Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 1 | //===- RSProfiling.cpp - Various profiling using random sampling ----------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file was developed by the LLVM research group and is distributed under |
| 6 | // the University of Illinois Open Source License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // These passes implement a random sampling based profiling. Different methods |
| 11 | // of choosing when to sample are supported, as well as different types of |
| 12 | // profiling. This is done as two passes. The first is a sequence of profiling |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 13 | // passes which insert profiling into the program, and remember what they |
| 14 | // inserted. |
| 15 | // |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 16 | // The second stage duplicates all instructions in a function, ignoring the |
| 17 | // profiling code, then connects the two versions togeather at the entry and at |
| 18 | // backedges. At each connection point a choice is made as to whether to jump |
| 19 | // to the profiled code (take a sample) or execute the unprofiled code. |
| 20 | // |
| 21 | // It is highly recommeneded that after this pass one runs mem2reg and adce |
| 22 | // (instcombine load-vn gdce dse also are good to run afterwards) |
| 23 | // |
| 24 | // This design is intended to make the profiling passes independent of the RS |
| 25 | // framework, but any profiling pass that implements the RSProfiling interface |
| 26 | // is compatible with the rs framework (and thus can be sampled) |
| 27 | // |
| 28 | // TODO: obviously the block and function profiling are almost identical to the |
| 29 | // existing ones, so they can be unified (esp since these passes are valid |
| 30 | // without the rs framework). |
| 31 | // TODO: Fix choice code so that frequency is not hard coded |
| 32 | // |
| 33 | //===----------------------------------------------------------------------===// |
| 34 | |
| 35 | #include "llvm/Pass.h" |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 36 | #include "llvm/Module.h" |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 37 | #include "llvm/Instructions.h" |
| 38 | #include "llvm/Constants.h" |
| 39 | #include "llvm/DerivedTypes.h" |
| 40 | #include "llvm/Transforms/Scalar.h" |
| 41 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 42 | #include "llvm/ADT/Statistic.h" |
| 43 | #include "llvm/Support/CommandLine.h" |
| 44 | #include "llvm/Support/Debug.h" |
| 45 | #include "llvm/Transforms/Instrumentation.h" |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 46 | //#include "ProfilingUtils.h" |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 47 | #include "RSProfiling.h" |
| 48 | |
| 49 | #include <set> |
| 50 | #include <map> |
| 51 | #include <queue> |
| 52 | #include <list> |
| 53 | #include <iostream> |
| 54 | |
| 55 | using namespace llvm; |
| 56 | |
| 57 | namespace { |
| 58 | Statistic<> NumBackEdges("bedge", "Number of BackEdges"); |
| 59 | |
| 60 | enum RandomMeth { |
| 61 | GBV, GBVO, HOSTCC |
| 62 | }; |
| 63 | |
| 64 | cl::opt<RandomMeth> RandomMethod("profile-randomness", |
| 65 | cl::desc("How to randomly choose to profile:"), |
| 66 | cl::values( |
| 67 | clEnumValN(GBV, "global", "global counter"), |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 68 | clEnumValN(GBVO, "ra_global", |
| 69 | "register allocated global counter"), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 70 | clEnumValN(HOSTCC, "rdcc", "cycle counter"), |
| 71 | clEnumValEnd)); |
| 72 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 73 | /// NullProfilerRS - The basic profiler that does nothing. It is the default |
| 74 | /// profiler and thus terminates RSProfiler chains. It is useful for |
| 75 | /// measuring framework overhead |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 76 | class NullProfilerRS : public RSProfilers { |
| 77 | public: |
| 78 | bool isProfiling(Value* v) { |
| 79 | return false; |
| 80 | } |
| 81 | bool runOnModule(Module &M) { |
| 82 | return false; |
| 83 | } |
| 84 | void getAnalysisUsage(AnalysisUsage &AU) const { |
| 85 | AU.setPreservesAll(); |
| 86 | } |
| 87 | }; |
| 88 | |
| 89 | static RegisterAnalysisGroup<RSProfilers> A("Profiling passes"); |
Chris Lattner | 7f8897f | 2006-08-27 22:42:52 +0000 | [diff] [blame^] | 90 | static RegisterPass<NullProfilerRS> NP("insert-null-profiling-rs", |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 91 | "Measure profiling framework overhead"); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 92 | static RegisterAnalysisGroup<RSProfilers, NullProfilerRS, true> NPT; |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 93 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 94 | /// Chooser - Something that chooses when to make a sample of the profiled code |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 95 | class Chooser { |
| 96 | public: |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 97 | /// ProcessChoicePoint - is called for each basic block inserted to choose |
| 98 | /// between normal and sample code |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 99 | virtual void ProcessChoicePoint(BasicBlock*) = 0; |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 100 | /// PrepFunction - is called once per function before other work is done. |
| 101 | /// This gives the opertunity to insert new allocas and such. |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 102 | virtual void PrepFunction(Function*) = 0; |
| 103 | virtual ~Chooser() {} |
| 104 | }; |
| 105 | |
| 106 | //Things that implement sampling policies |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 107 | //A global value that is read-mod-stored to choose when to sample. |
| 108 | //A sample is taken when the global counter hits 0 |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 109 | class GlobalRandomCounter : public Chooser { |
| 110 | GlobalVariable* Counter; |
| 111 | Value* ResetValue; |
| 112 | const Type* T; |
| 113 | public: |
| 114 | GlobalRandomCounter(Module& M, const Type* t, uint64_t resetval); |
| 115 | virtual ~GlobalRandomCounter(); |
| 116 | virtual void PrepFunction(Function* F); |
| 117 | virtual void ProcessChoicePoint(BasicBlock* bb); |
| 118 | }; |
| 119 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 120 | //Same is GRC, but allow register allocation of the global counter |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 121 | class GlobalRandomCounterOpt : public Chooser { |
| 122 | GlobalVariable* Counter; |
| 123 | Value* ResetValue; |
| 124 | AllocaInst* AI; |
| 125 | const Type* T; |
| 126 | public: |
| 127 | GlobalRandomCounterOpt(Module& M, const Type* t, uint64_t resetval); |
| 128 | virtual ~GlobalRandomCounterOpt(); |
| 129 | virtual void PrepFunction(Function* F); |
| 130 | virtual void ProcessChoicePoint(BasicBlock* bb); |
| 131 | }; |
| 132 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 133 | //Use the cycle counter intrinsic as a source of pseudo randomness when |
| 134 | //deciding when to sample. |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 135 | class CycleCounter : public Chooser { |
| 136 | uint64_t rm; |
| 137 | Function* F; |
| 138 | public: |
| 139 | CycleCounter(Module& m, uint64_t resetmask); |
| 140 | virtual ~CycleCounter(); |
| 141 | virtual void PrepFunction(Function* F); |
| 142 | virtual void ProcessChoicePoint(BasicBlock* bb); |
| 143 | }; |
| 144 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 145 | /// ProfilerRS - Insert the random sampling framework |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 146 | struct ProfilerRS : public FunctionPass { |
| 147 | std::map<Value*, Value*> TransCache; |
| 148 | std::set<BasicBlock*> ChoicePoints; |
| 149 | Chooser* c; |
| 150 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 151 | //Translate and duplicate values for the new profile free version of stuff |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 152 | Value* Translate(Value* v); |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 153 | //Duplicate an entire function (with out profiling) |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 154 | void Duplicate(Function& F, RSProfilers& LI); |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 155 | //Called once for each backedge, handle the insertion of choice points and |
| 156 | //the interconection of the two versions of the code |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 157 | void ProcessBackEdge(BasicBlock* src, BasicBlock* dst, Function& F); |
| 158 | bool runOnFunction(Function& F); |
| 159 | bool doInitialization(Module &M); |
| 160 | virtual void getAnalysisUsage(AnalysisUsage &AU) const; |
| 161 | }; |
| 162 | |
Chris Lattner | 7f8897f | 2006-08-27 22:42:52 +0000 | [diff] [blame^] | 163 | RegisterPass<ProfilerRS> X("insert-rs-profiling-framework", |
| 164 | "Insert random sampling instrumentation framework"); |
Chris Lattner | d74ea2b | 2006-05-24 17:04:05 +0000 | [diff] [blame] | 165 | } |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 166 | |
| 167 | //Local utilities |
| 168 | static void ReplacePhiPred(BasicBlock* btarget, |
| 169 | BasicBlock* bold, BasicBlock* bnew); |
| 170 | |
| 171 | static void CollapsePhi(BasicBlock* btarget, BasicBlock* bsrc); |
| 172 | |
| 173 | template<class T> |
| 174 | static void recBackEdge(BasicBlock* bb, T& BackEdges, |
| 175 | std::map<BasicBlock*, int>& color, |
| 176 | std::map<BasicBlock*, int>& depth, |
| 177 | std::map<BasicBlock*, int>& finish, |
| 178 | int& time); |
| 179 | |
| 180 | //find the back edges and where they go to |
| 181 | template<class T> |
| 182 | static void getBackEdges(Function& F, T& BackEdges); |
| 183 | |
| 184 | |
| 185 | /////////////////////////////////////// |
| 186 | // Methods of choosing when to profile |
| 187 | /////////////////////////////////////// |
| 188 | |
| 189 | GlobalRandomCounter::GlobalRandomCounter(Module& M, const Type* t, |
| 190 | uint64_t resetval) : T(t) { |
| 191 | Counter = new GlobalVariable(T, false, GlobalValue::InternalLinkage, |
| 192 | ConstantUInt::get(T, resetval), |
| 193 | "RandomSteeringCounter", &M); |
| 194 | ResetValue = ConstantUInt::get(T, resetval); |
| 195 | } |
| 196 | |
| 197 | GlobalRandomCounter::~GlobalRandomCounter() {} |
| 198 | |
| 199 | void GlobalRandomCounter::PrepFunction(Function* F) {} |
| 200 | |
| 201 | void GlobalRandomCounter::ProcessChoicePoint(BasicBlock* bb) { |
| 202 | BranchInst* t = cast<BranchInst>(bb->getTerminator()); |
| 203 | |
| 204 | //decrement counter |
| 205 | LoadInst* l = new LoadInst(Counter, "counter", t); |
| 206 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 207 | SetCondInst* s = new SetCondInst(Instruction::SetEQ, l, |
| 208 | ConstantUInt::get(T, 0), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 209 | "countercc", t); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 210 | Value* nv = BinaryOperator::createSub(l, ConstantInt::get(T, 1), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 211 | "counternew", t); |
| 212 | new StoreInst(nv, Counter, t); |
| 213 | t->setCondition(s); |
| 214 | |
| 215 | //reset counter |
| 216 | BasicBlock* oldnext = t->getSuccessor(0); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 217 | BasicBlock* resetblock = new BasicBlock("reset", oldnext->getParent(), |
| 218 | oldnext); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 219 | TerminatorInst* t2 = new BranchInst(oldnext, resetblock); |
| 220 | t->setSuccessor(0, resetblock); |
| 221 | new StoreInst(ResetValue, Counter, t2); |
| 222 | ReplacePhiPred(oldnext, bb, resetblock); |
| 223 | } |
| 224 | |
| 225 | GlobalRandomCounterOpt::GlobalRandomCounterOpt(Module& M, const Type* t, |
| 226 | uint64_t resetval) |
| 227 | : AI(0), T(t) { |
| 228 | Counter = new GlobalVariable(T, false, GlobalValue::InternalLinkage, |
| 229 | ConstantUInt::get(T, resetval), |
| 230 | "RandomSteeringCounter", &M); |
| 231 | ResetValue = ConstantUInt::get(T, resetval); |
| 232 | } |
| 233 | |
| 234 | GlobalRandomCounterOpt::~GlobalRandomCounterOpt() {} |
| 235 | |
| 236 | void GlobalRandomCounterOpt::PrepFunction(Function* F) { |
| 237 | //make a local temporary to cache the global |
| 238 | BasicBlock& bb = F->getEntryBlock(); |
| 239 | AI = new AllocaInst(T, 0, "localcounter", bb.begin()); |
| 240 | LoadInst* l = new LoadInst(Counter, "counterload", AI->getNext()); |
| 241 | new StoreInst(l, AI, l->getNext()); |
| 242 | |
Andrew Lenharth | 8dc2d50 | 2005-11-28 18:10:59 +0000 | [diff] [blame] | 243 | //modify all functions and return values to restore the local variable to/from |
| 244 | //the global variable |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 245 | for(Function::iterator fib = F->begin(), fie = F->end(); |
| 246 | fib != fie; ++fib) |
| 247 | for(BasicBlock::iterator bib = fib->begin(), bie = fib->end(); |
| 248 | bib != bie; ++bib) |
| 249 | if (isa<CallInst>(&*bib)) { |
| 250 | LoadInst* l = new LoadInst(AI, "counter", bib); |
| 251 | new StoreInst(l, Counter, bib); |
| 252 | l = new LoadInst(Counter, "counter", bib->getNext()); |
| 253 | new StoreInst(l, AI, l->getNext()); |
| 254 | } else if (isa<InvokeInst>(&*bib)) { |
| 255 | LoadInst* l = new LoadInst(AI, "counter", bib); |
| 256 | new StoreInst(l, Counter, bib); |
| 257 | |
| 258 | BasicBlock* bb = cast<InvokeInst>(&*bib)->getNormalDest(); |
| 259 | Instruction* i = bb->begin(); |
| 260 | while (isa<PHINode>(i)) i = i->getNext(); |
| 261 | l = new LoadInst(Counter, "counter", i); |
| 262 | |
| 263 | bb = cast<InvokeInst>(&*bib)->getUnwindDest(); |
| 264 | i = bb->begin(); |
| 265 | while (isa<PHINode>(i)) i = i->getNext(); |
| 266 | l = new LoadInst(Counter, "counter", i); |
| 267 | new StoreInst(l, AI, l->getNext()); |
| 268 | } else if (isa<UnwindInst>(&*bib) || isa<ReturnInst>(&*bib)) { |
| 269 | LoadInst* l = new LoadInst(AI, "counter", bib); |
| 270 | new StoreInst(l, Counter, bib); |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | void GlobalRandomCounterOpt::ProcessChoicePoint(BasicBlock* bb) { |
| 275 | BranchInst* t = cast<BranchInst>(bb->getTerminator()); |
| 276 | |
| 277 | //decrement counter |
| 278 | LoadInst* l = new LoadInst(AI, "counter", t); |
| 279 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 280 | SetCondInst* s = new SetCondInst(Instruction::SetEQ, l, |
| 281 | ConstantUInt::get(T, 0), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 282 | "countercc", t); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 283 | Value* nv = BinaryOperator::createSub(l, ConstantInt::get(T, 1), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 284 | "counternew", t); |
| 285 | new StoreInst(nv, AI, t); |
| 286 | t->setCondition(s); |
| 287 | |
| 288 | //reset counter |
| 289 | BasicBlock* oldnext = t->getSuccessor(0); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 290 | BasicBlock* resetblock = new BasicBlock("reset", oldnext->getParent(), |
| 291 | oldnext); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 292 | TerminatorInst* t2 = new BranchInst(oldnext, resetblock); |
| 293 | t->setSuccessor(0, resetblock); |
| 294 | new StoreInst(ResetValue, AI, t2); |
| 295 | ReplacePhiPred(oldnext, bb, resetblock); |
| 296 | } |
| 297 | |
| 298 | |
| 299 | CycleCounter::CycleCounter(Module& m, uint64_t resetmask) : rm(resetmask) { |
| 300 | F = m.getOrInsertFunction("llvm.readcyclecounter", Type::ULongTy, NULL); |
| 301 | } |
| 302 | |
| 303 | CycleCounter::~CycleCounter() {} |
| 304 | |
| 305 | void CycleCounter::PrepFunction(Function* F) {} |
| 306 | |
| 307 | void CycleCounter::ProcessChoicePoint(BasicBlock* bb) { |
| 308 | BranchInst* t = cast<BranchInst>(bb->getTerminator()); |
| 309 | |
| 310 | CallInst* c = new CallInst(F, "rdcc", t); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 311 | BinaryOperator* b = |
| 312 | BinaryOperator::createAnd(c, ConstantUInt::get(Type::ULongTy, rm), |
| 313 | "mrdcc", t); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 314 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 315 | SetCondInst* s = new SetCondInst(Instruction::SetEQ, b, |
| 316 | ConstantUInt::get(Type::ULongTy, 0), |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 317 | "mrdccc", t); |
| 318 | t->setCondition(s); |
| 319 | } |
| 320 | |
| 321 | /////////////////////////////////////// |
| 322 | // Profiling: |
| 323 | /////////////////////////////////////// |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 324 | bool RSProfilers_std::isProfiling(Value* v) { |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 325 | if (profcode.find(v) != profcode.end()) |
| 326 | return true; |
| 327 | //else |
| 328 | RSProfilers& LI = getAnalysis<RSProfilers>(); |
| 329 | return LI.isProfiling(v); |
| 330 | } |
| 331 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 332 | void RSProfilers_std::IncrementCounterInBlock(BasicBlock *BB, unsigned CounterNum, |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 333 | GlobalValue *CounterArray) { |
| 334 | // Insert the increment after any alloca or PHI instructions... |
| 335 | BasicBlock::iterator InsertPos = BB->begin(); |
| 336 | while (isa<AllocaInst>(InsertPos) || isa<PHINode>(InsertPos)) |
| 337 | ++InsertPos; |
| 338 | |
| 339 | // Create the getelementptr constant expression |
| 340 | std::vector<Constant*> Indices(2); |
| 341 | Indices[0] = Constant::getNullValue(Type::IntTy); |
| 342 | Indices[1] = ConstantSInt::get(Type::IntTy, CounterNum); |
| 343 | Constant *ElementPtr = ConstantExpr::getGetElementPtr(CounterArray, Indices); |
| 344 | |
| 345 | // Load, increment and store the value back. |
| 346 | Value *OldVal = new LoadInst(ElementPtr, "OldCounter", InsertPos); |
| 347 | profcode.insert(OldVal); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 348 | Value *NewVal = BinaryOperator::createAdd(OldVal, |
| 349 | ConstantInt::get(Type::UIntTy, 1), |
| 350 | "NewCounter", InsertPos); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 351 | profcode.insert(NewVal); |
| 352 | profcode.insert(new StoreInst(NewVal, ElementPtr, InsertPos)); |
| 353 | } |
| 354 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 355 | void RSProfilers_std::getAnalysisUsage(AnalysisUsage &AU) const { |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 356 | //grab any outstanding profiler, or get the null one |
| 357 | AU.addRequired<RSProfilers>(); |
| 358 | } |
| 359 | |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 360 | /////////////////////////////////////// |
| 361 | // RS Framework |
| 362 | /////////////////////////////////////// |
| 363 | |
| 364 | Value* ProfilerRS::Translate(Value* v) { |
| 365 | if(TransCache[v]) |
| 366 | return TransCache[v]; |
| 367 | |
| 368 | if (BasicBlock* bb = dyn_cast<BasicBlock>(v)) { |
| 369 | if (bb == &bb->getParent()->getEntryBlock()) |
| 370 | TransCache[bb] = bb; //don't translate entry block |
| 371 | else |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 372 | TransCache[bb] = new BasicBlock("dup_" + bb->getName(), bb->getParent(), |
| 373 | NULL); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 374 | return TransCache[bb]; |
| 375 | } else if (Instruction* i = dyn_cast<Instruction>(v)) { |
| 376 | //we have already translated this |
| 377 | //do not translate entry block allocas |
| 378 | if(&i->getParent()->getParent()->getEntryBlock() == i->getParent()) { |
| 379 | TransCache[i] = i; |
| 380 | return i; |
| 381 | } else { |
| 382 | //translate this |
| 383 | Instruction* i2 = i->clone(); |
| 384 | if (i->hasName()) |
| 385 | i2->setName("dup_" + i->getName()); |
| 386 | TransCache[i] = i2; |
| 387 | //NumNewInst++; |
| 388 | for (unsigned x = 0; x < i2->getNumOperands(); ++x) |
| 389 | i2->setOperand(x, Translate(i2->getOperand(x))); |
| 390 | return i2; |
| 391 | } |
| 392 | } else if (isa<Function>(v) || isa<Constant>(v) || isa<Argument>(v)) { |
| 393 | TransCache[v] = v; |
| 394 | return v; |
| 395 | } |
| 396 | assert(0 && "Value not handled"); |
Jeff Cohen | 3523f6e | 2005-11-28 06:45:57 +0000 | [diff] [blame] | 397 | return 0; |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 398 | } |
| 399 | |
| 400 | void ProfilerRS::Duplicate(Function& F, RSProfilers& LI) |
| 401 | { |
| 402 | //perform a breadth first search, building up a duplicate of the code |
| 403 | std::queue<BasicBlock*> worklist; |
| 404 | std::set<BasicBlock*> seen; |
| 405 | |
| 406 | //This loop ensures proper BB order, to help performance |
| 407 | for (Function::iterator fib = F.begin(), fie = F.end(); fib != fie; ++fib) |
| 408 | worklist.push(fib); |
| 409 | while (!worklist.empty()) { |
| 410 | Translate(worklist.front()); |
| 411 | worklist.pop(); |
| 412 | } |
| 413 | |
| 414 | //remember than reg2mem created a new entry block we don't want to duplicate |
| 415 | worklist.push(F.getEntryBlock().getTerminator()->getSuccessor(0)); |
| 416 | seen.insert(&F.getEntryBlock()); |
| 417 | |
| 418 | while (!worklist.empty()) { |
| 419 | BasicBlock* bb = worklist.front(); |
| 420 | worklist.pop(); |
| 421 | if(seen.find(bb) == seen.end()) { |
| 422 | BasicBlock* bbtarget = cast<BasicBlock>(Translate(bb)); |
| 423 | BasicBlock::InstListType& instlist = bbtarget->getInstList(); |
| 424 | for (BasicBlock::iterator iib = bb->begin(), iie = bb->end(); |
| 425 | iib != iie; ++iib) { |
| 426 | //NumOldInst++; |
| 427 | if (!LI.isProfiling(&*iib)) { |
| 428 | Instruction* i = cast<Instruction>(Translate(iib)); |
| 429 | instlist.insert(bbtarget->end(), i); |
| 430 | } |
| 431 | } |
| 432 | //updated search state; |
| 433 | seen.insert(bb); |
| 434 | TerminatorInst* ti = bb->getTerminator(); |
| 435 | for (unsigned x = 0; x < ti->getNumSuccessors(); ++x) { |
| 436 | BasicBlock* bbs = ti->getSuccessor(x); |
| 437 | if (seen.find(bbs) == seen.end()) { |
| 438 | worklist.push(bbs); |
| 439 | } |
| 440 | } |
| 441 | } |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | void ProfilerRS::ProcessBackEdge(BasicBlock* src, BasicBlock* dst, Function& F) { |
| 446 | //given a backedge from B -> A, and translations A' and B', |
| 447 | //a: insert C and C' |
| 448 | //b: add branches in C to A and A' and in C' to A and A' |
| 449 | //c: mod terminators@B, replace A with C |
| 450 | //d: mod terminators@B', replace A' with C' |
| 451 | //e: mod phis@A for pred B to be pred C |
| 452 | // if multiple entries, simplify to one |
| 453 | //f: mod phis@A' for pred B' to be pred C' |
| 454 | // if multiple entries, simplify to one |
| 455 | //g: for all phis@A with pred C using x |
| 456 | // add in edge from C' using x' |
| 457 | // add in edge from C using x in A' |
| 458 | |
| 459 | //a: |
| 460 | BasicBlock* bbC = new BasicBlock("choice", &F, src->getNext() ); |
| 461 | //ChoicePoints.insert(bbC); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 462 | BasicBlock* bbCp = |
| 463 | new BasicBlock("choice", &F, cast<BasicBlock>(Translate(src))->getNext() ); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 464 | ChoicePoints.insert(bbCp); |
| 465 | |
| 466 | //b: |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 467 | new BranchInst(cast<BasicBlock>(Translate(dst)), bbC); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 468 | new BranchInst(dst, cast<BasicBlock>(Translate(dst)), |
| 469 | ConstantBool::get(true), bbCp); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 470 | //c: |
| 471 | { |
| 472 | TerminatorInst* iB = src->getTerminator(); |
| 473 | for (unsigned x = 0; x < iB->getNumSuccessors(); ++x) |
| 474 | if (iB->getSuccessor(x) == dst) |
| 475 | iB->setSuccessor(x, bbC); |
| 476 | } |
| 477 | //d: |
| 478 | { |
| 479 | TerminatorInst* iBp = cast<TerminatorInst>(Translate(src->getTerminator())); |
| 480 | for (unsigned x = 0; x < iBp->getNumSuccessors(); ++x) |
| 481 | if (iBp->getSuccessor(x) == cast<BasicBlock>(Translate(dst))) |
| 482 | iBp->setSuccessor(x, bbCp); |
| 483 | } |
| 484 | //e: |
| 485 | ReplacePhiPred(dst, src, bbC); |
| 486 | //src could be a switch, in which case we are replacing several edges with one |
| 487 | //thus collapse those edges int the Phi |
| 488 | CollapsePhi(dst, bbC); |
| 489 | //f: |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 490 | ReplacePhiPred(cast<BasicBlock>(Translate(dst)), |
| 491 | cast<BasicBlock>(Translate(src)),bbCp); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 492 | CollapsePhi(cast<BasicBlock>(Translate(dst)), bbCp); |
| 493 | //g: |
| 494 | for(BasicBlock::iterator ib = dst->begin(), ie = dst->end(); ib != ie; |
| 495 | ++ib) |
| 496 | if (PHINode* phi = dyn_cast<PHINode>(&*ib)) { |
| 497 | for(unsigned x = 0; x < phi->getNumIncomingValues(); ++x) |
| 498 | if(bbC == phi->getIncomingBlock(x)) { |
| 499 | phi->addIncoming(Translate(phi->getIncomingValue(x)), bbCp); |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 500 | cast<PHINode>(Translate(phi))->addIncoming(phi->getIncomingValue(x), |
| 501 | bbC); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 502 | } |
| 503 | phi->removeIncomingValue(bbC); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | bool ProfilerRS::runOnFunction(Function& F) { |
| 508 | if (!F.isExternal()) { |
| 509 | std::set<std::pair<BasicBlock*, BasicBlock*> > BackEdges; |
| 510 | RSProfilers& LI = getAnalysis<RSProfilers>(); |
| 511 | |
| 512 | getBackEdges(F, BackEdges); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 513 | Duplicate(F, LI); |
| 514 | //assume that stuff worked. now connect the duplicated basic blocks |
| 515 | //with the originals in such a way as to preserve ssa. yuk! |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 516 | for (std::set<std::pair<BasicBlock*, BasicBlock*> >::iterator |
| 517 | ib = BackEdges.begin(), ie = BackEdges.end(); ib != ie; ++ib) |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 518 | ProcessBackEdge(ib->first, ib->second, F); |
| 519 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 520 | //oh, and add the edge from the reg2mem created entry node to the |
| 521 | //duplicated second node |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 522 | TerminatorInst* T = F.getEntryBlock().getTerminator(); |
| 523 | ReplaceInstWithInst(T, new BranchInst(T->getSuccessor(0), |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 524 | cast<BasicBlock>(Translate(T->getSuccessor(0))), |
| 525 | ConstantBool::get(true))); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 526 | |
| 527 | //do whatever is needed now that the function is duplicated |
| 528 | c->PrepFunction(&F); |
| 529 | |
| 530 | //add entry node to choice points |
| 531 | ChoicePoints.insert(&F.getEntryBlock()); |
| 532 | |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 533 | for (std::set<BasicBlock*>::iterator |
| 534 | ii = ChoicePoints.begin(), ie = ChoicePoints.end(); ii != ie; ++ii) |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 535 | c->ProcessChoicePoint(*ii); |
| 536 | |
| 537 | ChoicePoints.clear(); |
| 538 | TransCache.clear(); |
| 539 | |
| 540 | return true; |
| 541 | } |
| 542 | return false; |
| 543 | } |
| 544 | |
| 545 | bool ProfilerRS::doInitialization(Module &M) { |
| 546 | switch (RandomMethod) { |
| 547 | case GBV: |
| 548 | c = new GlobalRandomCounter(M, Type::UIntTy, (1 << 14) - 1); |
| 549 | break; |
| 550 | case GBVO: |
| 551 | c = new GlobalRandomCounterOpt(M, Type::UIntTy, (1 << 14) - 1); |
| 552 | break; |
| 553 | case HOSTCC: |
| 554 | c = new CycleCounter(M, (1 << 14) - 1); |
| 555 | break; |
| 556 | }; |
| 557 | return true; |
| 558 | } |
| 559 | |
| 560 | void ProfilerRS::getAnalysisUsage(AnalysisUsage &AU) const { |
| 561 | AU.addRequired<RSProfilers>(); |
| 562 | AU.addRequiredID(DemoteRegisterToMemoryID); |
| 563 | } |
| 564 | |
| 565 | /////////////////////////////////////// |
| 566 | // Utilities: |
| 567 | /////////////////////////////////////// |
| 568 | static void ReplacePhiPred(BasicBlock* btarget, |
| 569 | BasicBlock* bold, BasicBlock* bnew) { |
| 570 | for(BasicBlock::iterator ib = btarget->begin(), ie = btarget->end(); |
| 571 | ib != ie; ++ib) |
| 572 | if (PHINode* phi = dyn_cast<PHINode>(&*ib)) { |
| 573 | for(unsigned x = 0; x < phi->getNumIncomingValues(); ++x) |
| 574 | if(bold == phi->getIncomingBlock(x)) |
| 575 | phi->setIncomingBlock(x, bnew); |
| 576 | } |
| 577 | } |
| 578 | |
| 579 | static void CollapsePhi(BasicBlock* btarget, BasicBlock* bsrc) { |
| 580 | for(BasicBlock::iterator ib = btarget->begin(), ie = btarget->end(); |
| 581 | ib != ie; ++ib) |
| 582 | if (PHINode* phi = dyn_cast<PHINode>(&*ib)) { |
| 583 | unsigned total = phi->getNumIncomingValues(); |
| 584 | std::map<BasicBlock*, Value*> counter; |
| 585 | for(unsigned i = 0; i < phi->getNumIncomingValues(); ) { |
| 586 | if (counter[phi->getIncomingBlock(i)]) { |
Andrew Lenharth | bb227c1 | 2005-11-28 18:00:38 +0000 | [diff] [blame] | 587 | assert(phi->getIncomingValue(i) == counter[phi->getIncomingBlock(i)]); |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 588 | phi->removeIncomingValue(i, false); |
| 589 | } else { |
| 590 | counter[phi->getIncomingBlock(i)] = phi->getIncomingValue(i); |
| 591 | ++i; |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | } |
| 596 | |
| 597 | template<class T> |
| 598 | static void recBackEdge(BasicBlock* bb, T& BackEdges, |
| 599 | std::map<BasicBlock*, int>& color, |
| 600 | std::map<BasicBlock*, int>& depth, |
| 601 | std::map<BasicBlock*, int>& finish, |
| 602 | int& time) |
| 603 | { |
| 604 | color[bb] = 1; |
| 605 | ++time; |
| 606 | depth[bb] = time; |
| 607 | TerminatorInst* t= bb->getTerminator(); |
| 608 | for(unsigned i = 0; i < t->getNumSuccessors(); ++i) { |
| 609 | BasicBlock* bbnew = t->getSuccessor(i); |
| 610 | if (color[bbnew] == 0) |
| 611 | recBackEdge(bbnew, BackEdges, color, depth, finish, time); |
| 612 | else if (color[bbnew] == 1) { |
| 613 | BackEdges.insert(std::make_pair(bb, bbnew)); |
| 614 | //NumBackEdges++; |
| 615 | } |
| 616 | } |
| 617 | color[bb] = 2; |
| 618 | ++time; |
| 619 | finish[bb] = time; |
| 620 | } |
| 621 | |
| 622 | |
| 623 | |
| 624 | //find the back edges and where they go to |
| 625 | template<class T> |
| 626 | static void getBackEdges(Function& F, T& BackEdges) { |
| 627 | std::map<BasicBlock*, int> color; |
| 628 | std::map<BasicBlock*, int> depth; |
| 629 | std::map<BasicBlock*, int> finish; |
| 630 | int time = 0; |
| 631 | recBackEdge(&F.getEntryBlock(), BackEdges, color, depth, finish, time); |
| 632 | DEBUG(std::cerr << F.getName() << " " << BackEdges.size() << "\n"); |
| 633 | } |
| 634 | |
| 635 | |
| 636 | //Creation functions |
Andrew Lenharth | 701f5ac | 2005-11-28 00:58:09 +0000 | [diff] [blame] | 637 | ModulePass* llvm::createNullProfilerRSPass() { |
| 638 | return new NullProfilerRS(); |
| 639 | } |
| 640 | |
| 641 | FunctionPass* llvm::createRSProfilingPass() { |
| 642 | return new ProfilerRS(); |
| 643 | } |