annotate src/share/vm/runtime/deoptimization.cpp @ 3534:1d7922586cf6

7023639: JSR 292 method handle invocation needs a fast path for compiled code 6984705: JSR 292 method handle creation should not go through JNI Summary: remove assembly code for JDK 7 chained method handles Reviewed-by: jrose, twisti, kvn, mhaupt Contributed-by: John Rose <john.r.rose@oracle.com>, Christian Thalinger <christian.thalinger@oracle.com>, Michael Haupt <michael.haupt@oracle.com>
author twisti
date Tue, 24 Jul 2012 10:51:00 -0700
parents d2a62e0f25eb
children da91efe96a93
rev   line source
duke@0 1 /*
never@3064 2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1879 25 #include "precompiled.hpp"
stefank@1879 26 #include "classfile/systemDictionary.hpp"
stefank@1879 27 #include "code/debugInfoRec.hpp"
stefank@1879 28 #include "code/nmethod.hpp"
stefank@1879 29 #include "code/pcDesc.hpp"
stefank@1879 30 #include "code/scopeDesc.hpp"
stefank@1879 31 #include "interpreter/bytecode.hpp"
stefank@1879 32 #include "interpreter/interpreter.hpp"
stefank@1879 33 #include "interpreter/oopMapCache.hpp"
stefank@1879 34 #include "memory/allocation.inline.hpp"
stefank@1879 35 #include "memory/oopFactory.hpp"
stefank@1879 36 #include "memory/resourceArea.hpp"
stefank@1879 37 #include "oops/methodOop.hpp"
stefank@1879 38 #include "oops/oop.inline.hpp"
stefank@1879 39 #include "prims/jvmtiThreadState.hpp"
stefank@1879 40 #include "runtime/biasedLocking.hpp"
stefank@1879 41 #include "runtime/compilationPolicy.hpp"
stefank@1879 42 #include "runtime/deoptimization.hpp"
stefank@1879 43 #include "runtime/interfaceSupport.hpp"
stefank@1879 44 #include "runtime/sharedRuntime.hpp"
stefank@1879 45 #include "runtime/signature.hpp"
stefank@1879 46 #include "runtime/stubRoutines.hpp"
stefank@1879 47 #include "runtime/thread.hpp"
stefank@1879 48 #include "runtime/vframe.hpp"
stefank@1879 49 #include "runtime/vframeArray.hpp"
stefank@1879 50 #include "runtime/vframe_hp.hpp"
stefank@1879 51 #include "utilities/events.hpp"
stefank@1879 52 #include "utilities/xmlstream.hpp"
stefank@1879 53 #ifdef TARGET_ARCH_x86
stefank@1879 54 # include "vmreg_x86.inline.hpp"
stefank@1879 55 #endif
stefank@1879 56 #ifdef TARGET_ARCH_sparc
stefank@1879 57 # include "vmreg_sparc.inline.hpp"
stefank@1879 58 #endif
stefank@1879 59 #ifdef TARGET_ARCH_zero
stefank@1879 60 # include "vmreg_zero.inline.hpp"
stefank@1879 61 #endif
bobv@2073 62 #ifdef TARGET_ARCH_arm
bobv@2073 63 # include "vmreg_arm.inline.hpp"
bobv@2073 64 #endif
bobv@2073 65 #ifdef TARGET_ARCH_ppc
bobv@2073 66 # include "vmreg_ppc.inline.hpp"
bobv@2073 67 #endif
stefank@1879 68 #ifdef COMPILER2
stefank@1879 69 #ifdef TARGET_ARCH_MODEL_x86_32
stefank@1879 70 # include "adfiles/ad_x86_32.hpp"
stefank@1879 71 #endif
stefank@1879 72 #ifdef TARGET_ARCH_MODEL_x86_64
stefank@1879 73 # include "adfiles/ad_x86_64.hpp"
stefank@1879 74 #endif
stefank@1879 75 #ifdef TARGET_ARCH_MODEL_sparc
stefank@1879 76 # include "adfiles/ad_sparc.hpp"
stefank@1879 77 #endif
stefank@1879 78 #ifdef TARGET_ARCH_MODEL_zero
stefank@1879 79 # include "adfiles/ad_zero.hpp"
stefank@1879 80 #endif
bobv@2073 81 #ifdef TARGET_ARCH_MODEL_arm
bobv@2073 82 # include "adfiles/ad_arm.hpp"
bobv@2073 83 #endif
bobv@2073 84 #ifdef TARGET_ARCH_MODEL_ppc
bobv@2073 85 # include "adfiles/ad_ppc.hpp"
bobv@2073 86 #endif
stefank@1879 87 #endif
duke@0 88
duke@0 89 bool DeoptimizationMarker::_is_active = false;
duke@0 90
duke@0 91 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
duke@0 92 int caller_adjustment,
never@2466 93 int caller_actual_parameters,
duke@0 94 int number_of_frames,
duke@0 95 intptr_t* frame_sizes,
duke@0 96 address* frame_pcs,
duke@0 97 BasicType return_type) {
duke@0 98 _size_of_deoptimized_frame = size_of_deoptimized_frame;
duke@0 99 _caller_adjustment = caller_adjustment;
never@2466 100 _caller_actual_parameters = caller_actual_parameters;
duke@0 101 _number_of_frames = number_of_frames;
duke@0 102 _frame_sizes = frame_sizes;
duke@0 103 _frame_pcs = frame_pcs;
zgu@3465 104 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
duke@0 105 _return_type = return_type;
bdelsart@2695 106 _initial_info = 0;
duke@0 107 // PD (x86 only)
duke@0 108 _counter_temp = 0;
duke@0 109 _unpack_kind = 0;
duke@0 110 _sender_sp_temp = 0;
duke@0 111
duke@0 112 _total_frame_sizes = size_of_frames();
duke@0 113 }
duke@0 114
duke@0 115
duke@0 116 Deoptimization::UnrollBlock::~UnrollBlock() {
zgu@3465 117 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
zgu@3465 118 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
zgu@3465 119 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
duke@0 120 }
duke@0 121
duke@0 122
duke@0 123 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
duke@0 124 assert(register_number < RegisterMap::reg_count, "checking register number");
duke@0 125 return &_register_block[register_number * 2];
duke@0 126 }
duke@0 127
duke@0 128
duke@0 129
duke@0 130 int Deoptimization::UnrollBlock::size_of_frames() const {
duke@0 131 // Acount first for the adjustment of the initial frame
duke@0 132 int result = _caller_adjustment;
duke@0 133 for (int index = 0; index < number_of_frames(); index++) {
duke@0 134 result += frame_sizes()[index];
duke@0 135 }
duke@0 136 return result;
duke@0 137 }
duke@0 138
duke@0 139
duke@0 140 void Deoptimization::UnrollBlock::print() {
duke@0 141 ttyLocker ttyl;
duke@0 142 tty->print_cr("UnrollBlock");
duke@0 143 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
duke@0 144 tty->print( " frame_sizes: ");
duke@0 145 for (int index = 0; index < number_of_frames(); index++) {
duke@0 146 tty->print("%d ", frame_sizes()[index]);
duke@0 147 }
duke@0 148 tty->cr();
duke@0 149 }
duke@0 150
duke@0 151
duke@0 152 // In order to make fetch_unroll_info work properly with escape
duke@0 153 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
duke@0 154 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
duke@0 155 // of previously eliminated objects occurs in realloc_objects, which is
duke@0 156 // called from the method fetch_unroll_info_helper below.
duke@0 157 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
duke@0 158 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
duke@0 159 // but makes the entry a little slower. There is however a little dance we have to
duke@0 160 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
duke@0 161
duke@0 162 // fetch_unroll_info() is called at the beginning of the deoptimization
duke@0 163 // handler. Note this fact before we start generating temporary frames
duke@0 164 // that can confuse an asynchronous stack walker. This counter is
duke@0 165 // decremented at the end of unpack_frames().
duke@0 166 thread->inc_in_deopt_handler();
duke@0 167
duke@0 168 return fetch_unroll_info_helper(thread);
duke@0 169 JRT_END
duke@0 170
duke@0 171
duke@0 172 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
duke@0 173 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
duke@0 174
duke@0 175 // Note: there is a safepoint safety issue here. No matter whether we enter
duke@0 176 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
duke@0 177 // the vframeArray is created.
duke@0 178 //
duke@0 179
duke@0 180 // Allocate our special deoptimization ResourceMark
duke@0 181 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
duke@0 182 assert(thread->deopt_mark() == NULL, "Pending deopt!");
duke@0 183 thread->set_deopt_mark(dmark);
duke@0 184
duke@0 185 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
duke@0 186 RegisterMap map(thread, true);
duke@0 187 RegisterMap dummy_map(thread, false);
duke@0 188 // Now get the deoptee with a valid map
duke@0 189 frame deoptee = stub_frame.sender(&map);
iveresov@1734 190 // Set the deoptee nmethod
iveresov@1734 191 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
iveresov@1734 192 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
duke@0 193
never@2433 194 if (VerifyStack) {
never@2433 195 thread->validate_frame_layout();
never@2433 196 }
never@2433 197
duke@0 198 // Create a growable array of VFrames where each VFrame represents an inlined
duke@0 199 // Java frame. This storage is allocated with the usual system arena.
duke@0 200 assert(deoptee.is_compiled_frame(), "Wrong frame type");
duke@0 201 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
duke@0 202 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
duke@0 203 while (!vf->is_top()) {
duke@0 204 assert(vf->is_compiled_frame(), "Wrong frame type");
duke@0 205 chunk->push(compiledVFrame::cast(vf));
duke@0 206 vf = vf->sender();
duke@0 207 }
duke@0 208 assert(vf->is_compiled_frame(), "Wrong frame type");
duke@0 209 chunk->push(compiledVFrame::cast(vf));
duke@0 210
duke@0 211 #ifdef COMPILER2
duke@0 212 // Reallocate the non-escaping objects and restore their fields. Then
duke@0 213 // relock objects if synchronization on them was eliminated.
kvn@2971 214 if (DoEscapeAnalysis || EliminateNestedLocks) {
kvn@44 215 if (EliminateAllocations) {
kvn@83 216 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
kvn@44 217 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
kvn@1253 218
kvn@1253 219 // The flag return_oop() indicates call sites which return oop
kvn@1253 220 // in compiled code. Such sites include java method calls,
kvn@1253 221 // runtime calls (for example, used to allocate new objects/arrays
kvn@1253 222 // on slow code path) and any other calls generated in compiled code.
kvn@1253 223 // It is not guaranteed that we can get such information here only
kvn@1253 224 // by analyzing bytecode in deoptimized frames. This is why this flag
kvn@1253 225 // is set during method compilation (see Compile::Process_OopMap_Node()).
kvn@1253 226 bool save_oop_result = chunk->at(0)->scope()->return_oop();
kvn@1253 227 Handle return_value;
kvn@1253 228 if (save_oop_result) {
kvn@1253 229 // Reallocation may trigger GC. If deoptimization happened on return from
kvn@1253 230 // call which returns oop we need to save it since it is not in oopmap.
kvn@1253 231 oop result = deoptee.saved_oop_result(&map);
kvn@1253 232 assert(result == NULL || result->is_oop(), "must be oop");
kvn@1253 233 return_value = Handle(thread, result);
kvn@1253 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
kvn@1253 235 if (TraceDeoptimization) {
kvn@1253 236 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
kvn@1253 237 }
kvn@1253 238 }
kvn@44 239 bool reallocated = false;
kvn@44 240 if (objects != NULL) {
kvn@44 241 JRT_BLOCK
kvn@44 242 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
kvn@44 243 JRT_END
duke@0 244 }
kvn@44 245 if (reallocated) {
kvn@44 246 reassign_fields(&deoptee, &map, objects);
duke@0 247 #ifndef PRODUCT
duke@0 248 if (TraceDeoptimization) {
duke@0 249 ttyLocker ttyl;
kvn@44 250 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
kvn@44 251 print_objects(objects);
kvn@1253 252 }
kvn@1253 253 #endif
kvn@44 254 }
kvn@1253 255 if (save_oop_result) {
kvn@1253 256 // Restore result.
kvn@1253 257 deoptee.set_saved_oop_result(&map, return_value());
kvn@44 258 }
kvn@44 259 }
kvn@44 260 if (EliminateLocks) {
kvn@83 261 #ifndef PRODUCT
kvn@83 262 bool first = true;
kvn@83 263 #endif
kvn@44 264 for (int i = 0; i < chunk->length(); i++) {
kvn@83 265 compiledVFrame* cvf = chunk->at(i);
kvn@83 266 assert (cvf->scope() != NULL,"expect only compiled java frames");
kvn@83 267 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
kvn@83 268 if (monitors->is_nonempty()) {
kvn@83 269 relock_objects(monitors, thread);
kvn@44 270 #ifndef PRODUCT
kvn@44 271 if (TraceDeoptimization) {
kvn@44 272 ttyLocker ttyl;
kvn@44 273 for (int j = 0; j < monitors->length(); j++) {
kvn@83 274 MonitorInfo* mi = monitors->at(j);
kvn@83 275 if (mi->eliminated()) {
kvn@83 276 if (first) {
kvn@83 277 first = false;
kvn@83 278 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
kvn@83 279 }
kvn@83 280 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
kvn@44 281 }
duke@0 282 }
duke@0 283 }
kvn@44 284 #endif
duke@0 285 }
duke@0 286 }
duke@0 287 }
duke@0 288 }
duke@0 289 #endif // COMPILER2
duke@0 290 // Ensure that no safepoint is taken after pointers have been stored
duke@0 291 // in fields of rematerialized objects. If a safepoint occurs from here on
duke@0 292 // out the java state residing in the vframeArray will be missed.
duke@0 293 No_Safepoint_Verifier no_safepoint;
duke@0 294
duke@0 295 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
duke@0 296
duke@0 297 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
duke@0 298 thread->set_vframe_array_head(array);
duke@0 299
duke@0 300 // Now that the vframeArray has been created if we have any deferred local writes
duke@0 301 // added by jvmti then we can free up that structure as the data is now in the
duke@0 302 // vframeArray
duke@0 303
duke@0 304 if (thread->deferred_locals() != NULL) {
duke@0 305 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
duke@0 306 int i = 0;
duke@0 307 do {
duke@0 308 // Because of inlining we could have multiple vframes for a single frame
duke@0 309 // and several of the vframes could have deferred writes. Find them all.
duke@0 310 if (list->at(i)->id() == array->original().id()) {
duke@0 311 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
duke@0 312 list->remove_at(i);
duke@0 313 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
duke@0 314 delete dlv;
duke@0 315 } else {
duke@0 316 i++;
duke@0 317 }
duke@0 318 } while ( i < list->length() );
duke@0 319 if (list->length() == 0) {
duke@0 320 thread->set_deferred_locals(NULL);
duke@0 321 // free the list and elements back to C heap.
duke@0 322 delete list;
duke@0 323 }
duke@0 324
duke@0 325 }
duke@0 326
twisti@1612 327 #ifndef SHARK
duke@0 328 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
duke@0 329 CodeBlob* cb = stub_frame.cb();
duke@0 330 // Verify we have the right vframeArray
duke@0 331 assert(cb->frame_size() >= 0, "Unexpected frame size");
duke@0 332 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
duke@0 333
twisti@1204 334 // If the deopt call site is a MethodHandle invoke call site we have
twisti@1204 335 // to adjust the unpack_sp.
twisti@1204 336 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
twisti@1204 337 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
twisti@1204 338 unpack_sp = deoptee.unextended_sp();
twisti@1204 339
duke@0 340 #ifdef ASSERT
duke@0 341 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
duke@0 342 #endif
twisti@1612 343 #else
twisti@1612 344 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
twisti@1612 345 #endif // !SHARK
twisti@1612 346
duke@0 347 // This is a guarantee instead of an assert because if vframe doesn't match
duke@0 348 // we will unpack the wrong deoptimized frame and wind up in strange places
duke@0 349 // where it will be very difficult to figure out what went wrong. Better
duke@0 350 // to die an early death here than some very obscure death later when the
duke@0 351 // trail is cold.
duke@0 352 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
duke@0 353 // in that it will fail to detect a problem when there is one. This needs
duke@0 354 // more work in tiger timeframe.
duke@0 355 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
duke@0 356
duke@0 357 int number_of_frames = array->frames();
duke@0 358
duke@0 359 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
duke@0 360 // virtual activation, which is the reverse of the elements in the vframes array.
zgu@3465 361 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
duke@0 362 // +1 because we always have an interpreter return address for the final slot.
zgu@3465 363 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
duke@0 364 int popframe_extra_args = 0;
duke@0 365 // Create an interpreter return address for the stub to use as its return
duke@0 366 // address so the skeletal frames are perfectly walkable
duke@0 367 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
duke@0 368
duke@0 369 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
duke@0 370 // activation be put back on the expression stack of the caller for reexecution
duke@0 371 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
duke@0 372 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
duke@0 373 }
duke@0 374
never@2466 375 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
never@2466 376 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
never@2466 377 // than simply use array->sender.pc(). This requires us to walk the current set of frames
never@2466 378 //
never@2466 379 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
never@2466 380 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
never@2466 381
never@2466 382 // It's possible that the number of paramters at the call site is
never@2466 383 // different than number of arguments in the callee when method
never@2466 384 // handles are used. If the caller is interpreted get the real
never@2466 385 // value so that the proper amount of space can be added to it's
never@2466 386 // frame.
twisti@2803 387 bool caller_was_method_handle = false;
never@2466 388 if (deopt_sender.is_interpreted_frame()) {
never@2466 389 methodHandle method = deopt_sender.interpreter_frame_method();
twisti@2816 390 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
twisti@3534 391 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
twisti@2803 392 // Method handle invokes may involve fairly arbitrary chains of
twisti@2803 393 // calls so it's impossible to know how much actual space the
twisti@2803 394 // caller has for locals.
twisti@2803 395 caller_was_method_handle = true;
twisti@2803 396 }
never@2466 397 }
never@2466 398
duke@0 399 //
duke@0 400 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
duke@0 401 // frame_sizes/frame_pcs[1] next oldest frame (int)
duke@0 402 // frame_sizes/frame_pcs[n] youngest frame (int)
duke@0 403 //
duke@0 404 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
duke@0 405 // owns the space for the return address to it's caller). Confusing ain't it.
duke@0 406 //
duke@0 407 // The vframe array can address vframes with indices running from
duke@0 408 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
duke@0 409 // When we create the skeletal frames we need the oldest frame to be in the zero slot
duke@0 410 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
duke@0 411 // so things look a little strange in this loop.
duke@0 412 //
twisti@2803 413 int callee_parameters = 0;
twisti@2803 414 int callee_locals = 0;
duke@0 415 for (int index = 0; index < array->frames(); index++ ) {
duke@0 416 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
duke@0 417 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
duke@0 418 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
never@2466 419 int caller_parms = callee_parameters;
twisti@2803 420 if ((index == array->frames() - 1) && caller_was_method_handle) {
twisti@2803 421 caller_parms = 0;
never@2466 422 }
never@2466 423 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms,
never@2466 424 callee_parameters,
duke@0 425 callee_locals,
duke@0 426 index == 0,
duke@0 427 popframe_extra_args);
duke@0 428 // This pc doesn't have to be perfect just good enough to identify the frame
duke@0 429 // as interpreted so the skeleton frame will be walkable
duke@0 430 // The correct pc will be set when the skeleton frame is completely filled out
duke@0 431 // The final pc we store in the loop is wrong and will be overwritten below
duke@0 432 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
duke@0 433
duke@0 434 callee_parameters = array->element(index)->method()->size_of_parameters();
duke@0 435 callee_locals = array->element(index)->method()->max_locals();
duke@0 436 popframe_extra_args = 0;
duke@0 437 }
duke@0 438
duke@0 439 // Compute whether the root vframe returns a float or double value.
duke@0 440 BasicType return_type;
duke@0 441 {
duke@0 442 HandleMark hm;
duke@0 443 methodHandle method(thread, array->element(0)->method());
never@2027 444 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
coleenp@2062 445 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
duke@0 446 }
duke@0 447
duke@0 448 // Compute information for handling adapters and adjusting the frame size of the caller.
duke@0 449 int caller_adjustment = 0;
duke@0 450
duke@0 451 // Compute the amount the oldest interpreter frame will have to adjust
duke@0 452 // its caller's stack by. If the caller is a compiled frame then
duke@0 453 // we pretend that the callee has no parameters so that the
duke@0 454 // extension counts for the full amount of locals and not just
duke@0 455 // locals-parms. This is because without a c2i adapter the parm
duke@0 456 // area as created by the compiled frame will not be usable by
duke@0 457 // the interpreter. (Depending on the calling convention there
duke@0 458 // may not even be enough space).
duke@0 459
duke@0 460 // QQQ I'd rather see this pushed down into last_frame_adjust
duke@0 461 // and have it take the sender (aka caller).
duke@0 462
twisti@2803 463 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
duke@0 464 caller_adjustment = last_frame_adjust(0, callee_locals);
twisti@2803 465 } else if (callee_locals > callee_parameters) {
duke@0 466 // The caller frame may need extending to accommodate
duke@0 467 // non-parameter locals of the first unpacked interpreted frame.
duke@0 468 // Compute that adjustment.
twisti@2803 469 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
duke@0 470 }
duke@0 471
duke@0 472 // If the sender is deoptimized the we must retrieve the address of the handler
duke@0 473 // since the frame will "magically" show the original pc before the deopt
duke@0 474 // and we'd undo the deopt.
duke@0 475
duke@0 476 frame_pcs[0] = deopt_sender.raw_pc();
duke@0 477
twisti@1612 478 #ifndef SHARK
duke@0 479 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
twisti@1612 480 #endif // SHARK
duke@0 481
duke@0 482 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
duke@0 483 caller_adjustment * BytesPerWord,
twisti@2803 484 caller_was_method_handle ? 0 : callee_parameters,
duke@0 485 number_of_frames,
duke@0 486 frame_sizes,
duke@0 487 frame_pcs,
duke@0 488 return_type);
bdelsart@2695 489 // On some platforms, we need a way to pass some platform dependent
bdelsart@2695 490 // information to the unpacking code so the skeletal frames come out
bdelsart@2695 491 // correct (initial fp value, unextended sp, ...)
bdelsart@2695 492 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
duke@0 493
duke@0 494 if (array->frames() > 1) {
duke@0 495 if (VerifyStack && TraceDeoptimization) {
duke@0 496 tty->print_cr("Deoptimizing method containing inlining");
duke@0 497 }
duke@0 498 }
duke@0 499
duke@0 500 array->set_unroll_block(info);
duke@0 501 return info;
duke@0 502 }
duke@0 503
duke@0 504 // Called to cleanup deoptimization data structures in normal case
duke@0 505 // after unpacking to stack and when stack overflow error occurs
duke@0 506 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
duke@0 507 vframeArray *array) {
duke@0 508
duke@0 509 // Get array if coming from exception
duke@0 510 if (array == NULL) {
duke@0 511 array = thread->vframe_array_head();
duke@0 512 }
duke@0 513 thread->set_vframe_array_head(NULL);
duke@0 514
duke@0 515 // Free the previous UnrollBlock
duke@0 516 vframeArray* old_array = thread->vframe_array_last();
duke@0 517 thread->set_vframe_array_last(array);
duke@0 518
duke@0 519 if (old_array != NULL) {
duke@0 520 UnrollBlock* old_info = old_array->unroll_block();
duke@0 521 old_array->set_unroll_block(NULL);
duke@0 522 delete old_info;
duke@0 523 delete old_array;
duke@0 524 }
duke@0 525
duke@0 526 // Deallocate any resource creating in this routine and any ResourceObjs allocated
duke@0 527 // inside the vframeArray (StackValueCollections)
duke@0 528
duke@0 529 delete thread->deopt_mark();
duke@0 530 thread->set_deopt_mark(NULL);
iveresov@1734 531 thread->set_deopt_nmethod(NULL);
duke@0 532
duke@0 533
duke@0 534 if (JvmtiExport::can_pop_frame()) {
duke@0 535 #ifndef CC_INTERP
duke@0 536 // Regardless of whether we entered this routine with the pending
duke@0 537 // popframe condition bit set, we should always clear it now
duke@0 538 thread->clear_popframe_condition();
duke@0 539 #else
duke@0 540 // C++ interpeter will clear has_pending_popframe when it enters
duke@0 541 // with method_resume. For deopt_resume2 we clear it now.
duke@0 542 if (thread->popframe_forcing_deopt_reexecution())
duke@0 543 thread->clear_popframe_condition();
duke@0 544 #endif /* CC_INTERP */
duke@0 545 }
duke@0 546
duke@0 547 // unpack_frames() is called at the end of the deoptimization handler
duke@0 548 // and (in C2) at the end of the uncommon trap handler. Note this fact
duke@0 549 // so that an asynchronous stack walker can work again. This counter is
duke@0 550 // incremented at the beginning of fetch_unroll_info() and (in C2) at
duke@0 551 // the beginning of uncommon_trap().
duke@0 552 thread->dec_in_deopt_handler();
duke@0 553 }
duke@0 554
duke@0 555
duke@0 556 // Return BasicType of value being returned
duke@0 557 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
duke@0 558
duke@0 559 // We are already active int he special DeoptResourceMark any ResourceObj's we
duke@0 560 // allocate will be freed at the end of the routine.
duke@0 561
duke@0 562 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
duke@0 563 // but makes the entry a little slower. There is however a little dance we have to
duke@0 564 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
duke@0 565 ResetNoHandleMark rnhm; // No-op in release/product versions
duke@0 566 HandleMark hm;
duke@0 567
duke@0 568 frame stub_frame = thread->last_frame();
duke@0 569
duke@0 570 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
duke@0 571 // must point to the vframeArray for the unpack frame.
duke@0 572 vframeArray* array = thread->vframe_array_head();
duke@0 573
duke@0 574 #ifndef PRODUCT
duke@0 575 if (TraceDeoptimization) {
duke@0 576 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
duke@0 577 }
duke@0 578 #endif
never@3064 579 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
never@3064 580 stub_frame.pc(), stub_frame.sp(), exec_mode);
duke@0 581
duke@0 582 UnrollBlock* info = array->unroll_block();
duke@0 583
duke@0 584 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
never@2466 585 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
duke@0 586
duke@0 587 BasicType bt = info->return_type();
duke@0 588
duke@0 589 // If we have an exception pending, claim that the return type is an oop
duke@0 590 // so the deopt_blob does not overwrite the exception_oop.
duke@0 591
duke@0 592 if (exec_mode == Unpack_exception)
duke@0 593 bt = T_OBJECT;
duke@0 594
duke@0 595 // Cleanup thread deopt data
duke@0 596 cleanup_deopt_info(thread, array);
duke@0 597
duke@0 598 #ifndef PRODUCT
duke@0 599 if (VerifyStack) {
duke@0 600 ResourceMark res_mark;
duke@0 601
never@2433 602 thread->validate_frame_layout();
never@2433 603
duke@0 604 // Verify that the just-unpacked frames match the interpreter's
duke@0 605 // notions of expression stack and locals
duke@0 606 vframeArray* cur_array = thread->vframe_array_last();
duke@0 607 RegisterMap rm(thread, false);
duke@0 608 rm.set_include_argument_oops(false);
duke@0 609 bool is_top_frame = true;
duke@0 610 int callee_size_of_parameters = 0;
duke@0 611 int callee_max_locals = 0;
duke@0 612 for (int i = 0; i < cur_array->frames(); i++) {
duke@0 613 vframeArrayElement* el = cur_array->element(i);
duke@0 614 frame* iframe = el->iframe();
duke@0 615 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
duke@0 616
duke@0 617 // Get the oop map for this bci
duke@0 618 InterpreterOopMap mask;
duke@0 619 int cur_invoke_parameter_size = 0;
duke@0 620 bool try_next_mask = false;
duke@0 621 int next_mask_expression_stack_size = -1;
duke@0 622 int top_frame_expression_stack_adjustment = 0;
duke@0 623 methodHandle mh(thread, iframe->interpreter_frame_method());
duke@0 624 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
duke@0 625 BytecodeStream str(mh);
duke@0 626 str.set_start(iframe->interpreter_frame_bci());
duke@0 627 int max_bci = mh->code_size();
duke@0 628 // Get to the next bytecode if possible
duke@0 629 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
duke@0 630 // Check to see if we can grab the number of outgoing arguments
duke@0 631 // at an uncommon trap for an invoke (where the compiler
duke@0 632 // generates debug info before the invoke has executed)
duke@0 633 Bytecodes::Code cur_code = str.next();
duke@0 634 if (cur_code == Bytecodes::_invokevirtual ||
duke@0 635 cur_code == Bytecodes::_invokespecial ||
duke@0 636 cur_code == Bytecodes::_invokestatic ||
duke@0 637 cur_code == Bytecodes::_invokeinterface) {
never@2027 638 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
coleenp@2062 639 Symbol* signature = invoke.signature();
duke@0 640 ArgumentSizeComputer asc(signature);
duke@0 641 cur_invoke_parameter_size = asc.size();
duke@0 642 if (cur_code != Bytecodes::_invokestatic) {
duke@0 643 // Add in receiver
duke@0 644 ++cur_invoke_parameter_size;
duke@0 645 }
duke@0 646 }
duke@0 647 if (str.bci() < max_bci) {
duke@0 648 Bytecodes::Code bc = str.next();
duke@0 649 if (bc >= 0) {
duke@0 650 // The interpreter oop map generator reports results before
duke@0 651 // the current bytecode has executed except in the case of
duke@0 652 // calls. It seems to be hard to tell whether the compiler
duke@0 653 // has emitted debug information matching the "state before"
duke@0 654 // a given bytecode or the state after, so we try both
duke@0 655 switch (cur_code) {
duke@0 656 case Bytecodes::_invokevirtual:
duke@0 657 case Bytecodes::_invokespecial:
duke@0 658 case Bytecodes::_invokestatic:
duke@0 659 case Bytecodes::_invokeinterface:
duke@0 660 case Bytecodes::_athrow:
duke@0 661 break;
duke@0 662 default: {
duke@0 663 InterpreterOopMap next_mask;
duke@0 664 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
duke@0 665 next_mask_expression_stack_size = next_mask.expression_stack_size();
duke@0 666 // Need to subtract off the size of the result type of
duke@0 667 // the bytecode because this is not described in the
duke@0 668 // debug info but returned to the interpreter in the TOS
duke@0 669 // caching register
duke@0 670 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
duke@0 671 if (bytecode_result_type != T_ILLEGAL) {
duke@0 672 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
duke@0 673 }
duke@0 674 assert(top_frame_expression_stack_adjustment >= 0, "");
duke@0 675 try_next_mask = true;
duke@0 676 break;
duke@0 677 }
duke@0 678 }
duke@0 679 }
duke@0 680 }
duke@0 681
duke@0 682 // Verify stack depth and oops in frame
duke@0 683 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
duke@0 684 if (!(
duke@0 685 /* SPARC */
duke@0 686 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
duke@0 687 /* x86 */
duke@0 688 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
duke@0 689 (try_next_mask &&
duke@0 690 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
duke@0 691 top_frame_expression_stack_adjustment))) ||
duke@0 692 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
duke@0 693 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
duke@0 694 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
duke@0 695 )) {
duke@0 696 ttyLocker ttyl;
duke@0 697
duke@0 698 // Print out some information that will help us debug the problem
duke@0 699 tty->print_cr("Wrong number of expression stack elements during deoptimization");
duke@0 700 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
duke@0 701 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
duke@0 702 iframe->interpreter_frame_expression_stack_size());
duke@0 703 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
duke@0 704 tty->print_cr(" try_next_mask = %d", try_next_mask);
duke@0 705 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
duke@0 706 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
duke@0 707 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
duke@0 708 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
duke@0 709 tty->print_cr(" exec_mode = %d", exec_mode);
duke@0 710 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
duke@0 711 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
duke@0 712 tty->print_cr(" Interpreted frames:");
duke@0 713 for (int k = 0; k < cur_array->frames(); k++) {
duke@0 714 vframeArrayElement* el = cur_array->element(k);
duke@0 715 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
duke@0 716 }
duke@0 717 cur_array->print_on_2(tty);
duke@0 718 guarantee(false, "wrong number of expression stack elements during deopt");
duke@0 719 }
duke@0 720 VerifyOopClosure verify;
duke@0 721 iframe->oops_interpreted_do(&verify, &rm, false);
duke@0 722 callee_size_of_parameters = mh->size_of_parameters();
duke@0 723 callee_max_locals = mh->max_locals();
duke@0 724 is_top_frame = false;
duke@0 725 }
duke@0 726 }
duke@0 727 #endif /* !PRODUCT */
duke@0 728
duke@0 729
duke@0 730 return bt;
duke@0 731 JRT_END
duke@0 732
duke@0 733
duke@0 734 int Deoptimization::deoptimize_dependents() {
duke@0 735 Threads::deoptimized_wrt_marked_nmethods();
duke@0 736 return 0;
duke@0 737 }
duke@0 738
duke@0 739
duke@0 740 #ifdef COMPILER2
duke@0 741 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
duke@0 742 Handle pending_exception(thread->pending_exception());
duke@0 743 const char* exception_file = thread->exception_file();
duke@0 744 int exception_line = thread->exception_line();
duke@0 745 thread->clear_pending_exception();
duke@0 746
duke@0 747 for (int i = 0; i < objects->length(); i++) {
duke@0 748 assert(objects->at(i)->is_object(), "invalid debug information");
duke@0 749 ObjectValue* sv = (ObjectValue*) objects->at(i);
duke@0 750
duke@0 751 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
duke@0 752 oop obj = NULL;
duke@0 753
duke@0 754 if (k->oop_is_instance()) {
duke@0 755 instanceKlass* ik = instanceKlass::cast(k());
duke@0 756 obj = ik->allocate_instance(CHECK_(false));
duke@0 757 } else if (k->oop_is_typeArray()) {
duke@0 758 typeArrayKlass* ak = typeArrayKlass::cast(k());
duke@0 759 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
duke@0 760 int len = sv->field_size() / type2size[ak->element_type()];
duke@0 761 obj = ak->allocate(len, CHECK_(false));
duke@0 762 } else if (k->oop_is_objArray()) {
duke@0 763 objArrayKlass* ak = objArrayKlass::cast(k());
duke@0 764 obj = ak->allocate(sv->field_size(), CHECK_(false));
duke@0 765 }
duke@0 766
duke@0 767 assert(obj != NULL, "allocation failed");
duke@0 768 assert(sv->value().is_null(), "redundant reallocation");
duke@0 769 sv->set_value(obj);
duke@0 770 }
duke@0 771
duke@0 772 if (pending_exception.not_null()) {
duke@0 773 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
duke@0 774 }
duke@0 775
duke@0 776 return true;
duke@0 777 }
duke@0 778
duke@0 779 // This assumes that the fields are stored in ObjectValue in the same order
duke@0 780 // they are yielded by do_nonstatic_fields.
duke@0 781 class FieldReassigner: public FieldClosure {
duke@0 782 frame* _fr;
duke@0 783 RegisterMap* _reg_map;
duke@0 784 ObjectValue* _sv;
duke@0 785 instanceKlass* _ik;
duke@0 786 oop _obj;
duke@0 787
duke@0 788 int _i;
duke@0 789 public:
duke@0 790 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
duke@0 791 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
duke@0 792
duke@0 793 int i() const { return _i; }
duke@0 794
duke@0 795
duke@0 796 void do_field(fieldDescriptor* fd) {
kvn@44 797 intptr_t val;
duke@0 798 StackValue* value =
duke@0 799 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
duke@0 800 int offset = fd->offset();
duke@0 801 switch (fd->field_type()) {
duke@0 802 case T_OBJECT: case T_ARRAY:
duke@0 803 assert(value->type() == T_OBJECT, "Agreement.");
duke@0 804 _obj->obj_field_put(offset, value->get_obj()());
duke@0 805 break;
duke@0 806
duke@0 807 case T_LONG: case T_DOUBLE: {
duke@0 808 assert(value->type() == T_INT, "Agreement.");
duke@0 809 StackValue* low =
duke@0 810 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
kvn@44 811 #ifdef _LP64
kvn@44 812 jlong res = (jlong)low->get_int();
kvn@44 813 #else
kvn@44 814 #ifdef SPARC
kvn@44 815 // For SPARC we have to swap high and low words.
kvn@44 816 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
kvn@44 817 #else
duke@0 818 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
kvn@44 819 #endif //SPARC
kvn@44 820 #endif
duke@0 821 _obj->long_field_put(offset, res);
duke@0 822 break;
duke@0 823 }
kvn@44 824 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
duke@0 825 case T_INT: case T_FLOAT: // 4 bytes.
duke@0 826 assert(value->type() == T_INT, "Agreement.");
kvn@44 827 val = value->get_int();
kvn@44 828 _obj->int_field_put(offset, (jint)*((jint*)&val));
duke@0 829 break;
duke@0 830
duke@0 831 case T_SHORT: case T_CHAR: // 2 bytes
duke@0 832 assert(value->type() == T_INT, "Agreement.");
kvn@44 833 val = value->get_int();
kvn@44 834 _obj->short_field_put(offset, (jshort)*((jint*)&val));
duke@0 835 break;
duke@0 836
kvn@44 837 case T_BOOLEAN: case T_BYTE: // 1 byte
duke@0 838 assert(value->type() == T_INT, "Agreement.");
kvn@44 839 val = value->get_int();
kvn@44 840 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
duke@0 841 break;
duke@0 842
duke@0 843 default:
duke@0 844 ShouldNotReachHere();
duke@0 845 }
duke@0 846 _i++;
duke@0 847 }
duke@0 848 };
duke@0 849
duke@0 850 // restore elements of an eliminated type array
duke@0 851 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
duke@0 852 int index = 0;
kvn@44 853 intptr_t val;
duke@0 854
duke@0 855 for (int i = 0; i < sv->field_size(); i++) {
duke@0 856 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
duke@0 857 switch(type) {
kvn@44 858 case T_LONG: case T_DOUBLE: {
kvn@44 859 assert(value->type() == T_INT, "Agreement.");
kvn@44 860 StackValue* low =
kvn@44 861 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
kvn@44 862 #ifdef _LP64
kvn@44 863 jlong res = (jlong)low->get_int();
kvn@44 864 #else
kvn@44 865 #ifdef SPARC
kvn@44 866 // For SPARC we have to swap high and low words.
kvn@44 867 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
kvn@44 868 #else
kvn@44 869 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
kvn@44 870 #endif //SPARC
kvn@44 871 #endif
kvn@44 872 obj->long_at_put(index, res);
kvn@44 873 break;
kvn@44 874 }
kvn@44 875
kvn@44 876 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
kvn@44 877 case T_INT: case T_FLOAT: // 4 bytes.
kvn@44 878 assert(value->type() == T_INT, "Agreement.");
kvn@44 879 val = value->get_int();
kvn@44 880 obj->int_at_put(index, (jint)*((jint*)&val));
kvn@44 881 break;
kvn@44 882
kvn@44 883 case T_SHORT: case T_CHAR: // 2 bytes
kvn@44 884 assert(value->type() == T_INT, "Agreement.");
kvn@44 885 val = value->get_int();
kvn@44 886 obj->short_at_put(index, (jshort)*((jint*)&val));
kvn@44 887 break;
kvn@44 888
kvn@44 889 case T_BOOLEAN: case T_BYTE: // 1 byte
kvn@44 890 assert(value->type() == T_INT, "Agreement.");
kvn@44 891 val = value->get_int();
kvn@44 892 obj->bool_at_put(index, (jboolean)*((jint*)&val));
kvn@44 893 break;
kvn@44 894
duke@0 895 default:
duke@0 896 ShouldNotReachHere();
duke@0 897 }
duke@0 898 index++;
duke@0 899 }
duke@0 900 }
duke@0 901
duke@0 902
duke@0 903 // restore fields of an eliminated object array
duke@0 904 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
duke@0 905 for (int i = 0; i < sv->field_size(); i++) {
duke@0 906 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
duke@0 907 assert(value->type() == T_OBJECT, "object element expected");
duke@0 908 obj->obj_at_put(i, value->get_obj()());
duke@0 909 }
duke@0 910 }
duke@0 911
duke@0 912
duke@0 913 // restore fields of all eliminated objects and arrays
duke@0 914 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
duke@0 915 for (int i = 0; i < objects->length(); i++) {
duke@0 916 ObjectValue* sv = (ObjectValue*) objects->at(i);
duke@0 917 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
duke@0 918 Handle obj = sv->value();
duke@0 919 assert(obj.not_null(), "reallocation was missed");
duke@0 920
duke@0 921 if (k->oop_is_instance()) {
duke@0 922 instanceKlass* ik = instanceKlass::cast(k());
duke@0 923 FieldReassigner reassign(fr, reg_map, sv, obj());
duke@0 924 ik->do_nonstatic_fields(&reassign);
duke@0 925 } else if (k->oop_is_typeArray()) {
duke@0 926 typeArrayKlass* ak = typeArrayKlass::cast(k());
duke@0 927 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
duke@0 928 } else if (k->oop_is_objArray()) {
duke@0 929 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
duke@0 930 }
duke@0 931 }
duke@0 932 }
duke@0 933
duke@0 934
duke@0 935 // relock objects for which synchronization was eliminated
kvn@83 936 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
duke@0 937 for (int i = 0; i < monitors->length(); i++) {
kvn@83 938 MonitorInfo* mon_info = monitors->at(i);
kvn@83 939 if (mon_info->eliminated()) {
kvn@83 940 assert(mon_info->owner() != NULL, "reallocation was missed");
kvn@83 941 Handle obj = Handle(mon_info->owner());
kvn@83 942 markOop mark = obj->mark();
kvn@83 943 if (UseBiasedLocking && mark->has_bias_pattern()) {
kvn@83 944 // New allocated objects may have the mark set to anonymously biased.
kvn@83 945 // Also the deoptimized method may called methods with synchronization
kvn@83 946 // where the thread-local object is bias locked to the current thread.
kvn@83 947 assert(mark->is_biased_anonymously() ||
kvn@83 948 mark->biased_locker() == thread, "should be locked to current thread");
kvn@83 949 // Reset mark word to unbiased prototype.
kvn@83 950 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
kvn@83 951 obj->set_mark(unbiased_prototype);
kvn@83 952 }
kvn@83 953 BasicLock* lock = mon_info->lock();
kvn@83 954 ObjectSynchronizer::slow_enter(obj, lock, thread);
duke@0 955 }
kvn@83 956 assert(mon_info->owner()->is_locked(), "object must be locked now");
duke@0 957 }
duke@0 958 }
duke@0 959
duke@0 960
duke@0 961 #ifndef PRODUCT
duke@0 962 // print information about reallocated objects
duke@0 963 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
duke@0 964 fieldDescriptor fd;
duke@0 965
duke@0 966 for (int i = 0; i < objects->length(); i++) {
duke@0 967 ObjectValue* sv = (ObjectValue*) objects->at(i);
duke@0 968 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
duke@0 969 Handle obj = sv->value();
duke@0 970
duke@0 971 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
duke@0 972 k->as_klassOop()->print_value();
duke@0 973 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
duke@0 974 tty->cr();
duke@0 975
duke@0 976 if (Verbose) {
duke@0 977 k->oop_print_on(obj(), tty);
duke@0 978 }
duke@0 979 }
duke@0 980 }
duke@0 981 #endif
duke@0 982 #endif // COMPILER2
duke@0 983
duke@0 984 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
never@3064 985 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
duke@0 986
duke@0 987 #ifndef PRODUCT
duke@0 988 if (TraceDeoptimization) {
duke@0 989 ttyLocker ttyl;
duke@0 990 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
duke@0 991 fr.print_on(tty);
duke@0 992 tty->print_cr(" Virtual frames (innermost first):");
duke@0 993 for (int index = 0; index < chunk->length(); index++) {
duke@0 994 compiledVFrame* vf = chunk->at(index);
duke@0 995 tty->print(" %2d - ", index);
duke@0 996 vf->print_value();
duke@0 997 int bci = chunk->at(index)->raw_bci();
duke@0 998 const char* code_name;
duke@0 999 if (bci == SynchronizationEntryBCI) {
duke@0 1000 code_name = "sync entry";
duke@0 1001 } else {
never@2027 1002 Bytecodes::Code code = vf->method()->code_at(bci);
duke@0 1003 code_name = Bytecodes::name(code);
duke@0 1004 }
duke@0 1005 tty->print(" - %s", code_name);
duke@0 1006 tty->print_cr(" @ bci %d ", bci);
duke@0 1007 if (Verbose) {
duke@0 1008 vf->print();
duke@0 1009 tty->cr();
duke@0 1010 }
duke@0 1011 }
duke@0 1012 }
duke@0 1013 #endif
duke@0 1014
duke@0 1015 // Register map for next frame (used for stack crawl). We capture
duke@0 1016 // the state of the deopt'ing frame's caller. Thus if we need to
duke@0 1017 // stuff a C2I adapter we can properly fill in the callee-save
duke@0 1018 // register locations.
duke@0 1019 frame caller = fr.sender(reg_map);
duke@0 1020 int frame_size = caller.sp() - fr.sp();
duke@0 1021
duke@0 1022 frame sender = caller;
duke@0 1023
duke@0 1024 // Since the Java thread being deoptimized will eventually adjust it's own stack,
duke@0 1025 // the vframeArray containing the unpacking information is allocated in the C heap.
duke@0 1026 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
duke@0 1027 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
duke@0 1028
duke@0 1029 // Compare the vframeArray to the collected vframes
duke@0 1030 assert(array->structural_compare(thread, chunk), "just checking");
duke@0 1031
duke@0 1032 #ifndef PRODUCT
duke@0 1033 if (TraceDeoptimization) {
duke@0 1034 ttyLocker ttyl;
duke@0 1035 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
duke@0 1036 }
duke@0 1037 #endif // PRODUCT
duke@0 1038
duke@0 1039 return array;
duke@0 1040 }
duke@0 1041
duke@0 1042
duke@0 1043 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
duke@0 1044 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
duke@0 1045 for (int i = 0; i < monitors->length(); i++) {
duke@0 1046 MonitorInfo* mon_info = monitors->at(i);
kvn@818 1047 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
duke@0 1048 objects_to_revoke->append(Handle(mon_info->owner()));
duke@0 1049 }
duke@0 1050 }
duke@0 1051 }
duke@0 1052
duke@0 1053
duke@0 1054 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
duke@0 1055 if (!UseBiasedLocking) {
duke@0 1056 return;
duke@0 1057 }
duke@0 1058
duke@0 1059 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
duke@0 1060
duke@0 1061 // Unfortunately we don't have a RegisterMap available in most of
duke@0 1062 // the places we want to call this routine so we need to walk the
duke@0 1063 // stack again to update the register map.
duke@0 1064 if (map == NULL || !map->update_map()) {
duke@0 1065 StackFrameStream sfs(thread, true);
duke@0 1066 bool found = false;
duke@0 1067 while (!found && !sfs.is_done()) {
duke@0 1068 frame* cur = sfs.current();
duke@0 1069 sfs.next();
duke@0 1070 found = cur->id() == fr.id();
duke@0 1071 }
duke@0 1072 assert(found, "frame to be deoptimized not found on target thread's stack");
duke@0 1073 map = sfs.register_map();
duke@0 1074 }
duke@0 1075
duke@0 1076 vframe* vf = vframe::new_vframe(&fr, map, thread);
duke@0 1077 compiledVFrame* cvf = compiledVFrame::cast(vf);
duke@0 1078 // Revoke monitors' biases in all scopes
duke@0 1079 while (!cvf->is_top()) {
duke@0 1080 collect_monitors(cvf, objects_to_revoke);
duke@0 1081 cvf = compiledVFrame::cast(cvf->sender());
duke@0 1082 }
duke@0 1083 collect_monitors(cvf, objects_to_revoke);
duke@0 1084
duke@0 1085 if (SafepointSynchronize::is_at_safepoint()) {
duke@0 1086 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
duke@0 1087 } else {
duke@0 1088 BiasedLocking::revoke(objects_to_revoke);
duke@0 1089 }
duke@0 1090 }
duke@0 1091
duke@0 1092
duke@0 1093 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
duke@0 1094 if (!UseBiasedLocking) {
duke@0 1095 return;
duke@0 1096 }
duke@0 1097
duke@0 1098 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
duke@0 1099 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
duke@0 1100 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
duke@0 1101 if (jt->has_last_Java_frame()) {
duke@0 1102 StackFrameStream sfs(jt, true);
duke@0 1103 while (!sfs.is_done()) {
duke@0 1104 frame* cur = sfs.current();
duke@0 1105 if (cb->contains(cur->pc())) {
duke@0 1106 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
duke@0 1107 compiledVFrame* cvf = compiledVFrame::cast(vf);
duke@0 1108 // Revoke monitors' biases in all scopes
duke@0 1109 while (!cvf->is_top()) {
duke@0 1110 collect_monitors(cvf, objects_to_revoke);
duke@0 1111 cvf = compiledVFrame::cast(cvf->sender());
duke@0 1112 }
duke@0 1113 collect_monitors(cvf, objects_to_revoke);
duke@0 1114 }
duke@0 1115 sfs.next();
duke@0 1116 }
duke@0 1117 }
duke@0 1118 }
duke@0 1119 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
duke@0 1120 }
duke@0 1121
duke@0 1122
duke@0 1123 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
duke@0 1124 assert(fr.can_be_deoptimized(), "checking frame type");
duke@0 1125
duke@0 1126 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
duke@0 1127
duke@0 1128 // Patch the nmethod so that when execution returns to it we will
duke@0 1129 // deopt the execution state and return to the interpreter.
duke@0 1130 fr.deoptimize(thread);
duke@0 1131 }
duke@0 1132
duke@0 1133 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
duke@0 1134 // Deoptimize only if the frame comes from compile code.
duke@0 1135 // Do not deoptimize the frame which is already patched
duke@0 1136 // during the execution of the loops below.
duke@0 1137 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
duke@0 1138 return;
duke@0 1139 }
duke@0 1140 ResourceMark rm;
duke@0 1141 DeoptimizationMarker dm;
duke@0 1142 if (UseBiasedLocking) {
duke@0 1143 revoke_biases_of_monitors(thread, fr, map);
duke@0 1144 }
duke@0 1145 deoptimize_single_frame(thread, fr);
duke@0 1146
duke@0 1147 }
duke@0 1148
duke@0 1149
never@1825 1150 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
never@1825 1151 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
never@1825 1152 "can only deoptimize other thread at a safepoint");
duke@0 1153 // Compute frame and register map based on thread and sp.
duke@0 1154 RegisterMap reg_map(thread, UseBiasedLocking);
duke@0 1155 frame fr = thread->last_frame();
duke@0 1156 while (fr.id() != id) {
duke@0 1157 fr = fr.sender(&reg_map);
duke@0 1158 }
duke@0 1159 deoptimize(thread, fr, &reg_map);
duke@0 1160 }
duke@0 1161
duke@0 1162
never@1825 1163 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
never@1825 1164 if (thread == Thread::current()) {
never@1825 1165 Deoptimization::deoptimize_frame_internal(thread, id);
never@1825 1166 } else {
never@1825 1167 VM_DeoptimizeFrame deopt(thread, id);
never@1825 1168 VMThread::execute(&deopt);
never@1825 1169 }
never@1825 1170 }
never@1825 1171
never@1825 1172
duke@0 1173 // JVMTI PopFrame support
duke@0 1174 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
duke@0 1175 {
duke@0 1176 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
duke@0 1177 }
duke@0 1178 JRT_END
duke@0 1179
duke@0 1180
twisti@1612 1181 #if defined(COMPILER2) || defined(SHARK)
duke@0 1182 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
duke@0 1183 // in case of an unresolved klass entry, load the class.
duke@0 1184 if (constant_pool->tag_at(index).is_unresolved_klass()) {
duke@0 1185 klassOop tk = constant_pool->klass_at(index, CHECK);
duke@0 1186 return;
duke@0 1187 }
duke@0 1188
duke@0 1189 if (!constant_pool->tag_at(index).is_symbol()) return;
duke@0 1190
duke@0 1191 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
coleenp@2062 1192 Symbol* symbol = constant_pool->symbol_at(index);
duke@0 1193
duke@0 1194 // class name?
duke@0 1195 if (symbol->byte_at(0) != '(') {
duke@0 1196 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
duke@0 1197 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
duke@0 1198 return;
duke@0 1199 }
duke@0 1200
duke@0 1201 // then it must be a signature!
coleenp@2062 1202 ResourceMark rm(THREAD);
duke@0 1203 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
duke@0 1204 if (ss.is_object()) {
coleenp@2062 1205 Symbol* class_name = ss.as_symbol(CHECK);
duke@0 1206 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
duke@0 1207 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
duke@0 1208 }
duke@0 1209 }
duke@0 1210 }
duke@0 1211
duke@0 1212
duke@0 1213 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
duke@0 1214 EXCEPTION_MARK;
duke@0 1215 load_class_by_index(constant_pool, index, THREAD);
duke@0 1216 if (HAS_PENDING_EXCEPTION) {
duke@0 1217 // Exception happened during classloading. We ignore the exception here, since it
duke@0 1218 // is going to be rethrown since the current activation is going to be deoptimzied and
duke@0 1219 // the interpreter will re-execute the bytecode.
duke@0 1220 CLEAR_PENDING_EXCEPTION;
duke@0 1221 }
duke@0 1222 }
duke@0 1223
duke@0 1224 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
duke@0 1225 HandleMark hm;
duke@0 1226
duke@0 1227 // uncommon_trap() is called at the beginning of the uncommon trap
duke@0 1228 // handler. Note this fact before we start generating temporary frames
duke@0 1229 // that can confuse an asynchronous stack walker. This counter is
duke@0 1230 // decremented at the end of unpack_frames().
duke@0 1231 thread->inc_in_deopt_handler();
duke@0 1232
duke@0 1233 // We need to update the map if we have biased locking.
duke@0 1234 RegisterMap reg_map(thread, UseBiasedLocking);
duke@0 1235 frame stub_frame = thread->last_frame();
duke@0 1236 frame fr = stub_frame.sender(&reg_map);
duke@0 1237 // Make sure the calling nmethod is not getting deoptimized and removed
duke@0 1238 // before we are done with it.
duke@0 1239 nmethodLocker nl(fr.pc());
duke@0 1240
never@3064 1241 // Log a message
never@3064 1242 Events::log_deopt_message(thread, "Uncommon trap %d fr.pc " INTPTR_FORMAT,
never@3064 1243 trap_request, fr.pc());
never@3064 1244
duke@0 1245 {
duke@0 1246 ResourceMark rm;
duke@0 1247
duke@0 1248 // Revoke biases of any monitors in the frame to ensure we can migrate them
duke@0 1249 revoke_biases_of_monitors(thread, fr, &reg_map);
duke@0 1250
duke@0 1251 DeoptReason reason = trap_request_reason(trap_request);
duke@0 1252 DeoptAction action = trap_request_action(trap_request);
duke@0 1253 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
duke@0 1254
duke@0 1255 vframe* vf = vframe::new_vframe(&fr, &reg_map, thread);
duke@0 1256 compiledVFrame* cvf = compiledVFrame::cast(vf);
duke@0 1257
duke@0 1258 nmethod* nm = cvf->code();
duke@0 1259
duke@0 1260 ScopeDesc* trap_scope = cvf->scope();
duke@0 1261 methodHandle trap_method = trap_scope->method();
duke@0 1262 int trap_bci = trap_scope->bci();
never@2027 1263 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
duke@0 1264
duke@0 1265 // Record this event in the histogram.
duke@0 1266 gather_statistics(reason, action, trap_bc);
duke@0 1267
duke@0 1268 // Ensure that we can record deopt. history:
duke@0 1269 bool create_if_missing = ProfileTraps;
duke@0 1270
duke@0 1271 methodDataHandle trap_mdo
duke@0 1272 (THREAD, get_method_data(thread, trap_method, create_if_missing));
duke@0 1273
duke@0 1274 // Print a bunch of diagnostics, if requested.
duke@0 1275 if (TraceDeoptimization || LogCompilation) {
duke@0 1276 ResourceMark rm;
duke@0 1277 ttyLocker ttyl;
duke@0 1278 char buf[100];
duke@0 1279 if (xtty != NULL) {
duke@0 1280 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
duke@0 1281 os::current_thread_id(),
duke@0 1282 format_trap_request(buf, sizeof(buf), trap_request));
duke@0 1283 nm->log_identity(xtty);
duke@0 1284 }
coleenp@2062 1285 Symbol* class_name = NULL;
duke@0 1286 bool unresolved = false;
duke@0 1287 if (unloaded_class_index >= 0) {
duke@0 1288 constantPoolHandle constants (THREAD, trap_method->constants());
duke@0 1289 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
coleenp@2062 1290 class_name = constants->klass_name_at(unloaded_class_index);
duke@0 1291 unresolved = true;
duke@0 1292 if (xtty != NULL)
duke@0 1293 xtty->print(" unresolved='1'");
duke@0 1294 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
coleenp@2062 1295 class_name = constants->symbol_at(unloaded_class_index);
duke@0 1296 }
duke@0 1297 if (xtty != NULL)
duke@0 1298 xtty->name(class_name);
duke@0 1299 }
duke@0 1300 if (xtty != NULL && trap_mdo.not_null()) {
duke@0 1301 // Dump the relevant MDO state.
duke@0 1302 // This is the deopt count for the current reason, any previous
duke@0 1303 // reasons or recompiles seen at this point.
duke@0 1304 int dcnt = trap_mdo->trap_count(reason);
duke@0 1305 if (dcnt != 0)
duke@0 1306 xtty->print(" count='%d'", dcnt);
duke@0 1307 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
duke@0 1308 int dos = (pdata == NULL)? 0: pdata->trap_state();
duke@0 1309 if (dos != 0) {
duke@0 1310 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
duke@0 1311 if (trap_state_is_recompiled(dos)) {
duke@0 1312 int recnt2 = trap_mdo->overflow_recompile_count();
duke@0 1313 if (recnt2 != 0)
duke@0 1314 xtty->print(" recompiles2='%d'", recnt2);
duke@0 1315 }
duke@0 1316 }
duke@0 1317 }
duke@0 1318 if (xtty != NULL) {
duke@0 1319 xtty->stamp();
duke@0 1320 xtty->end_head();
duke@0 1321 }
duke@0 1322 if (TraceDeoptimization) { // make noise on the tty
duke@0 1323 tty->print("Uncommon trap occurred in");
duke@0 1324 nm->method()->print_short_name(tty);
duke@0 1325 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
duke@0 1326 fr.pc(),
duke@0 1327 (int) os::current_thread_id(),
duke@0 1328 trap_reason_name(reason),
duke@0 1329 trap_action_name(action),
duke@0 1330 unloaded_class_index);
coleenp@2062 1331 if (class_name != NULL) {
duke@0 1332 tty->print(unresolved ? " unresolved class: " : " symbol: ");
duke@0 1333 class_name->print_symbol_on(tty);
duke@0 1334 }
duke@0 1335 tty->cr();
duke@0 1336 }
duke@0 1337 if (xtty != NULL) {
duke@0 1338 // Log the precise location of the trap.
duke@0 1339 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
duke@0 1340 xtty->begin_elem("jvms bci='%d'", sd->bci());
duke@0 1341 xtty->method(sd->method());
duke@0 1342 xtty->end_elem();
duke@0 1343 if (sd->is_top()) break;
duke@0 1344 }
duke@0 1345 xtty->tail("uncommon_trap");
duke@0 1346 }
duke@0 1347 }
duke@0 1348 // (End diagnostic printout.)
duke@0 1349
duke@0 1350 // Load class if necessary
duke@0 1351 if (unloaded_class_index >= 0) {
duke@0 1352 constantPoolHandle constants(THREAD, trap_method->constants());
duke@0 1353 load_class_by_index(constants, unloaded_class_index);
duke@0 1354 }
duke@0 1355
duke@0 1356 // Flush the nmethod if necessary and desirable.
duke@0 1357 //
duke@0 1358 // We need to avoid situations where we are re-flushing the nmethod
duke@0 1359 // because of a hot deoptimization site. Repeated flushes at the same
duke@0 1360 // point need to be detected by the compiler and avoided. If the compiler
duke@0 1361 // cannot avoid them (or has a bug and "refuses" to avoid them), this
duke@0 1362 // module must take measures to avoid an infinite cycle of recompilation
duke@0 1363 // and deoptimization. There are several such measures:
duke@0 1364 //
duke@0 1365 // 1. If a recompilation is ordered a second time at some site X
duke@0 1366 // and for the same reason R, the action is adjusted to 'reinterpret',
duke@0 1367 // to give the interpreter time to exercise the method more thoroughly.
duke@0 1368 // If this happens, the method's overflow_recompile_count is incremented.
duke@0 1369 //
duke@0 1370 // 2. If the compiler fails to reduce the deoptimization rate, then
duke@0 1371 // the method's overflow_recompile_count will begin to exceed the set
duke@0 1372 // limit PerBytecodeRecompilationCutoff. If this happens, the action
duke@0 1373 // is adjusted to 'make_not_compilable', and the method is abandoned
duke@0 1374 // to the interpreter. This is a performance hit for hot methods,
duke@0 1375 // but is better than a disastrous infinite cycle of recompilations.
duke@0 1376 // (Actually, only the method containing the site X is abandoned.)
duke@0 1377 //
duke@0 1378 // 3. In parallel with the previous measures, if the total number of
duke@0 1379 // recompilations of a method exceeds the much larger set limit
duke@0 1380 // PerMethodRecompilationCutoff, the method is abandoned.
duke@0 1381 // This should only happen if the method is very large and has
duke@0 1382 // many "lukewarm" deoptimizations. The code which enforces this
duke@0 1383 // limit is elsewhere (class nmethod, class methodOopDesc).
duke@0 1384 //
duke@0 1385 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
duke@0 1386 // to recompile at each bytecode independently of the per-BCI cutoff.
duke@0 1387 //
duke@0 1388 // The decision to update code is up to the compiler, and is encoded
duke@0 1389 // in the Action_xxx code. If the compiler requests Action_none
duke@0 1390 // no trap state is changed, no compiled code is changed, and the
duke@0 1391 // computation suffers along in the interpreter.
duke@0 1392 //
duke@0 1393 // The other action codes specify various tactics for decompilation
duke@0 1394 // and recompilation. Action_maybe_recompile is the loosest, and
duke@0 1395 // allows the compiled code to stay around until enough traps are seen,
duke@0 1396 // and until the compiler gets around to recompiling the trapping method.
duke@0 1397 //
duke@0 1398 // The other actions cause immediate removal of the present code.
duke@0 1399
duke@0 1400 bool update_trap_state = true;
duke@0 1401 bool make_not_entrant = false;
duke@0 1402 bool make_not_compilable = false;
iveresov@1703 1403 bool reprofile = false;
duke@0 1404 switch (action) {
duke@0 1405 case Action_none:
duke@0 1406 // Keep the old code.
duke@0 1407 update_trap_state = false;
duke@0 1408 break;
duke@0 1409 case Action_maybe_recompile:
duke@0 1410 // Do not need to invalidate the present code, but we can
duke@0 1411 // initiate another
duke@0 1412 // Start compiler without (necessarily) invalidating the nmethod.
duke@0 1413 // The system will tolerate the old code, but new code should be
duke@0 1414 // generated when possible.
duke@0 1415 break;
duke@0 1416 case Action_reinterpret:
duke@0 1417 // Go back into the interpreter for a while, and then consider
duke@0 1418 // recompiling form scratch.
duke@0 1419 make_not_entrant = true;
duke@0 1420 // Reset invocation counter for outer most method.
duke@0 1421 // This will allow the interpreter to exercise the bytecodes
duke@0 1422 // for a while before recompiling.
duke@0 1423 // By contrast, Action_make_not_entrant is immediate.
duke@0 1424 //
duke@0 1425 // Note that the compiler will track null_check, null_assert,
duke@0 1426 // range_check, and class_check events and log them as if they
duke@0 1427 // had been traps taken from compiled code. This will update
duke@0 1428 // the MDO trap history so that the next compilation will
duke@0 1429 // properly detect hot trap sites.
iveresov@1703 1430 reprofile = true;
duke@0 1431 break;
duke@0 1432 case Action_make_not_entrant:
duke@0 1433 // Request immediate recompilation, and get rid of the old code.
duke@0 1434 // Make them not entrant, so next time they are called they get
duke@0 1435 // recompiled. Unloaded classes are loaded now so recompile before next
duke@0 1436 // time they are called. Same for uninitialized. The interpreter will
duke@0 1437 // link the missing class, if any.
duke@0 1438 make_not_entrant = true;
duke@0 1439 break;
duke@0 1440 case Action_make_not_compilable:
duke@0 1441 // Give up on compiling this method at all.
duke@0 1442 make_not_entrant = true;
duke@0 1443 make_not_compilable = true;
duke@0 1444 break;
duke@0 1445 default:
duke@0 1446 ShouldNotReachHere();
duke@0 1447 }
duke@0 1448
duke@0 1449 // Setting +ProfileTraps fixes the following, on all platforms:
duke@0 1450 // 4852688: ProfileInterpreter is off by default for ia64. The result is
duke@0 1451 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
duke@0 1452 // recompile relies on a methodDataOop to record heroic opt failures.
duke@0 1453
duke@0 1454 // Whether the interpreter is producing MDO data or not, we also need
duke@0 1455 // to use the MDO to detect hot deoptimization points and control
duke@0 1456 // aggressive optimization.
kvn@1206 1457 bool inc_recompile_count = false;
kvn@1206 1458 ProfileData* pdata = NULL;
duke@0 1459 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
duke@0 1460 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
duke@0 1461 uint this_trap_count = 0;
duke@0 1462 bool maybe_prior_trap = false;
duke@0 1463 bool maybe_prior_recompile = false;
kvn@1206 1464 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
duke@0 1465 //outputs:
duke@0 1466 this_trap_count,
duke@0 1467 maybe_prior_trap,
duke@0 1468 maybe_prior_recompile);
duke@0 1469 // Because the interpreter also counts null, div0, range, and class
duke@0 1470 // checks, these traps from compiled code are double-counted.
duke@0 1471 // This is harmless; it just means that the PerXTrapLimit values
duke@0 1472 // are in effect a little smaller than they look.
duke@0 1473
duke@0 1474 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
duke@0 1475 if (per_bc_reason != Reason_none) {
duke@0 1476 // Now take action based on the partially known per-BCI history.
duke@0 1477 if (maybe_prior_trap
duke@0 1478 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
duke@0 1479 // If there are too many traps at this BCI, force a recompile.
duke@0 1480 // This will allow the compiler to see the limit overflow, and
duke@0 1481 // take corrective action, if possible. The compiler generally
duke@0 1482 // does not use the exact PerBytecodeTrapLimit value, but instead
duke@0 1483 // changes its tactics if it sees any traps at all. This provides
duke@0 1484 // a little hysteresis, delaying a recompile until a trap happens
duke@0 1485 // several times.
duke@0 1486 //
duke@0 1487 // Actually, since there is only one bit of counter per BCI,
duke@0 1488 // the possible per-BCI counts are {0,1,(per-method count)}.
duke@0 1489 // This produces accurate results if in fact there is only
duke@0 1490 // one hot trap site, but begins to get fuzzy if there are
duke@0 1491 // many sites. For example, if there are ten sites each
duke@0 1492 // trapping two or more times, they each get the blame for
duke@0 1493 // all of their traps.
duke@0 1494 make_not_entrant = true;
duke@0 1495 }
duke@0 1496
duke@0 1497 // Detect repeated recompilation at the same BCI, and enforce a limit.
duke@0 1498 if (make_not_entrant && maybe_prior_recompile) {
duke@0 1499 // More than one recompile at this point.
kvn@1206 1500 inc_recompile_count = maybe_prior_trap;
duke@0 1501 }
duke@0 1502 } else {
duke@0 1503 // For reasons which are not recorded per-bytecode, we simply
duke@0 1504 // force recompiles unconditionally.
duke@0 1505 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
duke@0 1506 make_not_entrant = true;
duke@0 1507 }
duke@0 1508
duke@0 1509 // Go back to the compiler if there are too many traps in this method.
duke@0 1510 if (this_trap_count >= (uint)PerMethodTrapLimit) {
duke@0 1511 // If there are too many traps in this method, force a recompile.
duke@0 1512 // This will allow the compiler to see the limit overflow, and
duke@0 1513 // take corrective action, if possible.
duke@0 1514 // (This condition is an unlikely backstop only, because the
duke@0 1515 // PerBytecodeTrapLimit is more likely to take effect first,
duke@0 1516 // if it is applicable.)
duke@0 1517 make_not_entrant = true;
duke@0 1518 }
duke@0 1519
duke@0 1520 // Here's more hysteresis: If there has been a recompile at
duke@0 1521 // this trap point already, run the method in the interpreter
duke@0 1522 // for a while to exercise it more thoroughly.
duke@0 1523 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
iveresov@1703 1524 reprofile = true;
duke@0 1525 }
duke@0 1526
kvn@1206 1527 }
kvn@1206 1528
kvn@1206 1529 // Take requested actions on the method:
kvn@1206 1530
kvn@1206 1531 // Recompile
kvn@1206 1532 if (make_not_entrant) {
kvn@1206 1533 if (!nm->make_not_entrant()) {
kvn@1206 1534 return; // the call did not change nmethod's state
kvn@1206 1535 }
kvn@1206 1536
kvn@1206 1537 if (pdata != NULL) {
duke@0 1538 // Record the recompilation event, if any.
duke@0 1539 int tstate0 = pdata->trap_state();
duke@0 1540 int tstate1 = trap_state_set_recompiled(tstate0, true);
duke@0 1541 if (tstate1 != tstate0)
duke@0 1542 pdata->set_trap_state(tstate1);
duke@0 1543 }
duke@0 1544 }
duke@0 1545
kvn@1206 1546 if (inc_recompile_count) {
kvn@1206 1547 trap_mdo->inc_overflow_recompile_count();
kvn@1206 1548 if ((uint)trap_mdo->overflow_recompile_count() >
kvn@1206 1549 (uint)PerBytecodeRecompilationCutoff) {
kvn@1206 1550 // Give up on the method containing the bad BCI.
kvn@1206 1551 if (trap_method() == nm->method()) {
kvn@1206 1552 make_not_compilable = true;
kvn@1206 1553 } else {
iveresov@1703 1554 trap_method->set_not_compilable(CompLevel_full_optimization);
kvn@1206 1555 // But give grace to the enclosing nm->method().
kvn@1206 1556 }
kvn@1206 1557 }
kvn@1206 1558 }
duke@0 1559
iveresov@1703 1560 // Reprofile
iveresov@1703 1561 if (reprofile) {
iveresov@1703 1562 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
duke@0 1563 }
duke@0 1564
duke@0 1565 // Give up compiling
iveresov@1703 1566 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
duke@0 1567 assert(make_not_entrant, "consistent");
iveresov@1703 1568 nm->method()->set_not_compilable(CompLevel_full_optimization);
duke@0 1569 }
duke@0 1570
duke@0 1571 } // Free marked resources
duke@0 1572
duke@0 1573 }
duke@0 1574 JRT_END
duke@0 1575
duke@0 1576 methodDataOop
duke@0 1577 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
duke@0 1578 bool create_if_missing) {
duke@0 1579 Thread* THREAD = thread;
duke@0 1580 methodDataOop mdo = m()->method_data();
duke@0 1581 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
duke@0 1582 // Build an MDO. Ignore errors like OutOfMemory;
duke@0 1583 // that simply means we won't have an MDO to update.
duke@0 1584 methodOopDesc::build_interpreter_method_data(m, THREAD);
duke@0 1585 if (HAS_PENDING_EXCEPTION) {
duke@0 1586 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
duke@0 1587 CLEAR_PENDING_EXCEPTION;
duke@0 1588 }
duke@0 1589 mdo = m()->method_data();
duke@0 1590 }
duke@0 1591 return mdo;
duke@0 1592 }
duke@0 1593
duke@0 1594 ProfileData*
duke@0 1595 Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
duke@0 1596 int trap_bci,
duke@0 1597 Deoptimization::DeoptReason reason,
duke@0 1598 //outputs:
duke@0 1599 uint& ret_this_trap_count,
duke@0 1600 bool& ret_maybe_prior_trap,
duke@0 1601 bool& ret_maybe_prior_recompile) {
duke@0 1602 uint prior_trap_count = trap_mdo->trap_count(reason);
duke@0 1603 uint this_trap_count = trap_mdo->inc_trap_count(reason);
duke@0 1604
duke@0 1605 // If the runtime cannot find a place to store trap history,
duke@0 1606 // it is estimated based on the general condition of the method.
duke@0 1607 // If the method has ever been recompiled, or has ever incurred
duke@0 1608 // a trap with the present reason , then this BCI is assumed
duke@0 1609 // (pessimistically) to be the culprit.
duke@0 1610 bool maybe_prior_trap = (prior_trap_count != 0);
duke@0 1611 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
duke@0 1612 ProfileData* pdata = NULL;
duke@0 1613
duke@0 1614
duke@0 1615 // For reasons which are recorded per bytecode, we check per-BCI data.
duke@0 1616 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
duke@0 1617 if (per_bc_reason != Reason_none) {
duke@0 1618 // Find the profile data for this BCI. If there isn't one,
duke@0 1619 // try to allocate one from the MDO's set of spares.
duke@0 1620 // This will let us detect a repeated trap at this point.
duke@0 1621 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
duke@0 1622
duke@0 1623 if (pdata != NULL) {
duke@0 1624 // Query the trap state of this profile datum.
duke@0 1625 int tstate0 = pdata->trap_state();
duke@0 1626 if (!trap_state_has_reason(tstate0, per_bc_reason))
duke@0 1627 maybe_prior_trap = false;
duke@0 1628 if (!trap_state_is_recompiled(tstate0))
duke@0 1629 maybe_prior_recompile = false;
duke@0 1630
duke@0 1631 // Update the trap state of this profile datum.
duke@0 1632 int tstate1 = tstate0;
duke@0 1633 // Record the reason.
duke@0 1634 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
duke@0 1635 // Store the updated state on the MDO, for next time.
duke@0 1636 if (tstate1 != tstate0)
duke@0 1637 pdata->set_trap_state(tstate1);
duke@0 1638 } else {
kvn@1206 1639 if (LogCompilation && xtty != NULL) {
kvn@1206 1640 ttyLocker ttyl;
duke@0 1641 // Missing MDP? Leave a small complaint in the log.
duke@0 1642 xtty->elem("missing_mdp bci='%d'", trap_bci);
kvn@1206 1643 }
duke@0 1644 }
duke@0 1645 }
duke@0 1646
duke@0 1647 // Return results:
duke@0 1648 ret_this_trap_count = this_trap_count;
duke@0 1649 ret_maybe_prior_trap = maybe_prior_trap;
duke@0 1650 ret_maybe_prior_recompile = maybe_prior_recompile;
duke@0 1651 return pdata;
duke@0 1652 }
duke@0 1653
duke@0 1654 void
duke@0 1655 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
duke@0 1656 ResourceMark rm;
duke@0 1657 // Ignored outputs:
duke@0 1658 uint ignore_this_trap_count;
duke@0 1659 bool ignore_maybe_prior_trap;
duke@0 1660 bool ignore_maybe_prior_recompile;
duke@0 1661 query_update_method_data(trap_mdo, trap_bci,
duke@0 1662 (DeoptReason)reason,
duke@0 1663 ignore_this_trap_count,
duke@0 1664 ignore_maybe_prior_trap,
duke@0 1665 ignore_maybe_prior_recompile);
duke@0 1666 }
duke@0 1667
duke@0 1668 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
duke@0 1669
duke@0 1670 // Still in Java no safepoints
duke@0 1671 {
duke@0 1672 // This enters VM and may safepoint
duke@0 1673 uncommon_trap_inner(thread, trap_request);
duke@0 1674 }
duke@0 1675 return fetch_unroll_info_helper(thread);
duke@0 1676 }
duke@0 1677
duke@0 1678 // Local derived constants.
duke@0 1679 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
duke@0 1680 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
duke@0 1681 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
duke@0 1682
duke@0 1683 //---------------------------trap_state_reason---------------------------------
duke@0 1684 Deoptimization::DeoptReason
duke@0 1685 Deoptimization::trap_state_reason(int trap_state) {
duke@0 1686 // This assert provides the link between the width of DataLayout::trap_bits
duke@0 1687 // and the encoding of "recorded" reasons. It ensures there are enough
duke@0 1688 // bits to store all needed reasons in the per-BCI MDO profile.
duke@0 1689 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
duke@0 1690 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
duke@0 1691 trap_state -= recompile_bit;
duke@0 1692 if (trap_state == DS_REASON_MASK) {
duke@0 1693 return Reason_many;
duke@0 1694 } else {
duke@0 1695 assert((int)Reason_none == 0, "state=0 => Reason_none");
duke@0 1696 return (DeoptReason)trap_state;
duke@0 1697 }
duke@0 1698 }
duke@0 1699 //-------------------------trap_state_has_reason-------------------------------
duke@0 1700 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
duke@0 1701 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
duke@0 1702 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
duke@0 1703 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
duke@0 1704 trap_state -= recompile_bit;
duke@0 1705 if (trap_state == DS_REASON_MASK) {
duke@0 1706 return -1; // true, unspecifically (bottom of state lattice)
duke@0 1707 } else if (trap_state == reason) {
duke@0 1708 return 1; // true, definitely
duke@0 1709 } else if (trap_state == 0) {
duke@0 1710 return 0; // false, definitely (top of state lattice)
duke@0 1711 } else {
duke@0 1712 return 0; // false, definitely
duke@0 1713 }
duke@0 1714 }
duke@0 1715 //-------------------------trap_state_add_reason-------------------------------
duke@0 1716 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
duke@0 1717 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
duke@0 1718 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
duke@0 1719 trap_state -= recompile_bit;
duke@0 1720 if (trap_state == DS_REASON_MASK) {
duke@0 1721 return trap_state + recompile_bit; // already at state lattice bottom
duke@0 1722 } else if (trap_state == reason) {
duke@0 1723 return trap_state + recompile_bit; // the condition is already true
duke@0 1724 } else if (trap_state == 0) {
duke@0 1725 return reason + recompile_bit; // no condition has yet been true
duke@0 1726 } else {
duke@0 1727 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
duke@0 1728 }
duke@0 1729 }
duke@0 1730 //-----------------------trap_state_is_recompiled------------------------------
duke@0 1731 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
duke@0 1732 return (trap_state & DS_RECOMPILE_BIT) != 0;
duke@0 1733 }
duke@0 1734 //-----------------------trap_state_set_recompiled-----------------------------
duke@0 1735 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
duke@0 1736 if (z) return trap_state | DS_RECOMPILE_BIT;
duke@0 1737 else return trap_state & ~DS_RECOMPILE_BIT;
duke@0 1738 }
duke@0 1739 //---------------------------format_trap_state---------------------------------
duke@0 1740 // This is used for debugging and diagnostics, including hotspot.log output.
duke@0 1741 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
duke@0 1742 int trap_state) {
duke@0 1743 DeoptReason reason = trap_state_reason(trap_state);
duke@0 1744 bool recomp_flag = trap_state_is_recompiled(trap_state);
duke@0 1745 // Re-encode the state from its decoded components.
duke@0 1746 int decoded_state = 0;
duke@0 1747 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
duke@0 1748 decoded_state = trap_state_add_reason(decoded_state, reason);
duke@0 1749 if (recomp_flag)
duke@0 1750 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
duke@0 1751 // If the state re-encodes properly, format it symbolically.
duke@0 1752 // Because this routine is used for debugging and diagnostics,
duke@0 1753 // be robust even if the state is a strange value.
duke@0 1754 size_t len;
duke@0 1755 if (decoded_state != trap_state) {
duke@0 1756 // Random buggy state that doesn't decode??
duke@0 1757 len = jio_snprintf(buf, buflen, "#%d", trap_state);
duke@0 1758 } else {
duke@0 1759 len = jio_snprintf(buf, buflen, "%s%s",
duke@0 1760 trap_reason_name(reason),
duke@0 1761 recomp_flag ? " recompiled" : "");
duke@0 1762 }
duke@0 1763 if (len >= buflen)
duke@0 1764 buf[buflen-1] = '\0';
duke@0 1765 return buf;
duke@0 1766 }
duke@0 1767
duke@0 1768
duke@0 1769 //--------------------------------statics--------------------------------------
duke@0 1770 Deoptimization::DeoptAction Deoptimization::_unloaded_action
duke@0 1771 = Deoptimization::Action_reinterpret;
duke@0 1772 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
duke@0 1773 // Note: Keep this in sync. with enum DeoptReason.
duke@0 1774 "none",
duke@0 1775 "null_check",
duke@0 1776 "null_assert",
duke@0 1777 "range_check",
duke@0 1778 "class_check",
duke@0 1779 "array_check",
duke@0 1780 "intrinsic",
kvn@1206 1781 "bimorphic",
duke@0 1782 "unloaded",
duke@0 1783 "uninitialized",
duke@0 1784 "unreached",
duke@0 1785 "unhandled",
duke@0 1786 "constraint",
duke@0 1787 "div0_check",
cfang@1172 1788 "age",
kvn@2442 1789 "predicate",
kvn@2442 1790 "loop_limit_check"
duke@0 1791 };
duke@0 1792 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
duke@0 1793 // Note: Keep this in sync. with enum DeoptAction.
duke@0 1794 "none",
duke@0 1795 "maybe_recompile",
duke@0 1796 "reinterpret",
duke@0 1797 "make_not_entrant",
duke@0 1798 "make_not_compilable"
duke@0 1799 };
duke@0 1800
duke@0 1801 const char* Deoptimization::trap_reason_name(int reason) {
duke@0 1802 if (reason == Reason_many) return "many";
duke@0 1803 if ((uint)reason < Reason_LIMIT)
duke@0 1804 return _trap_reason_name[reason];
duke@0 1805 static char buf[20];
duke@0 1806 sprintf(buf, "reason%d", reason);
duke@0 1807 return buf;
duke@0 1808 }
duke@0 1809 const char* Deoptimization::trap_action_name(int action) {
duke@0 1810 if ((uint)action < Action_LIMIT)
duke@0 1811 return _trap_action_name[action];
duke@0 1812 static char buf[20];
duke@0 1813 sprintf(buf, "action%d", action);
duke@0 1814 return buf;
duke@0 1815 }
duke@0 1816
duke@0 1817 // This is used for debugging and diagnostics, including hotspot.log output.
duke@0 1818 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
duke@0 1819 int trap_request) {
duke@0 1820 jint unloaded_class_index = trap_request_index(trap_request);
duke@0 1821 const char* reason = trap_reason_name(trap_request_reason(trap_request));
duke@0 1822 const char* action = trap_action_name(trap_request_action(trap_request));
duke@0 1823 size_t len;
duke@0 1824 if (unloaded_class_index < 0) {
duke@0 1825 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
duke@0 1826 reason, action);
duke@0 1827 } else {
duke@0 1828 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
duke@0 1829 reason, action, unloaded_class_index);
duke@0 1830 }
duke@0 1831 if (len >= buflen)
duke@0 1832 buf[buflen-1] = '\0';
duke@0 1833 return buf;
duke@0 1834 }
duke@0 1835
duke@0 1836 juint Deoptimization::_deoptimization_hist
duke@0 1837 [Deoptimization::Reason_LIMIT]
duke@0 1838 [1 + Deoptimization::Action_LIMIT]
duke@0 1839 [Deoptimization::BC_CASE_LIMIT]
duke@0 1840 = {0};
duke@0 1841
duke@0 1842 enum {
duke@0 1843 LSB_BITS = 8,
duke@0 1844 LSB_MASK = right_n_bits(LSB_BITS)
duke@0 1845 };
duke@0 1846
duke@0 1847 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
duke@0 1848 Bytecodes::Code bc) {
duke@0 1849 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
duke@0 1850 assert(action >= 0 && action < Action_LIMIT, "oob");
duke@0 1851 _deoptimization_hist[Reason_none][0][0] += 1; // total
duke@0 1852 _deoptimization_hist[reason][0][0] += 1; // per-reason total
duke@0 1853 juint* cases = _deoptimization_hist[reason][1+action];
duke@0 1854 juint* bc_counter_addr = NULL;
duke@0 1855 juint bc_counter = 0;
duke@0 1856 // Look for an unused counter, or an exact match to this BC.
duke@0 1857 if (bc != Bytecodes::_illegal) {
duke@0 1858 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
duke@0 1859 juint* counter_addr = &cases[bc_case];
duke@0 1860 juint counter = *counter_addr;
duke@0 1861 if ((counter == 0 && bc_counter_addr == NULL)
duke@0 1862 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
duke@0 1863 // this counter is either free or is already devoted to this BC
duke@0 1864 bc_counter_addr = counter_addr;
duke@0 1865 bc_counter = counter | bc;
duke@0 1866 }
duke@0 1867 }
duke@0 1868 }
duke@0 1869 if (bc_counter_addr == NULL) {
duke@0 1870 // Overflow, or no given bytecode.
duke@0 1871 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
duke@0 1872 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
duke@0 1873 }
duke@0 1874 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
duke@0 1875 }
duke@0 1876
duke@0 1877 jint Deoptimization::total_deoptimization_count() {
duke@0 1878 return _deoptimization_hist[Reason_none][0][0];
duke@0 1879 }
duke@0 1880
duke@0 1881 jint Deoptimization::deoptimization_count(DeoptReason reason) {
duke@0 1882 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
duke@0 1883 return _deoptimization_hist[reason][0][0];
duke@0 1884 }
duke@0 1885
duke@0 1886 void Deoptimization::print_statistics() {
duke@0 1887 juint total = total_deoptimization_count();
duke@0 1888 juint account = total;
duke@0 1889 if (total != 0) {
duke@0 1890 ttyLocker ttyl;
duke@0 1891 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
duke@0 1892 tty->print_cr("Deoptimization traps recorded:");
duke@0 1893 #define PRINT_STAT_LINE(name, r) \
duke@0 1894 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
duke@0 1895 PRINT_STAT_LINE("total", total);
duke@0 1896 // For each non-zero entry in the histogram, print the reason,
duke@0 1897 // the action, and (if specifically known) the type of bytecode.
duke@0 1898 for (int reason = 0; reason < Reason_LIMIT; reason++) {
duke@0 1899 for (int action = 0; action < Action_LIMIT; action++) {
duke@0 1900 juint* cases = _deoptimization_hist[reason][1+action];
duke@0 1901 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
duke@0 1902 juint counter = cases[bc_case];
duke@0 1903 if (counter != 0) {
duke@0 1904 char name[1*K];
duke@0 1905 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
duke@0 1906 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
duke@0 1907 bc = Bytecodes::_illegal;
duke@0 1908 sprintf(name, "%s/%s/%s",
duke@0 1909 trap_reason_name(reason),
duke@0 1910 trap_action_name(action),
duke@0 1911 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
duke@0 1912 juint r = counter >> LSB_BITS;
duke@0 1913 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
duke@0 1914 account -= r;
duke@0 1915 }
duke@0 1916 }
duke@0 1917 }
duke@0 1918 }
duke@0 1919 if (account != 0) {
duke@0 1920 PRINT_STAT_LINE("unaccounted", account);
duke@0 1921 }
duke@0 1922 #undef PRINT_STAT_LINE
duke@0 1923 if (xtty != NULL) xtty->tail("statistics");
duke@0 1924 }
duke@0 1925 }
twisti@1612 1926 #else // COMPILER2 || SHARK
duke@0 1927
duke@0 1928
duke@0 1929 // Stubs for C1 only system.
duke@0 1930 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
duke@0 1931 return false;
duke@0 1932 }
duke@0 1933
duke@0 1934 const char* Deoptimization::trap_reason_name(int reason) {
duke@0 1935 return "unknown";
duke@0 1936 }
duke@0 1937
duke@0 1938 void Deoptimization::print_statistics() {
duke@0 1939 // no output
duke@0 1940 }
duke@0 1941
duke@0 1942 void
duke@0 1943 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
duke@0 1944 // no udpate
duke@0 1945 }
duke@0 1946
duke@0 1947 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
duke@0 1948 return 0;
duke@0 1949 }
duke@0 1950
duke@0 1951 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
duke@0 1952 Bytecodes::Code bc) {
duke@0 1953 // no update
duke@0 1954 }
duke@0 1955
duke@0 1956 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
duke@0 1957 int trap_state) {
duke@0 1958 jio_snprintf(buf, buflen, "#%d", trap_state);
duke@0 1959 return buf;
duke@0 1960 }
duke@0 1961
twisti@1612 1962 #endif // COMPILER2 || SHARK