1 //===------ PerfMonitor.cpp - Generate a run-time performance monitor. -======//
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 //===----------------------------------------------------------------------===//
10 
11 #include "polly/CodeGen/PerfMonitor.h"
12 #include "polly/CodeGen/RuntimeDebugBuilder.h"
13 #include "polly/ScopInfo.h"
14 #include "llvm/ADT/Triple.h"
15 #include "llvm/ADT/Twine.h"
16 #include "llvm/IR/IntrinsicsX86.h"
17 
18 using namespace llvm;
19 using namespace polly;
20 
21 Function *PerfMonitor::getAtExit() {
22   const char *Name = "atexit";
23   Function *F = M->getFunction(Name);
24 
25   if (!F) {
26     GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage;
27     FunctionType *Ty = FunctionType::get(Builder.getInt32Ty(),
28                                          {Builder.getInt8PtrTy()}, false);
29     F = Function::Create(Ty, Linkage, Name, M);
30   }
31 
32   return F;
33 }
34 
35 void PerfMonitor::addToGlobalConstructors(Function *Fn) {
36   const char *Name = "llvm.global_ctors";
37   GlobalVariable *GV = M->getGlobalVariable(Name);
38   std::vector<Constant *> V;
39 
40   if (GV) {
41     Constant *Array = GV->getInitializer();
42     for (Value *X : Array->operand_values())
43       V.push_back(cast<Constant>(X));
44     GV->eraseFromParent();
45   }
46 
47   StructType *ST = StructType::get(Builder.getInt32Ty(), Fn->getType(),
48                                    Builder.getInt8PtrTy());
49 
50   V.push_back(
51       ConstantStruct::get(ST, Builder.getInt32(10), Fn,
52                           ConstantPointerNull::get(Builder.getInt8PtrTy())));
53   ArrayType *Ty = ArrayType::get(ST, V.size());
54 
55   GV = new GlobalVariable(*M, Ty, true, GlobalValue::AppendingLinkage,
56                           ConstantArray::get(Ty, V), Name, nullptr,
57                           GlobalVariable::NotThreadLocal);
58 }
59 
60 Function *PerfMonitor::getRDTSCP() {
61   return Intrinsic::getDeclaration(M, Intrinsic::x86_rdtscp);
62 }
63 
64 PerfMonitor::PerfMonitor(const Scop &S, Module *M)
65     : M(M), Builder(M->getContext()), S(S) {
66   if (Triple(M->getTargetTriple()).getArch() == llvm::Triple::x86_64)
67     Supported = true;
68   else
69     Supported = false;
70 }
71 
72 static void TryRegisterGlobal(Module *M, const char *Name,
73                               Constant *InitialValue, Value **Location) {
74   *Location = M->getGlobalVariable(Name);
75 
76   if (!*Location)
77     *Location = new GlobalVariable(
78         *M, InitialValue->getType(), true, GlobalValue::WeakAnyLinkage,
79         InitialValue, Name, nullptr, GlobalVariable::InitialExecTLSModel);
80 }
81 
82 // Generate a unique name that is usable as a LLVM name for a scop to name its
83 // performance counter.
84 static std::string GetScopUniqueVarname(const Scop &S) {
85   std::string EntryString, ExitString;
86   std::tie(EntryString, ExitString) = S.getEntryExitStr();
87 
88   return (Twine("__polly_perf_in_") + S.getFunction().getName() + "_from__" +
89           EntryString + "__to__" + ExitString)
90       .str();
91 }
92 
93 void PerfMonitor::addScopCounter() {
94   const std::string varname = GetScopUniqueVarname(S);
95   TryRegisterGlobal(M, (varname + "_cycles").c_str(), Builder.getInt64(0),
96                     &CyclesInCurrentScopPtr);
97 
98   TryRegisterGlobal(M, (varname + "_trip_count").c_str(), Builder.getInt64(0),
99                     &TripCountForCurrentScopPtr);
100 }
101 
102 void PerfMonitor::addGlobalVariables() {
103   TryRegisterGlobal(M, "__polly_perf_cycles_total_start", Builder.getInt64(0),
104                     &CyclesTotalStartPtr);
105 
106   TryRegisterGlobal(M, "__polly_perf_initialized", Builder.getInt1(0),
107                     &AlreadyInitializedPtr);
108 
109   TryRegisterGlobal(M, "__polly_perf_cycles_in_scops", Builder.getInt64(0),
110                     &CyclesInScopsPtr);
111 
112   TryRegisterGlobal(M, "__polly_perf_cycles_in_scop_start", Builder.getInt64(0),
113                     &CyclesInScopStartPtr);
114 }
115 
116 static const char *InitFunctionName = "__polly_perf_init";
117 static const char *FinalReportingFunctionName = "__polly_perf_final";
118 
119 static BasicBlock *FinalStartBB = nullptr;
120 static ReturnInst *ReturnFromFinal = nullptr;
121 
122 Function *PerfMonitor::insertFinalReporting() {
123   // Create new function.
124   GlobalValue::LinkageTypes Linkage = Function::WeakODRLinkage;
125   FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), {}, false);
126   Function *ExitFn =
127       Function::Create(Ty, Linkage, FinalReportingFunctionName, M);
128   FinalStartBB = BasicBlock::Create(M->getContext(), "start", ExitFn);
129   Builder.SetInsertPoint(FinalStartBB);
130 
131   if (!Supported) {
132     RuntimeDebugBuilder::createCPUPrinter(
133         Builder, "Polly runtime information generation not supported\n");
134     Builder.CreateRetVoid();
135     return ExitFn;
136   }
137 
138   // Measure current cycles and compute final timings.
139   Function *RDTSCPFn = getRDTSCP();
140 
141   Value *CurrentCycles =
142       Builder.CreateExtractValue(Builder.CreateCall(RDTSCPFn), {0});
143   Value *CyclesStart = Builder.CreateLoad(CyclesTotalStartPtr, true);
144   Value *CyclesTotal = Builder.CreateSub(CurrentCycles, CyclesStart);
145   Value *CyclesInScops = Builder.CreateLoad(CyclesInScopsPtr, true);
146 
147   // Print the runtime information.
148   RuntimeDebugBuilder::createCPUPrinter(Builder, "Polly runtime information\n");
149   RuntimeDebugBuilder::createCPUPrinter(Builder, "-------------------------\n");
150   RuntimeDebugBuilder::createCPUPrinter(Builder, "Total: ", CyclesTotal, "\n");
151   RuntimeDebugBuilder::createCPUPrinter(Builder, "Scops: ", CyclesInScops,
152                                         "\n");
153 
154   // Print the preamble for per-scop information.
155   RuntimeDebugBuilder::createCPUPrinter(Builder, "\n");
156   RuntimeDebugBuilder::createCPUPrinter(Builder, "Per SCoP information\n");
157   RuntimeDebugBuilder::createCPUPrinter(Builder, "--------------------\n");
158 
159   RuntimeDebugBuilder::createCPUPrinter(
160       Builder, "scop function, "
161                "entry block name, exit block name, total time, trip count\n");
162   ReturnFromFinal = Builder.CreateRetVoid();
163   return ExitFn;
164 }
165 
166 void PerfMonitor::AppendScopReporting() {
167   if (!Supported)
168     return;
169 
170   assert(FinalStartBB && "Expected FinalStartBB to be initialized by "
171                          "PerfMonitor::insertFinalReporting.");
172   assert(ReturnFromFinal && "Expected ReturnFromFinal to be initialized by "
173                             "PerfMonitor::insertFinalReporting.");
174 
175   Builder.SetInsertPoint(FinalStartBB);
176   ReturnFromFinal->eraseFromParent();
177 
178   Value *CyclesInCurrentScop =
179       Builder.CreateLoad(this->CyclesInCurrentScopPtr, true);
180 
181   Value *TripCountForCurrentScop =
182       Builder.CreateLoad(this->TripCountForCurrentScopPtr, true);
183 
184   std::string EntryName, ExitName;
185   std::tie(EntryName, ExitName) = S.getEntryExitStr();
186 
187   // print in CSV for easy parsing with other tools.
188   RuntimeDebugBuilder::createCPUPrinter(
189       Builder, S.getFunction().getName(), ", ", EntryName, ", ", ExitName, ", ",
190       CyclesInCurrentScop, ", ", TripCountForCurrentScop, "\n");
191 
192   ReturnFromFinal = Builder.CreateRetVoid();
193 }
194 
195 static Function *FinalReporting = nullptr;
196 
197 void PerfMonitor::initialize() {
198   addGlobalVariables();
199   addScopCounter();
200 
201   // Ensure that we only add the final reporting function once.
202   // On later invocations, append to the reporting function.
203   if (!FinalReporting) {
204     FinalReporting = insertFinalReporting();
205 
206     Function *InitFn = insertInitFunction(FinalReporting);
207     addToGlobalConstructors(InitFn);
208   }
209 
210   AppendScopReporting();
211 }
212 
213 Function *PerfMonitor::insertInitFunction(Function *FinalReporting) {
214   // Insert function definition and BBs.
215   GlobalValue::LinkageTypes Linkage = Function::WeakODRLinkage;
216   FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), {}, false);
217   Function *InitFn = Function::Create(Ty, Linkage, InitFunctionName, M);
218   BasicBlock *Start = BasicBlock::Create(M->getContext(), "start", InitFn);
219   BasicBlock *EarlyReturn =
220       BasicBlock::Create(M->getContext(), "earlyreturn", InitFn);
221   BasicBlock *InitBB = BasicBlock::Create(M->getContext(), "initbb", InitFn);
222 
223   Builder.SetInsertPoint(Start);
224 
225   // Check if this function was already run. If yes, return.
226   //
227   // In case profiling has been enabled in multiple translation units, the
228   // initializer function will be added to the global constructors list of
229   // each translation unit. When merging translation units, the global
230   // constructor lists are just appended, such that the initializer will appear
231   // multiple times. To avoid initializations being run multiple times (and
232   // especially to avoid that atExitFn is called more than once), we bail
233   // out if the initializer is run more than once.
234   Value *HasRunBefore = Builder.CreateLoad(AlreadyInitializedPtr);
235   Builder.CreateCondBr(HasRunBefore, EarlyReturn, InitBB);
236   Builder.SetInsertPoint(EarlyReturn);
237   Builder.CreateRetVoid();
238 
239   // Keep track that this function has been run once.
240   Builder.SetInsertPoint(InitBB);
241   Value *True = Builder.getInt1(true);
242   Builder.CreateStore(True, AlreadyInitializedPtr);
243 
244   // Register the final reporting function with atexit().
245   Value *FinalReportingPtr =
246       Builder.CreatePointerCast(FinalReporting, Builder.getInt8PtrTy());
247   Function *AtExitFn = getAtExit();
248   Builder.CreateCall(AtExitFn, {FinalReportingPtr});
249 
250   if (Supported) {
251     // Read the currently cycle counter and store the result for later.
252     Function *RDTSCPFn = getRDTSCP();
253     Value *CurrentCycles =
254         Builder.CreateExtractValue(Builder.CreateCall(RDTSCPFn), {0});
255     Builder.CreateStore(CurrentCycles, CyclesTotalStartPtr, true);
256   }
257   Builder.CreateRetVoid();
258 
259   return InitFn;
260 }
261 
262 void PerfMonitor::insertRegionStart(Instruction *InsertBefore) {
263   if (!Supported)
264     return;
265 
266   Builder.SetInsertPoint(InsertBefore);
267   Function *RDTSCPFn = getRDTSCP();
268   Value *CurrentCycles =
269       Builder.CreateExtractValue(Builder.CreateCall(RDTSCPFn), {0});
270   Builder.CreateStore(CurrentCycles, CyclesInScopStartPtr, true);
271 }
272 
273 void PerfMonitor::insertRegionEnd(Instruction *InsertBefore) {
274   if (!Supported)
275     return;
276 
277   Builder.SetInsertPoint(InsertBefore);
278   Function *RDTSCPFn = getRDTSCP();
279   LoadInst *CyclesStart = Builder.CreateLoad(CyclesInScopStartPtr, true);
280   Value *CurrentCycles =
281       Builder.CreateExtractValue(Builder.CreateCall(RDTSCPFn), {0});
282   Value *CyclesInScop = Builder.CreateSub(CurrentCycles, CyclesStart);
283   Value *CyclesInScops = Builder.CreateLoad(CyclesInScopsPtr, true);
284   CyclesInScops = Builder.CreateAdd(CyclesInScops, CyclesInScop);
285   Builder.CreateStore(CyclesInScops, CyclesInScopsPtr, true);
286 
287   Value *CyclesInCurrentScop = Builder.CreateLoad(CyclesInCurrentScopPtr, true);
288   CyclesInCurrentScop = Builder.CreateAdd(CyclesInCurrentScop, CyclesInScop);
289   Builder.CreateStore(CyclesInCurrentScop, CyclesInCurrentScopPtr, true);
290 
291   Value *TripCountForCurrentScop =
292       Builder.CreateLoad(TripCountForCurrentScopPtr, true);
293   TripCountForCurrentScop =
294       Builder.CreateAdd(TripCountForCurrentScop, Builder.getInt64(1));
295   Builder.CreateStore(TripCountForCurrentScop, TripCountForCurrentScopPtr,
296                       true);
297 }
298