1 //===------ PerfMonitor.cpp - Generate a run-time performance monitor. -======//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 //===----------------------------------------------------------------------===//
11 
12 #include "polly/CodeGen/PerfMonitor.h"
13 #include "polly/CodeGen/RuntimeDebugBuilder.h"
14 #include "polly/ScopInfo.h"
15 #include "llvm/ADT/Triple.h"
16 #include "llvm/IR/Intrinsics.h"
17 #include <sstream>
18 
19 using namespace llvm;
20 using namespace polly;
21 
22 Function *PerfMonitor::getAtExit() {
23   const char *Name = "atexit";
24   Function *F = M->getFunction(Name);
25 
26   if (!F) {
27     GlobalValue::LinkageTypes Linkage = Function::ExternalLinkage;
28     FunctionType *Ty = FunctionType::get(Builder.getInt32Ty(),
29                                          {Builder.getInt8PtrTy()}, false);
30     F = Function::Create(Ty, Linkage, Name, M);
31   }
32 
33   return F;
34 }
35 
36 void PerfMonitor::addToGlobalConstructors(Function *Fn) {
37   const char *Name = "llvm.global_ctors";
38   GlobalVariable *GV = M->getGlobalVariable(Name);
39   std::vector<Constant *> V;
40 
41   if (GV) {
42     Constant *Array = GV->getInitializer();
43     for (Value *X : Array->operand_values())
44       V.push_back(cast<Constant>(X));
45     GV->eraseFromParent();
46   }
47 
48   StructType *ST = StructType::get(Builder.getInt32Ty(), Fn->getType(),
49                                    Builder.getInt8PtrTy());
50 
51   V.push_back(
52       ConstantStruct::get(ST, Builder.getInt32(10), Fn,
53                           ConstantPointerNull::get(Builder.getInt8PtrTy())));
54   ArrayType *Ty = ArrayType::get(ST, V.size());
55 
56   GV = new GlobalVariable(*M, Ty, true, GlobalValue::AppendingLinkage,
57                           ConstantArray::get(Ty, V), Name, nullptr,
58                           GlobalVariable::NotThreadLocal);
59 }
60 
61 Function *PerfMonitor::getRDTSCP() {
62   return Intrinsic::getDeclaration(M, Intrinsic::x86_rdtscp);
63 }
64 
65 PerfMonitor::PerfMonitor(const Scop &S, Module *M)
66     : M(M), Builder(M->getContext()), S(S) {
67   if (Triple(M->getTargetTriple()).getArch() == llvm::Triple::x86_64)
68     Supported = true;
69   else
70     Supported = false;
71 }
72 
73 static void TryRegisterGlobal(Module *M, const char *Name,
74                               Constant *InitialValue, Value **Location) {
75   *Location = M->getGlobalVariable(Name);
76 
77   if (!*Location)
78     *Location = new GlobalVariable(
79         *M, InitialValue->getType(), true, GlobalValue::WeakAnyLinkage,
80         InitialValue, Name, nullptr, GlobalVariable::InitialExecTLSModel);
81 };
82 
83 // Generate a unique name that is usable as a LLVM name for a scop to name its
84 // performance counter.
85 static std::string GetScopUniqueVarname(const Scop &S) {
86   std::stringstream Name;
87   std::string EntryString, ExitString;
88   std::tie(EntryString, ExitString) = S.getEntryExitStr();
89 
90   Name << "__polly_perf_cycles_in_" << std::string(S.getFunction().getName())
91        << "_from__" << EntryString << "__to__" << ExitString;
92   return Name.str();
93 }
94 
95 void PerfMonitor::addScopCounter() {
96   const std::string varname = GetScopUniqueVarname(S);
97   TryRegisterGlobal(M, varname.c_str(), Builder.getInt64(0),
98                     &CyclesInCurrentScopPtr);
99 }
100 
101 void PerfMonitor::addGlobalVariables() {
102   TryRegisterGlobal(M, "__polly_perf_cycles_total_start", Builder.getInt64(0),
103                     &CyclesTotalStartPtr);
104 
105   TryRegisterGlobal(M, "__polly_perf_initialized", Builder.getInt1(0),
106                     &AlreadyInitializedPtr);
107 
108   TryRegisterGlobal(M, "__polly_perf_cycles_in_scops", Builder.getInt64(0),
109                     &CyclesInScopsPtr);
110 
111   TryRegisterGlobal(M, "__polly_perf_cycles_in_scop_start", Builder.getInt64(0),
112                     &CyclesInScopStartPtr);
113 
114   TryRegisterGlobal(M, "__polly_perf_write_loation", Builder.getInt32(0),
115                     &RDTSCPWriteLocation);
116 }
117 
118 static const char *InitFunctionName = "__polly_perf_init";
119 static const char *FinalReportingFunctionName = "__polly_perf_final";
120 
121 static BasicBlock *FinalStartBB = nullptr;
122 static ReturnInst *ReturnFromFinal = nullptr;
123 
124 Function *PerfMonitor::insertFinalReporting() {
125   // Create new function.
126   GlobalValue::LinkageTypes Linkage = Function::WeakODRLinkage;
127   FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), {}, false);
128   Function *ExitFn =
129       Function::Create(Ty, Linkage, FinalReportingFunctionName, M);
130   FinalStartBB = BasicBlock::Create(M->getContext(), "start", ExitFn);
131   Builder.SetInsertPoint(FinalStartBB);
132 
133   if (!Supported) {
134     RuntimeDebugBuilder::createCPUPrinter(
135         Builder, "Polly runtime information generation not supported\n");
136     Builder.CreateRetVoid();
137     return ExitFn;
138   }
139 
140   // Measure current cycles and compute final timings.
141   Function *RDTSCPFn = getRDTSCP();
142   Value *CurrentCycles = Builder.CreateCall(
143       RDTSCPFn,
144       Builder.CreatePointerCast(RDTSCPWriteLocation, Builder.getInt8PtrTy()));
145   Value *CyclesStart = Builder.CreateLoad(CyclesTotalStartPtr, true);
146   Value *CyclesTotal = Builder.CreateSub(CurrentCycles, CyclesStart);
147   Value *CyclesInScops = Builder.CreateLoad(CyclesInScopsPtr, true);
148 
149   // Print the runtime information.
150   RuntimeDebugBuilder::createCPUPrinter(Builder, "Polly runtime information\n");
151   RuntimeDebugBuilder::createCPUPrinter(Builder, "-------------------------\n");
152   RuntimeDebugBuilder::createCPUPrinter(Builder, "Total: ", CyclesTotal, "\n");
153   RuntimeDebugBuilder::createCPUPrinter(Builder, "Scops: ", CyclesInScops,
154                                         "\n");
155 
156   // Print the preamble for per-scop information.
157   RuntimeDebugBuilder::createCPUPrinter(Builder, "\n");
158   RuntimeDebugBuilder::createCPUPrinter(Builder, "Per SCoP information\n");
159   RuntimeDebugBuilder::createCPUPrinter(Builder, "--------------------\n");
160 
161   RuntimeDebugBuilder::createCPUPrinter(
162       Builder, "scop function, "
163                "entry block name, exit block name, total time\n");
164   ReturnFromFinal = Builder.CreateRetVoid();
165   return ExitFn;
166 }
167 
168 void PerfMonitor::AppendScopReporting() {
169   if (!Supported)
170     return;
171 
172   assert(FinalStartBB && "Expected FinalStartBB to be initialized by "
173                          "PerfMonitor::insertFinalReporting.");
174   assert(ReturnFromFinal && "Expected ReturnFromFinal to be initialized by "
175                             "PerfMonitor::insertFinalReporting.");
176 
177   Builder.SetInsertPoint(FinalStartBB);
178   ReturnFromFinal->eraseFromParent();
179 
180   Value *CyclesInCurrentScop =
181       Builder.CreateLoad(this->CyclesInCurrentScopPtr, true);
182   std::string EntryName, ExitName;
183   std::tie(EntryName, ExitName) = S.getEntryExitStr();
184 
185   // print in CSV for easy parsing with other tools.
186   RuntimeDebugBuilder::createCPUPrinter(Builder, S.getFunction().getName(),
187                                         ", ", EntryName, ", ", ExitName, ", ",
188                                         CyclesInCurrentScop, "\n");
189 
190   ReturnFromFinal = Builder.CreateRetVoid();
191 }
192 
193 static Function *FinalReporting = nullptr;
194 
195 void PerfMonitor::initialize() {
196   addGlobalVariables();
197   addScopCounter();
198 
199   // Ensure that we only add the final reporting function once.
200   // On later invocations, append to the reporting function.
201   if (!FinalReporting) {
202     FinalReporting = insertFinalReporting();
203 
204     Function *InitFn = insertInitFunction(FinalReporting);
205     addToGlobalConstructors(InitFn);
206   }
207 
208   AppendScopReporting();
209 }
210 
211 Function *PerfMonitor::insertInitFunction(Function *FinalReporting) {
212   // Insert function definition and BBs.
213   GlobalValue::LinkageTypes Linkage = Function::WeakODRLinkage;
214   FunctionType *Ty = FunctionType::get(Builder.getVoidTy(), {}, false);
215   Function *InitFn = Function::Create(Ty, Linkage, InitFunctionName, M);
216   BasicBlock *Start = BasicBlock::Create(M->getContext(), "start", InitFn);
217   BasicBlock *EarlyReturn =
218       BasicBlock::Create(M->getContext(), "earlyreturn", InitFn);
219   BasicBlock *InitBB = BasicBlock::Create(M->getContext(), "initbb", InitFn);
220 
221   Builder.SetInsertPoint(Start);
222 
223   // Check if this function was already run. If yes, return.
224   //
225   // In case profiling has been enabled in multiple translation units, the
226   // initializer function will be added to the global constructors list of
227   // each translation unit. When merging translation units, the global
228   // constructor lists are just appended, such that the initializer will appear
229   // multiple times. To avoid initializations being run multiple times (and
230   // especially to avoid that atExitFn is called more than once), we bail
231   // out if the intializer is run more than once.
232   Value *HasRunBefore = Builder.CreateLoad(AlreadyInitializedPtr);
233   Builder.CreateCondBr(HasRunBefore, EarlyReturn, InitBB);
234   Builder.SetInsertPoint(EarlyReturn);
235   Builder.CreateRetVoid();
236 
237   // Keep track that this function has been run once.
238   Builder.SetInsertPoint(InitBB);
239   Value *True = Builder.getInt1(true);
240   Builder.CreateStore(True, AlreadyInitializedPtr);
241 
242   // Register the final reporting function with atexit().
243   Value *FinalReportingPtr =
244       Builder.CreatePointerCast(FinalReporting, Builder.getInt8PtrTy());
245   Function *AtExitFn = getAtExit();
246   Builder.CreateCall(AtExitFn, {FinalReportingPtr});
247 
248   if (Supported) {
249     // Read the currently cycle counter and store the result for later.
250     Function *RDTSCPFn = getRDTSCP();
251     Value *CurrentCycles = Builder.CreateCall(
252         RDTSCPFn,
253         Builder.CreatePointerCast(RDTSCPWriteLocation, Builder.getInt8PtrTy()));
254     Builder.CreateStore(CurrentCycles, CyclesTotalStartPtr, true);
255   }
256   Builder.CreateRetVoid();
257 
258   return InitFn;
259 }
260 
261 void PerfMonitor::insertRegionStart(Instruction *InsertBefore) {
262   if (!Supported)
263     return;
264 
265   Builder.SetInsertPoint(InsertBefore);
266   Function *RDTSCPFn = getRDTSCP();
267   Value *CurrentCycles = Builder.CreateCall(
268       RDTSCPFn,
269       Builder.CreatePointerCast(RDTSCPWriteLocation, Builder.getInt8PtrTy()));
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 = Builder.CreateCall(
281       RDTSCPFn,
282       Builder.CreatePointerCast(RDTSCPWriteLocation, Builder.getInt8PtrTy()));
283   Value *CyclesInScop = Builder.CreateSub(CurrentCycles, CyclesStart);
284   Value *CyclesInScops = Builder.CreateLoad(CyclesInScopsPtr, true);
285   CyclesInScops = Builder.CreateAdd(CyclesInScops, CyclesInScop);
286   Builder.CreateStore(CyclesInScops, CyclesInScopsPtr, true);
287 
288   Value *CyclesInCurrentScop = Builder.CreateLoad(CyclesInCurrentScopPtr, true);
289   CyclesInCurrentScop = Builder.CreateAdd(CyclesInCurrentScop, CyclesInScop);
290   Builder.CreateStore(CyclesInCurrentScop, CyclesInCurrentScopPtr, true);
291 }
292