1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/share/vm/runtime/deoptimization.cpp Wed Apr 27 01:25:04 2016 +0800 1.3 @@ -0,0 +1,2010 @@ 1.4 +/* 1.5 + * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 1.23 + * or visit www.oracle.com if you need additional information or have any 1.24 + * questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#include "precompiled.hpp" 1.29 +#include "classfile/systemDictionary.hpp" 1.30 +#include "code/debugInfoRec.hpp" 1.31 +#include "code/nmethod.hpp" 1.32 +#include "code/pcDesc.hpp" 1.33 +#include "code/scopeDesc.hpp" 1.34 +#include "interpreter/bytecode.hpp" 1.35 +#include "interpreter/interpreter.hpp" 1.36 +#include "interpreter/oopMapCache.hpp" 1.37 +#include "memory/allocation.inline.hpp" 1.38 +#include "memory/oopFactory.hpp" 1.39 +#include "memory/resourceArea.hpp" 1.40 +#include "oops/method.hpp" 1.41 +#include "oops/oop.inline.hpp" 1.42 +#include "prims/jvmtiThreadState.hpp" 1.43 +#include "runtime/biasedLocking.hpp" 1.44 +#include "runtime/compilationPolicy.hpp" 1.45 +#include "runtime/deoptimization.hpp" 1.46 +#include "runtime/interfaceSupport.hpp" 1.47 +#include "runtime/sharedRuntime.hpp" 1.48 +#include "runtime/signature.hpp" 1.49 +#include "runtime/stubRoutines.hpp" 1.50 +#include "runtime/thread.hpp" 1.51 +#include "runtime/vframe.hpp" 1.52 +#include "runtime/vframeArray.hpp" 1.53 +#include "runtime/vframe_hp.hpp" 1.54 +#include "utilities/events.hpp" 1.55 +#include "utilities/xmlstream.hpp" 1.56 +#ifdef TARGET_ARCH_x86 1.57 +# include "vmreg_x86.inline.hpp" 1.58 +#endif 1.59 +#ifdef TARGET_ARCH_sparc 1.60 +# include "vmreg_sparc.inline.hpp" 1.61 +#endif 1.62 +#ifdef TARGET_ARCH_zero 1.63 +# include "vmreg_zero.inline.hpp" 1.64 +#endif 1.65 +#ifdef TARGET_ARCH_arm 1.66 +# include "vmreg_arm.inline.hpp" 1.67 +#endif 1.68 +#ifdef TARGET_ARCH_ppc 1.69 +# include "vmreg_ppc.inline.hpp" 1.70 +#endif 1.71 +#ifdef COMPILER2 1.72 +#ifdef TARGET_ARCH_MODEL_x86_32 1.73 +# include "adfiles/ad_x86_32.hpp" 1.74 +#endif 1.75 +#ifdef TARGET_ARCH_MODEL_x86_64 1.76 +# include "adfiles/ad_x86_64.hpp" 1.77 +#endif 1.78 +#ifdef TARGET_ARCH_MODEL_sparc 1.79 +# include "adfiles/ad_sparc.hpp" 1.80 +#endif 1.81 +#ifdef TARGET_ARCH_MODEL_zero 1.82 +# include "adfiles/ad_zero.hpp" 1.83 +#endif 1.84 +#ifdef TARGET_ARCH_MODEL_arm 1.85 +# include "adfiles/ad_arm.hpp" 1.86 +#endif 1.87 +#ifdef TARGET_ARCH_MODEL_ppc_32 1.88 +# include "adfiles/ad_ppc_32.hpp" 1.89 +#endif 1.90 +#ifdef TARGET_ARCH_MODEL_ppc_64 1.91 +# include "adfiles/ad_ppc_64.hpp" 1.92 +#endif 1.93 +#endif // COMPILER2 1.94 + 1.95 +PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 1.96 + 1.97 +bool DeoptimizationMarker::_is_active = false; 1.98 + 1.99 +Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 1.100 + int caller_adjustment, 1.101 + int caller_actual_parameters, 1.102 + int number_of_frames, 1.103 + intptr_t* frame_sizes, 1.104 + address* frame_pcs, 1.105 + BasicType return_type) { 1.106 + _size_of_deoptimized_frame = size_of_deoptimized_frame; 1.107 + _caller_adjustment = caller_adjustment; 1.108 + _caller_actual_parameters = caller_actual_parameters; 1.109 + _number_of_frames = number_of_frames; 1.110 + _frame_sizes = frame_sizes; 1.111 + _frame_pcs = frame_pcs; 1.112 + _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 1.113 + _return_type = return_type; 1.114 + _initial_info = 0; 1.115 + // PD (x86 only) 1.116 + _counter_temp = 0; 1.117 + _unpack_kind = 0; 1.118 + _sender_sp_temp = 0; 1.119 + 1.120 + _total_frame_sizes = size_of_frames(); 1.121 +} 1.122 + 1.123 + 1.124 +Deoptimization::UnrollBlock::~UnrollBlock() { 1.125 + FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); 1.126 + FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); 1.127 + FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); 1.128 +} 1.129 + 1.130 + 1.131 +intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 1.132 + assert(register_number < RegisterMap::reg_count, "checking register number"); 1.133 + return &_register_block[register_number * 2]; 1.134 +} 1.135 + 1.136 + 1.137 + 1.138 +int Deoptimization::UnrollBlock::size_of_frames() const { 1.139 + // Acount first for the adjustment of the initial frame 1.140 + int result = _caller_adjustment; 1.141 + for (int index = 0; index < number_of_frames(); index++) { 1.142 + result += frame_sizes()[index]; 1.143 + } 1.144 + return result; 1.145 +} 1.146 + 1.147 + 1.148 +void Deoptimization::UnrollBlock::print() { 1.149 + ttyLocker ttyl; 1.150 + tty->print_cr("UnrollBlock"); 1.151 + tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 1.152 + tty->print( " frame_sizes: "); 1.153 + for (int index = 0; index < number_of_frames(); index++) { 1.154 + tty->print("%d ", frame_sizes()[index]); 1.155 + } 1.156 + tty->cr(); 1.157 +} 1.158 + 1.159 + 1.160 +// In order to make fetch_unroll_info work properly with escape 1.161 +// analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 1.162 +// ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 1.163 +// of previously eliminated objects occurs in realloc_objects, which is 1.164 +// called from the method fetch_unroll_info_helper below. 1.165 +JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 1.166 + // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 1.167 + // but makes the entry a little slower. There is however a little dance we have to 1.168 + // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 1.169 + 1.170 + // fetch_unroll_info() is called at the beginning of the deoptimization 1.171 + // handler. Note this fact before we start generating temporary frames 1.172 + // that can confuse an asynchronous stack walker. This counter is 1.173 + // decremented at the end of unpack_frames(). 1.174 + thread->inc_in_deopt_handler(); 1.175 + 1.176 + return fetch_unroll_info_helper(thread); 1.177 +JRT_END 1.178 + 1.179 + 1.180 +// This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 1.181 +Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 1.182 + 1.183 + // Note: there is a safepoint safety issue here. No matter whether we enter 1.184 + // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 1.185 + // the vframeArray is created. 1.186 + // 1.187 + 1.188 + // Allocate our special deoptimization ResourceMark 1.189 + DeoptResourceMark* dmark = new DeoptResourceMark(thread); 1.190 + assert(thread->deopt_mark() == NULL, "Pending deopt!"); 1.191 + thread->set_deopt_mark(dmark); 1.192 + 1.193 + frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 1.194 + RegisterMap map(thread, true); 1.195 + RegisterMap dummy_map(thread, false); 1.196 + // Now get the deoptee with a valid map 1.197 + frame deoptee = stub_frame.sender(&map); 1.198 + // Set the deoptee nmethod 1.199 + assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 1.200 + thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 1.201 + 1.202 + if (VerifyStack) { 1.203 + thread->validate_frame_layout(); 1.204 + } 1.205 + 1.206 + // Create a growable array of VFrames where each VFrame represents an inlined 1.207 + // Java frame. This storage is allocated with the usual system arena. 1.208 + assert(deoptee.is_compiled_frame(), "Wrong frame type"); 1.209 + GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 1.210 + vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 1.211 + while (!vf->is_top()) { 1.212 + assert(vf->is_compiled_frame(), "Wrong frame type"); 1.213 + chunk->push(compiledVFrame::cast(vf)); 1.214 + vf = vf->sender(); 1.215 + } 1.216 + assert(vf->is_compiled_frame(), "Wrong frame type"); 1.217 + chunk->push(compiledVFrame::cast(vf)); 1.218 + 1.219 +#ifdef COMPILER2 1.220 + // Reallocate the non-escaping objects and restore their fields. Then 1.221 + // relock objects if synchronization on them was eliminated. 1.222 + if (DoEscapeAnalysis || EliminateNestedLocks) { 1.223 + if (EliminateAllocations) { 1.224 + assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 1.225 + GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 1.226 + 1.227 + // The flag return_oop() indicates call sites which return oop 1.228 + // in compiled code. Such sites include java method calls, 1.229 + // runtime calls (for example, used to allocate new objects/arrays 1.230 + // on slow code path) and any other calls generated in compiled code. 1.231 + // It is not guaranteed that we can get such information here only 1.232 + // by analyzing bytecode in deoptimized frames. This is why this flag 1.233 + // is set during method compilation (see Compile::Process_OopMap_Node()). 1.234 + bool save_oop_result = chunk->at(0)->scope()->return_oop(); 1.235 + Handle return_value; 1.236 + if (save_oop_result) { 1.237 + // Reallocation may trigger GC. If deoptimization happened on return from 1.238 + // call which returns oop we need to save it since it is not in oopmap. 1.239 + oop result = deoptee.saved_oop_result(&map); 1.240 + assert(result == NULL || result->is_oop(), "must be oop"); 1.241 + return_value = Handle(thread, result); 1.242 + assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 1.243 + if (TraceDeoptimization) { 1.244 + ttyLocker ttyl; 1.245 + tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread); 1.246 + } 1.247 + } 1.248 + bool reallocated = false; 1.249 + if (objects != NULL) { 1.250 + JRT_BLOCK 1.251 + reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 1.252 + JRT_END 1.253 + } 1.254 + if (reallocated) { 1.255 + reassign_fields(&deoptee, &map, objects); 1.256 +#ifndef PRODUCT 1.257 + if (TraceDeoptimization) { 1.258 + ttyLocker ttyl; 1.259 + tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 1.260 + print_objects(objects); 1.261 + } 1.262 +#endif 1.263 + } 1.264 + if (save_oop_result) { 1.265 + // Restore result. 1.266 + deoptee.set_saved_oop_result(&map, return_value()); 1.267 + } 1.268 + } 1.269 + if (EliminateLocks) { 1.270 +#ifndef PRODUCT 1.271 + bool first = true; 1.272 +#endif 1.273 + for (int i = 0; i < chunk->length(); i++) { 1.274 + compiledVFrame* cvf = chunk->at(i); 1.275 + assert (cvf->scope() != NULL,"expect only compiled java frames"); 1.276 + GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1.277 + if (monitors->is_nonempty()) { 1.278 + relock_objects(monitors, thread); 1.279 +#ifndef PRODUCT 1.280 + if (TraceDeoptimization) { 1.281 + ttyLocker ttyl; 1.282 + for (int j = 0; j < monitors->length(); j++) { 1.283 + MonitorInfo* mi = monitors->at(j); 1.284 + if (mi->eliminated()) { 1.285 + if (first) { 1.286 + first = false; 1.287 + tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 1.288 + } 1.289 + tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner()); 1.290 + } 1.291 + } 1.292 + } 1.293 +#endif 1.294 + } 1.295 + } 1.296 + } 1.297 + } 1.298 +#endif // COMPILER2 1.299 + // Ensure that no safepoint is taken after pointers have been stored 1.300 + // in fields of rematerialized objects. If a safepoint occurs from here on 1.301 + // out the java state residing in the vframeArray will be missed. 1.302 + No_Safepoint_Verifier no_safepoint; 1.303 + 1.304 + vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 1.305 + 1.306 + assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 1.307 + thread->set_vframe_array_head(array); 1.308 + 1.309 + // Now that the vframeArray has been created if we have any deferred local writes 1.310 + // added by jvmti then we can free up that structure as the data is now in the 1.311 + // vframeArray 1.312 + 1.313 + if (thread->deferred_locals() != NULL) { 1.314 + GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 1.315 + int i = 0; 1.316 + do { 1.317 + // Because of inlining we could have multiple vframes for a single frame 1.318 + // and several of the vframes could have deferred writes. Find them all. 1.319 + if (list->at(i)->id() == array->original().id()) { 1.320 + jvmtiDeferredLocalVariableSet* dlv = list->at(i); 1.321 + list->remove_at(i); 1.322 + // individual jvmtiDeferredLocalVariableSet are CHeapObj's 1.323 + delete dlv; 1.324 + } else { 1.325 + i++; 1.326 + } 1.327 + } while ( i < list->length() ); 1.328 + if (list->length() == 0) { 1.329 + thread->set_deferred_locals(NULL); 1.330 + // free the list and elements back to C heap. 1.331 + delete list; 1.332 + } 1.333 + 1.334 + } 1.335 + 1.336 +#ifndef SHARK 1.337 + // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 1.338 + CodeBlob* cb = stub_frame.cb(); 1.339 + // Verify we have the right vframeArray 1.340 + assert(cb->frame_size() >= 0, "Unexpected frame size"); 1.341 + intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 1.342 + 1.343 + // If the deopt call site is a MethodHandle invoke call site we have 1.344 + // to adjust the unpack_sp. 1.345 + nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 1.346 + if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 1.347 + unpack_sp = deoptee.unextended_sp(); 1.348 + 1.349 +#ifdef ASSERT 1.350 + assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 1.351 +#endif 1.352 +#else 1.353 + intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 1.354 +#endif // !SHARK 1.355 + 1.356 + // This is a guarantee instead of an assert because if vframe doesn't match 1.357 + // we will unpack the wrong deoptimized frame and wind up in strange places 1.358 + // where it will be very difficult to figure out what went wrong. Better 1.359 + // to die an early death here than some very obscure death later when the 1.360 + // trail is cold. 1.361 + // Note: on ia64 this guarantee can be fooled by frames with no memory stack 1.362 + // in that it will fail to detect a problem when there is one. This needs 1.363 + // more work in tiger timeframe. 1.364 + guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 1.365 + 1.366 + int number_of_frames = array->frames(); 1.367 + 1.368 + // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 1.369 + // virtual activation, which is the reverse of the elements in the vframes array. 1.370 + intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 1.371 + // +1 because we always have an interpreter return address for the final slot. 1.372 + address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 1.373 + int popframe_extra_args = 0; 1.374 + // Create an interpreter return address for the stub to use as its return 1.375 + // address so the skeletal frames are perfectly walkable 1.376 + frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 1.377 + 1.378 + // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 1.379 + // activation be put back on the expression stack of the caller for reexecution 1.380 + if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 1.381 + popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 1.382 + } 1.383 + 1.384 + // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 1.385 + // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 1.386 + // than simply use array->sender.pc(). This requires us to walk the current set of frames 1.387 + // 1.388 + frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 1.389 + deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 1.390 + 1.391 + // It's possible that the number of paramters at the call site is 1.392 + // different than number of arguments in the callee when method 1.393 + // handles are used. If the caller is interpreted get the real 1.394 + // value so that the proper amount of space can be added to it's 1.395 + // frame. 1.396 + bool caller_was_method_handle = false; 1.397 + if (deopt_sender.is_interpreted_frame()) { 1.398 + methodHandle method = deopt_sender.interpreter_frame_method(); 1.399 + Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 1.400 + if (cur.is_invokedynamic() || cur.is_invokehandle()) { 1.401 + // Method handle invokes may involve fairly arbitrary chains of 1.402 + // calls so it's impossible to know how much actual space the 1.403 + // caller has for locals. 1.404 + caller_was_method_handle = true; 1.405 + } 1.406 + } 1.407 + 1.408 + // 1.409 + // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 1.410 + // frame_sizes/frame_pcs[1] next oldest frame (int) 1.411 + // frame_sizes/frame_pcs[n] youngest frame (int) 1.412 + // 1.413 + // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 1.414 + // owns the space for the return address to it's caller). Confusing ain't it. 1.415 + // 1.416 + // The vframe array can address vframes with indices running from 1.417 + // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 1.418 + // When we create the skeletal frames we need the oldest frame to be in the zero slot 1.419 + // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 1.420 + // so things look a little strange in this loop. 1.421 + // 1.422 + int callee_parameters = 0; 1.423 + int callee_locals = 0; 1.424 + for (int index = 0; index < array->frames(); index++ ) { 1.425 + // frame[number_of_frames - 1 ] = on_stack_size(youngest) 1.426 + // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 1.427 + // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 1.428 + frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 1.429 + callee_locals, 1.430 + index == 0, 1.431 + popframe_extra_args); 1.432 + // This pc doesn't have to be perfect just good enough to identify the frame 1.433 + // as interpreted so the skeleton frame will be walkable 1.434 + // The correct pc will be set when the skeleton frame is completely filled out 1.435 + // The final pc we store in the loop is wrong and will be overwritten below 1.436 + frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 1.437 + 1.438 + callee_parameters = array->element(index)->method()->size_of_parameters(); 1.439 + callee_locals = array->element(index)->method()->max_locals(); 1.440 + popframe_extra_args = 0; 1.441 + } 1.442 + 1.443 + // Compute whether the root vframe returns a float or double value. 1.444 + BasicType return_type; 1.445 + { 1.446 + HandleMark hm; 1.447 + methodHandle method(thread, array->element(0)->method()); 1.448 + Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 1.449 + return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 1.450 + } 1.451 + 1.452 + // Compute information for handling adapters and adjusting the frame size of the caller. 1.453 + int caller_adjustment = 0; 1.454 + 1.455 + // Compute the amount the oldest interpreter frame will have to adjust 1.456 + // its caller's stack by. If the caller is a compiled frame then 1.457 + // we pretend that the callee has no parameters so that the 1.458 + // extension counts for the full amount of locals and not just 1.459 + // locals-parms. This is because without a c2i adapter the parm 1.460 + // area as created by the compiled frame will not be usable by 1.461 + // the interpreter. (Depending on the calling convention there 1.462 + // may not even be enough space). 1.463 + 1.464 + // QQQ I'd rather see this pushed down into last_frame_adjust 1.465 + // and have it take the sender (aka caller). 1.466 + 1.467 + if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 1.468 + caller_adjustment = last_frame_adjust(0, callee_locals); 1.469 + } else if (callee_locals > callee_parameters) { 1.470 + // The caller frame may need extending to accommodate 1.471 + // non-parameter locals of the first unpacked interpreted frame. 1.472 + // Compute that adjustment. 1.473 + caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 1.474 + } 1.475 + 1.476 + // If the sender is deoptimized the we must retrieve the address of the handler 1.477 + // since the frame will "magically" show the original pc before the deopt 1.478 + // and we'd undo the deopt. 1.479 + 1.480 + frame_pcs[0] = deopt_sender.raw_pc(); 1.481 + 1.482 +#ifndef SHARK 1.483 + assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 1.484 +#endif // SHARK 1.485 + 1.486 + UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 1.487 + caller_adjustment * BytesPerWord, 1.488 + caller_was_method_handle ? 0 : callee_parameters, 1.489 + number_of_frames, 1.490 + frame_sizes, 1.491 + frame_pcs, 1.492 + return_type); 1.493 + // On some platforms, we need a way to pass some platform dependent 1.494 + // information to the unpacking code so the skeletal frames come out 1.495 + // correct (initial fp value, unextended sp, ...) 1.496 + info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 1.497 + 1.498 + if (array->frames() > 1) { 1.499 + if (VerifyStack && TraceDeoptimization) { 1.500 + ttyLocker ttyl; 1.501 + tty->print_cr("Deoptimizing method containing inlining"); 1.502 + } 1.503 + } 1.504 + 1.505 + array->set_unroll_block(info); 1.506 + return info; 1.507 +} 1.508 + 1.509 +// Called to cleanup deoptimization data structures in normal case 1.510 +// after unpacking to stack and when stack overflow error occurs 1.511 +void Deoptimization::cleanup_deopt_info(JavaThread *thread, 1.512 + vframeArray *array) { 1.513 + 1.514 + // Get array if coming from exception 1.515 + if (array == NULL) { 1.516 + array = thread->vframe_array_head(); 1.517 + } 1.518 + thread->set_vframe_array_head(NULL); 1.519 + 1.520 + // Free the previous UnrollBlock 1.521 + vframeArray* old_array = thread->vframe_array_last(); 1.522 + thread->set_vframe_array_last(array); 1.523 + 1.524 + if (old_array != NULL) { 1.525 + UnrollBlock* old_info = old_array->unroll_block(); 1.526 + old_array->set_unroll_block(NULL); 1.527 + delete old_info; 1.528 + delete old_array; 1.529 + } 1.530 + 1.531 + // Deallocate any resource creating in this routine and any ResourceObjs allocated 1.532 + // inside the vframeArray (StackValueCollections) 1.533 + 1.534 + delete thread->deopt_mark(); 1.535 + thread->set_deopt_mark(NULL); 1.536 + thread->set_deopt_nmethod(NULL); 1.537 + 1.538 + 1.539 + if (JvmtiExport::can_pop_frame()) { 1.540 +#ifndef CC_INTERP 1.541 + // Regardless of whether we entered this routine with the pending 1.542 + // popframe condition bit set, we should always clear it now 1.543 + thread->clear_popframe_condition(); 1.544 +#else 1.545 + // C++ interpeter will clear has_pending_popframe when it enters 1.546 + // with method_resume. For deopt_resume2 we clear it now. 1.547 + if (thread->popframe_forcing_deopt_reexecution()) 1.548 + thread->clear_popframe_condition(); 1.549 +#endif /* CC_INTERP */ 1.550 + } 1.551 + 1.552 + // unpack_frames() is called at the end of the deoptimization handler 1.553 + // and (in C2) at the end of the uncommon trap handler. Note this fact 1.554 + // so that an asynchronous stack walker can work again. This counter is 1.555 + // incremented at the beginning of fetch_unroll_info() and (in C2) at 1.556 + // the beginning of uncommon_trap(). 1.557 + thread->dec_in_deopt_handler(); 1.558 +} 1.559 + 1.560 + 1.561 +// Return BasicType of value being returned 1.562 +JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 1.563 + 1.564 + // We are already active int he special DeoptResourceMark any ResourceObj's we 1.565 + // allocate will be freed at the end of the routine. 1.566 + 1.567 + // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 1.568 + // but makes the entry a little slower. There is however a little dance we have to 1.569 + // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 1.570 + ResetNoHandleMark rnhm; // No-op in release/product versions 1.571 + HandleMark hm; 1.572 + 1.573 + frame stub_frame = thread->last_frame(); 1.574 + 1.575 + // Since the frame to unpack is the top frame of this thread, the vframe_array_head 1.576 + // must point to the vframeArray for the unpack frame. 1.577 + vframeArray* array = thread->vframe_array_head(); 1.578 + 1.579 +#ifndef PRODUCT 1.580 + if (TraceDeoptimization) { 1.581 + ttyLocker ttyl; 1.582 + tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 1.583 + } 1.584 +#endif 1.585 + Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 1.586 + stub_frame.pc(), stub_frame.sp(), exec_mode); 1.587 + 1.588 + UnrollBlock* info = array->unroll_block(); 1.589 + 1.590 + // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 1.591 + array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 1.592 + 1.593 + BasicType bt = info->return_type(); 1.594 + 1.595 + // If we have an exception pending, claim that the return type is an oop 1.596 + // so the deopt_blob does not overwrite the exception_oop. 1.597 + 1.598 + if (exec_mode == Unpack_exception) 1.599 + bt = T_OBJECT; 1.600 + 1.601 + // Cleanup thread deopt data 1.602 + cleanup_deopt_info(thread, array); 1.603 + 1.604 +#ifndef PRODUCT 1.605 + if (VerifyStack) { 1.606 + ResourceMark res_mark; 1.607 + 1.608 + thread->validate_frame_layout(); 1.609 + 1.610 + // Verify that the just-unpacked frames match the interpreter's 1.611 + // notions of expression stack and locals 1.612 + vframeArray* cur_array = thread->vframe_array_last(); 1.613 + RegisterMap rm(thread, false); 1.614 + rm.set_include_argument_oops(false); 1.615 + bool is_top_frame = true; 1.616 + int callee_size_of_parameters = 0; 1.617 + int callee_max_locals = 0; 1.618 + for (int i = 0; i < cur_array->frames(); i++) { 1.619 + vframeArrayElement* el = cur_array->element(i); 1.620 + frame* iframe = el->iframe(); 1.621 + guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 1.622 + 1.623 + // Get the oop map for this bci 1.624 + InterpreterOopMap mask; 1.625 + int cur_invoke_parameter_size = 0; 1.626 + bool try_next_mask = false; 1.627 + int next_mask_expression_stack_size = -1; 1.628 + int top_frame_expression_stack_adjustment = 0; 1.629 + methodHandle mh(thread, iframe->interpreter_frame_method()); 1.630 + OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 1.631 + BytecodeStream str(mh); 1.632 + str.set_start(iframe->interpreter_frame_bci()); 1.633 + int max_bci = mh->code_size(); 1.634 + // Get to the next bytecode if possible 1.635 + assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 1.636 + // Check to see if we can grab the number of outgoing arguments 1.637 + // at an uncommon trap for an invoke (where the compiler 1.638 + // generates debug info before the invoke has executed) 1.639 + Bytecodes::Code cur_code = str.next(); 1.640 + if (cur_code == Bytecodes::_invokevirtual || 1.641 + cur_code == Bytecodes::_invokespecial || 1.642 + cur_code == Bytecodes::_invokestatic || 1.643 + cur_code == Bytecodes::_invokeinterface || 1.644 + cur_code == Bytecodes::_invokedynamic) { 1.645 + Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 1.646 + Symbol* signature = invoke.signature(); 1.647 + ArgumentSizeComputer asc(signature); 1.648 + cur_invoke_parameter_size = asc.size(); 1.649 + if (invoke.has_receiver()) { 1.650 + // Add in receiver 1.651 + ++cur_invoke_parameter_size; 1.652 + } 1.653 + if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 1.654 + callee_size_of_parameters++; 1.655 + } 1.656 + } 1.657 + if (str.bci() < max_bci) { 1.658 + Bytecodes::Code bc = str.next(); 1.659 + if (bc >= 0) { 1.660 + // The interpreter oop map generator reports results before 1.661 + // the current bytecode has executed except in the case of 1.662 + // calls. It seems to be hard to tell whether the compiler 1.663 + // has emitted debug information matching the "state before" 1.664 + // a given bytecode or the state after, so we try both 1.665 + switch (cur_code) { 1.666 + case Bytecodes::_invokevirtual: 1.667 + case Bytecodes::_invokespecial: 1.668 + case Bytecodes::_invokestatic: 1.669 + case Bytecodes::_invokeinterface: 1.670 + case Bytecodes::_invokedynamic: 1.671 + case Bytecodes::_athrow: 1.672 + break; 1.673 + default: { 1.674 + InterpreterOopMap next_mask; 1.675 + OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 1.676 + next_mask_expression_stack_size = next_mask.expression_stack_size(); 1.677 + // Need to subtract off the size of the result type of 1.678 + // the bytecode because this is not described in the 1.679 + // debug info but returned to the interpreter in the TOS 1.680 + // caching register 1.681 + BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 1.682 + if (bytecode_result_type != T_ILLEGAL) { 1.683 + top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 1.684 + } 1.685 + assert(top_frame_expression_stack_adjustment >= 0, ""); 1.686 + try_next_mask = true; 1.687 + break; 1.688 + } 1.689 + } 1.690 + } 1.691 + } 1.692 + 1.693 + // Verify stack depth and oops in frame 1.694 + // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 1.695 + if (!( 1.696 + /* SPARC */ 1.697 + (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 1.698 + /* x86 */ 1.699 + (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 1.700 + (try_next_mask && 1.701 + (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 1.702 + top_frame_expression_stack_adjustment))) || 1.703 + (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 1.704 + (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 1.705 + (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 1.706 + )) { 1.707 + ttyLocker ttyl; 1.708 + 1.709 + // Print out some information that will help us debug the problem 1.710 + tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1.711 + tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 1.712 + tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1.713 + iframe->interpreter_frame_expression_stack_size()); 1.714 + tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1.715 + tty->print_cr(" try_next_mask = %d", try_next_mask); 1.716 + tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 1.717 + tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1.718 + tty->print_cr(" callee_max_locals = %d", callee_max_locals); 1.719 + tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1.720 + tty->print_cr(" exec_mode = %d", exec_mode); 1.721 + tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1.722 + tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 1.723 + tty->print_cr(" Interpreted frames:"); 1.724 + for (int k = 0; k < cur_array->frames(); k++) { 1.725 + vframeArrayElement* el = cur_array->element(k); 1.726 + tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1.727 + } 1.728 + cur_array->print_on_2(tty); 1.729 + guarantee(false, "wrong number of expression stack elements during deopt"); 1.730 + } 1.731 + VerifyOopClosure verify; 1.732 + iframe->oops_interpreted_do(&verify, NULL, &rm, false); 1.733 + callee_size_of_parameters = mh->size_of_parameters(); 1.734 + callee_max_locals = mh->max_locals(); 1.735 + is_top_frame = false; 1.736 + } 1.737 + } 1.738 +#endif /* !PRODUCT */ 1.739 + 1.740 + 1.741 + return bt; 1.742 +JRT_END 1.743 + 1.744 + 1.745 +int Deoptimization::deoptimize_dependents() { 1.746 + Threads::deoptimized_wrt_marked_nmethods(); 1.747 + return 0; 1.748 +} 1.749 + 1.750 + 1.751 +#ifdef COMPILER2 1.752 +bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 1.753 + Handle pending_exception(thread->pending_exception()); 1.754 + const char* exception_file = thread->exception_file(); 1.755 + int exception_line = thread->exception_line(); 1.756 + thread->clear_pending_exception(); 1.757 + 1.758 + for (int i = 0; i < objects->length(); i++) { 1.759 + assert(objects->at(i)->is_object(), "invalid debug information"); 1.760 + ObjectValue* sv = (ObjectValue*) objects->at(i); 1.761 + 1.762 + KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1.763 + oop obj = NULL; 1.764 + 1.765 + if (k->oop_is_instance()) { 1.766 + InstanceKlass* ik = InstanceKlass::cast(k()); 1.767 + obj = ik->allocate_instance(CHECK_(false)); 1.768 + } else if (k->oop_is_typeArray()) { 1.769 + TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 1.770 + assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1.771 + int len = sv->field_size() / type2size[ak->element_type()]; 1.772 + obj = ak->allocate(len, CHECK_(false)); 1.773 + } else if (k->oop_is_objArray()) { 1.774 + ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 1.775 + obj = ak->allocate(sv->field_size(), CHECK_(false)); 1.776 + } 1.777 + 1.778 + assert(obj != NULL, "allocation failed"); 1.779 + assert(sv->value().is_null(), "redundant reallocation"); 1.780 + sv->set_value(obj); 1.781 + } 1.782 + 1.783 + if (pending_exception.not_null()) { 1.784 + thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1.785 + } 1.786 + 1.787 + return true; 1.788 +} 1.789 + 1.790 +// This assumes that the fields are stored in ObjectValue in the same order 1.791 +// they are yielded by do_nonstatic_fields. 1.792 +class FieldReassigner: public FieldClosure { 1.793 + frame* _fr; 1.794 + RegisterMap* _reg_map; 1.795 + ObjectValue* _sv; 1.796 + InstanceKlass* _ik; 1.797 + oop _obj; 1.798 + 1.799 + int _i; 1.800 +public: 1.801 + FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 1.802 + _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 1.803 + 1.804 + int i() const { return _i; } 1.805 + 1.806 + 1.807 + void do_field(fieldDescriptor* fd) { 1.808 + intptr_t val; 1.809 + StackValue* value = 1.810 + StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 1.811 + int offset = fd->offset(); 1.812 + switch (fd->field_type()) { 1.813 + case T_OBJECT: case T_ARRAY: 1.814 + assert(value->type() == T_OBJECT, "Agreement."); 1.815 + _obj->obj_field_put(offset, value->get_obj()()); 1.816 + break; 1.817 + 1.818 + case T_LONG: case T_DOUBLE: { 1.819 + assert(value->type() == T_INT, "Agreement."); 1.820 + StackValue* low = 1.821 + StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 1.822 +#ifdef _LP64 1.823 + jlong res = (jlong)low->get_int(); 1.824 +#else 1.825 +#ifdef SPARC 1.826 + // For SPARC we have to swap high and low words. 1.827 + jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1.828 +#else 1.829 + jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1.830 +#endif //SPARC 1.831 +#endif 1.832 + _obj->long_field_put(offset, res); 1.833 + break; 1.834 + } 1.835 + // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1.836 + case T_INT: case T_FLOAT: // 4 bytes. 1.837 + assert(value->type() == T_INT, "Agreement."); 1.838 + val = value->get_int(); 1.839 + _obj->int_field_put(offset, (jint)*((jint*)&val)); 1.840 + break; 1.841 + 1.842 + case T_SHORT: case T_CHAR: // 2 bytes 1.843 + assert(value->type() == T_INT, "Agreement."); 1.844 + val = value->get_int(); 1.845 + _obj->short_field_put(offset, (jshort)*((jint*)&val)); 1.846 + break; 1.847 + 1.848 + case T_BOOLEAN: case T_BYTE: // 1 byte 1.849 + assert(value->type() == T_INT, "Agreement."); 1.850 + val = value->get_int(); 1.851 + _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1.852 + break; 1.853 + 1.854 + default: 1.855 + ShouldNotReachHere(); 1.856 + } 1.857 + _i++; 1.858 + } 1.859 +}; 1.860 + 1.861 +// restore elements of an eliminated type array 1.862 +void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1.863 + int index = 0; 1.864 + intptr_t val; 1.865 + 1.866 + for (int i = 0; i < sv->field_size(); i++) { 1.867 + StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1.868 + switch(type) { 1.869 + case T_LONG: case T_DOUBLE: { 1.870 + assert(value->type() == T_INT, "Agreement."); 1.871 + StackValue* low = 1.872 + StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1.873 +#ifdef _LP64 1.874 + jlong res = (jlong)low->get_int(); 1.875 +#else 1.876 +#ifdef SPARC 1.877 + // For SPARC we have to swap high and low words. 1.878 + jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1.879 +#else 1.880 + jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1.881 +#endif //SPARC 1.882 +#endif 1.883 + obj->long_at_put(index, res); 1.884 + break; 1.885 + } 1.886 + 1.887 + // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1.888 + case T_INT: case T_FLOAT: // 4 bytes. 1.889 + assert(value->type() == T_INT, "Agreement."); 1.890 + val = value->get_int(); 1.891 + obj->int_at_put(index, (jint)*((jint*)&val)); 1.892 + break; 1.893 + 1.894 + case T_SHORT: case T_CHAR: // 2 bytes 1.895 + assert(value->type() == T_INT, "Agreement."); 1.896 + val = value->get_int(); 1.897 + obj->short_at_put(index, (jshort)*((jint*)&val)); 1.898 + break; 1.899 + 1.900 + case T_BOOLEAN: case T_BYTE: // 1 byte 1.901 + assert(value->type() == T_INT, "Agreement."); 1.902 + val = value->get_int(); 1.903 + obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1.904 + break; 1.905 + 1.906 + default: 1.907 + ShouldNotReachHere(); 1.908 + } 1.909 + index++; 1.910 + } 1.911 +} 1.912 + 1.913 + 1.914 +// restore fields of an eliminated object array 1.915 +void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1.916 + for (int i = 0; i < sv->field_size(); i++) { 1.917 + StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1.918 + assert(value->type() == T_OBJECT, "object element expected"); 1.919 + obj->obj_at_put(i, value->get_obj()()); 1.920 + } 1.921 +} 1.922 + 1.923 + 1.924 +// restore fields of all eliminated objects and arrays 1.925 +void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 1.926 + for (int i = 0; i < objects->length(); i++) { 1.927 + ObjectValue* sv = (ObjectValue*) objects->at(i); 1.928 + KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1.929 + Handle obj = sv->value(); 1.930 + assert(obj.not_null(), "reallocation was missed"); 1.931 + 1.932 + if (k->oop_is_instance()) { 1.933 + InstanceKlass* ik = InstanceKlass::cast(k()); 1.934 + FieldReassigner reassign(fr, reg_map, sv, obj()); 1.935 + ik->do_nonstatic_fields(&reassign); 1.936 + } else if (k->oop_is_typeArray()) { 1.937 + TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 1.938 + reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1.939 + } else if (k->oop_is_objArray()) { 1.940 + reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1.941 + } 1.942 + } 1.943 +} 1.944 + 1.945 + 1.946 +// relock objects for which synchronization was eliminated 1.947 +void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 1.948 + for (int i = 0; i < monitors->length(); i++) { 1.949 + MonitorInfo* mon_info = monitors->at(i); 1.950 + if (mon_info->eliminated()) { 1.951 + assert(mon_info->owner() != NULL, "reallocation was missed"); 1.952 + Handle obj = Handle(mon_info->owner()); 1.953 + markOop mark = obj->mark(); 1.954 + if (UseBiasedLocking && mark->has_bias_pattern()) { 1.955 + // New allocated objects may have the mark set to anonymously biased. 1.956 + // Also the deoptimized method may called methods with synchronization 1.957 + // where the thread-local object is bias locked to the current thread. 1.958 + assert(mark->is_biased_anonymously() || 1.959 + mark->biased_locker() == thread, "should be locked to current thread"); 1.960 + // Reset mark word to unbiased prototype. 1.961 + markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1.962 + obj->set_mark(unbiased_prototype); 1.963 + } 1.964 + BasicLock* lock = mon_info->lock(); 1.965 + ObjectSynchronizer::slow_enter(obj, lock, thread); 1.966 + } 1.967 + assert(mon_info->owner()->is_locked(), "object must be locked now"); 1.968 + } 1.969 +} 1.970 + 1.971 + 1.972 +#ifndef PRODUCT 1.973 +// print information about reallocated objects 1.974 +void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 1.975 + fieldDescriptor fd; 1.976 + 1.977 + for (int i = 0; i < objects->length(); i++) { 1.978 + ObjectValue* sv = (ObjectValue*) objects->at(i); 1.979 + KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1.980 + Handle obj = sv->value(); 1.981 + 1.982 + tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()()); 1.983 + k->print_value(); 1.984 + tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1.985 + tty->cr(); 1.986 + 1.987 + if (Verbose) { 1.988 + k->oop_print_on(obj(), tty); 1.989 + } 1.990 + } 1.991 +} 1.992 +#endif 1.993 +#endif // COMPILER2 1.994 + 1.995 +vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 1.996 + Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 1.997 + 1.998 +#ifndef PRODUCT 1.999 + if (TraceDeoptimization) { 1.1000 + ttyLocker ttyl; 1.1001 + tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 1.1002 + fr.print_on(tty); 1.1003 + tty->print_cr(" Virtual frames (innermost first):"); 1.1004 + for (int index = 0; index < chunk->length(); index++) { 1.1005 + compiledVFrame* vf = chunk->at(index); 1.1006 + tty->print(" %2d - ", index); 1.1007 + vf->print_value(); 1.1008 + int bci = chunk->at(index)->raw_bci(); 1.1009 + const char* code_name; 1.1010 + if (bci == SynchronizationEntryBCI) { 1.1011 + code_name = "sync entry"; 1.1012 + } else { 1.1013 + Bytecodes::Code code = vf->method()->code_at(bci); 1.1014 + code_name = Bytecodes::name(code); 1.1015 + } 1.1016 + tty->print(" - %s", code_name); 1.1017 + tty->print_cr(" @ bci %d ", bci); 1.1018 + if (Verbose) { 1.1019 + vf->print(); 1.1020 + tty->cr(); 1.1021 + } 1.1022 + } 1.1023 + } 1.1024 +#endif 1.1025 + 1.1026 + // Register map for next frame (used for stack crawl). We capture 1.1027 + // the state of the deopt'ing frame's caller. Thus if we need to 1.1028 + // stuff a C2I adapter we can properly fill in the callee-save 1.1029 + // register locations. 1.1030 + frame caller = fr.sender(reg_map); 1.1031 + int frame_size = caller.sp() - fr.sp(); 1.1032 + 1.1033 + frame sender = caller; 1.1034 + 1.1035 + // Since the Java thread being deoptimized will eventually adjust it's own stack, 1.1036 + // the vframeArray containing the unpacking information is allocated in the C heap. 1.1037 + // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1.1038 + vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1.1039 + 1.1040 + // Compare the vframeArray to the collected vframes 1.1041 + assert(array->structural_compare(thread, chunk), "just checking"); 1.1042 + 1.1043 +#ifndef PRODUCT 1.1044 + if (TraceDeoptimization) { 1.1045 + ttyLocker ttyl; 1.1046 + tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1.1047 + } 1.1048 +#endif // PRODUCT 1.1049 + 1.1050 + return array; 1.1051 +} 1.1052 + 1.1053 + 1.1054 +static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1.1055 + GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1.1056 + for (int i = 0; i < monitors->length(); i++) { 1.1057 + MonitorInfo* mon_info = monitors->at(i); 1.1058 + if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1.1059 + objects_to_revoke->append(Handle(mon_info->owner())); 1.1060 + } 1.1061 + } 1.1062 +} 1.1063 + 1.1064 + 1.1065 +void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1.1066 + if (!UseBiasedLocking) { 1.1067 + return; 1.1068 + } 1.1069 + 1.1070 + GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1.1071 + 1.1072 + // Unfortunately we don't have a RegisterMap available in most of 1.1073 + // the places we want to call this routine so we need to walk the 1.1074 + // stack again to update the register map. 1.1075 + if (map == NULL || !map->update_map()) { 1.1076 + StackFrameStream sfs(thread, true); 1.1077 + bool found = false; 1.1078 + while (!found && !sfs.is_done()) { 1.1079 + frame* cur = sfs.current(); 1.1080 + sfs.next(); 1.1081 + found = cur->id() == fr.id(); 1.1082 + } 1.1083 + assert(found, "frame to be deoptimized not found on target thread's stack"); 1.1084 + map = sfs.register_map(); 1.1085 + } 1.1086 + 1.1087 + vframe* vf = vframe::new_vframe(&fr, map, thread); 1.1088 + compiledVFrame* cvf = compiledVFrame::cast(vf); 1.1089 + // Revoke monitors' biases in all scopes 1.1090 + while (!cvf->is_top()) { 1.1091 + collect_monitors(cvf, objects_to_revoke); 1.1092 + cvf = compiledVFrame::cast(cvf->sender()); 1.1093 + } 1.1094 + collect_monitors(cvf, objects_to_revoke); 1.1095 + 1.1096 + if (SafepointSynchronize::is_at_safepoint()) { 1.1097 + BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1.1098 + } else { 1.1099 + BiasedLocking::revoke(objects_to_revoke); 1.1100 + } 1.1101 +} 1.1102 + 1.1103 + 1.1104 +void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1.1105 + if (!UseBiasedLocking) { 1.1106 + return; 1.1107 + } 1.1108 + 1.1109 + assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1.1110 + GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1.1111 + for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1.1112 + if (jt->has_last_Java_frame()) { 1.1113 + StackFrameStream sfs(jt, true); 1.1114 + while (!sfs.is_done()) { 1.1115 + frame* cur = sfs.current(); 1.1116 + if (cb->contains(cur->pc())) { 1.1117 + vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1.1118 + compiledVFrame* cvf = compiledVFrame::cast(vf); 1.1119 + // Revoke monitors' biases in all scopes 1.1120 + while (!cvf->is_top()) { 1.1121 + collect_monitors(cvf, objects_to_revoke); 1.1122 + cvf = compiledVFrame::cast(cvf->sender()); 1.1123 + } 1.1124 + collect_monitors(cvf, objects_to_revoke); 1.1125 + } 1.1126 + sfs.next(); 1.1127 + } 1.1128 + } 1.1129 + } 1.1130 + BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1.1131 +} 1.1132 + 1.1133 + 1.1134 +void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1.1135 + assert(fr.can_be_deoptimized(), "checking frame type"); 1.1136 + 1.1137 + gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1.1138 + 1.1139 + // Patch the nmethod so that when execution returns to it we will 1.1140 + // deopt the execution state and return to the interpreter. 1.1141 + fr.deoptimize(thread); 1.1142 +} 1.1143 + 1.1144 +void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1.1145 + // Deoptimize only if the frame comes from compile code. 1.1146 + // Do not deoptimize the frame which is already patched 1.1147 + // during the execution of the loops below. 1.1148 + if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1.1149 + return; 1.1150 + } 1.1151 + ResourceMark rm; 1.1152 + DeoptimizationMarker dm; 1.1153 + if (UseBiasedLocking) { 1.1154 + revoke_biases_of_monitors(thread, fr, map); 1.1155 + } 1.1156 + deoptimize_single_frame(thread, fr); 1.1157 + 1.1158 +} 1.1159 + 1.1160 + 1.1161 +void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1.1162 + assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1.1163 + "can only deoptimize other thread at a safepoint"); 1.1164 + // Compute frame and register map based on thread and sp. 1.1165 + RegisterMap reg_map(thread, UseBiasedLocking); 1.1166 + frame fr = thread->last_frame(); 1.1167 + while (fr.id() != id) { 1.1168 + fr = fr.sender(®_map); 1.1169 + } 1.1170 + deoptimize(thread, fr, ®_map); 1.1171 +} 1.1172 + 1.1173 + 1.1174 +void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1.1175 + if (thread == Thread::current()) { 1.1176 + Deoptimization::deoptimize_frame_internal(thread, id); 1.1177 + } else { 1.1178 + VM_DeoptimizeFrame deopt(thread, id); 1.1179 + VMThread::execute(&deopt); 1.1180 + } 1.1181 +} 1.1182 + 1.1183 + 1.1184 +// JVMTI PopFrame support 1.1185 +JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1.1186 +{ 1.1187 + thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1.1188 +} 1.1189 +JRT_END 1.1190 + 1.1191 + 1.1192 +#if defined(COMPILER2) || defined(SHARK) 1.1193 +void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1.1194 + // in case of an unresolved klass entry, load the class. 1.1195 + if (constant_pool->tag_at(index).is_unresolved_klass()) { 1.1196 + Klass* tk = constant_pool->klass_at(index, CHECK); 1.1197 + return; 1.1198 + } 1.1199 + 1.1200 + if (!constant_pool->tag_at(index).is_symbol()) return; 1.1201 + 1.1202 + Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1.1203 + Symbol* symbol = constant_pool->symbol_at(index); 1.1204 + 1.1205 + // class name? 1.1206 + if (symbol->byte_at(0) != '(') { 1.1207 + Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1.1208 + SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1.1209 + return; 1.1210 + } 1.1211 + 1.1212 + // then it must be a signature! 1.1213 + ResourceMark rm(THREAD); 1.1214 + for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1.1215 + if (ss.is_object()) { 1.1216 + Symbol* class_name = ss.as_symbol(CHECK); 1.1217 + Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1.1218 + SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1.1219 + } 1.1220 + } 1.1221 +} 1.1222 + 1.1223 + 1.1224 +void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1.1225 + EXCEPTION_MARK; 1.1226 + load_class_by_index(constant_pool, index, THREAD); 1.1227 + if (HAS_PENDING_EXCEPTION) { 1.1228 + // Exception happened during classloading. We ignore the exception here, since it 1.1229 + // is going to be rethrown since the current activation is going to be deoptimized and 1.1230 + // the interpreter will re-execute the bytecode. 1.1231 + CLEAR_PENDING_EXCEPTION; 1.1232 + // Class loading called java code which may have caused a stack 1.1233 + // overflow. If the exception was thrown right before the return 1.1234 + // to the runtime the stack is no longer guarded. Reguard the 1.1235 + // stack otherwise if we return to the uncommon trap blob and the 1.1236 + // stack bang causes a stack overflow we crash. 1.1237 + assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1.1238 + JavaThread* thread = (JavaThread*)THREAD; 1.1239 + bool guard_pages_enabled = thread->stack_yellow_zone_enabled(); 1.1240 + if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1.1241 + assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1.1242 + } 1.1243 +} 1.1244 + 1.1245 +JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1.1246 + HandleMark hm; 1.1247 + 1.1248 + // uncommon_trap() is called at the beginning of the uncommon trap 1.1249 + // handler. Note this fact before we start generating temporary frames 1.1250 + // that can confuse an asynchronous stack walker. This counter is 1.1251 + // decremented at the end of unpack_frames(). 1.1252 + thread->inc_in_deopt_handler(); 1.1253 + 1.1254 + // We need to update the map if we have biased locking. 1.1255 + RegisterMap reg_map(thread, UseBiasedLocking); 1.1256 + frame stub_frame = thread->last_frame(); 1.1257 + frame fr = stub_frame.sender(®_map); 1.1258 + // Make sure the calling nmethod is not getting deoptimized and removed 1.1259 + // before we are done with it. 1.1260 + nmethodLocker nl(fr.pc()); 1.1261 + 1.1262 + // Log a message 1.1263 + Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT, 1.1264 + trap_request, fr.pc()); 1.1265 + 1.1266 + { 1.1267 + ResourceMark rm; 1.1268 + 1.1269 + // Revoke biases of any monitors in the frame to ensure we can migrate them 1.1270 + revoke_biases_of_monitors(thread, fr, ®_map); 1.1271 + 1.1272 + DeoptReason reason = trap_request_reason(trap_request); 1.1273 + DeoptAction action = trap_request_action(trap_request); 1.1274 + jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1.1275 + 1.1276 + vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1.1277 + compiledVFrame* cvf = compiledVFrame::cast(vf); 1.1278 + 1.1279 + nmethod* nm = cvf->code(); 1.1280 + 1.1281 + ScopeDesc* trap_scope = cvf->scope(); 1.1282 + methodHandle trap_method = trap_scope->method(); 1.1283 + int trap_bci = trap_scope->bci(); 1.1284 + Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1.1285 + 1.1286 + // Record this event in the histogram. 1.1287 + gather_statistics(reason, action, trap_bc); 1.1288 + 1.1289 + // Ensure that we can record deopt. history: 1.1290 + // Need MDO to record RTM code generation state. 1.1291 + bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 1.1292 + 1.1293 + MethodData* trap_mdo = 1.1294 + get_method_data(thread, trap_method, create_if_missing); 1.1295 + 1.1296 + // Log a message 1.1297 + Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1.1298 + trap_reason_name(reason), trap_action_name(action), fr.pc(), 1.1299 + trap_method->name_and_sig_as_C_string(), trap_bci); 1.1300 + 1.1301 + // Print a bunch of diagnostics, if requested. 1.1302 + if (TraceDeoptimization || LogCompilation) { 1.1303 + ResourceMark rm; 1.1304 + ttyLocker ttyl; 1.1305 + char buf[100]; 1.1306 + if (xtty != NULL) { 1.1307 + xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1.1308 + os::current_thread_id(), 1.1309 + format_trap_request(buf, sizeof(buf), trap_request)); 1.1310 + nm->log_identity(xtty); 1.1311 + } 1.1312 + Symbol* class_name = NULL; 1.1313 + bool unresolved = false; 1.1314 + if (unloaded_class_index >= 0) { 1.1315 + constantPoolHandle constants (THREAD, trap_method->constants()); 1.1316 + if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1.1317 + class_name = constants->klass_name_at(unloaded_class_index); 1.1318 + unresolved = true; 1.1319 + if (xtty != NULL) 1.1320 + xtty->print(" unresolved='1'"); 1.1321 + } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1.1322 + class_name = constants->symbol_at(unloaded_class_index); 1.1323 + } 1.1324 + if (xtty != NULL) 1.1325 + xtty->name(class_name); 1.1326 + } 1.1327 + if (xtty != NULL && trap_mdo != NULL) { 1.1328 + // Dump the relevant MDO state. 1.1329 + // This is the deopt count for the current reason, any previous 1.1330 + // reasons or recompiles seen at this point. 1.1331 + int dcnt = trap_mdo->trap_count(reason); 1.1332 + if (dcnt != 0) 1.1333 + xtty->print(" count='%d'", dcnt); 1.1334 + ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1.1335 + int dos = (pdata == NULL)? 0: pdata->trap_state(); 1.1336 + if (dos != 0) { 1.1337 + xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1.1338 + if (trap_state_is_recompiled(dos)) { 1.1339 + int recnt2 = trap_mdo->overflow_recompile_count(); 1.1340 + if (recnt2 != 0) 1.1341 + xtty->print(" recompiles2='%d'", recnt2); 1.1342 + } 1.1343 + } 1.1344 + } 1.1345 + if (xtty != NULL) { 1.1346 + xtty->stamp(); 1.1347 + xtty->end_head(); 1.1348 + } 1.1349 + if (TraceDeoptimization) { // make noise on the tty 1.1350 + tty->print("Uncommon trap occurred in"); 1.1351 + nm->method()->print_short_name(tty); 1.1352 + tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1.1353 + fr.pc(), 1.1354 + os::current_thread_id(), 1.1355 + trap_reason_name(reason), 1.1356 + trap_action_name(action), 1.1357 + unloaded_class_index); 1.1358 + if (class_name != NULL) { 1.1359 + tty->print(unresolved ? " unresolved class: " : " symbol: "); 1.1360 + class_name->print_symbol_on(tty); 1.1361 + } 1.1362 + tty->cr(); 1.1363 + } 1.1364 + if (xtty != NULL) { 1.1365 + // Log the precise location of the trap. 1.1366 + for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1.1367 + xtty->begin_elem("jvms bci='%d'", sd->bci()); 1.1368 + xtty->method(sd->method()); 1.1369 + xtty->end_elem(); 1.1370 + if (sd->is_top()) break; 1.1371 + } 1.1372 + xtty->tail("uncommon_trap"); 1.1373 + } 1.1374 + } 1.1375 + // (End diagnostic printout.) 1.1376 + 1.1377 + // Load class if necessary 1.1378 + if (unloaded_class_index >= 0) { 1.1379 + constantPoolHandle constants(THREAD, trap_method->constants()); 1.1380 + load_class_by_index(constants, unloaded_class_index); 1.1381 + } 1.1382 + 1.1383 + // Flush the nmethod if necessary and desirable. 1.1384 + // 1.1385 + // We need to avoid situations where we are re-flushing the nmethod 1.1386 + // because of a hot deoptimization site. Repeated flushes at the same 1.1387 + // point need to be detected by the compiler and avoided. If the compiler 1.1388 + // cannot avoid them (or has a bug and "refuses" to avoid them), this 1.1389 + // module must take measures to avoid an infinite cycle of recompilation 1.1390 + // and deoptimization. There are several such measures: 1.1391 + // 1.1392 + // 1. If a recompilation is ordered a second time at some site X 1.1393 + // and for the same reason R, the action is adjusted to 'reinterpret', 1.1394 + // to give the interpreter time to exercise the method more thoroughly. 1.1395 + // If this happens, the method's overflow_recompile_count is incremented. 1.1396 + // 1.1397 + // 2. If the compiler fails to reduce the deoptimization rate, then 1.1398 + // the method's overflow_recompile_count will begin to exceed the set 1.1399 + // limit PerBytecodeRecompilationCutoff. If this happens, the action 1.1400 + // is adjusted to 'make_not_compilable', and the method is abandoned 1.1401 + // to the interpreter. This is a performance hit for hot methods, 1.1402 + // but is better than a disastrous infinite cycle of recompilations. 1.1403 + // (Actually, only the method containing the site X is abandoned.) 1.1404 + // 1.1405 + // 3. In parallel with the previous measures, if the total number of 1.1406 + // recompilations of a method exceeds the much larger set limit 1.1407 + // PerMethodRecompilationCutoff, the method is abandoned. 1.1408 + // This should only happen if the method is very large and has 1.1409 + // many "lukewarm" deoptimizations. The code which enforces this 1.1410 + // limit is elsewhere (class nmethod, class Method). 1.1411 + // 1.1412 + // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1.1413 + // to recompile at each bytecode independently of the per-BCI cutoff. 1.1414 + // 1.1415 + // The decision to update code is up to the compiler, and is encoded 1.1416 + // in the Action_xxx code. If the compiler requests Action_none 1.1417 + // no trap state is changed, no compiled code is changed, and the 1.1418 + // computation suffers along in the interpreter. 1.1419 + // 1.1420 + // The other action codes specify various tactics for decompilation 1.1421 + // and recompilation. Action_maybe_recompile is the loosest, and 1.1422 + // allows the compiled code to stay around until enough traps are seen, 1.1423 + // and until the compiler gets around to recompiling the trapping method. 1.1424 + // 1.1425 + // The other actions cause immediate removal of the present code. 1.1426 + 1.1427 + bool update_trap_state = true; 1.1428 + bool make_not_entrant = false; 1.1429 + bool make_not_compilable = false; 1.1430 + bool reprofile = false; 1.1431 + switch (action) { 1.1432 + case Action_none: 1.1433 + // Keep the old code. 1.1434 + update_trap_state = false; 1.1435 + break; 1.1436 + case Action_maybe_recompile: 1.1437 + // Do not need to invalidate the present code, but we can 1.1438 + // initiate another 1.1439 + // Start compiler without (necessarily) invalidating the nmethod. 1.1440 + // The system will tolerate the old code, but new code should be 1.1441 + // generated when possible. 1.1442 + break; 1.1443 + case Action_reinterpret: 1.1444 + // Go back into the interpreter for a while, and then consider 1.1445 + // recompiling form scratch. 1.1446 + make_not_entrant = true; 1.1447 + // Reset invocation counter for outer most method. 1.1448 + // This will allow the interpreter to exercise the bytecodes 1.1449 + // for a while before recompiling. 1.1450 + // By contrast, Action_make_not_entrant is immediate. 1.1451 + // 1.1452 + // Note that the compiler will track null_check, null_assert, 1.1453 + // range_check, and class_check events and log them as if they 1.1454 + // had been traps taken from compiled code. This will update 1.1455 + // the MDO trap history so that the next compilation will 1.1456 + // properly detect hot trap sites. 1.1457 + reprofile = true; 1.1458 + break; 1.1459 + case Action_make_not_entrant: 1.1460 + // Request immediate recompilation, and get rid of the old code. 1.1461 + // Make them not entrant, so next time they are called they get 1.1462 + // recompiled. Unloaded classes are loaded now so recompile before next 1.1463 + // time they are called. Same for uninitialized. The interpreter will 1.1464 + // link the missing class, if any. 1.1465 + make_not_entrant = true; 1.1466 + break; 1.1467 + case Action_make_not_compilable: 1.1468 + // Give up on compiling this method at all. 1.1469 + make_not_entrant = true; 1.1470 + make_not_compilable = true; 1.1471 + break; 1.1472 + default: 1.1473 + ShouldNotReachHere(); 1.1474 + } 1.1475 + 1.1476 + // Setting +ProfileTraps fixes the following, on all platforms: 1.1477 + // 4852688: ProfileInterpreter is off by default for ia64. The result is 1.1478 + // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1.1479 + // recompile relies on a MethodData* to record heroic opt failures. 1.1480 + 1.1481 + // Whether the interpreter is producing MDO data or not, we also need 1.1482 + // to use the MDO to detect hot deoptimization points and control 1.1483 + // aggressive optimization. 1.1484 + bool inc_recompile_count = false; 1.1485 + ProfileData* pdata = NULL; 1.1486 + if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1.1487 + assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1.1488 + uint this_trap_count = 0; 1.1489 + bool maybe_prior_trap = false; 1.1490 + bool maybe_prior_recompile = false; 1.1491 + pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1.1492 + nm->method(), 1.1493 + //outputs: 1.1494 + this_trap_count, 1.1495 + maybe_prior_trap, 1.1496 + maybe_prior_recompile); 1.1497 + // Because the interpreter also counts null, div0, range, and class 1.1498 + // checks, these traps from compiled code are double-counted. 1.1499 + // This is harmless; it just means that the PerXTrapLimit values 1.1500 + // are in effect a little smaller than they look. 1.1501 + 1.1502 + DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1.1503 + if (per_bc_reason != Reason_none) { 1.1504 + // Now take action based on the partially known per-BCI history. 1.1505 + if (maybe_prior_trap 1.1506 + && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1.1507 + // If there are too many traps at this BCI, force a recompile. 1.1508 + // This will allow the compiler to see the limit overflow, and 1.1509 + // take corrective action, if possible. The compiler generally 1.1510 + // does not use the exact PerBytecodeTrapLimit value, but instead 1.1511 + // changes its tactics if it sees any traps at all. This provides 1.1512 + // a little hysteresis, delaying a recompile until a trap happens 1.1513 + // several times. 1.1514 + // 1.1515 + // Actually, since there is only one bit of counter per BCI, 1.1516 + // the possible per-BCI counts are {0,1,(per-method count)}. 1.1517 + // This produces accurate results if in fact there is only 1.1518 + // one hot trap site, but begins to get fuzzy if there are 1.1519 + // many sites. For example, if there are ten sites each 1.1520 + // trapping two or more times, they each get the blame for 1.1521 + // all of their traps. 1.1522 + make_not_entrant = true; 1.1523 + } 1.1524 + 1.1525 + // Detect repeated recompilation at the same BCI, and enforce a limit. 1.1526 + if (make_not_entrant && maybe_prior_recompile) { 1.1527 + // More than one recompile at this point. 1.1528 + inc_recompile_count = maybe_prior_trap; 1.1529 + } 1.1530 + } else { 1.1531 + // For reasons which are not recorded per-bytecode, we simply 1.1532 + // force recompiles unconditionally. 1.1533 + // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1.1534 + make_not_entrant = true; 1.1535 + } 1.1536 + 1.1537 + // Go back to the compiler if there are too many traps in this method. 1.1538 + if (this_trap_count >= per_method_trap_limit(reason)) { 1.1539 + // If there are too many traps in this method, force a recompile. 1.1540 + // This will allow the compiler to see the limit overflow, and 1.1541 + // take corrective action, if possible. 1.1542 + // (This condition is an unlikely backstop only, because the 1.1543 + // PerBytecodeTrapLimit is more likely to take effect first, 1.1544 + // if it is applicable.) 1.1545 + make_not_entrant = true; 1.1546 + } 1.1547 + 1.1548 + // Here's more hysteresis: If there has been a recompile at 1.1549 + // this trap point already, run the method in the interpreter 1.1550 + // for a while to exercise it more thoroughly. 1.1551 + if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1.1552 + reprofile = true; 1.1553 + } 1.1554 + 1.1555 + } 1.1556 + 1.1557 + // Take requested actions on the method: 1.1558 + 1.1559 + // Recompile 1.1560 + if (make_not_entrant) { 1.1561 + if (!nm->make_not_entrant()) { 1.1562 + return; // the call did not change nmethod's state 1.1563 + } 1.1564 + 1.1565 + if (pdata != NULL) { 1.1566 + // Record the recompilation event, if any. 1.1567 + int tstate0 = pdata->trap_state(); 1.1568 + int tstate1 = trap_state_set_recompiled(tstate0, true); 1.1569 + if (tstate1 != tstate0) 1.1570 + pdata->set_trap_state(tstate1); 1.1571 + } 1.1572 + 1.1573 +#if INCLUDE_RTM_OPT 1.1574 + // Restart collecting RTM locking abort statistic if the method 1.1575 + // is recompiled for a reason other than RTM state change. 1.1576 + // Assume that in new recompiled code the statistic could be different, 1.1577 + // for example, due to different inlining. 1.1578 + if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1.1579 + UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1.1580 + trap_mdo->atomic_set_rtm_state(ProfileRTM); 1.1581 + } 1.1582 +#endif 1.1583 + } 1.1584 + 1.1585 + if (inc_recompile_count) { 1.1586 + trap_mdo->inc_overflow_recompile_count(); 1.1587 + if ((uint)trap_mdo->overflow_recompile_count() > 1.1588 + (uint)PerBytecodeRecompilationCutoff) { 1.1589 + // Give up on the method containing the bad BCI. 1.1590 + if (trap_method() == nm->method()) { 1.1591 + make_not_compilable = true; 1.1592 + } else { 1.1593 + trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1.1594 + // But give grace to the enclosing nm->method(). 1.1595 + } 1.1596 + } 1.1597 + } 1.1598 + 1.1599 + // Reprofile 1.1600 + if (reprofile) { 1.1601 + CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1.1602 + } 1.1603 + 1.1604 + // Give up compiling 1.1605 + if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1.1606 + assert(make_not_entrant, "consistent"); 1.1607 + nm->method()->set_not_compilable(CompLevel_full_optimization); 1.1608 + } 1.1609 + 1.1610 + } // Free marked resources 1.1611 + 1.1612 +} 1.1613 +JRT_END 1.1614 + 1.1615 +MethodData* 1.1616 +Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1.1617 + bool create_if_missing) { 1.1618 + Thread* THREAD = thread; 1.1619 + MethodData* mdo = m()->method_data(); 1.1620 + if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1.1621 + // Build an MDO. Ignore errors like OutOfMemory; 1.1622 + // that simply means we won't have an MDO to update. 1.1623 + Method::build_interpreter_method_data(m, THREAD); 1.1624 + if (HAS_PENDING_EXCEPTION) { 1.1625 + assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1.1626 + CLEAR_PENDING_EXCEPTION; 1.1627 + } 1.1628 + mdo = m()->method_data(); 1.1629 + } 1.1630 + return mdo; 1.1631 +} 1.1632 + 1.1633 +ProfileData* 1.1634 +Deoptimization::query_update_method_data(MethodData* trap_mdo, 1.1635 + int trap_bci, 1.1636 + Deoptimization::DeoptReason reason, 1.1637 + Method* compiled_method, 1.1638 + //outputs: 1.1639 + uint& ret_this_trap_count, 1.1640 + bool& ret_maybe_prior_trap, 1.1641 + bool& ret_maybe_prior_recompile) { 1.1642 + uint prior_trap_count = trap_mdo->trap_count(reason); 1.1643 + uint this_trap_count = trap_mdo->inc_trap_count(reason); 1.1644 + 1.1645 + // If the runtime cannot find a place to store trap history, 1.1646 + // it is estimated based on the general condition of the method. 1.1647 + // If the method has ever been recompiled, or has ever incurred 1.1648 + // a trap with the present reason , then this BCI is assumed 1.1649 + // (pessimistically) to be the culprit. 1.1650 + bool maybe_prior_trap = (prior_trap_count != 0); 1.1651 + bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1.1652 + ProfileData* pdata = NULL; 1.1653 + 1.1654 + 1.1655 + // For reasons which are recorded per bytecode, we check per-BCI data. 1.1656 + DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1.1657 + if (per_bc_reason != Reason_none) { 1.1658 + // Find the profile data for this BCI. If there isn't one, 1.1659 + // try to allocate one from the MDO's set of spares. 1.1660 + // This will let us detect a repeated trap at this point. 1.1661 + pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1.1662 + 1.1663 + if (pdata != NULL) { 1.1664 + if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1.1665 + if (LogCompilation && xtty != NULL) { 1.1666 + ttyLocker ttyl; 1.1667 + // no more room for speculative traps in this MDO 1.1668 + xtty->elem("speculative_traps_oom"); 1.1669 + } 1.1670 + } 1.1671 + // Query the trap state of this profile datum. 1.1672 + int tstate0 = pdata->trap_state(); 1.1673 + if (!trap_state_has_reason(tstate0, per_bc_reason)) 1.1674 + maybe_prior_trap = false; 1.1675 + if (!trap_state_is_recompiled(tstate0)) 1.1676 + maybe_prior_recompile = false; 1.1677 + 1.1678 + // Update the trap state of this profile datum. 1.1679 + int tstate1 = tstate0; 1.1680 + // Record the reason. 1.1681 + tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1.1682 + // Store the updated state on the MDO, for next time. 1.1683 + if (tstate1 != tstate0) 1.1684 + pdata->set_trap_state(tstate1); 1.1685 + } else { 1.1686 + if (LogCompilation && xtty != NULL) { 1.1687 + ttyLocker ttyl; 1.1688 + // Missing MDP? Leave a small complaint in the log. 1.1689 + xtty->elem("missing_mdp bci='%d'", trap_bci); 1.1690 + } 1.1691 + } 1.1692 + } 1.1693 + 1.1694 + // Return results: 1.1695 + ret_this_trap_count = this_trap_count; 1.1696 + ret_maybe_prior_trap = maybe_prior_trap; 1.1697 + ret_maybe_prior_recompile = maybe_prior_recompile; 1.1698 + return pdata; 1.1699 +} 1.1700 + 1.1701 +void 1.1702 +Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1.1703 + ResourceMark rm; 1.1704 + // Ignored outputs: 1.1705 + uint ignore_this_trap_count; 1.1706 + bool ignore_maybe_prior_trap; 1.1707 + bool ignore_maybe_prior_recompile; 1.1708 + assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 1.1709 + query_update_method_data(trap_mdo, trap_bci, 1.1710 + (DeoptReason)reason, 1.1711 + NULL, 1.1712 + ignore_this_trap_count, 1.1713 + ignore_maybe_prior_trap, 1.1714 + ignore_maybe_prior_recompile); 1.1715 +} 1.1716 + 1.1717 +Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1.1718 + 1.1719 + // Still in Java no safepoints 1.1720 + { 1.1721 + // This enters VM and may safepoint 1.1722 + uncommon_trap_inner(thread, trap_request); 1.1723 + } 1.1724 + return fetch_unroll_info_helper(thread); 1.1725 +} 1.1726 + 1.1727 +// Local derived constants. 1.1728 +// Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1.1729 +const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1.1730 +const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1.1731 + 1.1732 +//---------------------------trap_state_reason--------------------------------- 1.1733 +Deoptimization::DeoptReason 1.1734 +Deoptimization::trap_state_reason(int trap_state) { 1.1735 + // This assert provides the link between the width of DataLayout::trap_bits 1.1736 + // and the encoding of "recorded" reasons. It ensures there are enough 1.1737 + // bits to store all needed reasons in the per-BCI MDO profile. 1.1738 + assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1.1739 + int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1.1740 + trap_state -= recompile_bit; 1.1741 + if (trap_state == DS_REASON_MASK) { 1.1742 + return Reason_many; 1.1743 + } else { 1.1744 + assert((int)Reason_none == 0, "state=0 => Reason_none"); 1.1745 + return (DeoptReason)trap_state; 1.1746 + } 1.1747 +} 1.1748 +//-------------------------trap_state_has_reason------------------------------- 1.1749 +int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1.1750 + assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1.1751 + assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1.1752 + int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1.1753 + trap_state -= recompile_bit; 1.1754 + if (trap_state == DS_REASON_MASK) { 1.1755 + return -1; // true, unspecifically (bottom of state lattice) 1.1756 + } else if (trap_state == reason) { 1.1757 + return 1; // true, definitely 1.1758 + } else if (trap_state == 0) { 1.1759 + return 0; // false, definitely (top of state lattice) 1.1760 + } else { 1.1761 + return 0; // false, definitely 1.1762 + } 1.1763 +} 1.1764 +//-------------------------trap_state_add_reason------------------------------- 1.1765 +int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1.1766 + assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1.1767 + int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1.1768 + trap_state -= recompile_bit; 1.1769 + if (trap_state == DS_REASON_MASK) { 1.1770 + return trap_state + recompile_bit; // already at state lattice bottom 1.1771 + } else if (trap_state == reason) { 1.1772 + return trap_state + recompile_bit; // the condition is already true 1.1773 + } else if (trap_state == 0) { 1.1774 + return reason + recompile_bit; // no condition has yet been true 1.1775 + } else { 1.1776 + return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1.1777 + } 1.1778 +} 1.1779 +//-----------------------trap_state_is_recompiled------------------------------ 1.1780 +bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1.1781 + return (trap_state & DS_RECOMPILE_BIT) != 0; 1.1782 +} 1.1783 +//-----------------------trap_state_set_recompiled----------------------------- 1.1784 +int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1.1785 + if (z) return trap_state | DS_RECOMPILE_BIT; 1.1786 + else return trap_state & ~DS_RECOMPILE_BIT; 1.1787 +} 1.1788 +//---------------------------format_trap_state--------------------------------- 1.1789 +// This is used for debugging and diagnostics, including LogFile output. 1.1790 +const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1.1791 + int trap_state) { 1.1792 + DeoptReason reason = trap_state_reason(trap_state); 1.1793 + bool recomp_flag = trap_state_is_recompiled(trap_state); 1.1794 + // Re-encode the state from its decoded components. 1.1795 + int decoded_state = 0; 1.1796 + if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1.1797 + decoded_state = trap_state_add_reason(decoded_state, reason); 1.1798 + if (recomp_flag) 1.1799 + decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1.1800 + // If the state re-encodes properly, format it symbolically. 1.1801 + // Because this routine is used for debugging and diagnostics, 1.1802 + // be robust even if the state is a strange value. 1.1803 + size_t len; 1.1804 + if (decoded_state != trap_state) { 1.1805 + // Random buggy state that doesn't decode?? 1.1806 + len = jio_snprintf(buf, buflen, "#%d", trap_state); 1.1807 + } else { 1.1808 + len = jio_snprintf(buf, buflen, "%s%s", 1.1809 + trap_reason_name(reason), 1.1810 + recomp_flag ? " recompiled" : ""); 1.1811 + } 1.1812 + if (len >= buflen) 1.1813 + buf[buflen-1] = '\0'; 1.1814 + return buf; 1.1815 +} 1.1816 + 1.1817 + 1.1818 +//--------------------------------statics-------------------------------------- 1.1819 +Deoptimization::DeoptAction Deoptimization::_unloaded_action 1.1820 + = Deoptimization::Action_reinterpret; 1.1821 +const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1.1822 + // Note: Keep this in sync. with enum DeoptReason. 1.1823 + "none", 1.1824 + "null_check", 1.1825 + "null_assert", 1.1826 + "range_check", 1.1827 + "class_check", 1.1828 + "array_check", 1.1829 + "intrinsic", 1.1830 + "bimorphic", 1.1831 + "unloaded", 1.1832 + "uninitialized", 1.1833 + "unreached", 1.1834 + "unhandled", 1.1835 + "constraint", 1.1836 + "div0_check", 1.1837 + "age", 1.1838 + "predicate", 1.1839 + "loop_limit_check", 1.1840 + "speculate_class_check", 1.1841 + "rtm_state_change" 1.1842 +}; 1.1843 +const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1.1844 + // Note: Keep this in sync. with enum DeoptAction. 1.1845 + "none", 1.1846 + "maybe_recompile", 1.1847 + "reinterpret", 1.1848 + "make_not_entrant", 1.1849 + "make_not_compilable" 1.1850 +}; 1.1851 + 1.1852 +const char* Deoptimization::trap_reason_name(int reason) { 1.1853 + if (reason == Reason_many) return "many"; 1.1854 + if ((uint)reason < Reason_LIMIT) 1.1855 + return _trap_reason_name[reason]; 1.1856 + static char buf[20]; 1.1857 + sprintf(buf, "reason%d", reason); 1.1858 + return buf; 1.1859 +} 1.1860 +const char* Deoptimization::trap_action_name(int action) { 1.1861 + if ((uint)action < Action_LIMIT) 1.1862 + return _trap_action_name[action]; 1.1863 + static char buf[20]; 1.1864 + sprintf(buf, "action%d", action); 1.1865 + return buf; 1.1866 +} 1.1867 + 1.1868 +// This is used for debugging and diagnostics, including LogFile output. 1.1869 +const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1.1870 + int trap_request) { 1.1871 + jint unloaded_class_index = trap_request_index(trap_request); 1.1872 + const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1.1873 + const char* action = trap_action_name(trap_request_action(trap_request)); 1.1874 + size_t len; 1.1875 + if (unloaded_class_index < 0) { 1.1876 + len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1.1877 + reason, action); 1.1878 + } else { 1.1879 + len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1.1880 + reason, action, unloaded_class_index); 1.1881 + } 1.1882 + if (len >= buflen) 1.1883 + buf[buflen-1] = '\0'; 1.1884 + return buf; 1.1885 +} 1.1886 + 1.1887 +juint Deoptimization::_deoptimization_hist 1.1888 + [Deoptimization::Reason_LIMIT] 1.1889 + [1 + Deoptimization::Action_LIMIT] 1.1890 + [Deoptimization::BC_CASE_LIMIT] 1.1891 + = {0}; 1.1892 + 1.1893 +enum { 1.1894 + LSB_BITS = 8, 1.1895 + LSB_MASK = right_n_bits(LSB_BITS) 1.1896 +}; 1.1897 + 1.1898 +void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1.1899 + Bytecodes::Code bc) { 1.1900 + assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1.1901 + assert(action >= 0 && action < Action_LIMIT, "oob"); 1.1902 + _deoptimization_hist[Reason_none][0][0] += 1; // total 1.1903 + _deoptimization_hist[reason][0][0] += 1; // per-reason total 1.1904 + juint* cases = _deoptimization_hist[reason][1+action]; 1.1905 + juint* bc_counter_addr = NULL; 1.1906 + juint bc_counter = 0; 1.1907 + // Look for an unused counter, or an exact match to this BC. 1.1908 + if (bc != Bytecodes::_illegal) { 1.1909 + for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1.1910 + juint* counter_addr = &cases[bc_case]; 1.1911 + juint counter = *counter_addr; 1.1912 + if ((counter == 0 && bc_counter_addr == NULL) 1.1913 + || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1.1914 + // this counter is either free or is already devoted to this BC 1.1915 + bc_counter_addr = counter_addr; 1.1916 + bc_counter = counter | bc; 1.1917 + } 1.1918 + } 1.1919 + } 1.1920 + if (bc_counter_addr == NULL) { 1.1921 + // Overflow, or no given bytecode. 1.1922 + bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1.1923 + bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1.1924 + } 1.1925 + *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1.1926 +} 1.1927 + 1.1928 +jint Deoptimization::total_deoptimization_count() { 1.1929 + return _deoptimization_hist[Reason_none][0][0]; 1.1930 +} 1.1931 + 1.1932 +jint Deoptimization::deoptimization_count(DeoptReason reason) { 1.1933 + assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1.1934 + return _deoptimization_hist[reason][0][0]; 1.1935 +} 1.1936 + 1.1937 +void Deoptimization::print_statistics() { 1.1938 + juint total = total_deoptimization_count(); 1.1939 + juint account = total; 1.1940 + if (total != 0) { 1.1941 + ttyLocker ttyl; 1.1942 + if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1.1943 + tty->print_cr("Deoptimization traps recorded:"); 1.1944 + #define PRINT_STAT_LINE(name, r) \ 1.1945 + tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1.1946 + PRINT_STAT_LINE("total", total); 1.1947 + // For each non-zero entry in the histogram, print the reason, 1.1948 + // the action, and (if specifically known) the type of bytecode. 1.1949 + for (int reason = 0; reason < Reason_LIMIT; reason++) { 1.1950 + for (int action = 0; action < Action_LIMIT; action++) { 1.1951 + juint* cases = _deoptimization_hist[reason][1+action]; 1.1952 + for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1.1953 + juint counter = cases[bc_case]; 1.1954 + if (counter != 0) { 1.1955 + char name[1*K]; 1.1956 + Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1.1957 + if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1.1958 + bc = Bytecodes::_illegal; 1.1959 + sprintf(name, "%s/%s/%s", 1.1960 + trap_reason_name(reason), 1.1961 + trap_action_name(action), 1.1962 + Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1.1963 + juint r = counter >> LSB_BITS; 1.1964 + tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1.1965 + account -= r; 1.1966 + } 1.1967 + } 1.1968 + } 1.1969 + } 1.1970 + if (account != 0) { 1.1971 + PRINT_STAT_LINE("unaccounted", account); 1.1972 + } 1.1973 + #undef PRINT_STAT_LINE 1.1974 + if (xtty != NULL) xtty->tail("statistics"); 1.1975 + } 1.1976 +} 1.1977 +#else // COMPILER2 || SHARK 1.1978 + 1.1979 + 1.1980 +// Stubs for C1 only system. 1.1981 +bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1.1982 + return false; 1.1983 +} 1.1984 + 1.1985 +const char* Deoptimization::trap_reason_name(int reason) { 1.1986 + return "unknown"; 1.1987 +} 1.1988 + 1.1989 +void Deoptimization::print_statistics() { 1.1990 + // no output 1.1991 +} 1.1992 + 1.1993 +void 1.1994 +Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1.1995 + // no udpate 1.1996 +} 1.1997 + 1.1998 +int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1.1999 + return 0; 1.2000 +} 1.2001 + 1.2002 +void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1.2003 + Bytecodes::Code bc) { 1.2004 + // no update 1.2005 +} 1.2006 + 1.2007 +const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1.2008 + int trap_state) { 1.2009 + jio_snprintf(buf, buflen, "#%d", trap_state); 1.2010 + return buf; 1.2011 +} 1.2012 + 1.2013 +#endif // COMPILER2 || SHARK