1 // Copyright 2012 the V8 project authors. All rights reserved. 2 // Redistribution and use in source and binary forms, with or without 3 // modification, are permitted provided that the following conditions are 4 // met: 5 // 6 // * Redistributions of source code must retain the above copyright 7 // notice, this list of conditions and the following disclaimer. 8 // * Redistributions in binary form must reproduce the above 9 // copyright notice, this list of conditions and the following 10 // disclaimer in the documentation and/or other materials provided 11 // with the distribution. 12 // * Neither the name of Google Inc. nor the names of its 13 // contributors may be used to endorse or promote products derived 14 // from this software without specific prior written permission. 15 // 16 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 28 #include "v8.h" 29 30 #include "runtime-profiler.h" 31 32 #include "assembler.h" 33 #include "bootstrapper.h" 34 #include "code-stubs.h" 35 #include "compilation-cache.h" 36 #include "deoptimizer.h" 37 #include "execution.h" 38 #include "full-codegen.h" 39 #include "global-handles.h" 40 #include "isolate-inl.h" 41 #include "mark-compact.h" 42 #include "platform.h" 43 #include "scopeinfo.h" 44 45 namespace v8 { 46 namespace internal { 47 48 49 // Optimization sampler constants. 50 static const int kSamplerFrameCount = 2; 51 52 // Constants for statistical profiler. 53 static const int kSamplerFrameWeight[kSamplerFrameCount] = { 2, 1 }; 54 55 static const int kSamplerTicksBetweenThresholdAdjustment = 32; 56 57 static const int kSamplerThresholdInit = 3; 58 static const int kSamplerThresholdMin = 1; 59 static const int kSamplerThresholdDelta = 1; 60 61 static const int kSamplerThresholdSizeFactorInit = 3; 62 63 static const int kSizeLimit = 1500; 64 65 // Constants for counter based profiler. 66 67 // Number of times a function has to be seen on the stack before it is 68 // optimized. 69 static const int kProfilerTicksBeforeOptimization = 2; 70 // If the function optimization was disabled due to high deoptimization count, 71 // but the function is hot and has been seen on the stack this number of times, 72 // then we try to reenable optimization for this function. 73 static const int kProfilerTicksBeforeReenablingOptimization = 250; 74 // If a function does not have enough type info (according to 75 // FLAG_type_info_threshold), but has seen a huge number of ticks, 76 // optimize it as it is. 77 static const int kTicksWhenNotEnoughTypeInfo = 100; 78 // We only have one byte to store the number of ticks. 79 STATIC_ASSERT(kProfilerTicksBeforeOptimization < 256); 80 STATIC_ASSERT(kProfilerTicksBeforeReenablingOptimization < 256); 81 STATIC_ASSERT(kTicksWhenNotEnoughTypeInfo < 256); 82 83 // Maximum size in bytes of generate code for a function to allow OSR. 84 static const int kOSRCodeSizeAllowanceBase = 85 100 * FullCodeGenerator::kCodeSizeMultiplier; 86 87 static const int kOSRCodeSizeAllowancePerTick = 88 3 * FullCodeGenerator::kCodeSizeMultiplier; 89 90 // Maximum size in bytes of generated code for a function to be optimized 91 // the very first time it is seen on the stack. 92 static const int kMaxSizeEarlyOpt = 93 5 * FullCodeGenerator::kCodeSizeMultiplier; 94 95 96 RuntimeProfiler::RuntimeProfiler(Isolate* isolate) 97 : isolate_(isolate), 98 sampler_threshold_(kSamplerThresholdInit), 99 sampler_threshold_size_factor_(kSamplerThresholdSizeFactorInit), 100 sampler_ticks_until_threshold_adjustment_( 101 kSamplerTicksBetweenThresholdAdjustment), 102 sampler_window_position_(0), 103 any_ic_changed_(false), 104 code_generated_(false) { 105 ClearSampleBuffer(); 106 } 107 108 109 static void GetICCounts(Code* shared_code, 110 int* ic_with_type_info_count, 111 int* ic_total_count, 112 int* percentage) { 113 *ic_total_count = 0; 114 *ic_with_type_info_count = 0; 115 Object* raw_info = shared_code->type_feedback_info(); 116 if (raw_info->IsTypeFeedbackInfo()) { 117 TypeFeedbackInfo* info = TypeFeedbackInfo::cast(raw_info); 118 *ic_with_type_info_count = info->ic_with_type_info_count(); 119 *ic_total_count = info->ic_total_count(); 120 } 121 *percentage = *ic_total_count > 0 122 ? 100 * *ic_with_type_info_count / *ic_total_count 123 : 100; 124 } 125 126 127 void RuntimeProfiler::Optimize(JSFunction* function, const char* reason) { 128 ASSERT(function->IsOptimizable()); 129 130 if (FLAG_trace_opt && function->PassesHydrogenFilter()) { 131 PrintF("[marking "); 132 function->ShortPrint(); 133 PrintF(" for recompilation, reason: %s", reason); 134 if (FLAG_type_info_threshold > 0) { 135 int typeinfo, total, percentage; 136 GetICCounts(function->shared()->code(), &typeinfo, &total, &percentage); 137 PrintF(", ICs with typeinfo: %d/%d (%d%%)", typeinfo, total, percentage); 138 } 139 PrintF("]\n"); 140 } 141 142 if (FLAG_parallel_recompilation && !isolate_->bootstrapper()->IsActive()) { 143 ASSERT(!function->IsMarkedForInstallingRecompiledCode()); 144 ASSERT(!function->IsInRecompileQueue()); 145 function->MarkForParallelRecompilation(); 146 } else { 147 // The next call to the function will trigger optimization. 148 function->MarkForLazyRecompilation(); 149 } 150 } 151 152 153 void RuntimeProfiler::AttemptOnStackReplacement(JSFunction* function) { 154 // See AlwaysFullCompiler (in compiler.cc) comment on why we need 155 // Debug::has_break_points(). 156 if (!FLAG_use_osr || 157 isolate_->DebuggerHasBreakPoints() || 158 function->IsBuiltin()) { 159 return; 160 } 161 162 SharedFunctionInfo* shared = function->shared(); 163 // If the code is not optimizable, don't try OSR. 164 if (!shared->code()->optimizable()) return; 165 166 // We are not prepared to do OSR for a function that already has an 167 // allocated arguments object. The optimized code would bypass it for 168 // arguments accesses, which is unsound. Don't try OSR. 169 if (shared->uses_arguments()) return; 170 171 // We're using on-stack replacement: patch the unoptimized code so that 172 // any back edge in any unoptimized frame will trigger on-stack 173 // replacement for that frame. 174 if (FLAG_trace_osr) { 175 PrintF("[patching back edges in "); 176 function->PrintName(); 177 PrintF(" for on-stack replacement]\n"); 178 } 179 180 // Get the interrupt stub code object to match against. We aren't 181 // prepared to generate it, but we don't expect to have to. 182 Code* interrupt_code = NULL; 183 InterruptStub interrupt_stub; 184 bool found_code = interrupt_stub.FindCodeInCache(&interrupt_code, isolate_); 185 if (found_code) { 186 Code* replacement_code = 187 isolate_->builtins()->builtin(Builtins::kOnStackReplacement); 188 Code* unoptimized_code = shared->code(); 189 Deoptimizer::PatchInterruptCode( 190 unoptimized_code, interrupt_code, replacement_code); 191 } 192 } 193 194 195 void RuntimeProfiler::ClearSampleBuffer() { 196 memset(sampler_window_, 0, sizeof(sampler_window_)); 197 memset(sampler_window_weight_, 0, sizeof(sampler_window_weight_)); 198 } 199 200 201 int RuntimeProfiler::LookupSample(JSFunction* function) { 202 int weight = 0; 203 for (int i = 0; i < kSamplerWindowSize; i++) { 204 Object* sample = sampler_window_[i]; 205 if (sample != NULL) { 206 bool fits = FLAG_lookup_sample_by_shared 207 ? (function->shared() == JSFunction::cast(sample)->shared()) 208 : (function == JSFunction::cast(sample)); 209 if (fits) { 210 weight += sampler_window_weight_[i]; 211 } 212 } 213 } 214 return weight; 215 } 216 217 218 void RuntimeProfiler::AddSample(JSFunction* function, int weight) { 219 ASSERT(IsPowerOf2(kSamplerWindowSize)); 220 sampler_window_[sampler_window_position_] = function; 221 sampler_window_weight_[sampler_window_position_] = weight; 222 sampler_window_position_ = (sampler_window_position_ + 1) & 223 (kSamplerWindowSize - 1); 224 } 225 226 227 void RuntimeProfiler::OptimizeNow() { 228 HandleScope scope(isolate_); 229 230 if (isolate_->DebuggerHasBreakPoints()) return; 231 232 if (FLAG_parallel_recompilation) { 233 // Take this as opportunity to process the optimizing compiler thread's 234 // output queue so that it does not unnecessarily keep objects alive. 235 isolate_->optimizing_compiler_thread()->InstallOptimizedFunctions(); 236 } 237 238 // Run through the JavaScript frames and collect them. If we already 239 // have a sample of the function, we mark it for optimizations 240 // (eagerly or lazily). 241 JSFunction* samples[kSamplerFrameCount]; 242 int sample_count = 0; 243 int frame_count = 0; 244 int frame_count_limit = FLAG_watch_ic_patching ? FLAG_frame_count 245 : kSamplerFrameCount; 246 for (JavaScriptFrameIterator it(isolate_); 247 frame_count++ < frame_count_limit && !it.done(); 248 it.Advance()) { 249 JavaScriptFrame* frame = it.frame(); 250 JSFunction* function = frame->function(); 251 252 if (!FLAG_watch_ic_patching) { 253 // Adjust threshold each time we have processed 254 // a certain number of ticks. 255 if (sampler_ticks_until_threshold_adjustment_ > 0) { 256 sampler_ticks_until_threshold_adjustment_--; 257 if (sampler_ticks_until_threshold_adjustment_ <= 0) { 258 // If the threshold is not already at the minimum 259 // modify and reset the ticks until next adjustment. 260 if (sampler_threshold_ > kSamplerThresholdMin) { 261 sampler_threshold_ -= kSamplerThresholdDelta; 262 sampler_ticks_until_threshold_adjustment_ = 263 kSamplerTicksBetweenThresholdAdjustment; 264 } 265 } 266 } 267 } 268 269 SharedFunctionInfo* shared = function->shared(); 270 Code* shared_code = shared->code(); 271 272 if (shared_code->kind() != Code::FUNCTION) continue; 273 if (function->IsInRecompileQueue()) continue; 274 275 if (FLAG_always_osr && 276 shared_code->allow_osr_at_loop_nesting_level() == 0) { 277 // Testing mode: always try an OSR compile for every function. 278 for (int i = 0; i < Code::kMaxLoopNestingMarker; i++) { 279 // TODO(titzer): fix AttemptOnStackReplacement to avoid this dumb loop. 280 shared_code->set_allow_osr_at_loop_nesting_level(i); 281 AttemptOnStackReplacement(function); 282 } 283 // Fall through and do a normal optimized compile as well. 284 } else if (!frame->is_optimized() && 285 (function->IsMarkedForLazyRecompilation() || 286 function->IsMarkedForParallelRecompilation() || 287 function->IsOptimized())) { 288 // Attempt OSR if we are still running unoptimized code even though the 289 // the function has long been marked or even already been optimized. 290 int ticks = shared_code->profiler_ticks(); 291 int allowance = kOSRCodeSizeAllowanceBase + 292 ticks * kOSRCodeSizeAllowancePerTick; 293 if (shared_code->CodeSize() > allowance) { 294 if (ticks < 255) shared_code->set_profiler_ticks(ticks + 1); 295 } else { 296 int nesting = shared_code->allow_osr_at_loop_nesting_level(); 297 if (nesting < Code::kMaxLoopNestingMarker) { 298 int new_nesting = nesting + 1; 299 shared_code->set_allow_osr_at_loop_nesting_level(new_nesting); 300 AttemptOnStackReplacement(function); 301 } 302 } 303 continue; 304 } 305 306 // Only record top-level code on top of the execution stack and 307 // avoid optimizing excessively large scripts since top-level code 308 // will be executed only once. 309 const int kMaxToplevelSourceSize = 10 * 1024; 310 if (shared->is_toplevel() && 311 (frame_count > 1 || shared->SourceSize() > kMaxToplevelSourceSize)) { 312 continue; 313 } 314 315 // Do not record non-optimizable functions. 316 if (shared->optimization_disabled()) { 317 if (shared->deopt_count() >= FLAG_max_opt_count) { 318 // If optimization was disabled due to many deoptimizations, 319 // then check if the function is hot and try to reenable optimization. 320 int ticks = shared_code->profiler_ticks(); 321 if (ticks >= kProfilerTicksBeforeReenablingOptimization) { 322 shared_code->set_profiler_ticks(0); 323 shared->TryReenableOptimization(); 324 } else { 325 shared_code->set_profiler_ticks(ticks + 1); 326 } 327 } 328 continue; 329 } 330 if (!function->IsOptimizable()) continue; 331 332 if (FLAG_watch_ic_patching) { 333 int ticks = shared_code->profiler_ticks(); 334 335 if (ticks >= kProfilerTicksBeforeOptimization) { 336 int typeinfo, total, percentage; 337 GetICCounts(shared_code, &typeinfo, &total, &percentage); 338 if (percentage >= FLAG_type_info_threshold) { 339 // If this particular function hasn't had any ICs patched for enough 340 // ticks, optimize it now. 341 Optimize(function, "hot and stable"); 342 } else if (ticks >= kTicksWhenNotEnoughTypeInfo) { 343 Optimize(function, "not much type info but very hot"); 344 } else { 345 shared_code->set_profiler_ticks(ticks + 1); 346 if (FLAG_trace_opt_verbose) { 347 PrintF("[not yet optimizing "); 348 function->PrintName(); 349 PrintF(", not enough type info: %d/%d (%d%%)]\n", 350 typeinfo, total, percentage); 351 } 352 } 353 } else if (!any_ic_changed_ && 354 shared_code->instruction_size() < kMaxSizeEarlyOpt) { 355 // If no IC was patched since the last tick and this function is very 356 // small, optimistically optimize it now. 357 Optimize(function, "small function"); 358 } else { 359 shared_code->set_profiler_ticks(ticks + 1); 360 } 361 } else { // !FLAG_watch_ic_patching 362 samples[sample_count++] = function; 363 364 int function_size = function->shared()->SourceSize(); 365 int threshold_size_factor = (function_size > kSizeLimit) 366 ? sampler_threshold_size_factor_ 367 : 1; 368 369 int threshold = sampler_threshold_ * threshold_size_factor; 370 371 if (LookupSample(function) >= threshold) { 372 Optimize(function, "sampler window lookup"); 373 } 374 } 375 } 376 if (FLAG_watch_ic_patching) { 377 any_ic_changed_ = false; 378 } else { // !FLAG_watch_ic_patching 379 // Add the collected functions as samples. It's important not to do 380 // this as part of collecting them because this will interfere with 381 // the sample lookup in case of recursive functions. 382 for (int i = 0; i < sample_count; i++) { 383 AddSample(samples[i], kSamplerFrameWeight[i]); 384 } 385 } 386 } 387 388 389 void RuntimeProfiler::SetUp() { 390 if (!FLAG_watch_ic_patching) { 391 ClearSampleBuffer(); 392 } 393 } 394 395 396 void RuntimeProfiler::Reset() { 397 if (!FLAG_watch_ic_patching) { 398 sampler_threshold_ = kSamplerThresholdInit; 399 sampler_threshold_size_factor_ = kSamplerThresholdSizeFactorInit; 400 sampler_ticks_until_threshold_adjustment_ = 401 kSamplerTicksBetweenThresholdAdjustment; 402 } 403 } 404 405 406 void RuntimeProfiler::TearDown() { 407 // Nothing to do. 408 } 409 410 411 // Update the pointers in the sampler window after a GC. 412 void RuntimeProfiler::UpdateSamplesAfterScavenge() { 413 for (int i = 0; i < kSamplerWindowSize; i++) { 414 Object* function = sampler_window_[i]; 415 if (function != NULL && isolate_->heap()->InNewSpace(function)) { 416 MapWord map_word = HeapObject::cast(function)->map_word(); 417 if (map_word.IsForwardingAddress()) { 418 sampler_window_[i] = map_word.ToForwardingAddress(); 419 } else { 420 sampler_window_[i] = NULL; 421 } 422 } 423 } 424 } 425 426 427 void RuntimeProfiler::RemoveDeadSamples() { 428 for (int i = 0; i < kSamplerWindowSize; i++) { 429 Object* function = sampler_window_[i]; 430 if (function != NULL && 431 !Marking::MarkBitFrom(HeapObject::cast(function)).Get()) { 432 sampler_window_[i] = NULL; 433 } 434 } 435 } 436 437 438 void RuntimeProfiler::UpdateSamplesAfterCompact(ObjectVisitor* visitor) { 439 for (int i = 0; i < kSamplerWindowSize; i++) { 440 visitor->VisitPointer(&sampler_window_[i]); 441 } 442 } 443 444 445 } } // namespace v8::internal 446