LLVM 19.0.0git
TargetSchedule.cpp
Go to the documentation of this file.
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
20#include "llvm/MC/MCInstrDesc.h"
22#include "llvm/MC/MCSchedule.h"
26#include <algorithm>
27#include <cassert>
28#include <cstdint>
29#include <numeric>
30
31using namespace llvm;
32
34 cl::desc("Use TargetSchedModel for latency lookup"));
35
37 cl::desc("Use InstrItineraryData for latency lookup"));
38
40 "sched-model-force-enable-intervals", cl::Hidden, cl::init(false),
41 cl::desc("Force the use of resource intervals in the schedule model"));
42
44 return EnableSchedModel && SchedModel.hasInstrSchedModel();
45}
46
48 return EnableSchedItins && !InstrItins.isEmpty();
49}
50
52 STI = TSInfo;
53 SchedModel = TSInfo->getSchedModel();
54 TII = TSInfo->getInstrInfo();
55 STI->initInstrItins(InstrItins);
56
57 unsigned NumRes = SchedModel.getNumProcResourceKinds();
58 ResourceFactors.resize(NumRes);
59 ResourceLCM = SchedModel.IssueWidth;
60 for (unsigned Idx = 0; Idx < NumRes; ++Idx) {
61 unsigned NumUnits = SchedModel.getProcResource(Idx)->NumUnits;
62 if (NumUnits > 0)
63 ResourceLCM = std::lcm(ResourceLCM, NumUnits);
64 }
65 MicroOpFactor = ResourceLCM / SchedModel.IssueWidth;
66 for (unsigned Idx = 0; Idx < NumRes; ++Idx) {
67 unsigned NumUnits = SchedModel.getProcResource(Idx)->NumUnits;
68 ResourceFactors[Idx] = NumUnits ? (ResourceLCM / NumUnits) : 0;
69 }
70}
71
72/// Returns true only if instruction is specified as single issue.
74 const MCSchedClassDesc *SC) const {
75 if (hasInstrSchedModel()) {
76 if (!SC)
78 if (SC->isValid())
79 return SC->BeginGroup;
80 }
81 return false;
82}
83
85 const MCSchedClassDesc *SC) const {
86 if (hasInstrSchedModel()) {
87 if (!SC)
89 if (SC->isValid())
90 return SC->EndGroup;
91 }
92 return false;
93}
94
96 const MCSchedClassDesc *SC) const {
97 if (hasInstrItineraries()) {
98 int UOps = InstrItins.getNumMicroOps(MI->getDesc().getSchedClass());
99 return (UOps >= 0) ? UOps : TII->getNumMicroOps(&InstrItins, *MI);
100 }
101 if (hasInstrSchedModel()) {
102 if (!SC)
103 SC = resolveSchedClass(MI);
104 if (SC->isValid())
105 return SC->NumMicroOps;
106 }
107 return MI->isTransient() ? 0 : 1;
108}
109
110// The machine model may explicitly specify an invalid latency, which
111// effectively means infinite latency. Since users of the TargetSchedule API
112// don't know how to handle this, we convert it to a very large latency that is
113// easy to distinguish when debugging the DAG but won't induce overflow.
114static unsigned capLatency(int Cycles) {
115 return Cycles >= 0 ? Cycles : 1000;
116}
117
118/// Return the MCSchedClassDesc for this instruction. Some SchedClasses require
119/// evaluation of predicates that depend on instruction operands or flags.
121resolveSchedClass(const MachineInstr *MI) const {
122 // Get the definition's scheduling class descriptor from this machine model.
123 unsigned SchedClass = MI->getDesc().getSchedClass();
124 const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClass);
125 if (!SCDesc->isValid())
126 return SCDesc;
127
128#ifndef NDEBUG
129 unsigned NIter = 0;
130#endif
131 while (SCDesc->isVariant()) {
132 assert(++NIter < 6 && "Variants are nested deeper than the magic number");
133
134 SchedClass = STI->resolveSchedClass(SchedClass, MI, this);
135 SCDesc = SchedModel.getSchedClassDesc(SchedClass);
136 }
137 return SCDesc;
138}
139
140/// Find the def index of this operand. This index maps to the machine model and
141/// is independent of use operands. Def operands may be reordered with uses or
142/// merged with uses without affecting the def index (e.g. before/after
143/// regalloc). However, an instruction's def operands must never be reordered
144/// with respect to each other.
145static unsigned findDefIdx(const MachineInstr *MI, unsigned DefOperIdx) {
146 unsigned DefIdx = 0;
147 for (unsigned i = 0; i != DefOperIdx; ++i) {
148 const MachineOperand &MO = MI->getOperand(i);
149 if (MO.isReg() && MO.isDef())
150 ++DefIdx;
151 }
152 return DefIdx;
153}
154
155/// Find the use index of this operand. This is independent of the instruction's
156/// def operands.
157///
158/// Note that uses are not determined by the operand's isUse property, which
159/// is simply the inverse of isDef. Here we consider any readsReg operand to be
160/// a "use". The machine model allows an operand to be both a Def and Use.
161static unsigned findUseIdx(const MachineInstr *MI, unsigned UseOperIdx) {
162 unsigned UseIdx = 0;
163 for (unsigned i = 0; i != UseOperIdx; ++i) {
164 const MachineOperand &MO = MI->getOperand(i);
165 if (MO.isReg() && MO.readsReg() && !MO.isDef())
166 ++UseIdx;
167 }
168 return UseIdx;
169}
170
171// Top-level API for clients that know the operand indices. This doesn't need to
172// return std::optional<unsigned>, as it always returns a valid latency.
174 const MachineInstr *DefMI, unsigned DefOperIdx,
175 const MachineInstr *UseMI, unsigned UseOperIdx) const {
176
177 const unsigned InstrLatency = computeInstrLatency(DefMI);
178 const unsigned DefaultDefLatency = TII->defaultDefLatency(SchedModel, *DefMI);
179
181 return DefaultDefLatency;
182
183 if (hasInstrItineraries()) {
184 std::optional<unsigned> OperLatency;
185 if (UseMI) {
186 OperLatency = TII->getOperandLatency(&InstrItins, *DefMI, DefOperIdx,
187 *UseMI, UseOperIdx);
188 }
189 else {
190 unsigned DefClass = DefMI->getDesc().getSchedClass();
191 OperLatency = InstrItins.getOperandCycle(DefClass, DefOperIdx);
192 }
193
194 // Expected latency is the max of InstrLatency and DefaultDefLatency, if we
195 // didn't find an operand latency.
196 return OperLatency ? *OperLatency
197 : std::max(InstrLatency, DefaultDefLatency);
198 }
199
200 // hasInstrSchedModel()
202 unsigned DefIdx = findDefIdx(DefMI, DefOperIdx);
203 if (DefIdx < SCDesc->NumWriteLatencyEntries) {
204 // Lookup the definition's write latency in SubtargetInfo.
205 const MCWriteLatencyEntry *WLEntry =
206 STI->getWriteLatencyEntry(SCDesc, DefIdx);
207 unsigned WriteID = WLEntry->WriteResourceID;
208 unsigned Latency = capLatency(WLEntry->Cycles);
209 if (!UseMI)
210 return Latency;
211
212 // Lookup the use's latency adjustment in SubtargetInfo.
213 const MCSchedClassDesc *UseDesc = resolveSchedClass(UseMI);
214 if (UseDesc->NumReadAdvanceEntries == 0)
215 return Latency;
216 unsigned UseIdx = findUseIdx(UseMI, UseOperIdx);
217 int Advance = STI->getReadAdvanceCycles(UseDesc, UseIdx, WriteID);
218 if (Advance > 0 && (unsigned)Advance > Latency) // unsigned wrap
219 return 0;
220 return Latency - Advance;
221 }
222 // If DefIdx does not exist in the model (e.g. implicit defs), then return
223 // unit latency (defaultDefLatency may be too conservative).
224#ifndef NDEBUG
225 if (SCDesc->isValid() && !DefMI->getOperand(DefOperIdx).isImplicit() &&
226 !DefMI->getDesc().operands()[DefOperIdx].isOptionalDef() &&
227 SchedModel.isComplete()) {
228 errs() << "DefIdx " << DefIdx << " exceeds machine model writes for "
229 << *DefMI << " (Try with MCSchedModel.CompleteModel set to false)";
230 llvm_unreachable("incomplete machine model");
231 }
232#endif
233 // FIXME: Automatically giving all implicit defs defaultDefLatency is
234 // undesirable. We should only do it for defs that are known to the MC
235 // desc like flags. Truly implicit defs should get 1 cycle latency.
236 return DefMI->isTransient() ? 0 : DefaultDefLatency;
237}
238
239unsigned
240TargetSchedModel::computeInstrLatency(const MCSchedClassDesc &SCDesc) const {
241 return capLatency(MCSchedModel::computeInstrLatency(*STI, SCDesc));
242}
243
244unsigned TargetSchedModel::computeInstrLatency(unsigned Opcode) const {
245 assert(hasInstrSchedModel() && "Only call this function with a SchedModel");
246 unsigned SCIdx = TII->get(Opcode).getSchedClass();
247 return capLatency(SchedModel.computeInstrLatency(*STI, SCIdx));
248}
249
250unsigned TargetSchedModel::computeInstrLatency(const MCInst &Inst) const {
251 if (hasInstrSchedModel())
252 return capLatency(SchedModel.computeInstrLatency(*STI, *TII, Inst));
253 return computeInstrLatency(Inst.getOpcode());
254}
255
256unsigned
257TargetSchedModel::computeInstrLatency(const MachineInstr *MI,
258 bool UseDefaultDefLatency) const {
259 // For the itinerary model, fall back to the old subtarget hook.
260 // Allow subtargets to compute Bundle latencies outside the machine model.
261 if (hasInstrItineraries() || MI->isBundle() ||
262 (!hasInstrSchedModel() && !UseDefaultDefLatency))
263 return TII->getInstrLatency(&InstrItins, *MI);
264
265 if (hasInstrSchedModel()) {
266 const MCSchedClassDesc *SCDesc = resolveSchedClass(MI);
267 if (SCDesc->isValid())
268 return computeInstrLatency(*SCDesc);
269 }
270 return TII->defaultDefLatency(SchedModel, *MI);
271}
272
274computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx,
275 const MachineInstr *DepMI) const {
276 if (!SchedModel.isOutOfOrder())
277 return 1;
278
279 // Out-of-order processor can dispatch WAW dependencies in the same cycle.
280
281 // Treat predication as a data dependency for out-of-order cpus. In-order
282 // cpus do not need to treat predicated writes specially.
283 //
284 // TODO: The following hack exists because predication passes do not
285 // correctly append imp-use operands, and readsReg() strangely returns false
286 // for predicated defs.
287 Register Reg = DefMI->getOperand(DefOperIdx).getReg();
288 const MachineFunction &MF = *DefMI->getMF();
290 if (!DepMI->readsRegister(Reg, TRI) && TII->isPredicated(*DepMI))
291 return computeInstrLatency(DefMI);
292
293 // If we have a per operand scheduling model, check if this def is writing
294 // an unbuffered resource. If so, it treated like an in-order cpu.
295 if (hasInstrSchedModel()) {
297 if (SCDesc->isValid()) {
298 for (const MCWriteProcResEntry *PRI = STI->getWriteProcResBegin(SCDesc),
299 *PRE = STI->getWriteProcResEnd(SCDesc); PRI != PRE; ++PRI) {
300 if (!SchedModel.getProcResource(PRI->ProcResourceIdx)->BufferSize)
301 return 1;
302 }
303 }
304 }
305 return 0;
306}
307
308double
310 if (hasInstrItineraries()) {
311 unsigned SchedClass = MI->getDesc().getSchedClass();
314 }
315
316 if (hasInstrSchedModel())
318
319 return 0.0;
320}
321
322double
324 unsigned SchedClass = TII->get(Opcode).getSchedClass();
328 if (hasInstrSchedModel()) {
329 const MCSchedClassDesc &SCDesc = *SchedModel.getSchedClassDesc(SchedClass);
330 if (SCDesc.isValid() && !SCDesc.isVariant())
331 return MCSchedModel::getReciprocalThroughput(*STI, SCDesc);
332 }
333
334 return 0.0;
335}
336
337double
339 if (hasInstrSchedModel())
340 return SchedModel.getReciprocalThroughput(*STI, *TII, MI);
341 return computeReciprocalThroughput(MI.getOpcode());
342}
343
346 return true;
347
348 return SchedModel.EnableIntervals;
349}
MachineInstrBuilder & UseMI
MachineInstrBuilder MachineInstrBuilder & DefMI
Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx
IRTranslator LLVM IR MI
unsigned const TargetRegisterInfo * TRI
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static unsigned findUseIdx(const MachineInstr *MI, unsigned UseOperIdx)
Find the use index of this operand.
static unsigned capLatency(int Cycles)
static cl::opt< bool > EnableSchedModel("schedmodel", cl::Hidden, cl::init(true), cl::desc("Use TargetSchedModel for latency lookup"))
static unsigned findDefIdx(const MachineInstr *MI, unsigned DefOperIdx)
Find the def index of this operand.
static cl::opt< bool > ForceEnableIntervals("sched-model-force-enable-intervals", cl::Hidden, cl::init(false), cl::desc("Force the use of resource intervals in the schedule model"))
static cl::opt< bool > EnableSchedItins("scheditins", cl::Hidden, cl::init(true), cl::desc("Use InstrItineraryData for latency lookup"))
int getNumMicroOps(unsigned ItinClassIndx) const
Return the number of micro-ops that the given class decodes to.
std::optional< unsigned > getOperandCycle(unsigned ItinClassIndx, unsigned OperandIdx) const
Return the cycle for the given class and operand.
bool isEmpty() const
Returns true if there are no itineraries.
Instances of this class represent a single low-level machine instruction.
Definition: MCInst.h:184
unsigned getOpcode() const
Definition: MCInst.h:198
unsigned getSchedClass() const
Return the scheduling class for this instruction.
Definition: MCInstrDesc.h:600
ArrayRef< MCOperandInfo > operands() const
Definition: MCInstrDesc.h:239
const MCInstrDesc & get(unsigned Opcode) const
Return the machine instruction descriptor that corresponds to the specified instruction opcode.
Definition: MCInstrInfo.h:63
const MCWriteProcResEntry * getWriteProcResEnd(const MCSchedClassDesc *SC) const
int getReadAdvanceCycles(const MCSchedClassDesc *SC, unsigned UseIdx, unsigned WriteResID) const
const MCWriteLatencyEntry * getWriteLatencyEntry(const MCSchedClassDesc *SC, unsigned DefIdx) const
void initInstrItins(InstrItineraryData &InstrItins) const
Initialize an InstrItineraryData instance.
const MCWriteProcResEntry * getWriteProcResBegin(const MCSchedClassDesc *SC) const
Return an iterator at the first process resource consumed by the given scheduling class.
const MCSchedModel & getSchedModel() const
Get the machine model for this subtarget's CPU.
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
const MCInstrDesc & getDesc() const
Returns the target instruction descriptor of this MachineInstr.
Definition: MachineInstr.h:541
const MachineFunction * getMF() const
Return the function that contains the basic block that this instruction belongs to.
bool readsRegister(Register Reg, const TargetRegisterInfo *TRI=nullptr) const
Return true if the MachineInstr reads the specified register.
bool isTransient() const
Return true if this is a transient instruction that is either very likely to be eliminated during reg...
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:554
MachineOperand class - Representation of each machine instruction operand.
bool readsReg() const
readsReg - Returns true if this operand reads the previous value of its register.
bool isImplicit() const
bool isReg() const
isReg - Tests if this is a MO_Register operand.
Register getReg() const
getReg - Returns the register number.
Wrapper class representing virtual and physical registers.
Definition: Register.h:19
void resize(size_type N)
Definition: SmallVector.h:651
virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData, const MachineInstr &MI) const
Return the number of u-operations the given machine instruction will be decoded to on the target cpu.
virtual std::optional< unsigned > getOperandLatency(const InstrItineraryData *ItinData, SDNode *DefNode, unsigned DefIdx, SDNode *UseNode, unsigned UseIdx) const
virtual bool isPredicated(const MachineInstr &MI) const
Returns true if the instruction is already predicated.
virtual unsigned getInstrLatency(const InstrItineraryData *ItinData, const MachineInstr &MI, unsigned *PredCost=nullptr) const
Compute the instruction latency of a given instruction.
unsigned defaultDefLatency(const MCSchedModel &SchedModel, const MachineInstr &DefMI) const
Return the default expected latency for a def based on its opcode.
TargetRegisterInfo base class - We assume that the target defines a static array of TargetRegisterDes...
bool mustEndGroup(const MachineInstr *MI, const MCSchedClassDesc *SC=nullptr) const
Return true if current group must end.
bool hasInstrSchedModel() const
Return true if this machine model includes an instruction-level scheduling model.
unsigned computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx, const MachineInstr *DepMI) const
Output dependency latency of a pair of defs of the same register.
bool mustBeginGroup(const MachineInstr *MI, const MCSchedClassDesc *SC=nullptr) const
Return true if new group must begin.
const MCSchedClassDesc * resolveSchedClass(const MachineInstr *MI) const
Return the MCSchedClassDesc for this instruction.
void init(const TargetSubtargetInfo *TSInfo)
Initialize the machine model for instruction scheduling.
unsigned computeOperandLatency(const MachineInstr *DefMI, unsigned DefOperIdx, const MachineInstr *UseMI, unsigned UseOperIdx) const
Compute operand latency based on the available machine model.
double computeReciprocalThroughput(const MachineInstr *MI) const
Compute the reciprocal throughput of the given instruction.
unsigned getNumMicroOps(const MachineInstr *MI, const MCSchedClassDesc *SC=nullptr) const
Return the number of issue slots required for this MI.
const InstrItineraryData * getInstrItineraries() const
bool hasInstrItineraries() const
Return true if this machine model includes cycle-to-cycle itinerary data.
TargetSubtargetInfo - Generic base class for all target subtargets.
virtual const TargetRegisterInfo * getRegisterInfo() const
getRegisterInfo - If register information is available, return it.
virtual const TargetInstrInfo * getInstrInfo() const
virtual unsigned resolveSchedClass(unsigned SchedClass, const MachineInstr *MI, const TargetSchedModel *SchedModel) const
Resolve a SchedClass at runtime, where SchedClass identifies an MCSchedClassDesc with the isVariant p...
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:450
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
raw_fd_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Summarize the scheduling resources required for an instruction of a particular scheduling class.
Definition: MCSchedule.h:118
bool isValid() const
Definition: MCSchedule.h:136
bool isVariant() const
Definition: MCSchedule.h:139
uint16_t NumReadAdvanceEntries
Definition: MCSchedule.h:134
bool isOutOfOrder() const
Return true if machine supports out of order execution.
Definition: MCSchedule.h:347
const MCSchedClassDesc * getSchedClassDesc(unsigned SchedClassIdx) const
Definition: MCSchedule.h:360
unsigned getNumProcResourceKinds() const
Definition: MCSchedule.h:349
bool hasInstrSchedModel() const
Does this machine model include instruction-level scheduling.
Definition: MCSchedule.h:334
static int computeInstrLatency(const MCSubtargetInfo &STI, const MCSchedClassDesc &SCDesc)
Returns the latency value for the scheduling class.
Definition: MCSchedule.cpp:42
unsigned IssueWidth
Definition: MCSchedule.h:265
const MCProcResourceDesc * getProcResource(unsigned ProcResourceIdx) const
Definition: MCSchedule.h:353
static double getReciprocalThroughput(const MCSubtargetInfo &STI, const MCSchedClassDesc &SCDesc)
Definition: MCSchedule.cpp:90
bool isComplete() const
Return true if this machine model data for all instructions with a scheduling class (itinerary class ...
Definition: MCSchedule.h:344
Specify the latency in cpu cycles for a particular scheduling class and def index.
Definition: MCSchedule.h:86
Identify one of the processor resource kinds consumed by a particular scheduling class for the specif...
Definition: MCSchedule.h:63