1//===-- Target.h ------------------------------------------------*- 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/// \file
10///
11/// Classes that handle the creation of target-specific objects. This is
12/// similar to Target/TargetRegistry.
13///
14//===----------------------------------------------------------------------===//
15
16#ifndef LLVM_TOOLS_LLVM_EXEGESIS_TARGET_H
17#define LLVM_TOOLS_LLVM_EXEGESIS_TARGET_H
18
19#include "BenchmarkResult.h"
20#include "BenchmarkRunner.h"
21#include "Error.h"
22#include "LlvmState.h"
23#include "PerfHelper.h"
24#include "SnippetGenerator.h"
25#include "ValidationEvent.h"
26#include "llvm/CodeGen/TargetPassConfig.h"
27#include "llvm/IR/CallingConv.h"
28#include "llvm/IR/LegacyPassManager.h"
29#include "llvm/MC/MCInst.h"
30#include "llvm/MC/MCRegisterInfo.h"
31#include "llvm/Support/CommandLine.h"
32#include "llvm/Support/Error.h"
33#include "llvm/TargetParser/SubtargetFeature.h"
34#include "llvm/TargetParser/Triple.h"
35
36namespace llvm {
37namespace exegesis {
38
39extern cl::OptionCategory Options;
40extern cl::OptionCategory BenchmarkOptions;
41extern cl::OptionCategory AnalysisOptions;
42
43struct PfmCountersInfo {
44 // An optional name of a performance counter that can be used to measure
45 // cycles.
46 const char *CycleCounter;
47
48 // An optional name of a performance counter that can be used to measure
49 // uops.
50 const char *UopsCounter;
51
52 // An IssueCounter specifies how to measure uops issued to specific proc
53 // resources.
54 struct IssueCounter {
55 const char *Counter;
56 // The name of the ProcResource that this counter measures.
57 const char *ProcResName;
58 };
59 // An optional list of IssueCounters.
60 const IssueCounter *IssueCounters;
61 unsigned NumIssueCounters;
62
63 const std::pair<ValidationEvent, const char *> *ValidationEvents;
64 unsigned NumValidationEvents;
65
66 static const PfmCountersInfo Default;
67 static const PfmCountersInfo Dummy;
68};
69
70struct CpuAndPfmCounters {
71 const char *CpuName;
72 const PfmCountersInfo *PCI;
73 bool operator<(StringRef S) const { return StringRef(CpuName) < S; }
74};
75
76class ExegesisTarget {
77public:
78 typedef bool (*OpcodeAvailabilityChecker)(unsigned, const FeatureBitset &);
79 ExegesisTarget(ArrayRef<CpuAndPfmCounters> CpuPfmCounters,
80 OpcodeAvailabilityChecker IsOpcodeAvailable)
81 : CpuPfmCounters(CpuPfmCounters), IsOpcodeAvailable(IsOpcodeAvailable) {}
82
83 // Targets can use this to create target-specific perf counters.
84 virtual Expected<std::unique_ptr<pfm::CounterGroup>>
85 createCounter(StringRef CounterName, const LLVMState &State,
86 ArrayRef<const char *> ValidationCounters,
87 const pid_t ProcessID = 0) const;
88
89 // Targets can use this to add target-specific passes in assembleToStream();
90 virtual void addTargetSpecificPasses(PassManagerBase &PM) const {}
91
92 // Generates code to move a constant into a the given register.
93 // Precondition: Value must fit into Reg.
94 virtual std::vector<MCInst> setRegTo(const MCSubtargetInfo &STI, unsigned Reg,
95 const APInt &Value) const = 0;
96
97 // Generates the code for the lower munmap call. The code generated by this
98 // function may clobber registers.
99 virtual void generateLowerMunmap(std::vector<MCInst> &GeneratedCode) const {
100 report_fatal_error(
101 reason: "generateLowerMunmap is not implemented on the current architecture");
102 }
103
104 // Generates the upper munmap call. The code generated by this function may
105 // clobber registers.
106 virtual void generateUpperMunmap(std::vector<MCInst> &GeneratedCode) const {
107 report_fatal_error(
108 reason: "generateUpperMunmap is not implemented on the current architecture");
109 }
110
111 // Generates the code for an exit syscall. The code generated by this function
112 // may clobber registers.
113 virtual std::vector<MCInst> generateExitSyscall(unsigned ExitCode) const {
114 report_fatal_error(
115 reason: "generateExitSyscall is not implemented on the current architecture");
116 }
117
118 // Generates the code to mmap a region of code. The code generated by this
119 // function may clobber registers.
120 virtual std::vector<MCInst>
121 generateMmap(intptr_t Address, size_t Length,
122 intptr_t FileDescriptorAddress) const {
123 report_fatal_error(
124 reason: "generateMmap is not implemented on the current architecture");
125 }
126
127 // Generates the mmap code for the aux memory. The code generated by this
128 // function may clobber registers.
129 virtual void generateMmapAuxMem(std::vector<MCInst> &GeneratedCode) const {
130 report_fatal_error(
131 reason: "generateMmapAuxMem is not implemented on the current architecture\n");
132 }
133
134 // Moves argument registers into other registers that won't get clobbered
135 // while making syscalls. The code generated by this function may clobber
136 // registers.
137 virtual void moveArgumentRegisters(std::vector<MCInst> &GeneratedCode) const {
138 report_fatal_error(reason: "moveArgumentRegisters is not implemented on the "
139 "current architecture\n");
140 }
141
142 // Generates code to move argument registers, unmap memory above and below the
143 // snippet, and map the auxiliary memory into the subprocess. The code
144 // generated by this function may clobber registers.
145 virtual std::vector<MCInst> generateMemoryInitialSetup() const {
146 report_fatal_error(reason: "generateMemoryInitialSetup is not supported on the "
147 "current architecture\n");
148 }
149
150 // Returns true if all features are available that are required by Opcode.
151 virtual bool isOpcodeAvailable(unsigned Opcode,
152 const FeatureBitset &Features) const {
153 return IsOpcodeAvailable(Opcode, Features);
154 }
155
156 // Sets the stack register to the auxiliary memory so that operations
157 // requiring the stack can be formed (e.g., setting large registers). The code
158 // generated by this function may clobber registers.
159 virtual std::vector<MCInst> setStackRegisterToAuxMem() const {
160 report_fatal_error(reason: "setStackRegisterToAuxMem is not implemented on the "
161 "current architectures");
162 }
163
164 virtual intptr_t getAuxiliaryMemoryStartAddress() const {
165 report_fatal_error(reason: "getAuxiliaryMemoryStartAddress is not implemented on "
166 "the current architecture");
167 }
168
169 // Generates the necessary ioctl system calls to configure the perf counters.
170 // The code generated by this function preserves all registers if the
171 // parameter SaveRegisters is set to true.
172 virtual std::vector<MCInst> configurePerfCounter(long Request,
173 bool SaveRegisters) const {
174 report_fatal_error(
175 reason: "configurePerfCounter is not implemented on the current architecture");
176 }
177
178 // Gets the ABI dependent registers that are used to pass arguments in a
179 // function call.
180 virtual std::vector<unsigned> getArgumentRegisters() const {
181 report_fatal_error(
182 reason: "getArgumentRegisters is not implemented on the current architecture");
183 };
184
185 // Gets the registers that might potentially need to be saved by while
186 // the setup in the test harness executes.
187 virtual std::vector<unsigned> getRegistersNeedSaving() const {
188 report_fatal_error(reason: "getRegistersNeedSaving is not implemented on the "
189 "current architecture");
190 };
191
192 // Returns the register pointing to scratch memory, or 0 if this target
193 // does not support memory operands. The benchmark function uses the
194 // default calling convention.
195 virtual unsigned getScratchMemoryRegister(const Triple &) const { return 0; }
196
197 // Fills memory operands with references to the address at [Reg] + Offset.
198 virtual void fillMemoryOperands(InstructionTemplate &IT, unsigned Reg,
199 unsigned Offset) const {
200 llvm_unreachable(
201 "fillMemoryOperands() requires getScratchMemoryRegister() > 0");
202 }
203
204 // Returns a counter usable as a loop counter.
205 virtual unsigned getDefaultLoopCounterRegister(const Triple &) const {
206 return 0;
207 }
208
209 // Adds the code to decrement the loop counter and
210 virtual void decrementLoopCounterAndJump(MachineBasicBlock &MBB,
211 MachineBasicBlock &TargetMBB,
212 const MCInstrInfo &MII,
213 unsigned LoopRegister) const {
214 llvm_unreachable("decrementLoopCounterAndBranch() requires "
215 "getLoopCounterRegister() > 0");
216 }
217
218 // Returns a list of unavailable registers.
219 // Targets can use this to prevent some registers to be automatically selected
220 // for use in snippets.
221 virtual ArrayRef<unsigned> getUnavailableRegisters() const { return {}; }
222
223 // Returns the maximum number of bytes a load/store instruction can access at
224 // once. This is typically the size of the largest register available on the
225 // processor. Note that this only used as a hint to generate independant
226 // load/stores to/from memory, so the exact returned value does not really
227 // matter as long as it's large enough.
228 virtual unsigned getMaxMemoryAccessSize() const { return 0; }
229
230 // Assigns a random operand of the right type to variable Var.
231 // The target is responsible for handling any operand starting from
232 // OPERAND_FIRST_TARGET.
233 virtual Error randomizeTargetMCOperand(const Instruction &Instr,
234 const Variable &Var,
235 MCOperand &AssignedValue,
236 const BitVector &ForbiddenRegs) const {
237 return make_error<Failure>(
238 Args: "targets with target-specific operands should implement this");
239 }
240
241 // Returns true if this instruction is supported as a back-to-back
242 // instructions.
243 // FIXME: Eventually we should discover this dynamically.
244 virtual bool allowAsBackToBack(const Instruction &Instr) const {
245 return true;
246 }
247
248 // For some instructions, it is interesting to measure how it's performance
249 // characteristics differ depending on it's operands.
250 // This allows us to produce all the interesting variants.
251 virtual std::vector<InstructionTemplate>
252 generateInstructionVariants(const Instruction &Instr,
253 unsigned MaxConfigsPerOpcode) const {
254 // By default, we're happy with whatever randomizer will give us.
255 return {&Instr};
256 }
257
258 // Checks hardware and software support for current benchmark mode.
259 // Returns an error if the target host does not have support to run the
260 // benchmark.
261 virtual Error checkFeatureSupport() const { return Error::success(); }
262
263 // Creates a snippet generator for the given mode.
264 std::unique_ptr<SnippetGenerator>
265 createSnippetGenerator(Benchmark::ModeE Mode,
266 const LLVMState &State,
267 const SnippetGenerator::Options &Opts) const;
268 // Creates a benchmark runner for the given mode.
269 Expected<std::unique_ptr<BenchmarkRunner>> createBenchmarkRunner(
270 Benchmark::ModeE Mode, const LLVMState &State,
271 BenchmarkPhaseSelectorE BenchmarkPhaseSelector,
272 BenchmarkRunner::ExecutionModeE ExecutionMode,
273 unsigned BenchmarkRepeatCount,
274 ArrayRef<ValidationEvent> ValidationCounters,
275 Benchmark::ResultAggregationModeE ResultAggMode = Benchmark::Min) const;
276
277 // Returns the ExegesisTarget for the given triple or nullptr if the target
278 // does not exist.
279 static const ExegesisTarget *lookup(Triple TT);
280 // Returns the default (unspecialized) ExegesisTarget.
281 static const ExegesisTarget &getDefault();
282 // Registers a target. Not thread safe.
283 static void registerTarget(ExegesisTarget *T);
284
285 virtual ~ExegesisTarget();
286
287 // Returns the Pfm counters for the given CPU (or the default if no pfm
288 // counters are defined for this CPU).
289 const PfmCountersInfo &getPfmCounters(StringRef CpuName) const;
290
291 // Returns dummy Pfm counters which can be used to execute generated snippet
292 // without access to performance counters.
293 const PfmCountersInfo &getDummyPfmCounters() const;
294
295 // Saves the CPU state that needs to be preserved when running a benchmark,
296 // and returns and RAII object that restores the state on destruction.
297 // By default no state is preserved.
298 struct SavedState {
299 virtual ~SavedState();
300 };
301 virtual std::unique_ptr<SavedState> withSavedState() const {
302 return std::make_unique<SavedState>();
303 }
304
305private:
306 virtual bool matchesArch(Triple::ArchType Arch) const = 0;
307
308 // Targets can implement their own snippet generators/benchmarks runners by
309 // implementing these.
310 std::unique_ptr<SnippetGenerator> virtual createSerialSnippetGenerator(
311 const LLVMState &State, const SnippetGenerator::Options &Opts) const;
312 std::unique_ptr<SnippetGenerator> virtual createParallelSnippetGenerator(
313 const LLVMState &State, const SnippetGenerator::Options &Opts) const;
314 std::unique_ptr<BenchmarkRunner> virtual createLatencyBenchmarkRunner(
315 const LLVMState &State, Benchmark::ModeE Mode,
316 BenchmarkPhaseSelectorE BenchmarkPhaseSelector,
317 Benchmark::ResultAggregationModeE ResultAggMode,
318 BenchmarkRunner::ExecutionModeE ExecutionMode,
319 ArrayRef<ValidationEvent> ValidationCounters,
320 unsigned BenchmarkRepeatCount) const;
321 std::unique_ptr<BenchmarkRunner> virtual createUopsBenchmarkRunner(
322 const LLVMState &State, BenchmarkPhaseSelectorE BenchmarkPhaseSelector,
323 Benchmark::ResultAggregationModeE ResultAggMode,
324 BenchmarkRunner::ExecutionModeE ExecutionMode,
325 ArrayRef<ValidationEvent> ValidationCounters) const;
326
327 const ExegesisTarget *Next = nullptr;
328 const ArrayRef<CpuAndPfmCounters> CpuPfmCounters;
329 const OpcodeAvailabilityChecker IsOpcodeAvailable;
330};
331
332} // namespace exegesis
333} // namespace llvm
334
335#endif // LLVM_TOOLS_LLVM_EXEGESIS_TARGET_H
336

source code of llvm/tools/llvm-exegesis/lib/Target.h