 0) ? UOps : TII->getNumMicroOps(ItinData: &InstrItins, MI: *MI); |
97 | } |
98 | if (hasInstrSchedModel()) { |
99 | if (!SC) |
100 | SC = resolveSchedClass(MI); |
101 | if (SC->isValid()) |
102 | return SC->NumMicroOps; |
103 | } |
104 | return MI->isTransient() ? 0 : 1; |
105 | } |
106 | |
107 | // The machine model may explicitly specify an invalid latency, which |
108 | // effectively means infinite latency. Since users of the TargetSchedule API |
109 | // don't know how to handle this, we convert it to a very large latency that is |
110 | // easy to distinguish when debugging the DAG but won't induce overflow. |
111 | static unsigned capLatency(int Cycles) { |
112 | return Cycles >= 0 ? Cycles : 1000; |
113 | } |
114 | |
115 | /// Return the MCSchedClassDesc for this instruction. Some SchedClasses require |
116 | /// evaluation of predicates that depend on instruction operands or flags. |
117 | const MCSchedClassDesc *TargetSchedModel:: |
118 | resolveSchedClass(const MachineInstr *MI) const { |
119 | // Get the definition's scheduling class descriptor from this machine model. |
120 | unsigned SchedClass = MI->getDesc().getSchedClass(); |
121 | const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClassIdx: SchedClass); |
122 | if (!SCDesc->isValid()) |
123 | return SCDesc; |
124 | |
125 | #ifndef NDEBUG |
126 | unsigned NIter = 0; |
127 | #endif |
128 | while (SCDesc->isVariant()) { |
129 | assert(++NIter < 6 && "Variants are nested deeper than the magic number" ); |
130 | |
131 | SchedClass = STI->resolveSchedClass(SchedClass, MI, SchedModel: this); |
132 | SCDesc = SchedModel.getSchedClassDesc(SchedClassIdx: SchedClass); |
133 | } |
134 | return SCDesc; |
135 | } |
136 | |
137 | /// Find the def index of this operand. This index maps to the machine model and |
138 | /// is independent of use operands. Def operands may be reordered with uses or |
139 | /// merged with uses without affecting the def index (e.g. before/after |
140 | /// regalloc). However, an instruction's def operands must never be reordered |
141 | /// with respect to each other. |
142 | static unsigned findDefIdx(const MachineInstr *MI, unsigned DefOperIdx) { |
143 | unsigned DefIdx = 0; |
144 | for (unsigned i = 0; i != DefOperIdx; ++i) { |
145 | const MachineOperand &MO = MI->getOperand(i); |
146 | if (MO.isReg() && MO.isDef()) |
147 | ++DefIdx; |
148 | } |
149 | return DefIdx; |
150 | } |
151 | |
152 | /// Find the use index of this operand. This is independent of the instruction's |
153 | /// def operands. |
154 | /// |
155 | /// Note that uses are not determined by the operand's isUse property, which |
156 | /// is simply the inverse of isDef. Here we consider any readsReg operand to be |
157 | /// a "use". The machine model allows an operand to be both a Def and Use. |
158 | static unsigned findUseIdx(const MachineInstr *MI, unsigned UseOperIdx) { |
159 | unsigned UseIdx = 0; |
160 | for (unsigned i = 0; i != UseOperIdx; ++i) { |
161 | const MachineOperand &MO = MI->getOperand(i); |
162 | if (MO.isReg() && MO.readsReg() && !MO.isDef()) |
163 | ++UseIdx; |
164 | } |
165 | return UseIdx; |
166 | } |
167 | |
168 | // Top-level API for clients that know the operand indices. This doesn't need to |
169 | // return std::optional<unsigned>, as it always returns a valid latency. |
170 | unsigned TargetSchedModel::computeOperandLatency( |
171 | const MachineInstr *DefMI, unsigned DefOperIdx, |
172 | const MachineInstr *UseMI, unsigned UseOperIdx) const { |
173 | |
174 | const unsigned InstrLatency = computeInstrLatency(MI: DefMI); |
175 | const unsigned DefaultDefLatency = TII->defaultDefLatency(SchedModel, DefMI: *DefMI); |
176 | |
177 | if (!hasInstrSchedModel() && !hasInstrItineraries()) |
178 | return DefaultDefLatency; |
179 | |
180 | if (hasInstrItineraries()) { |
181 | std::optional<unsigned> OperLatency; |
182 | if (UseMI) { |
183 | OperLatency = TII->getOperandLatency(ItinData: &InstrItins, DefMI: *DefMI, DefIdx: DefOperIdx, |
184 | UseMI: *UseMI, UseIdx: UseOperIdx); |
185 | } |
186 | else { |
187 | unsigned DefClass = DefMI->getDesc().getSchedClass(); |
188 | OperLatency = InstrItins.getOperandCycle(ItinClassIndx: DefClass, OperandIdx: DefOperIdx); |
189 | } |
190 | |
191 | // Expected latency is the max of InstrLatency and DefaultDefLatency, if we |
192 | // didn't find an operand latency. |
193 | return OperLatency ? *OperLatency |
194 | : std::max(a: InstrLatency, b: DefaultDefLatency); |
195 | } |
196 | |
197 | // hasInstrSchedModel() |
198 | const MCSchedClassDesc *SCDesc = resolveSchedClass(MI: DefMI); |
199 | unsigned DefIdx = findDefIdx(MI: DefMI, DefOperIdx); |
200 | if (DefIdx < SCDesc->NumWriteLatencyEntries) { |
201 | // Lookup the definition's write latency in SubtargetInfo. |
202 | const MCWriteLatencyEntry *WLEntry = |
203 | STI->getWriteLatencyEntry(SC: SCDesc, DefIdx); |
204 | unsigned WriteID = WLEntry->WriteResourceID; |
205 | unsigned Latency = capLatency(Cycles: WLEntry->Cycles); |
206 | if (!UseMI) |
207 | return Latency; |
208 | |
209 | // Lookup the use's latency adjustment in SubtargetInfo. |
210 | const MCSchedClassDesc *UseDesc = resolveSchedClass(MI: UseMI); |
211 | if (UseDesc->NumReadAdvanceEntries == 0) |
212 | return Latency; |
213 | unsigned UseIdx = findUseIdx(MI: UseMI, UseOperIdx); |
214 | int Advance = STI->getReadAdvanceCycles(SC: UseDesc, UseIdx, WriteResID: WriteID); |
215 | if (Advance > 0 && (unsigned)Advance > Latency) // unsigned wrap |
216 | return 0; |
217 | return Latency - Advance; |
218 | } |
219 | // If DefIdx does not exist in the model (e.g. implicit defs), then return |
220 | // unit latency (defaultDefLatency may be too conservative). |
221 | #ifndef NDEBUG |
222 | if (SCDesc->isValid() && !DefMI->getOperand(DefOperIdx).isImplicit() && |
223 | !DefMI->getDesc().operands()[DefOperIdx].isOptionalDef() && |
224 | SchedModel.isComplete()) { |
225 | errs() << "DefIdx " << DefIdx << " exceeds machine model writes for " |
226 | << *DefMI << " (Try with MCSchedModel.CompleteModel set to false)" ; |
227 | llvm_unreachable("incomplete machine model" ); |
228 | } |
229 | #endif |
230 | // FIXME: Automatically giving all implicit defs defaultDefLatency is |
231 | // undesirable. We should only do it for defs that are known to the MC |
232 | // desc like flags. Truly implicit defs should get 1 cycle latency. |
233 | return DefMI->isTransient() ? 0 : DefaultDefLatency; |
234 | } |
235 | |
236 | unsigned |
237 | TargetSchedModel::computeInstrLatency(const MCSchedClassDesc &SCDesc) const { |
238 | return capLatency(Cycles: MCSchedModel::computeInstrLatency(STI: *STI, SCDesc)); |
239 | } |
240 | |
241 | unsigned TargetSchedModel::computeInstrLatency(unsigned Opcode) const { |
242 | assert(hasInstrSchedModel() && "Only call this function with a SchedModel" ); |
243 | unsigned SCIdx = TII->get(Opcode).getSchedClass(); |
244 | return capLatency(Cycles: SchedModel.computeInstrLatency(STI: *STI, SClass: SCIdx)); |
245 | } |
246 | |
247 | unsigned TargetSchedModel::computeInstrLatency(const MCInst &Inst) const { |
248 | if (hasInstrSchedModel()) |
249 | return capLatency(Cycles: SchedModel.computeInstrLatency(STI: *STI, MCII: *TII, Inst)); |
250 | return computeInstrLatency(Opcode: Inst.getOpcode()); |
251 | } |
252 | |
253 | unsigned |
254 | TargetSchedModel::computeInstrLatency(const MachineInstr *MI, |
255 | bool UseDefaultDefLatency) const { |
256 | // For the itinerary model, fall back to the old subtarget hook. |
257 | // Allow subtargets to compute Bundle latencies outside the machine model. |
258 | if (hasInstrItineraries() || MI->isBundle() || |
259 | (!hasInstrSchedModel() && !UseDefaultDefLatency)) |
260 | return TII->getInstrLatency(ItinData: &InstrItins, MI: *MI); |
261 | |
262 | if (hasInstrSchedModel()) { |
263 | const MCSchedClassDesc *SCDesc = resolveSchedClass(MI); |
264 | if (SCDesc->isValid()) |
265 | return computeInstrLatency(SCDesc: *SCDesc); |
266 | } |
267 | return TII->defaultDefLatency(SchedModel, DefMI: *MI); |
268 | } |
269 | |
270 | unsigned TargetSchedModel:: |
271 | computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx, |
272 | const MachineInstr *DepMI) const { |
273 | if (!SchedModel.isOutOfOrder()) |
274 | return 1; |
275 | |
276 | // Out-of-order processor can dispatch WAW dependencies in the same cycle. |
277 | |
278 | // Treat predication as a data dependency for out-of-order cpus. In-order |
279 | // cpus do not need to treat predicated writes specially. |
280 | // |
281 | // TODO: The following hack exists because predication passes do not |
282 | // correctly append imp-use operands, and readsReg() strangely returns false |
283 | // for predicated defs. |
284 | Register Reg = DefMI->getOperand(i: DefOperIdx).getReg(); |
285 | const MachineFunction &MF = *DefMI->getMF(); |
286 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
287 | if (!DepMI->readsRegister(Reg, TRI) && TII->isPredicated(MI: *DepMI)) |
288 | return computeInstrLatency(MI: DefMI); |
289 | |
290 | // If we have a per operand scheduling model, check if this def is writing |
291 | // an unbuffered resource. If so, it treated like an in-order cpu. |
292 | if (hasInstrSchedModel()) { |
293 | const MCSchedClassDesc *SCDesc = resolveSchedClass(MI: DefMI); |
294 | if (SCDesc->isValid()) { |
295 | for (const MCWriteProcResEntry *PRI = STI->getWriteProcResBegin(SC: SCDesc), |
296 | *PRE = STI->getWriteProcResEnd(SC: SCDesc); PRI != PRE; ++PRI) { |
297 | if (!SchedModel.getProcResource(ProcResourceIdx: PRI->ProcResourceIdx)->BufferSize) |
298 | return 1; |
299 | } |
300 | } |
301 | } |
302 | return 0; |
303 | } |
304 | |
305 | double |
306 | TargetSchedModel::computeReciprocalThroughput(const MachineInstr *MI) const { |
307 | if (hasInstrItineraries()) { |
308 | unsigned SchedClass = MI->getDesc().getSchedClass(); |
309 | return MCSchedModel::getReciprocalThroughput(SchedClass, |
310 | IID: *getInstrItineraries()); |
311 | } |
312 | |
313 | if (hasInstrSchedModel()) |
314 | return MCSchedModel::getReciprocalThroughput(STI: *STI, SCDesc: *resolveSchedClass(MI)); |
315 | |
316 | return 0.0; |
317 | } |
318 | |
319 | double |
320 | TargetSchedModel::computeReciprocalThroughput(unsigned Opcode) const { |
321 | unsigned SchedClass = TII->get(Opcode).getSchedClass(); |
322 | if (hasInstrItineraries()) |
323 | return MCSchedModel::getReciprocalThroughput(SchedClass, |
324 | IID: *getInstrItineraries()); |
325 | if (hasInstrSchedModel()) { |
326 | const MCSchedClassDesc &SCDesc = *SchedModel.getSchedClassDesc(SchedClassIdx: SchedClass); |
327 | if (SCDesc.isValid() && !SCDesc.isVariant()) |
328 | return MCSchedModel::getReciprocalThroughput(STI: *STI, SCDesc); |
329 | } |
330 | |
331 | return 0.0; |
332 | } |
333 | |
334 | double |
335 | TargetSchedModel::computeReciprocalThroughput(const MCInst &MI) const { |
336 | if (hasInstrSchedModel()) |
337 | return SchedModel.getReciprocalThroughput(STI: *STI, MCII: *TII, Inst: MI); |
338 | return computeReciprocalThroughput(Opcode: MI.getOpcode()); |
339 | } |
340 | |
341 | bool TargetSchedModel::enableIntervals() const { |
342 | if (ForceEnableIntervals) |
343 | return true; |
344 | |
345 | return SchedModel.EnableIntervals; |
346 | } |
347 | |