Sat, 28 Mar 2009 15:47:29 -0700
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
Summary: When using compressed oops, rather than chaining the overflowed grey objects' pre-images through their klass words, we use GC-worker thread-local overflow stacks.
Reviewed-by: jcoomes, jmasa
duke@435 | 1 | /* |
xdono@631 | 2 | * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_deoptimization.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | bool DeoptimizationMarker::_is_active = false; |
duke@435 | 29 | |
duke@435 | 30 | Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, |
duke@435 | 31 | int caller_adjustment, |
duke@435 | 32 | int number_of_frames, |
duke@435 | 33 | intptr_t* frame_sizes, |
duke@435 | 34 | address* frame_pcs, |
duke@435 | 35 | BasicType return_type) { |
duke@435 | 36 | _size_of_deoptimized_frame = size_of_deoptimized_frame; |
duke@435 | 37 | _caller_adjustment = caller_adjustment; |
duke@435 | 38 | _number_of_frames = number_of_frames; |
duke@435 | 39 | _frame_sizes = frame_sizes; |
duke@435 | 40 | _frame_pcs = frame_pcs; |
duke@435 | 41 | _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); |
duke@435 | 42 | _return_type = return_type; |
duke@435 | 43 | // PD (x86 only) |
duke@435 | 44 | _counter_temp = 0; |
duke@435 | 45 | _initial_fp = 0; |
duke@435 | 46 | _unpack_kind = 0; |
duke@435 | 47 | _sender_sp_temp = 0; |
duke@435 | 48 | |
duke@435 | 49 | _total_frame_sizes = size_of_frames(); |
duke@435 | 50 | } |
duke@435 | 51 | |
duke@435 | 52 | |
duke@435 | 53 | Deoptimization::UnrollBlock::~UnrollBlock() { |
duke@435 | 54 | FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); |
duke@435 | 55 | FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); |
duke@435 | 56 | FREE_C_HEAP_ARRAY(intptr_t, _register_block); |
duke@435 | 57 | } |
duke@435 | 58 | |
duke@435 | 59 | |
duke@435 | 60 | intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { |
duke@435 | 61 | assert(register_number < RegisterMap::reg_count, "checking register number"); |
duke@435 | 62 | return &_register_block[register_number * 2]; |
duke@435 | 63 | } |
duke@435 | 64 | |
duke@435 | 65 | |
duke@435 | 66 | |
duke@435 | 67 | int Deoptimization::UnrollBlock::size_of_frames() const { |
duke@435 | 68 | // Acount first for the adjustment of the initial frame |
duke@435 | 69 | int result = _caller_adjustment; |
duke@435 | 70 | for (int index = 0; index < number_of_frames(); index++) { |
duke@435 | 71 | result += frame_sizes()[index]; |
duke@435 | 72 | } |
duke@435 | 73 | return result; |
duke@435 | 74 | } |
duke@435 | 75 | |
duke@435 | 76 | |
duke@435 | 77 | void Deoptimization::UnrollBlock::print() { |
duke@435 | 78 | ttyLocker ttyl; |
duke@435 | 79 | tty->print_cr("UnrollBlock"); |
duke@435 | 80 | tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); |
duke@435 | 81 | tty->print( " frame_sizes: "); |
duke@435 | 82 | for (int index = 0; index < number_of_frames(); index++) { |
duke@435 | 83 | tty->print("%d ", frame_sizes()[index]); |
duke@435 | 84 | } |
duke@435 | 85 | tty->cr(); |
duke@435 | 86 | } |
duke@435 | 87 | |
duke@435 | 88 | |
duke@435 | 89 | // In order to make fetch_unroll_info work properly with escape |
duke@435 | 90 | // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and |
duke@435 | 91 | // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation |
duke@435 | 92 | // of previously eliminated objects occurs in realloc_objects, which is |
duke@435 | 93 | // called from the method fetch_unroll_info_helper below. |
duke@435 | 94 | JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) |
duke@435 | 95 | // It is actually ok to allocate handles in a leaf method. It causes no safepoints, |
duke@435 | 96 | // but makes the entry a little slower. There is however a little dance we have to |
duke@435 | 97 | // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro |
duke@435 | 98 | |
duke@435 | 99 | // fetch_unroll_info() is called at the beginning of the deoptimization |
duke@435 | 100 | // handler. Note this fact before we start generating temporary frames |
duke@435 | 101 | // that can confuse an asynchronous stack walker. This counter is |
duke@435 | 102 | // decremented at the end of unpack_frames(). |
duke@435 | 103 | thread->inc_in_deopt_handler(); |
duke@435 | 104 | |
duke@435 | 105 | return fetch_unroll_info_helper(thread); |
duke@435 | 106 | JRT_END |
duke@435 | 107 | |
duke@435 | 108 | |
duke@435 | 109 | // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) |
duke@435 | 110 | Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { |
duke@435 | 111 | |
duke@435 | 112 | // Note: there is a safepoint safety issue here. No matter whether we enter |
duke@435 | 113 | // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once |
duke@435 | 114 | // the vframeArray is created. |
duke@435 | 115 | // |
duke@435 | 116 | |
duke@435 | 117 | // Allocate our special deoptimization ResourceMark |
duke@435 | 118 | DeoptResourceMark* dmark = new DeoptResourceMark(thread); |
duke@435 | 119 | assert(thread->deopt_mark() == NULL, "Pending deopt!"); |
duke@435 | 120 | thread->set_deopt_mark(dmark); |
duke@435 | 121 | |
duke@435 | 122 | frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect |
duke@435 | 123 | RegisterMap map(thread, true); |
duke@435 | 124 | RegisterMap dummy_map(thread, false); |
duke@435 | 125 | // Now get the deoptee with a valid map |
duke@435 | 126 | frame deoptee = stub_frame.sender(&map); |
duke@435 | 127 | |
duke@435 | 128 | // Create a growable array of VFrames where each VFrame represents an inlined |
duke@435 | 129 | // Java frame. This storage is allocated with the usual system arena. |
duke@435 | 130 | assert(deoptee.is_compiled_frame(), "Wrong frame type"); |
duke@435 | 131 | GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); |
duke@435 | 132 | vframe* vf = vframe::new_vframe(&deoptee, &map, thread); |
duke@435 | 133 | while (!vf->is_top()) { |
duke@435 | 134 | assert(vf->is_compiled_frame(), "Wrong frame type"); |
duke@435 | 135 | chunk->push(compiledVFrame::cast(vf)); |
duke@435 | 136 | vf = vf->sender(); |
duke@435 | 137 | } |
duke@435 | 138 | assert(vf->is_compiled_frame(), "Wrong frame type"); |
duke@435 | 139 | chunk->push(compiledVFrame::cast(vf)); |
duke@435 | 140 | |
duke@435 | 141 | #ifdef COMPILER2 |
duke@435 | 142 | // Reallocate the non-escaping objects and restore their fields. Then |
duke@435 | 143 | // relock objects if synchronization on them was eliminated. |
kvn@479 | 144 | if (DoEscapeAnalysis) { |
kvn@479 | 145 | if (EliminateAllocations) { |
kvn@518 | 146 | assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); |
kvn@479 | 147 | GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); |
kvn@479 | 148 | bool reallocated = false; |
kvn@479 | 149 | if (objects != NULL) { |
kvn@479 | 150 | JRT_BLOCK |
kvn@479 | 151 | reallocated = realloc_objects(thread, &deoptee, objects, THREAD); |
kvn@479 | 152 | JRT_END |
duke@435 | 153 | } |
kvn@479 | 154 | if (reallocated) { |
kvn@479 | 155 | reassign_fields(&deoptee, &map, objects); |
duke@435 | 156 | #ifndef PRODUCT |
duke@435 | 157 | if (TraceDeoptimization) { |
duke@435 | 158 | ttyLocker ttyl; |
kvn@479 | 159 | tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); |
kvn@479 | 160 | print_objects(objects); |
kvn@479 | 161 | } |
kvn@479 | 162 | #endif |
kvn@479 | 163 | } |
kvn@479 | 164 | } |
kvn@479 | 165 | if (EliminateLocks) { |
kvn@518 | 166 | #ifndef PRODUCT |
kvn@518 | 167 | bool first = true; |
kvn@518 | 168 | #endif |
kvn@479 | 169 | for (int i = 0; i < chunk->length(); i++) { |
kvn@518 | 170 | compiledVFrame* cvf = chunk->at(i); |
kvn@518 | 171 | assert (cvf->scope() != NULL,"expect only compiled java frames"); |
kvn@518 | 172 | GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
kvn@518 | 173 | if (monitors->is_nonempty()) { |
kvn@518 | 174 | relock_objects(monitors, thread); |
kvn@479 | 175 | #ifndef PRODUCT |
kvn@479 | 176 | if (TraceDeoptimization) { |
kvn@479 | 177 | ttyLocker ttyl; |
kvn@479 | 178 | for (int j = 0; j < monitors->length(); j++) { |
kvn@518 | 179 | MonitorInfo* mi = monitors->at(j); |
kvn@518 | 180 | if (mi->eliminated()) { |
kvn@518 | 181 | if (first) { |
kvn@518 | 182 | first = false; |
kvn@518 | 183 | tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); |
kvn@518 | 184 | } |
kvn@518 | 185 | tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); |
kvn@479 | 186 | } |
duke@435 | 187 | } |
duke@435 | 188 | } |
kvn@479 | 189 | #endif |
duke@435 | 190 | } |
duke@435 | 191 | } |
duke@435 | 192 | } |
duke@435 | 193 | } |
duke@435 | 194 | #endif // COMPILER2 |
duke@435 | 195 | // Ensure that no safepoint is taken after pointers have been stored |
duke@435 | 196 | // in fields of rematerialized objects. If a safepoint occurs from here on |
duke@435 | 197 | // out the java state residing in the vframeArray will be missed. |
duke@435 | 198 | No_Safepoint_Verifier no_safepoint; |
duke@435 | 199 | |
duke@435 | 200 | vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); |
duke@435 | 201 | |
duke@435 | 202 | assert(thread->vframe_array_head() == NULL, "Pending deopt!");; |
duke@435 | 203 | thread->set_vframe_array_head(array); |
duke@435 | 204 | |
duke@435 | 205 | // Now that the vframeArray has been created if we have any deferred local writes |
duke@435 | 206 | // added by jvmti then we can free up that structure as the data is now in the |
duke@435 | 207 | // vframeArray |
duke@435 | 208 | |
duke@435 | 209 | if (thread->deferred_locals() != NULL) { |
duke@435 | 210 | GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); |
duke@435 | 211 | int i = 0; |
duke@435 | 212 | do { |
duke@435 | 213 | // Because of inlining we could have multiple vframes for a single frame |
duke@435 | 214 | // and several of the vframes could have deferred writes. Find them all. |
duke@435 | 215 | if (list->at(i)->id() == array->original().id()) { |
duke@435 | 216 | jvmtiDeferredLocalVariableSet* dlv = list->at(i); |
duke@435 | 217 | list->remove_at(i); |
duke@435 | 218 | // individual jvmtiDeferredLocalVariableSet are CHeapObj's |
duke@435 | 219 | delete dlv; |
duke@435 | 220 | } else { |
duke@435 | 221 | i++; |
duke@435 | 222 | } |
duke@435 | 223 | } while ( i < list->length() ); |
duke@435 | 224 | if (list->length() == 0) { |
duke@435 | 225 | thread->set_deferred_locals(NULL); |
duke@435 | 226 | // free the list and elements back to C heap. |
duke@435 | 227 | delete list; |
duke@435 | 228 | } |
duke@435 | 229 | |
duke@435 | 230 | } |
duke@435 | 231 | |
duke@435 | 232 | // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. |
duke@435 | 233 | CodeBlob* cb = stub_frame.cb(); |
duke@435 | 234 | // Verify we have the right vframeArray |
duke@435 | 235 | assert(cb->frame_size() >= 0, "Unexpected frame size"); |
duke@435 | 236 | intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); |
duke@435 | 237 | |
duke@435 | 238 | #ifdef ASSERT |
duke@435 | 239 | assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); |
duke@435 | 240 | Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); |
duke@435 | 241 | #endif |
duke@435 | 242 | // This is a guarantee instead of an assert because if vframe doesn't match |
duke@435 | 243 | // we will unpack the wrong deoptimized frame and wind up in strange places |
duke@435 | 244 | // where it will be very difficult to figure out what went wrong. Better |
duke@435 | 245 | // to die an early death here than some very obscure death later when the |
duke@435 | 246 | // trail is cold. |
duke@435 | 247 | // Note: on ia64 this guarantee can be fooled by frames with no memory stack |
duke@435 | 248 | // in that it will fail to detect a problem when there is one. This needs |
duke@435 | 249 | // more work in tiger timeframe. |
duke@435 | 250 | guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); |
duke@435 | 251 | |
duke@435 | 252 | int number_of_frames = array->frames(); |
duke@435 | 253 | |
duke@435 | 254 | // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost |
duke@435 | 255 | // virtual activation, which is the reverse of the elements in the vframes array. |
duke@435 | 256 | intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); |
duke@435 | 257 | // +1 because we always have an interpreter return address for the final slot. |
duke@435 | 258 | address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); |
duke@435 | 259 | int callee_parameters = 0; |
duke@435 | 260 | int callee_locals = 0; |
duke@435 | 261 | int popframe_extra_args = 0; |
duke@435 | 262 | // Create an interpreter return address for the stub to use as its return |
duke@435 | 263 | // address so the skeletal frames are perfectly walkable |
duke@435 | 264 | frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); |
duke@435 | 265 | |
duke@435 | 266 | // PopFrame requires that the preserved incoming arguments from the recently-popped topmost |
duke@435 | 267 | // activation be put back on the expression stack of the caller for reexecution |
duke@435 | 268 | if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { |
duke@435 | 269 | popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); |
duke@435 | 270 | } |
duke@435 | 271 | |
duke@435 | 272 | // |
duke@435 | 273 | // frame_sizes/frame_pcs[0] oldest frame (int or c2i) |
duke@435 | 274 | // frame_sizes/frame_pcs[1] next oldest frame (int) |
duke@435 | 275 | // frame_sizes/frame_pcs[n] youngest frame (int) |
duke@435 | 276 | // |
duke@435 | 277 | // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame |
duke@435 | 278 | // owns the space for the return address to it's caller). Confusing ain't it. |
duke@435 | 279 | // |
duke@435 | 280 | // The vframe array can address vframes with indices running from |
duke@435 | 281 | // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. |
duke@435 | 282 | // When we create the skeletal frames we need the oldest frame to be in the zero slot |
duke@435 | 283 | // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. |
duke@435 | 284 | // so things look a little strange in this loop. |
duke@435 | 285 | // |
duke@435 | 286 | for (int index = 0; index < array->frames(); index++ ) { |
duke@435 | 287 | // frame[number_of_frames - 1 ] = on_stack_size(youngest) |
duke@435 | 288 | // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) |
duke@435 | 289 | // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) |
duke@435 | 290 | frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, |
duke@435 | 291 | callee_locals, |
duke@435 | 292 | index == 0, |
duke@435 | 293 | popframe_extra_args); |
duke@435 | 294 | // This pc doesn't have to be perfect just good enough to identify the frame |
duke@435 | 295 | // as interpreted so the skeleton frame will be walkable |
duke@435 | 296 | // The correct pc will be set when the skeleton frame is completely filled out |
duke@435 | 297 | // The final pc we store in the loop is wrong and will be overwritten below |
duke@435 | 298 | frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; |
duke@435 | 299 | |
duke@435 | 300 | callee_parameters = array->element(index)->method()->size_of_parameters(); |
duke@435 | 301 | callee_locals = array->element(index)->method()->max_locals(); |
duke@435 | 302 | popframe_extra_args = 0; |
duke@435 | 303 | } |
duke@435 | 304 | |
duke@435 | 305 | // Compute whether the root vframe returns a float or double value. |
duke@435 | 306 | BasicType return_type; |
duke@435 | 307 | { |
duke@435 | 308 | HandleMark hm; |
duke@435 | 309 | methodHandle method(thread, array->element(0)->method()); |
duke@435 | 310 | Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci()); |
duke@435 | 311 | return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL; |
duke@435 | 312 | } |
duke@435 | 313 | |
duke@435 | 314 | // Compute information for handling adapters and adjusting the frame size of the caller. |
duke@435 | 315 | int caller_adjustment = 0; |
duke@435 | 316 | |
duke@435 | 317 | // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized |
duke@435 | 318 | // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather |
duke@435 | 319 | // than simply use array->sender.pc(). This requires us to walk the current set of frames |
duke@435 | 320 | // |
duke@435 | 321 | frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame |
duke@435 | 322 | deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller |
duke@435 | 323 | |
duke@435 | 324 | // Compute the amount the oldest interpreter frame will have to adjust |
duke@435 | 325 | // its caller's stack by. If the caller is a compiled frame then |
duke@435 | 326 | // we pretend that the callee has no parameters so that the |
duke@435 | 327 | // extension counts for the full amount of locals and not just |
duke@435 | 328 | // locals-parms. This is because without a c2i adapter the parm |
duke@435 | 329 | // area as created by the compiled frame will not be usable by |
duke@435 | 330 | // the interpreter. (Depending on the calling convention there |
duke@435 | 331 | // may not even be enough space). |
duke@435 | 332 | |
duke@435 | 333 | // QQQ I'd rather see this pushed down into last_frame_adjust |
duke@435 | 334 | // and have it take the sender (aka caller). |
duke@435 | 335 | |
duke@435 | 336 | if (deopt_sender.is_compiled_frame()) { |
duke@435 | 337 | caller_adjustment = last_frame_adjust(0, callee_locals); |
duke@435 | 338 | } else if (callee_locals > callee_parameters) { |
duke@435 | 339 | // The caller frame may need extending to accommodate |
duke@435 | 340 | // non-parameter locals of the first unpacked interpreted frame. |
duke@435 | 341 | // Compute that adjustment. |
duke@435 | 342 | caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); |
duke@435 | 343 | } |
duke@435 | 344 | |
duke@435 | 345 | |
duke@435 | 346 | // If the sender is deoptimized the we must retrieve the address of the handler |
duke@435 | 347 | // since the frame will "magically" show the original pc before the deopt |
duke@435 | 348 | // and we'd undo the deopt. |
duke@435 | 349 | |
duke@435 | 350 | frame_pcs[0] = deopt_sender.raw_pc(); |
duke@435 | 351 | |
duke@435 | 352 | assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); |
duke@435 | 353 | |
duke@435 | 354 | UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, |
duke@435 | 355 | caller_adjustment * BytesPerWord, |
duke@435 | 356 | number_of_frames, |
duke@435 | 357 | frame_sizes, |
duke@435 | 358 | frame_pcs, |
duke@435 | 359 | return_type); |
duke@435 | 360 | #if defined(IA32) || defined(AMD64) |
duke@435 | 361 | // We need a way to pass fp to the unpacking code so the skeletal frames |
duke@435 | 362 | // come out correct. This is only needed for x86 because of c2 using ebp |
duke@435 | 363 | // as an allocatable register. So this update is useless (and harmless) |
duke@435 | 364 | // on the other platforms. It would be nice to do this in a different |
duke@435 | 365 | // way but even the old style deoptimization had a problem with deriving |
duke@435 | 366 | // this value. NEEDS_CLEANUP |
duke@435 | 367 | // Note: now that c1 is using c2's deopt blob we must do this on all |
duke@435 | 368 | // x86 based platforms |
duke@435 | 369 | intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); |
duke@435 | 370 | *fp_addr = array->sender().fp(); // was adapter_caller |
duke@435 | 371 | #endif /* IA32 || AMD64 */ |
duke@435 | 372 | |
duke@435 | 373 | if (array->frames() > 1) { |
duke@435 | 374 | if (VerifyStack && TraceDeoptimization) { |
duke@435 | 375 | tty->print_cr("Deoptimizing method containing inlining"); |
duke@435 | 376 | } |
duke@435 | 377 | } |
duke@435 | 378 | |
duke@435 | 379 | array->set_unroll_block(info); |
duke@435 | 380 | return info; |
duke@435 | 381 | } |
duke@435 | 382 | |
duke@435 | 383 | // Called to cleanup deoptimization data structures in normal case |
duke@435 | 384 | // after unpacking to stack and when stack overflow error occurs |
duke@435 | 385 | void Deoptimization::cleanup_deopt_info(JavaThread *thread, |
duke@435 | 386 | vframeArray *array) { |
duke@435 | 387 | |
duke@435 | 388 | // Get array if coming from exception |
duke@435 | 389 | if (array == NULL) { |
duke@435 | 390 | array = thread->vframe_array_head(); |
duke@435 | 391 | } |
duke@435 | 392 | thread->set_vframe_array_head(NULL); |
duke@435 | 393 | |
duke@435 | 394 | // Free the previous UnrollBlock |
duke@435 | 395 | vframeArray* old_array = thread->vframe_array_last(); |
duke@435 | 396 | thread->set_vframe_array_last(array); |
duke@435 | 397 | |
duke@435 | 398 | if (old_array != NULL) { |
duke@435 | 399 | UnrollBlock* old_info = old_array->unroll_block(); |
duke@435 | 400 | old_array->set_unroll_block(NULL); |
duke@435 | 401 | delete old_info; |
duke@435 | 402 | delete old_array; |
duke@435 | 403 | } |
duke@435 | 404 | |
duke@435 | 405 | // Deallocate any resource creating in this routine and any ResourceObjs allocated |
duke@435 | 406 | // inside the vframeArray (StackValueCollections) |
duke@435 | 407 | |
duke@435 | 408 | delete thread->deopt_mark(); |
duke@435 | 409 | thread->set_deopt_mark(NULL); |
duke@435 | 410 | |
duke@435 | 411 | |
duke@435 | 412 | if (JvmtiExport::can_pop_frame()) { |
duke@435 | 413 | #ifndef CC_INTERP |
duke@435 | 414 | // Regardless of whether we entered this routine with the pending |
duke@435 | 415 | // popframe condition bit set, we should always clear it now |
duke@435 | 416 | thread->clear_popframe_condition(); |
duke@435 | 417 | #else |
duke@435 | 418 | // C++ interpeter will clear has_pending_popframe when it enters |
duke@435 | 419 | // with method_resume. For deopt_resume2 we clear it now. |
duke@435 | 420 | if (thread->popframe_forcing_deopt_reexecution()) |
duke@435 | 421 | thread->clear_popframe_condition(); |
duke@435 | 422 | #endif /* CC_INTERP */ |
duke@435 | 423 | } |
duke@435 | 424 | |
duke@435 | 425 | // unpack_frames() is called at the end of the deoptimization handler |
duke@435 | 426 | // and (in C2) at the end of the uncommon trap handler. Note this fact |
duke@435 | 427 | // so that an asynchronous stack walker can work again. This counter is |
duke@435 | 428 | // incremented at the beginning of fetch_unroll_info() and (in C2) at |
duke@435 | 429 | // the beginning of uncommon_trap(). |
duke@435 | 430 | thread->dec_in_deopt_handler(); |
duke@435 | 431 | } |
duke@435 | 432 | |
duke@435 | 433 | |
duke@435 | 434 | // Return BasicType of value being returned |
duke@435 | 435 | JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) |
duke@435 | 436 | |
duke@435 | 437 | // We are already active int he special DeoptResourceMark any ResourceObj's we |
duke@435 | 438 | // allocate will be freed at the end of the routine. |
duke@435 | 439 | |
duke@435 | 440 | // It is actually ok to allocate handles in a leaf method. It causes no safepoints, |
duke@435 | 441 | // but makes the entry a little slower. There is however a little dance we have to |
duke@435 | 442 | // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro |
duke@435 | 443 | ResetNoHandleMark rnhm; // No-op in release/product versions |
duke@435 | 444 | HandleMark hm; |
duke@435 | 445 | |
duke@435 | 446 | frame stub_frame = thread->last_frame(); |
duke@435 | 447 | |
duke@435 | 448 | // Since the frame to unpack is the top frame of this thread, the vframe_array_head |
duke@435 | 449 | // must point to the vframeArray for the unpack frame. |
duke@435 | 450 | vframeArray* array = thread->vframe_array_head(); |
duke@435 | 451 | |
duke@435 | 452 | #ifndef PRODUCT |
duke@435 | 453 | if (TraceDeoptimization) { |
duke@435 | 454 | tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); |
duke@435 | 455 | } |
duke@435 | 456 | #endif |
duke@435 | 457 | |
duke@435 | 458 | UnrollBlock* info = array->unroll_block(); |
duke@435 | 459 | |
duke@435 | 460 | // Unpack the interpreter frames and any adapter frame (c2 only) we might create. |
duke@435 | 461 | array->unpack_to_stack(stub_frame, exec_mode); |
duke@435 | 462 | |
duke@435 | 463 | BasicType bt = info->return_type(); |
duke@435 | 464 | |
duke@435 | 465 | // If we have an exception pending, claim that the return type is an oop |
duke@435 | 466 | // so the deopt_blob does not overwrite the exception_oop. |
duke@435 | 467 | |
duke@435 | 468 | if (exec_mode == Unpack_exception) |
duke@435 | 469 | bt = T_OBJECT; |
duke@435 | 470 | |
duke@435 | 471 | // Cleanup thread deopt data |
duke@435 | 472 | cleanup_deopt_info(thread, array); |
duke@435 | 473 | |
duke@435 | 474 | #ifndef PRODUCT |
duke@435 | 475 | if (VerifyStack) { |
duke@435 | 476 | ResourceMark res_mark; |
duke@435 | 477 | |
duke@435 | 478 | // Verify that the just-unpacked frames match the interpreter's |
duke@435 | 479 | // notions of expression stack and locals |
duke@435 | 480 | vframeArray* cur_array = thread->vframe_array_last(); |
duke@435 | 481 | RegisterMap rm(thread, false); |
duke@435 | 482 | rm.set_include_argument_oops(false); |
duke@435 | 483 | bool is_top_frame = true; |
duke@435 | 484 | int callee_size_of_parameters = 0; |
duke@435 | 485 | int callee_max_locals = 0; |
duke@435 | 486 | for (int i = 0; i < cur_array->frames(); i++) { |
duke@435 | 487 | vframeArrayElement* el = cur_array->element(i); |
duke@435 | 488 | frame* iframe = el->iframe(); |
duke@435 | 489 | guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); |
duke@435 | 490 | |
duke@435 | 491 | // Get the oop map for this bci |
duke@435 | 492 | InterpreterOopMap mask; |
duke@435 | 493 | int cur_invoke_parameter_size = 0; |
duke@435 | 494 | bool try_next_mask = false; |
duke@435 | 495 | int next_mask_expression_stack_size = -1; |
duke@435 | 496 | int top_frame_expression_stack_adjustment = 0; |
duke@435 | 497 | methodHandle mh(thread, iframe->interpreter_frame_method()); |
duke@435 | 498 | OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); |
duke@435 | 499 | BytecodeStream str(mh); |
duke@435 | 500 | str.set_start(iframe->interpreter_frame_bci()); |
duke@435 | 501 | int max_bci = mh->code_size(); |
duke@435 | 502 | // Get to the next bytecode if possible |
duke@435 | 503 | assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); |
duke@435 | 504 | // Check to see if we can grab the number of outgoing arguments |
duke@435 | 505 | // at an uncommon trap for an invoke (where the compiler |
duke@435 | 506 | // generates debug info before the invoke has executed) |
duke@435 | 507 | Bytecodes::Code cur_code = str.next(); |
duke@435 | 508 | if (cur_code == Bytecodes::_invokevirtual || |
duke@435 | 509 | cur_code == Bytecodes::_invokespecial || |
duke@435 | 510 | cur_code == Bytecodes::_invokestatic || |
duke@435 | 511 | cur_code == Bytecodes::_invokeinterface) { |
duke@435 | 512 | Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci()); |
duke@435 | 513 | symbolHandle signature(thread, invoke->signature()); |
duke@435 | 514 | ArgumentSizeComputer asc(signature); |
duke@435 | 515 | cur_invoke_parameter_size = asc.size(); |
duke@435 | 516 | if (cur_code != Bytecodes::_invokestatic) { |
duke@435 | 517 | // Add in receiver |
duke@435 | 518 | ++cur_invoke_parameter_size; |
duke@435 | 519 | } |
duke@435 | 520 | } |
duke@435 | 521 | if (str.bci() < max_bci) { |
duke@435 | 522 | Bytecodes::Code bc = str.next(); |
duke@435 | 523 | if (bc >= 0) { |
duke@435 | 524 | // The interpreter oop map generator reports results before |
duke@435 | 525 | // the current bytecode has executed except in the case of |
duke@435 | 526 | // calls. It seems to be hard to tell whether the compiler |
duke@435 | 527 | // has emitted debug information matching the "state before" |
duke@435 | 528 | // a given bytecode or the state after, so we try both |
duke@435 | 529 | switch (cur_code) { |
duke@435 | 530 | case Bytecodes::_invokevirtual: |
duke@435 | 531 | case Bytecodes::_invokespecial: |
duke@435 | 532 | case Bytecodes::_invokestatic: |
duke@435 | 533 | case Bytecodes::_invokeinterface: |
duke@435 | 534 | case Bytecodes::_athrow: |
duke@435 | 535 | break; |
duke@435 | 536 | default: { |
duke@435 | 537 | InterpreterOopMap next_mask; |
duke@435 | 538 | OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); |
duke@435 | 539 | next_mask_expression_stack_size = next_mask.expression_stack_size(); |
duke@435 | 540 | // Need to subtract off the size of the result type of |
duke@435 | 541 | // the bytecode because this is not described in the |
duke@435 | 542 | // debug info but returned to the interpreter in the TOS |
duke@435 | 543 | // caching register |
duke@435 | 544 | BasicType bytecode_result_type = Bytecodes::result_type(cur_code); |
duke@435 | 545 | if (bytecode_result_type != T_ILLEGAL) { |
duke@435 | 546 | top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; |
duke@435 | 547 | } |
duke@435 | 548 | assert(top_frame_expression_stack_adjustment >= 0, ""); |
duke@435 | 549 | try_next_mask = true; |
duke@435 | 550 | break; |
duke@435 | 551 | } |
duke@435 | 552 | } |
duke@435 | 553 | } |
duke@435 | 554 | } |
duke@435 | 555 | |
duke@435 | 556 | // Verify stack depth and oops in frame |
duke@435 | 557 | // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) |
duke@435 | 558 | if (!( |
duke@435 | 559 | /* SPARC */ |
duke@435 | 560 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || |
duke@435 | 561 | /* x86 */ |
duke@435 | 562 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || |
duke@435 | 563 | (try_next_mask && |
duke@435 | 564 | (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - |
duke@435 | 565 | top_frame_expression_stack_adjustment))) || |
duke@435 | 566 | (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || |
duke@435 | 567 | (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && |
duke@435 | 568 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) |
duke@435 | 569 | )) { |
duke@435 | 570 | ttyLocker ttyl; |
duke@435 | 571 | |
duke@435 | 572 | // Print out some information that will help us debug the problem |
duke@435 | 573 | tty->print_cr("Wrong number of expression stack elements during deoptimization"); |
duke@435 | 574 | tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); |
duke@435 | 575 | tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", |
duke@435 | 576 | iframe->interpreter_frame_expression_stack_size()); |
duke@435 | 577 | tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); |
duke@435 | 578 | tty->print_cr(" try_next_mask = %d", try_next_mask); |
duke@435 | 579 | tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); |
duke@435 | 580 | tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); |
duke@435 | 581 | tty->print_cr(" callee_max_locals = %d", callee_max_locals); |
duke@435 | 582 | tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); |
duke@435 | 583 | tty->print_cr(" exec_mode = %d", exec_mode); |
duke@435 | 584 | tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); |
duke@435 | 585 | tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); |
duke@435 | 586 | tty->print_cr(" Interpreted frames:"); |
duke@435 | 587 | for (int k = 0; k < cur_array->frames(); k++) { |
duke@435 | 588 | vframeArrayElement* el = cur_array->element(k); |
duke@435 | 589 | tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); |
duke@435 | 590 | } |
duke@435 | 591 | cur_array->print_on_2(tty); |
duke@435 | 592 | guarantee(false, "wrong number of expression stack elements during deopt"); |
duke@435 | 593 | } |
duke@435 | 594 | VerifyOopClosure verify; |
duke@435 | 595 | iframe->oops_interpreted_do(&verify, &rm, false); |
duke@435 | 596 | callee_size_of_parameters = mh->size_of_parameters(); |
duke@435 | 597 | callee_max_locals = mh->max_locals(); |
duke@435 | 598 | is_top_frame = false; |
duke@435 | 599 | } |
duke@435 | 600 | } |
duke@435 | 601 | #endif /* !PRODUCT */ |
duke@435 | 602 | |
duke@435 | 603 | |
duke@435 | 604 | return bt; |
duke@435 | 605 | JRT_END |
duke@435 | 606 | |
duke@435 | 607 | |
duke@435 | 608 | int Deoptimization::deoptimize_dependents() { |
duke@435 | 609 | Threads::deoptimized_wrt_marked_nmethods(); |
duke@435 | 610 | return 0; |
duke@435 | 611 | } |
duke@435 | 612 | |
duke@435 | 613 | |
duke@435 | 614 | #ifdef COMPILER2 |
duke@435 | 615 | bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { |
duke@435 | 616 | Handle pending_exception(thread->pending_exception()); |
duke@435 | 617 | const char* exception_file = thread->exception_file(); |
duke@435 | 618 | int exception_line = thread->exception_line(); |
duke@435 | 619 | thread->clear_pending_exception(); |
duke@435 | 620 | |
duke@435 | 621 | for (int i = 0; i < objects->length(); i++) { |
duke@435 | 622 | assert(objects->at(i)->is_object(), "invalid debug information"); |
duke@435 | 623 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
duke@435 | 624 | |
duke@435 | 625 | KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); |
duke@435 | 626 | oop obj = NULL; |
duke@435 | 627 | |
duke@435 | 628 | if (k->oop_is_instance()) { |
duke@435 | 629 | instanceKlass* ik = instanceKlass::cast(k()); |
duke@435 | 630 | obj = ik->allocate_instance(CHECK_(false)); |
duke@435 | 631 | } else if (k->oop_is_typeArray()) { |
duke@435 | 632 | typeArrayKlass* ak = typeArrayKlass::cast(k()); |
duke@435 | 633 | assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); |
duke@435 | 634 | int len = sv->field_size() / type2size[ak->element_type()]; |
duke@435 | 635 | obj = ak->allocate(len, CHECK_(false)); |
duke@435 | 636 | } else if (k->oop_is_objArray()) { |
duke@435 | 637 | objArrayKlass* ak = objArrayKlass::cast(k()); |
duke@435 | 638 | obj = ak->allocate(sv->field_size(), CHECK_(false)); |
duke@435 | 639 | } |
duke@435 | 640 | |
duke@435 | 641 | assert(obj != NULL, "allocation failed"); |
duke@435 | 642 | assert(sv->value().is_null(), "redundant reallocation"); |
duke@435 | 643 | sv->set_value(obj); |
duke@435 | 644 | } |
duke@435 | 645 | |
duke@435 | 646 | if (pending_exception.not_null()) { |
duke@435 | 647 | thread->set_pending_exception(pending_exception(), exception_file, exception_line); |
duke@435 | 648 | } |
duke@435 | 649 | |
duke@435 | 650 | return true; |
duke@435 | 651 | } |
duke@435 | 652 | |
duke@435 | 653 | // This assumes that the fields are stored in ObjectValue in the same order |
duke@435 | 654 | // they are yielded by do_nonstatic_fields. |
duke@435 | 655 | class FieldReassigner: public FieldClosure { |
duke@435 | 656 | frame* _fr; |
duke@435 | 657 | RegisterMap* _reg_map; |
duke@435 | 658 | ObjectValue* _sv; |
duke@435 | 659 | instanceKlass* _ik; |
duke@435 | 660 | oop _obj; |
duke@435 | 661 | |
duke@435 | 662 | int _i; |
duke@435 | 663 | public: |
duke@435 | 664 | FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : |
duke@435 | 665 | _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} |
duke@435 | 666 | |
duke@435 | 667 | int i() const { return _i; } |
duke@435 | 668 | |
duke@435 | 669 | |
duke@435 | 670 | void do_field(fieldDescriptor* fd) { |
kvn@479 | 671 | intptr_t val; |
duke@435 | 672 | StackValue* value = |
duke@435 | 673 | StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); |
duke@435 | 674 | int offset = fd->offset(); |
duke@435 | 675 | switch (fd->field_type()) { |
duke@435 | 676 | case T_OBJECT: case T_ARRAY: |
duke@435 | 677 | assert(value->type() == T_OBJECT, "Agreement."); |
duke@435 | 678 | _obj->obj_field_put(offset, value->get_obj()()); |
duke@435 | 679 | break; |
duke@435 | 680 | |
duke@435 | 681 | case T_LONG: case T_DOUBLE: { |
duke@435 | 682 | assert(value->type() == T_INT, "Agreement."); |
duke@435 | 683 | StackValue* low = |
duke@435 | 684 | StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); |
kvn@479 | 685 | #ifdef _LP64 |
kvn@479 | 686 | jlong res = (jlong)low->get_int(); |
kvn@479 | 687 | #else |
kvn@479 | 688 | #ifdef SPARC |
kvn@479 | 689 | // For SPARC we have to swap high and low words. |
kvn@479 | 690 | jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
kvn@479 | 691 | #else |
duke@435 | 692 | jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
kvn@479 | 693 | #endif //SPARC |
kvn@479 | 694 | #endif |
duke@435 | 695 | _obj->long_field_put(offset, res); |
duke@435 | 696 | break; |
duke@435 | 697 | } |
kvn@479 | 698 | // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
duke@435 | 699 | case T_INT: case T_FLOAT: // 4 bytes. |
duke@435 | 700 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 701 | val = value->get_int(); |
kvn@479 | 702 | _obj->int_field_put(offset, (jint)*((jint*)&val)); |
duke@435 | 703 | break; |
duke@435 | 704 | |
duke@435 | 705 | case T_SHORT: case T_CHAR: // 2 bytes |
duke@435 | 706 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 707 | val = value->get_int(); |
kvn@479 | 708 | _obj->short_field_put(offset, (jshort)*((jint*)&val)); |
duke@435 | 709 | break; |
duke@435 | 710 | |
kvn@479 | 711 | case T_BOOLEAN: case T_BYTE: // 1 byte |
duke@435 | 712 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 713 | val = value->get_int(); |
kvn@479 | 714 | _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); |
duke@435 | 715 | break; |
duke@435 | 716 | |
duke@435 | 717 | default: |
duke@435 | 718 | ShouldNotReachHere(); |
duke@435 | 719 | } |
duke@435 | 720 | _i++; |
duke@435 | 721 | } |
duke@435 | 722 | }; |
duke@435 | 723 | |
duke@435 | 724 | // restore elements of an eliminated type array |
duke@435 | 725 | void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { |
duke@435 | 726 | int index = 0; |
kvn@479 | 727 | intptr_t val; |
duke@435 | 728 | |
duke@435 | 729 | for (int i = 0; i < sv->field_size(); i++) { |
duke@435 | 730 | StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); |
duke@435 | 731 | switch(type) { |
kvn@479 | 732 | case T_LONG: case T_DOUBLE: { |
kvn@479 | 733 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 734 | StackValue* low = |
kvn@479 | 735 | StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); |
kvn@479 | 736 | #ifdef _LP64 |
kvn@479 | 737 | jlong res = (jlong)low->get_int(); |
kvn@479 | 738 | #else |
kvn@479 | 739 | #ifdef SPARC |
kvn@479 | 740 | // For SPARC we have to swap high and low words. |
kvn@479 | 741 | jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
kvn@479 | 742 | #else |
kvn@479 | 743 | jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
kvn@479 | 744 | #endif //SPARC |
kvn@479 | 745 | #endif |
kvn@479 | 746 | obj->long_at_put(index, res); |
kvn@479 | 747 | break; |
kvn@479 | 748 | } |
kvn@479 | 749 | |
kvn@479 | 750 | // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
kvn@479 | 751 | case T_INT: case T_FLOAT: // 4 bytes. |
kvn@479 | 752 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 753 | val = value->get_int(); |
kvn@479 | 754 | obj->int_at_put(index, (jint)*((jint*)&val)); |
kvn@479 | 755 | break; |
kvn@479 | 756 | |
kvn@479 | 757 | case T_SHORT: case T_CHAR: // 2 bytes |
kvn@479 | 758 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 759 | val = value->get_int(); |
kvn@479 | 760 | obj->short_at_put(index, (jshort)*((jint*)&val)); |
kvn@479 | 761 | break; |
kvn@479 | 762 | |
kvn@479 | 763 | case T_BOOLEAN: case T_BYTE: // 1 byte |
kvn@479 | 764 | assert(value->type() == T_INT, "Agreement."); |
kvn@479 | 765 | val = value->get_int(); |
kvn@479 | 766 | obj->bool_at_put(index, (jboolean)*((jint*)&val)); |
kvn@479 | 767 | break; |
kvn@479 | 768 | |
duke@435 | 769 | default: |
duke@435 | 770 | ShouldNotReachHere(); |
duke@435 | 771 | } |
duke@435 | 772 | index++; |
duke@435 | 773 | } |
duke@435 | 774 | } |
duke@435 | 775 | |
duke@435 | 776 | |
duke@435 | 777 | // restore fields of an eliminated object array |
duke@435 | 778 | void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { |
duke@435 | 779 | for (int i = 0; i < sv->field_size(); i++) { |
duke@435 | 780 | StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); |
duke@435 | 781 | assert(value->type() == T_OBJECT, "object element expected"); |
duke@435 | 782 | obj->obj_at_put(i, value->get_obj()()); |
duke@435 | 783 | } |
duke@435 | 784 | } |
duke@435 | 785 | |
duke@435 | 786 | |
duke@435 | 787 | // restore fields of all eliminated objects and arrays |
duke@435 | 788 | void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { |
duke@435 | 789 | for (int i = 0; i < objects->length(); i++) { |
duke@435 | 790 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
duke@435 | 791 | KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); |
duke@435 | 792 | Handle obj = sv->value(); |
duke@435 | 793 | assert(obj.not_null(), "reallocation was missed"); |
duke@435 | 794 | |
duke@435 | 795 | if (k->oop_is_instance()) { |
duke@435 | 796 | instanceKlass* ik = instanceKlass::cast(k()); |
duke@435 | 797 | FieldReassigner reassign(fr, reg_map, sv, obj()); |
duke@435 | 798 | ik->do_nonstatic_fields(&reassign); |
duke@435 | 799 | } else if (k->oop_is_typeArray()) { |
duke@435 | 800 | typeArrayKlass* ak = typeArrayKlass::cast(k()); |
duke@435 | 801 | reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); |
duke@435 | 802 | } else if (k->oop_is_objArray()) { |
duke@435 | 803 | reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); |
duke@435 | 804 | } |
duke@435 | 805 | } |
duke@435 | 806 | } |
duke@435 | 807 | |
duke@435 | 808 | |
duke@435 | 809 | // relock objects for which synchronization was eliminated |
kvn@518 | 810 | void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { |
duke@435 | 811 | for (int i = 0; i < monitors->length(); i++) { |
kvn@518 | 812 | MonitorInfo* mon_info = monitors->at(i); |
kvn@518 | 813 | if (mon_info->eliminated()) { |
kvn@518 | 814 | assert(mon_info->owner() != NULL, "reallocation was missed"); |
kvn@518 | 815 | Handle obj = Handle(mon_info->owner()); |
kvn@518 | 816 | markOop mark = obj->mark(); |
kvn@518 | 817 | if (UseBiasedLocking && mark->has_bias_pattern()) { |
kvn@518 | 818 | // New allocated objects may have the mark set to anonymously biased. |
kvn@518 | 819 | // Also the deoptimized method may called methods with synchronization |
kvn@518 | 820 | // where the thread-local object is bias locked to the current thread. |
kvn@518 | 821 | assert(mark->is_biased_anonymously() || |
kvn@518 | 822 | mark->biased_locker() == thread, "should be locked to current thread"); |
kvn@518 | 823 | // Reset mark word to unbiased prototype. |
kvn@518 | 824 | markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); |
kvn@518 | 825 | obj->set_mark(unbiased_prototype); |
kvn@518 | 826 | } |
kvn@518 | 827 | BasicLock* lock = mon_info->lock(); |
kvn@518 | 828 | ObjectSynchronizer::slow_enter(obj, lock, thread); |
duke@435 | 829 | } |
kvn@518 | 830 | assert(mon_info->owner()->is_locked(), "object must be locked now"); |
duke@435 | 831 | } |
duke@435 | 832 | } |
duke@435 | 833 | |
duke@435 | 834 | |
duke@435 | 835 | #ifndef PRODUCT |
duke@435 | 836 | // print information about reallocated objects |
duke@435 | 837 | void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { |
duke@435 | 838 | fieldDescriptor fd; |
duke@435 | 839 | |
duke@435 | 840 | for (int i = 0; i < objects->length(); i++) { |
duke@435 | 841 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
duke@435 | 842 | KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); |
duke@435 | 843 | Handle obj = sv->value(); |
duke@435 | 844 | |
duke@435 | 845 | tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); |
duke@435 | 846 | k->as_klassOop()->print_value(); |
duke@435 | 847 | tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); |
duke@435 | 848 | tty->cr(); |
duke@435 | 849 | |
duke@435 | 850 | if (Verbose) { |
duke@435 | 851 | k->oop_print_on(obj(), tty); |
duke@435 | 852 | } |
duke@435 | 853 | } |
duke@435 | 854 | } |
duke@435 | 855 | #endif |
duke@435 | 856 | #endif // COMPILER2 |
duke@435 | 857 | |
duke@435 | 858 | vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { |
duke@435 | 859 | |
duke@435 | 860 | #ifndef PRODUCT |
duke@435 | 861 | if (TraceDeoptimization) { |
duke@435 | 862 | ttyLocker ttyl; |
duke@435 | 863 | tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); |
duke@435 | 864 | fr.print_on(tty); |
duke@435 | 865 | tty->print_cr(" Virtual frames (innermost first):"); |
duke@435 | 866 | for (int index = 0; index < chunk->length(); index++) { |
duke@435 | 867 | compiledVFrame* vf = chunk->at(index); |
duke@435 | 868 | tty->print(" %2d - ", index); |
duke@435 | 869 | vf->print_value(); |
duke@435 | 870 | int bci = chunk->at(index)->raw_bci(); |
duke@435 | 871 | const char* code_name; |
duke@435 | 872 | if (bci == SynchronizationEntryBCI) { |
duke@435 | 873 | code_name = "sync entry"; |
duke@435 | 874 | } else { |
duke@435 | 875 | Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci); |
duke@435 | 876 | code_name = Bytecodes::name(code); |
duke@435 | 877 | } |
duke@435 | 878 | tty->print(" - %s", code_name); |
duke@435 | 879 | tty->print_cr(" @ bci %d ", bci); |
duke@435 | 880 | if (Verbose) { |
duke@435 | 881 | vf->print(); |
duke@435 | 882 | tty->cr(); |
duke@435 | 883 | } |
duke@435 | 884 | } |
duke@435 | 885 | } |
duke@435 | 886 | #endif |
duke@435 | 887 | |
duke@435 | 888 | // Register map for next frame (used for stack crawl). We capture |
duke@435 | 889 | // the state of the deopt'ing frame's caller. Thus if we need to |
duke@435 | 890 | // stuff a C2I adapter we can properly fill in the callee-save |
duke@435 | 891 | // register locations. |
duke@435 | 892 | frame caller = fr.sender(reg_map); |
duke@435 | 893 | int frame_size = caller.sp() - fr.sp(); |
duke@435 | 894 | |
duke@435 | 895 | frame sender = caller; |
duke@435 | 896 | |
duke@435 | 897 | // Since the Java thread being deoptimized will eventually adjust it's own stack, |
duke@435 | 898 | // the vframeArray containing the unpacking information is allocated in the C heap. |
duke@435 | 899 | // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). |
duke@435 | 900 | vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); |
duke@435 | 901 | |
duke@435 | 902 | // Compare the vframeArray to the collected vframes |
duke@435 | 903 | assert(array->structural_compare(thread, chunk), "just checking"); |
duke@435 | 904 | Events::log("# vframes = %d", (intptr_t)chunk->length()); |
duke@435 | 905 | |
duke@435 | 906 | #ifndef PRODUCT |
duke@435 | 907 | if (TraceDeoptimization) { |
duke@435 | 908 | ttyLocker ttyl; |
duke@435 | 909 | tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); |
duke@435 | 910 | if (Verbose) { |
duke@435 | 911 | int count = 0; |
duke@435 | 912 | // this used to leak deoptimizedVFrame like it was going out of style!!! |
duke@435 | 913 | for (int index = 0; index < array->frames(); index++ ) { |
duke@435 | 914 | vframeArrayElement* e = array->element(index); |
duke@435 | 915 | e->print(tty); |
duke@435 | 916 | |
duke@435 | 917 | /* |
duke@435 | 918 | No printing yet. |
duke@435 | 919 | array->vframe_at(index)->print_activation(count++); |
duke@435 | 920 | // better as... |
duke@435 | 921 | array->print_activation_for(index, count++); |
duke@435 | 922 | */ |
duke@435 | 923 | } |
duke@435 | 924 | } |
duke@435 | 925 | } |
duke@435 | 926 | #endif // PRODUCT |
duke@435 | 927 | |
duke@435 | 928 | return array; |
duke@435 | 929 | } |
duke@435 | 930 | |
duke@435 | 931 | |
duke@435 | 932 | static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { |
duke@435 | 933 | GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
duke@435 | 934 | for (int i = 0; i < monitors->length(); i++) { |
duke@435 | 935 | MonitorInfo* mon_info = monitors->at(i); |
kvn@518 | 936 | if (mon_info->owner() != NULL && !mon_info->eliminated()) { |
duke@435 | 937 | objects_to_revoke->append(Handle(mon_info->owner())); |
duke@435 | 938 | } |
duke@435 | 939 | } |
duke@435 | 940 | } |
duke@435 | 941 | |
duke@435 | 942 | |
duke@435 | 943 | void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { |
duke@435 | 944 | if (!UseBiasedLocking) { |
duke@435 | 945 | return; |
duke@435 | 946 | } |
duke@435 | 947 | |
duke@435 | 948 | GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); |
duke@435 | 949 | |
duke@435 | 950 | // Unfortunately we don't have a RegisterMap available in most of |
duke@435 | 951 | // the places we want to call this routine so we need to walk the |
duke@435 | 952 | // stack again to update the register map. |
duke@435 | 953 | if (map == NULL || !map->update_map()) { |
duke@435 | 954 | StackFrameStream sfs(thread, true); |
duke@435 | 955 | bool found = false; |
duke@435 | 956 | while (!found && !sfs.is_done()) { |
duke@435 | 957 | frame* cur = sfs.current(); |
duke@435 | 958 | sfs.next(); |
duke@435 | 959 | found = cur->id() == fr.id(); |
duke@435 | 960 | } |
duke@435 | 961 | assert(found, "frame to be deoptimized not found on target thread's stack"); |
duke@435 | 962 | map = sfs.register_map(); |
duke@435 | 963 | } |
duke@435 | 964 | |
duke@435 | 965 | vframe* vf = vframe::new_vframe(&fr, map, thread); |
duke@435 | 966 | compiledVFrame* cvf = compiledVFrame::cast(vf); |
duke@435 | 967 | // Revoke monitors' biases in all scopes |
duke@435 | 968 | while (!cvf->is_top()) { |
duke@435 | 969 | collect_monitors(cvf, objects_to_revoke); |
duke@435 | 970 | cvf = compiledVFrame::cast(cvf->sender()); |
duke@435 | 971 | } |
duke@435 | 972 | collect_monitors(cvf, objects_to_revoke); |
duke@435 | 973 | |
duke@435 | 974 | if (SafepointSynchronize::is_at_safepoint()) { |
duke@435 | 975 | BiasedLocking::revoke_at_safepoint(objects_to_revoke); |
duke@435 | 976 | } else { |
duke@435 | 977 | BiasedLocking::revoke(objects_to_revoke); |
duke@435 | 978 | } |
duke@435 | 979 | } |
duke@435 | 980 | |
duke@435 | 981 | |
duke@435 | 982 | void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { |
duke@435 | 983 | if (!UseBiasedLocking) { |
duke@435 | 984 | return; |
duke@435 | 985 | } |
duke@435 | 986 | |
duke@435 | 987 | assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); |
duke@435 | 988 | GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); |
duke@435 | 989 | for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { |
duke@435 | 990 | if (jt->has_last_Java_frame()) { |
duke@435 | 991 | StackFrameStream sfs(jt, true); |
duke@435 | 992 | while (!sfs.is_done()) { |
duke@435 | 993 | frame* cur = sfs.current(); |
duke@435 | 994 | if (cb->contains(cur->pc())) { |
duke@435 | 995 | vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); |
duke@435 | 996 | compiledVFrame* cvf = compiledVFrame::cast(vf); |
duke@435 | 997 | // Revoke monitors' biases in all scopes |
duke@435 | 998 | while (!cvf->is_top()) { |
duke@435 | 999 | collect_monitors(cvf, objects_to_revoke); |
duke@435 | 1000 | cvf = compiledVFrame::cast(cvf->sender()); |
duke@435 | 1001 | } |
duke@435 | 1002 | collect_monitors(cvf, objects_to_revoke); |
duke@435 | 1003 | } |
duke@435 | 1004 | sfs.next(); |
duke@435 | 1005 | } |
duke@435 | 1006 | } |
duke@435 | 1007 | } |
duke@435 | 1008 | BiasedLocking::revoke_at_safepoint(objects_to_revoke); |
duke@435 | 1009 | } |
duke@435 | 1010 | |
duke@435 | 1011 | |
duke@435 | 1012 | void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { |
duke@435 | 1013 | assert(fr.can_be_deoptimized(), "checking frame type"); |
duke@435 | 1014 | |
duke@435 | 1015 | gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); |
duke@435 | 1016 | |
duke@435 | 1017 | EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); |
duke@435 | 1018 | |
duke@435 | 1019 | // Patch the nmethod so that when execution returns to it we will |
duke@435 | 1020 | // deopt the execution state and return to the interpreter. |
duke@435 | 1021 | fr.deoptimize(thread); |
duke@435 | 1022 | } |
duke@435 | 1023 | |
duke@435 | 1024 | void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { |
duke@435 | 1025 | // Deoptimize only if the frame comes from compile code. |
duke@435 | 1026 | // Do not deoptimize the frame which is already patched |
duke@435 | 1027 | // during the execution of the loops below. |
duke@435 | 1028 | if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { |
duke@435 | 1029 | return; |
duke@435 | 1030 | } |
duke@435 | 1031 | ResourceMark rm; |
duke@435 | 1032 | DeoptimizationMarker dm; |
duke@435 | 1033 | if (UseBiasedLocking) { |
duke@435 | 1034 | revoke_biases_of_monitors(thread, fr, map); |
duke@435 | 1035 | } |
duke@435 | 1036 | deoptimize_single_frame(thread, fr); |
duke@435 | 1037 | |
duke@435 | 1038 | } |
duke@435 | 1039 | |
duke@435 | 1040 | |
duke@435 | 1041 | void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { |
duke@435 | 1042 | // Compute frame and register map based on thread and sp. |
duke@435 | 1043 | RegisterMap reg_map(thread, UseBiasedLocking); |
duke@435 | 1044 | frame fr = thread->last_frame(); |
duke@435 | 1045 | while (fr.id() != id) { |
duke@435 | 1046 | fr = fr.sender(®_map); |
duke@435 | 1047 | } |
duke@435 | 1048 | deoptimize(thread, fr, ®_map); |
duke@435 | 1049 | } |
duke@435 | 1050 | |
duke@435 | 1051 | |
duke@435 | 1052 | // JVMTI PopFrame support |
duke@435 | 1053 | JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) |
duke@435 | 1054 | { |
duke@435 | 1055 | thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); |
duke@435 | 1056 | } |
duke@435 | 1057 | JRT_END |
duke@435 | 1058 | |
duke@435 | 1059 | |
duke@435 | 1060 | #ifdef COMPILER2 |
duke@435 | 1061 | void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { |
duke@435 | 1062 | // in case of an unresolved klass entry, load the class. |
duke@435 | 1063 | if (constant_pool->tag_at(index).is_unresolved_klass()) { |
duke@435 | 1064 | klassOop tk = constant_pool->klass_at(index, CHECK); |
duke@435 | 1065 | return; |
duke@435 | 1066 | } |
duke@435 | 1067 | |
duke@435 | 1068 | if (!constant_pool->tag_at(index).is_symbol()) return; |
duke@435 | 1069 | |
duke@435 | 1070 | Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); |
duke@435 | 1071 | symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); |
duke@435 | 1072 | |
duke@435 | 1073 | // class name? |
duke@435 | 1074 | if (symbol->byte_at(0) != '(') { |
duke@435 | 1075 | Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); |
duke@435 | 1076 | SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); |
duke@435 | 1077 | return; |
duke@435 | 1078 | } |
duke@435 | 1079 | |
duke@435 | 1080 | // then it must be a signature! |
duke@435 | 1081 | for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { |
duke@435 | 1082 | if (ss.is_object()) { |
duke@435 | 1083 | symbolOop s = ss.as_symbol(CHECK); |
duke@435 | 1084 | symbolHandle class_name (THREAD, s); |
duke@435 | 1085 | Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); |
duke@435 | 1086 | SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); |
duke@435 | 1087 | } |
duke@435 | 1088 | } |
duke@435 | 1089 | } |
duke@435 | 1090 | |
duke@435 | 1091 | |
duke@435 | 1092 | void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { |
duke@435 | 1093 | EXCEPTION_MARK; |
duke@435 | 1094 | load_class_by_index(constant_pool, index, THREAD); |
duke@435 | 1095 | if (HAS_PENDING_EXCEPTION) { |
duke@435 | 1096 | // Exception happened during classloading. We ignore the exception here, since it |
duke@435 | 1097 | // is going to be rethrown since the current activation is going to be deoptimzied and |
duke@435 | 1098 | // the interpreter will re-execute the bytecode. |
duke@435 | 1099 | CLEAR_PENDING_EXCEPTION; |
duke@435 | 1100 | } |
duke@435 | 1101 | } |
duke@435 | 1102 | |
duke@435 | 1103 | JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { |
duke@435 | 1104 | HandleMark hm; |
duke@435 | 1105 | |
duke@435 | 1106 | // uncommon_trap() is called at the beginning of the uncommon trap |
duke@435 | 1107 | // handler. Note this fact before we start generating temporary frames |
duke@435 | 1108 | // that can confuse an asynchronous stack walker. This counter is |
duke@435 | 1109 | // decremented at the end of unpack_frames(). |
duke@435 | 1110 | thread->inc_in_deopt_handler(); |
duke@435 | 1111 | |
duke@435 | 1112 | // We need to update the map if we have biased locking. |
duke@435 | 1113 | RegisterMap reg_map(thread, UseBiasedLocking); |
duke@435 | 1114 | frame stub_frame = thread->last_frame(); |
duke@435 | 1115 | frame fr = stub_frame.sender(®_map); |
duke@435 | 1116 | // Make sure the calling nmethod is not getting deoptimized and removed |
duke@435 | 1117 | // before we are done with it. |
duke@435 | 1118 | nmethodLocker nl(fr.pc()); |
duke@435 | 1119 | |
duke@435 | 1120 | { |
duke@435 | 1121 | ResourceMark rm; |
duke@435 | 1122 | |
duke@435 | 1123 | // Revoke biases of any monitors in the frame to ensure we can migrate them |
duke@435 | 1124 | revoke_biases_of_monitors(thread, fr, ®_map); |
duke@435 | 1125 | |
duke@435 | 1126 | DeoptReason reason = trap_request_reason(trap_request); |
duke@435 | 1127 | DeoptAction action = trap_request_action(trap_request); |
duke@435 | 1128 | jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 |
duke@435 | 1129 | |
duke@435 | 1130 | Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); |
duke@435 | 1131 | vframe* vf = vframe::new_vframe(&fr, ®_map, thread); |
duke@435 | 1132 | compiledVFrame* cvf = compiledVFrame::cast(vf); |
duke@435 | 1133 | |
duke@435 | 1134 | nmethod* nm = cvf->code(); |
duke@435 | 1135 | |
duke@435 | 1136 | ScopeDesc* trap_scope = cvf->scope(); |
duke@435 | 1137 | methodHandle trap_method = trap_scope->method(); |
duke@435 | 1138 | int trap_bci = trap_scope->bci(); |
duke@435 | 1139 | Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code(); |
duke@435 | 1140 | |
duke@435 | 1141 | // Record this event in the histogram. |
duke@435 | 1142 | gather_statistics(reason, action, trap_bc); |
duke@435 | 1143 | |
duke@435 | 1144 | // Ensure that we can record deopt. history: |
duke@435 | 1145 | bool create_if_missing = ProfileTraps; |
duke@435 | 1146 | |
duke@435 | 1147 | methodDataHandle trap_mdo |
duke@435 | 1148 | (THREAD, get_method_data(thread, trap_method, create_if_missing)); |
duke@435 | 1149 | |
duke@435 | 1150 | // Print a bunch of diagnostics, if requested. |
duke@435 | 1151 | if (TraceDeoptimization || LogCompilation) { |
duke@435 | 1152 | ResourceMark rm; |
duke@435 | 1153 | ttyLocker ttyl; |
duke@435 | 1154 | char buf[100]; |
duke@435 | 1155 | if (xtty != NULL) { |
duke@435 | 1156 | xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", |
duke@435 | 1157 | os::current_thread_id(), |
duke@435 | 1158 | format_trap_request(buf, sizeof(buf), trap_request)); |
duke@435 | 1159 | nm->log_identity(xtty); |
duke@435 | 1160 | } |
duke@435 | 1161 | symbolHandle class_name; |
duke@435 | 1162 | bool unresolved = false; |
duke@435 | 1163 | if (unloaded_class_index >= 0) { |
duke@435 | 1164 | constantPoolHandle constants (THREAD, trap_method->constants()); |
duke@435 | 1165 | if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { |
duke@435 | 1166 | class_name = symbolHandle(THREAD, |
duke@435 | 1167 | constants->klass_name_at(unloaded_class_index)); |
duke@435 | 1168 | unresolved = true; |
duke@435 | 1169 | if (xtty != NULL) |
duke@435 | 1170 | xtty->print(" unresolved='1'"); |
duke@435 | 1171 | } else if (constants->tag_at(unloaded_class_index).is_symbol()) { |
duke@435 | 1172 | class_name = symbolHandle(THREAD, |
duke@435 | 1173 | constants->symbol_at(unloaded_class_index)); |
duke@435 | 1174 | } |
duke@435 | 1175 | if (xtty != NULL) |
duke@435 | 1176 | xtty->name(class_name); |
duke@435 | 1177 | } |
duke@435 | 1178 | if (xtty != NULL && trap_mdo.not_null()) { |
duke@435 | 1179 | // Dump the relevant MDO state. |
duke@435 | 1180 | // This is the deopt count for the current reason, any previous |
duke@435 | 1181 | // reasons or recompiles seen at this point. |
duke@435 | 1182 | int dcnt = trap_mdo->trap_count(reason); |
duke@435 | 1183 | if (dcnt != 0) |
duke@435 | 1184 | xtty->print(" count='%d'", dcnt); |
duke@435 | 1185 | ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); |
duke@435 | 1186 | int dos = (pdata == NULL)? 0: pdata->trap_state(); |
duke@435 | 1187 | if (dos != 0) { |
duke@435 | 1188 | xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); |
duke@435 | 1189 | if (trap_state_is_recompiled(dos)) { |
duke@435 | 1190 | int recnt2 = trap_mdo->overflow_recompile_count(); |
duke@435 | 1191 | if (recnt2 != 0) |
duke@435 | 1192 | xtty->print(" recompiles2='%d'", recnt2); |
duke@435 | 1193 | } |
duke@435 | 1194 | } |
duke@435 | 1195 | } |
duke@435 | 1196 | if (xtty != NULL) { |
duke@435 | 1197 | xtty->stamp(); |
duke@435 | 1198 | xtty->end_head(); |
duke@435 | 1199 | } |
duke@435 | 1200 | if (TraceDeoptimization) { // make noise on the tty |
duke@435 | 1201 | tty->print("Uncommon trap occurred in"); |
duke@435 | 1202 | nm->method()->print_short_name(tty); |
duke@435 | 1203 | tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", |
duke@435 | 1204 | fr.pc(), |
duke@435 | 1205 | (int) os::current_thread_id(), |
duke@435 | 1206 | trap_reason_name(reason), |
duke@435 | 1207 | trap_action_name(action), |
duke@435 | 1208 | unloaded_class_index); |
duke@435 | 1209 | if (class_name.not_null()) { |
duke@435 | 1210 | tty->print(unresolved ? " unresolved class: " : " symbol: "); |
duke@435 | 1211 | class_name->print_symbol_on(tty); |
duke@435 | 1212 | } |
duke@435 | 1213 | tty->cr(); |
duke@435 | 1214 | } |
duke@435 | 1215 | if (xtty != NULL) { |
duke@435 | 1216 | // Log the precise location of the trap. |
duke@435 | 1217 | for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { |
duke@435 | 1218 | xtty->begin_elem("jvms bci='%d'", sd->bci()); |
duke@435 | 1219 | xtty->method(sd->method()); |
duke@435 | 1220 | xtty->end_elem(); |
duke@435 | 1221 | if (sd->is_top()) break; |
duke@435 | 1222 | } |
duke@435 | 1223 | xtty->tail("uncommon_trap"); |
duke@435 | 1224 | } |
duke@435 | 1225 | } |
duke@435 | 1226 | // (End diagnostic printout.) |
duke@435 | 1227 | |
duke@435 | 1228 | // Load class if necessary |
duke@435 | 1229 | if (unloaded_class_index >= 0) { |
duke@435 | 1230 | constantPoolHandle constants(THREAD, trap_method->constants()); |
duke@435 | 1231 | load_class_by_index(constants, unloaded_class_index); |
duke@435 | 1232 | } |
duke@435 | 1233 | |
duke@435 | 1234 | // Flush the nmethod if necessary and desirable. |
duke@435 | 1235 | // |
duke@435 | 1236 | // We need to avoid situations where we are re-flushing the nmethod |
duke@435 | 1237 | // because of a hot deoptimization site. Repeated flushes at the same |
duke@435 | 1238 | // point need to be detected by the compiler and avoided. If the compiler |
duke@435 | 1239 | // cannot avoid them (or has a bug and "refuses" to avoid them), this |
duke@435 | 1240 | // module must take measures to avoid an infinite cycle of recompilation |
duke@435 | 1241 | // and deoptimization. There are several such measures: |
duke@435 | 1242 | // |
duke@435 | 1243 | // 1. If a recompilation is ordered a second time at some site X |
duke@435 | 1244 | // and for the same reason R, the action is adjusted to 'reinterpret', |
duke@435 | 1245 | // to give the interpreter time to exercise the method more thoroughly. |
duke@435 | 1246 | // If this happens, the method's overflow_recompile_count is incremented. |
duke@435 | 1247 | // |
duke@435 | 1248 | // 2. If the compiler fails to reduce the deoptimization rate, then |
duke@435 | 1249 | // the method's overflow_recompile_count will begin to exceed the set |
duke@435 | 1250 | // limit PerBytecodeRecompilationCutoff. If this happens, the action |
duke@435 | 1251 | // is adjusted to 'make_not_compilable', and the method is abandoned |
duke@435 | 1252 | // to the interpreter. This is a performance hit for hot methods, |
duke@435 | 1253 | // but is better than a disastrous infinite cycle of recompilations. |
duke@435 | 1254 | // (Actually, only the method containing the site X is abandoned.) |
duke@435 | 1255 | // |
duke@435 | 1256 | // 3. In parallel with the previous measures, if the total number of |
duke@435 | 1257 | // recompilations of a method exceeds the much larger set limit |
duke@435 | 1258 | // PerMethodRecompilationCutoff, the method is abandoned. |
duke@435 | 1259 | // This should only happen if the method is very large and has |
duke@435 | 1260 | // many "lukewarm" deoptimizations. The code which enforces this |
duke@435 | 1261 | // limit is elsewhere (class nmethod, class methodOopDesc). |
duke@435 | 1262 | // |
duke@435 | 1263 | // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance |
duke@435 | 1264 | // to recompile at each bytecode independently of the per-BCI cutoff. |
duke@435 | 1265 | // |
duke@435 | 1266 | // The decision to update code is up to the compiler, and is encoded |
duke@435 | 1267 | // in the Action_xxx code. If the compiler requests Action_none |
duke@435 | 1268 | // no trap state is changed, no compiled code is changed, and the |
duke@435 | 1269 | // computation suffers along in the interpreter. |
duke@435 | 1270 | // |
duke@435 | 1271 | // The other action codes specify various tactics for decompilation |
duke@435 | 1272 | // and recompilation. Action_maybe_recompile is the loosest, and |
duke@435 | 1273 | // allows the compiled code to stay around until enough traps are seen, |
duke@435 | 1274 | // and until the compiler gets around to recompiling the trapping method. |
duke@435 | 1275 | // |
duke@435 | 1276 | // The other actions cause immediate removal of the present code. |
duke@435 | 1277 | |
duke@435 | 1278 | bool update_trap_state = true; |
duke@435 | 1279 | bool make_not_entrant = false; |
duke@435 | 1280 | bool make_not_compilable = false; |
duke@435 | 1281 | bool reset_counters = false; |
duke@435 | 1282 | switch (action) { |
duke@435 | 1283 | case Action_none: |
duke@435 | 1284 | // Keep the old code. |
duke@435 | 1285 | update_trap_state = false; |
duke@435 | 1286 | break; |
duke@435 | 1287 | case Action_maybe_recompile: |
duke@435 | 1288 | // Do not need to invalidate the present code, but we can |
duke@435 | 1289 | // initiate another |
duke@435 | 1290 | // Start compiler without (necessarily) invalidating the nmethod. |
duke@435 | 1291 | // The system will tolerate the old code, but new code should be |
duke@435 | 1292 | // generated when possible. |
duke@435 | 1293 | break; |
duke@435 | 1294 | case Action_reinterpret: |
duke@435 | 1295 | // Go back into the interpreter for a while, and then consider |
duke@435 | 1296 | // recompiling form scratch. |
duke@435 | 1297 | make_not_entrant = true; |
duke@435 | 1298 | // Reset invocation counter for outer most method. |
duke@435 | 1299 | // This will allow the interpreter to exercise the bytecodes |
duke@435 | 1300 | // for a while before recompiling. |
duke@435 | 1301 | // By contrast, Action_make_not_entrant is immediate. |
duke@435 | 1302 | // |
duke@435 | 1303 | // Note that the compiler will track null_check, null_assert, |
duke@435 | 1304 | // range_check, and class_check events and log them as if they |
duke@435 | 1305 | // had been traps taken from compiled code. This will update |
duke@435 | 1306 | // the MDO trap history so that the next compilation will |
duke@435 | 1307 | // properly detect hot trap sites. |
duke@435 | 1308 | reset_counters = true; |
duke@435 | 1309 | break; |
duke@435 | 1310 | case Action_make_not_entrant: |
duke@435 | 1311 | // Request immediate recompilation, and get rid of the old code. |
duke@435 | 1312 | // Make them not entrant, so next time they are called they get |
duke@435 | 1313 | // recompiled. Unloaded classes are loaded now so recompile before next |
duke@435 | 1314 | // time they are called. Same for uninitialized. The interpreter will |
duke@435 | 1315 | // link the missing class, if any. |
duke@435 | 1316 | make_not_entrant = true; |
duke@435 | 1317 | break; |
duke@435 | 1318 | case Action_make_not_compilable: |
duke@435 | 1319 | // Give up on compiling this method at all. |
duke@435 | 1320 | make_not_entrant = true; |
duke@435 | 1321 | make_not_compilable = true; |
duke@435 | 1322 | break; |
duke@435 | 1323 | default: |
duke@435 | 1324 | ShouldNotReachHere(); |
duke@435 | 1325 | } |
duke@435 | 1326 | |
duke@435 | 1327 | // Setting +ProfileTraps fixes the following, on all platforms: |
duke@435 | 1328 | // 4852688: ProfileInterpreter is off by default for ia64. The result is |
duke@435 | 1329 | // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the |
duke@435 | 1330 | // recompile relies on a methodDataOop to record heroic opt failures. |
duke@435 | 1331 | |
duke@435 | 1332 | // Whether the interpreter is producing MDO data or not, we also need |
duke@435 | 1333 | // to use the MDO to detect hot deoptimization points and control |
duke@435 | 1334 | // aggressive optimization. |
duke@435 | 1335 | if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { |
duke@435 | 1336 | assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); |
duke@435 | 1337 | uint this_trap_count = 0; |
duke@435 | 1338 | bool maybe_prior_trap = false; |
duke@435 | 1339 | bool maybe_prior_recompile = false; |
duke@435 | 1340 | ProfileData* pdata |
duke@435 | 1341 | = query_update_method_data(trap_mdo, trap_bci, reason, |
duke@435 | 1342 | //outputs: |
duke@435 | 1343 | this_trap_count, |
duke@435 | 1344 | maybe_prior_trap, |
duke@435 | 1345 | maybe_prior_recompile); |
duke@435 | 1346 | // Because the interpreter also counts null, div0, range, and class |
duke@435 | 1347 | // checks, these traps from compiled code are double-counted. |
duke@435 | 1348 | // This is harmless; it just means that the PerXTrapLimit values |
duke@435 | 1349 | // are in effect a little smaller than they look. |
duke@435 | 1350 | |
duke@435 | 1351 | DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); |
duke@435 | 1352 | if (per_bc_reason != Reason_none) { |
duke@435 | 1353 | // Now take action based on the partially known per-BCI history. |
duke@435 | 1354 | if (maybe_prior_trap |
duke@435 | 1355 | && this_trap_count >= (uint)PerBytecodeTrapLimit) { |
duke@435 | 1356 | // If there are too many traps at this BCI, force a recompile. |
duke@435 | 1357 | // This will allow the compiler to see the limit overflow, and |
duke@435 | 1358 | // take corrective action, if possible. The compiler generally |
duke@435 | 1359 | // does not use the exact PerBytecodeTrapLimit value, but instead |
duke@435 | 1360 | // changes its tactics if it sees any traps at all. This provides |
duke@435 | 1361 | // a little hysteresis, delaying a recompile until a trap happens |
duke@435 | 1362 | // several times. |
duke@435 | 1363 | // |
duke@435 | 1364 | // Actually, since there is only one bit of counter per BCI, |
duke@435 | 1365 | // the possible per-BCI counts are {0,1,(per-method count)}. |
duke@435 | 1366 | // This produces accurate results if in fact there is only |
duke@435 | 1367 | // one hot trap site, but begins to get fuzzy if there are |
duke@435 | 1368 | // many sites. For example, if there are ten sites each |
duke@435 | 1369 | // trapping two or more times, they each get the blame for |
duke@435 | 1370 | // all of their traps. |
duke@435 | 1371 | make_not_entrant = true; |
duke@435 | 1372 | } |
duke@435 | 1373 | |
duke@435 | 1374 | // Detect repeated recompilation at the same BCI, and enforce a limit. |
duke@435 | 1375 | if (make_not_entrant && maybe_prior_recompile) { |
duke@435 | 1376 | // More than one recompile at this point. |
duke@435 | 1377 | trap_mdo->inc_overflow_recompile_count(); |
duke@435 | 1378 | if (maybe_prior_trap |
duke@435 | 1379 | && ((uint)trap_mdo->overflow_recompile_count() |
duke@435 | 1380 | > (uint)PerBytecodeRecompilationCutoff)) { |
duke@435 | 1381 | // Give up on the method containing the bad BCI. |
duke@435 | 1382 | if (trap_method() == nm->method()) { |
duke@435 | 1383 | make_not_compilable = true; |
duke@435 | 1384 | } else { |
duke@435 | 1385 | trap_method->set_not_compilable(); |
duke@435 | 1386 | // But give grace to the enclosing nm->method(). |
duke@435 | 1387 | } |
duke@435 | 1388 | } |
duke@435 | 1389 | } |
duke@435 | 1390 | } else { |
duke@435 | 1391 | // For reasons which are not recorded per-bytecode, we simply |
duke@435 | 1392 | // force recompiles unconditionally. |
duke@435 | 1393 | // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) |
duke@435 | 1394 | make_not_entrant = true; |
duke@435 | 1395 | } |
duke@435 | 1396 | |
duke@435 | 1397 | // Go back to the compiler if there are too many traps in this method. |
duke@435 | 1398 | if (this_trap_count >= (uint)PerMethodTrapLimit) { |
duke@435 | 1399 | // If there are too many traps in this method, force a recompile. |
duke@435 | 1400 | // This will allow the compiler to see the limit overflow, and |
duke@435 | 1401 | // take corrective action, if possible. |
duke@435 | 1402 | // (This condition is an unlikely backstop only, because the |
duke@435 | 1403 | // PerBytecodeTrapLimit is more likely to take effect first, |
duke@435 | 1404 | // if it is applicable.) |
duke@435 | 1405 | make_not_entrant = true; |
duke@435 | 1406 | } |
duke@435 | 1407 | |
duke@435 | 1408 | // Here's more hysteresis: If there has been a recompile at |
duke@435 | 1409 | // this trap point already, run the method in the interpreter |
duke@435 | 1410 | // for a while to exercise it more thoroughly. |
duke@435 | 1411 | if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { |
duke@435 | 1412 | reset_counters = true; |
duke@435 | 1413 | } |
duke@435 | 1414 | |
duke@435 | 1415 | if (make_not_entrant && pdata != NULL) { |
duke@435 | 1416 | // Record the recompilation event, if any. |
duke@435 | 1417 | int tstate0 = pdata->trap_state(); |
duke@435 | 1418 | int tstate1 = trap_state_set_recompiled(tstate0, true); |
duke@435 | 1419 | if (tstate1 != tstate0) |
duke@435 | 1420 | pdata->set_trap_state(tstate1); |
duke@435 | 1421 | } |
duke@435 | 1422 | } |
duke@435 | 1423 | |
duke@435 | 1424 | // Take requested actions on the method: |
duke@435 | 1425 | |
duke@435 | 1426 | // Reset invocation counters |
duke@435 | 1427 | if (reset_counters) { |
duke@435 | 1428 | if (nm->is_osr_method()) |
duke@435 | 1429 | reset_invocation_counter(trap_scope, CompileThreshold); |
duke@435 | 1430 | else |
duke@435 | 1431 | reset_invocation_counter(trap_scope); |
duke@435 | 1432 | } |
duke@435 | 1433 | |
duke@435 | 1434 | // Recompile |
duke@435 | 1435 | if (make_not_entrant) { |
duke@435 | 1436 | nm->make_not_entrant(); |
duke@435 | 1437 | } |
duke@435 | 1438 | |
duke@435 | 1439 | // Give up compiling |
duke@435 | 1440 | if (make_not_compilable) { |
duke@435 | 1441 | assert(make_not_entrant, "consistent"); |
duke@435 | 1442 | nm->method()->set_not_compilable(); |
duke@435 | 1443 | } |
duke@435 | 1444 | |
duke@435 | 1445 | } // Free marked resources |
duke@435 | 1446 | |
duke@435 | 1447 | } |
duke@435 | 1448 | JRT_END |
duke@435 | 1449 | |
duke@435 | 1450 | methodDataOop |
duke@435 | 1451 | Deoptimization::get_method_data(JavaThread* thread, methodHandle m, |
duke@435 | 1452 | bool create_if_missing) { |
duke@435 | 1453 | Thread* THREAD = thread; |
duke@435 | 1454 | methodDataOop mdo = m()->method_data(); |
duke@435 | 1455 | if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { |
duke@435 | 1456 | // Build an MDO. Ignore errors like OutOfMemory; |
duke@435 | 1457 | // that simply means we won't have an MDO to update. |
duke@435 | 1458 | methodOopDesc::build_interpreter_method_data(m, THREAD); |
duke@435 | 1459 | if (HAS_PENDING_EXCEPTION) { |
duke@435 | 1460 | assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); |
duke@435 | 1461 | CLEAR_PENDING_EXCEPTION; |
duke@435 | 1462 | } |
duke@435 | 1463 | mdo = m()->method_data(); |
duke@435 | 1464 | } |
duke@435 | 1465 | return mdo; |
duke@435 | 1466 | } |
duke@435 | 1467 | |
duke@435 | 1468 | ProfileData* |
duke@435 | 1469 | Deoptimization::query_update_method_data(methodDataHandle trap_mdo, |
duke@435 | 1470 | int trap_bci, |
duke@435 | 1471 | Deoptimization::DeoptReason reason, |
duke@435 | 1472 | //outputs: |
duke@435 | 1473 | uint& ret_this_trap_count, |
duke@435 | 1474 | bool& ret_maybe_prior_trap, |
duke@435 | 1475 | bool& ret_maybe_prior_recompile) { |
duke@435 | 1476 | uint prior_trap_count = trap_mdo->trap_count(reason); |
duke@435 | 1477 | uint this_trap_count = trap_mdo->inc_trap_count(reason); |
duke@435 | 1478 | |
duke@435 | 1479 | // If the runtime cannot find a place to store trap history, |
duke@435 | 1480 | // it is estimated based on the general condition of the method. |
duke@435 | 1481 | // If the method has ever been recompiled, or has ever incurred |
duke@435 | 1482 | // a trap with the present reason , then this BCI is assumed |
duke@435 | 1483 | // (pessimistically) to be the culprit. |
duke@435 | 1484 | bool maybe_prior_trap = (prior_trap_count != 0); |
duke@435 | 1485 | bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); |
duke@435 | 1486 | ProfileData* pdata = NULL; |
duke@435 | 1487 | |
duke@435 | 1488 | |
duke@435 | 1489 | // For reasons which are recorded per bytecode, we check per-BCI data. |
duke@435 | 1490 | DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); |
duke@435 | 1491 | if (per_bc_reason != Reason_none) { |
duke@435 | 1492 | // Find the profile data for this BCI. If there isn't one, |
duke@435 | 1493 | // try to allocate one from the MDO's set of spares. |
duke@435 | 1494 | // This will let us detect a repeated trap at this point. |
duke@435 | 1495 | pdata = trap_mdo->allocate_bci_to_data(trap_bci); |
duke@435 | 1496 | |
duke@435 | 1497 | if (pdata != NULL) { |
duke@435 | 1498 | // Query the trap state of this profile datum. |
duke@435 | 1499 | int tstate0 = pdata->trap_state(); |
duke@435 | 1500 | if (!trap_state_has_reason(tstate0, per_bc_reason)) |
duke@435 | 1501 | maybe_prior_trap = false; |
duke@435 | 1502 | if (!trap_state_is_recompiled(tstate0)) |
duke@435 | 1503 | maybe_prior_recompile = false; |
duke@435 | 1504 | |
duke@435 | 1505 | // Update the trap state of this profile datum. |
duke@435 | 1506 | int tstate1 = tstate0; |
duke@435 | 1507 | // Record the reason. |
duke@435 | 1508 | tstate1 = trap_state_add_reason(tstate1, per_bc_reason); |
duke@435 | 1509 | // Store the updated state on the MDO, for next time. |
duke@435 | 1510 | if (tstate1 != tstate0) |
duke@435 | 1511 | pdata->set_trap_state(tstate1); |
duke@435 | 1512 | } else { |
duke@435 | 1513 | if (LogCompilation && xtty != NULL) |
duke@435 | 1514 | // Missing MDP? Leave a small complaint in the log. |
duke@435 | 1515 | xtty->elem("missing_mdp bci='%d'", trap_bci); |
duke@435 | 1516 | } |
duke@435 | 1517 | } |
duke@435 | 1518 | |
duke@435 | 1519 | // Return results: |
duke@435 | 1520 | ret_this_trap_count = this_trap_count; |
duke@435 | 1521 | ret_maybe_prior_trap = maybe_prior_trap; |
duke@435 | 1522 | ret_maybe_prior_recompile = maybe_prior_recompile; |
duke@435 | 1523 | return pdata; |
duke@435 | 1524 | } |
duke@435 | 1525 | |
duke@435 | 1526 | void |
duke@435 | 1527 | Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { |
duke@435 | 1528 | ResourceMark rm; |
duke@435 | 1529 | // Ignored outputs: |
duke@435 | 1530 | uint ignore_this_trap_count; |
duke@435 | 1531 | bool ignore_maybe_prior_trap; |
duke@435 | 1532 | bool ignore_maybe_prior_recompile; |
duke@435 | 1533 | query_update_method_data(trap_mdo, trap_bci, |
duke@435 | 1534 | (DeoptReason)reason, |
duke@435 | 1535 | ignore_this_trap_count, |
duke@435 | 1536 | ignore_maybe_prior_trap, |
duke@435 | 1537 | ignore_maybe_prior_recompile); |
duke@435 | 1538 | } |
duke@435 | 1539 | |
duke@435 | 1540 | void Deoptimization::reset_invocation_counter(ScopeDesc* trap_scope, jint top_count) { |
duke@435 | 1541 | ScopeDesc* sd = trap_scope; |
duke@435 | 1542 | for (; !sd->is_top(); sd = sd->sender()) { |
duke@435 | 1543 | // Reset ICs of inlined methods, since they can trigger compilations also. |
duke@435 | 1544 | sd->method()->invocation_counter()->reset(); |
duke@435 | 1545 | } |
duke@435 | 1546 | InvocationCounter* c = sd->method()->invocation_counter(); |
duke@435 | 1547 | if (top_count != _no_count) { |
duke@435 | 1548 | // It was an OSR method, so bump the count higher. |
duke@435 | 1549 | c->set(c->state(), top_count); |
duke@435 | 1550 | } else { |
duke@435 | 1551 | c->reset(); |
duke@435 | 1552 | } |
duke@435 | 1553 | sd->method()->backedge_counter()->reset(); |
duke@435 | 1554 | } |
duke@435 | 1555 | |
duke@435 | 1556 | Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { |
duke@435 | 1557 | |
duke@435 | 1558 | // Still in Java no safepoints |
duke@435 | 1559 | { |
duke@435 | 1560 | // This enters VM and may safepoint |
duke@435 | 1561 | uncommon_trap_inner(thread, trap_request); |
duke@435 | 1562 | } |
duke@435 | 1563 | return fetch_unroll_info_helper(thread); |
duke@435 | 1564 | } |
duke@435 | 1565 | |
duke@435 | 1566 | // Local derived constants. |
duke@435 | 1567 | // Further breakdown of DataLayout::trap_state, as promised by DataLayout. |
duke@435 | 1568 | const int DS_REASON_MASK = DataLayout::trap_mask >> 1; |
duke@435 | 1569 | const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; |
duke@435 | 1570 | |
duke@435 | 1571 | //---------------------------trap_state_reason--------------------------------- |
duke@435 | 1572 | Deoptimization::DeoptReason |
duke@435 | 1573 | Deoptimization::trap_state_reason(int trap_state) { |
duke@435 | 1574 | // This assert provides the link between the width of DataLayout::trap_bits |
duke@435 | 1575 | // and the encoding of "recorded" reasons. It ensures there are enough |
duke@435 | 1576 | // bits to store all needed reasons in the per-BCI MDO profile. |
duke@435 | 1577 | assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); |
duke@435 | 1578 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
duke@435 | 1579 | trap_state -= recompile_bit; |
duke@435 | 1580 | if (trap_state == DS_REASON_MASK) { |
duke@435 | 1581 | return Reason_many; |
duke@435 | 1582 | } else { |
duke@435 | 1583 | assert((int)Reason_none == 0, "state=0 => Reason_none"); |
duke@435 | 1584 | return (DeoptReason)trap_state; |
duke@435 | 1585 | } |
duke@435 | 1586 | } |
duke@435 | 1587 | //-------------------------trap_state_has_reason------------------------------- |
duke@435 | 1588 | int Deoptimization::trap_state_has_reason(int trap_state, int reason) { |
duke@435 | 1589 | assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); |
duke@435 | 1590 | assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); |
duke@435 | 1591 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
duke@435 | 1592 | trap_state -= recompile_bit; |
duke@435 | 1593 | if (trap_state == DS_REASON_MASK) { |
duke@435 | 1594 | return -1; // true, unspecifically (bottom of state lattice) |
duke@435 | 1595 | } else if (trap_state == reason) { |
duke@435 | 1596 | return 1; // true, definitely |
duke@435 | 1597 | } else if (trap_state == 0) { |
duke@435 | 1598 | return 0; // false, definitely (top of state lattice) |
duke@435 | 1599 | } else { |
duke@435 | 1600 | return 0; // false, definitely |
duke@435 | 1601 | } |
duke@435 | 1602 | } |
duke@435 | 1603 | //-------------------------trap_state_add_reason------------------------------- |
duke@435 | 1604 | int Deoptimization::trap_state_add_reason(int trap_state, int reason) { |
duke@435 | 1605 | assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); |
duke@435 | 1606 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
duke@435 | 1607 | trap_state -= recompile_bit; |
duke@435 | 1608 | if (trap_state == DS_REASON_MASK) { |
duke@435 | 1609 | return trap_state + recompile_bit; // already at state lattice bottom |
duke@435 | 1610 | } else if (trap_state == reason) { |
duke@435 | 1611 | return trap_state + recompile_bit; // the condition is already true |
duke@435 | 1612 | } else if (trap_state == 0) { |
duke@435 | 1613 | return reason + recompile_bit; // no condition has yet been true |
duke@435 | 1614 | } else { |
duke@435 | 1615 | return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom |
duke@435 | 1616 | } |
duke@435 | 1617 | } |
duke@435 | 1618 | //-----------------------trap_state_is_recompiled------------------------------ |
duke@435 | 1619 | bool Deoptimization::trap_state_is_recompiled(int trap_state) { |
duke@435 | 1620 | return (trap_state & DS_RECOMPILE_BIT) != 0; |
duke@435 | 1621 | } |
duke@435 | 1622 | //-----------------------trap_state_set_recompiled----------------------------- |
duke@435 | 1623 | int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { |
duke@435 | 1624 | if (z) return trap_state | DS_RECOMPILE_BIT; |
duke@435 | 1625 | else return trap_state & ~DS_RECOMPILE_BIT; |
duke@435 | 1626 | } |
duke@435 | 1627 | //---------------------------format_trap_state--------------------------------- |
duke@435 | 1628 | // This is used for debugging and diagnostics, including hotspot.log output. |
duke@435 | 1629 | const char* Deoptimization::format_trap_state(char* buf, size_t buflen, |
duke@435 | 1630 | int trap_state) { |
duke@435 | 1631 | DeoptReason reason = trap_state_reason(trap_state); |
duke@435 | 1632 | bool recomp_flag = trap_state_is_recompiled(trap_state); |
duke@435 | 1633 | // Re-encode the state from its decoded components. |
duke@435 | 1634 | int decoded_state = 0; |
duke@435 | 1635 | if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) |
duke@435 | 1636 | decoded_state = trap_state_add_reason(decoded_state, reason); |
duke@435 | 1637 | if (recomp_flag) |
duke@435 | 1638 | decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); |
duke@435 | 1639 | // If the state re-encodes properly, format it symbolically. |
duke@435 | 1640 | // Because this routine is used for debugging and diagnostics, |
duke@435 | 1641 | // be robust even if the state is a strange value. |
duke@435 | 1642 | size_t len; |
duke@435 | 1643 | if (decoded_state != trap_state) { |
duke@435 | 1644 | // Random buggy state that doesn't decode?? |
duke@435 | 1645 | len = jio_snprintf(buf, buflen, "#%d", trap_state); |
duke@435 | 1646 | } else { |
duke@435 | 1647 | len = jio_snprintf(buf, buflen, "%s%s", |
duke@435 | 1648 | trap_reason_name(reason), |
duke@435 | 1649 | recomp_flag ? " recompiled" : ""); |
duke@435 | 1650 | } |
duke@435 | 1651 | if (len >= buflen) |
duke@435 | 1652 | buf[buflen-1] = '\0'; |
duke@435 | 1653 | return buf; |
duke@435 | 1654 | } |
duke@435 | 1655 | |
duke@435 | 1656 | |
duke@435 | 1657 | //--------------------------------statics-------------------------------------- |
duke@435 | 1658 | Deoptimization::DeoptAction Deoptimization::_unloaded_action |
duke@435 | 1659 | = Deoptimization::Action_reinterpret; |
duke@435 | 1660 | const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { |
duke@435 | 1661 | // Note: Keep this in sync. with enum DeoptReason. |
duke@435 | 1662 | "none", |
duke@435 | 1663 | "null_check", |
duke@435 | 1664 | "null_assert", |
duke@435 | 1665 | "range_check", |
duke@435 | 1666 | "class_check", |
duke@435 | 1667 | "array_check", |
duke@435 | 1668 | "intrinsic", |
duke@435 | 1669 | "unloaded", |
duke@435 | 1670 | "uninitialized", |
duke@435 | 1671 | "unreached", |
duke@435 | 1672 | "unhandled", |
duke@435 | 1673 | "constraint", |
duke@435 | 1674 | "div0_check", |
duke@435 | 1675 | "age" |
duke@435 | 1676 | }; |
duke@435 | 1677 | const char* Deoptimization::_trap_action_name[Action_LIMIT] = { |
duke@435 | 1678 | // Note: Keep this in sync. with enum DeoptAction. |
duke@435 | 1679 | "none", |
duke@435 | 1680 | "maybe_recompile", |
duke@435 | 1681 | "reinterpret", |
duke@435 | 1682 | "make_not_entrant", |
duke@435 | 1683 | "make_not_compilable" |
duke@435 | 1684 | }; |
duke@435 | 1685 | |
duke@435 | 1686 | const char* Deoptimization::trap_reason_name(int reason) { |
duke@435 | 1687 | if (reason == Reason_many) return "many"; |
duke@435 | 1688 | if ((uint)reason < Reason_LIMIT) |
duke@435 | 1689 | return _trap_reason_name[reason]; |
duke@435 | 1690 | static char buf[20]; |
duke@435 | 1691 | sprintf(buf, "reason%d", reason); |
duke@435 | 1692 | return buf; |
duke@435 | 1693 | } |
duke@435 | 1694 | const char* Deoptimization::trap_action_name(int action) { |
duke@435 | 1695 | if ((uint)action < Action_LIMIT) |
duke@435 | 1696 | return _trap_action_name[action]; |
duke@435 | 1697 | static char buf[20]; |
duke@435 | 1698 | sprintf(buf, "action%d", action); |
duke@435 | 1699 | return buf; |
duke@435 | 1700 | } |
duke@435 | 1701 | |
duke@435 | 1702 | // This is used for debugging and diagnostics, including hotspot.log output. |
duke@435 | 1703 | const char* Deoptimization::format_trap_request(char* buf, size_t buflen, |
duke@435 | 1704 | int trap_request) { |
duke@435 | 1705 | jint unloaded_class_index = trap_request_index(trap_request); |
duke@435 | 1706 | const char* reason = trap_reason_name(trap_request_reason(trap_request)); |
duke@435 | 1707 | const char* action = trap_action_name(trap_request_action(trap_request)); |
duke@435 | 1708 | size_t len; |
duke@435 | 1709 | if (unloaded_class_index < 0) { |
duke@435 | 1710 | len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", |
duke@435 | 1711 | reason, action); |
duke@435 | 1712 | } else { |
duke@435 | 1713 | len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", |
duke@435 | 1714 | reason, action, unloaded_class_index); |
duke@435 | 1715 | } |
duke@435 | 1716 | if (len >= buflen) |
duke@435 | 1717 | buf[buflen-1] = '\0'; |
duke@435 | 1718 | return buf; |
duke@435 | 1719 | } |
duke@435 | 1720 | |
duke@435 | 1721 | juint Deoptimization::_deoptimization_hist |
duke@435 | 1722 | [Deoptimization::Reason_LIMIT] |
duke@435 | 1723 | [1 + Deoptimization::Action_LIMIT] |
duke@435 | 1724 | [Deoptimization::BC_CASE_LIMIT] |
duke@435 | 1725 | = {0}; |
duke@435 | 1726 | |
duke@435 | 1727 | enum { |
duke@435 | 1728 | LSB_BITS = 8, |
duke@435 | 1729 | LSB_MASK = right_n_bits(LSB_BITS) |
duke@435 | 1730 | }; |
duke@435 | 1731 | |
duke@435 | 1732 | void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, |
duke@435 | 1733 | Bytecodes::Code bc) { |
duke@435 | 1734 | assert(reason >= 0 && reason < Reason_LIMIT, "oob"); |
duke@435 | 1735 | assert(action >= 0 && action < Action_LIMIT, "oob"); |
duke@435 | 1736 | _deoptimization_hist[Reason_none][0][0] += 1; // total |
duke@435 | 1737 | _deoptimization_hist[reason][0][0] += 1; // per-reason total |
duke@435 | 1738 | juint* cases = _deoptimization_hist[reason][1+action]; |
duke@435 | 1739 | juint* bc_counter_addr = NULL; |
duke@435 | 1740 | juint bc_counter = 0; |
duke@435 | 1741 | // Look for an unused counter, or an exact match to this BC. |
duke@435 | 1742 | if (bc != Bytecodes::_illegal) { |
duke@435 | 1743 | for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { |
duke@435 | 1744 | juint* counter_addr = &cases[bc_case]; |
duke@435 | 1745 | juint counter = *counter_addr; |
duke@435 | 1746 | if ((counter == 0 && bc_counter_addr == NULL) |
duke@435 | 1747 | || (Bytecodes::Code)(counter & LSB_MASK) == bc) { |
duke@435 | 1748 | // this counter is either free or is already devoted to this BC |
duke@435 | 1749 | bc_counter_addr = counter_addr; |
duke@435 | 1750 | bc_counter = counter | bc; |
duke@435 | 1751 | } |
duke@435 | 1752 | } |
duke@435 | 1753 | } |
duke@435 | 1754 | if (bc_counter_addr == NULL) { |
duke@435 | 1755 | // Overflow, or no given bytecode. |
duke@435 | 1756 | bc_counter_addr = &cases[BC_CASE_LIMIT-1]; |
duke@435 | 1757 | bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB |
duke@435 | 1758 | } |
duke@435 | 1759 | *bc_counter_addr = bc_counter + (1 << LSB_BITS); |
duke@435 | 1760 | } |
duke@435 | 1761 | |
duke@435 | 1762 | jint Deoptimization::total_deoptimization_count() { |
duke@435 | 1763 | return _deoptimization_hist[Reason_none][0][0]; |
duke@435 | 1764 | } |
duke@435 | 1765 | |
duke@435 | 1766 | jint Deoptimization::deoptimization_count(DeoptReason reason) { |
duke@435 | 1767 | assert(reason >= 0 && reason < Reason_LIMIT, "oob"); |
duke@435 | 1768 | return _deoptimization_hist[reason][0][0]; |
duke@435 | 1769 | } |
duke@435 | 1770 | |
duke@435 | 1771 | void Deoptimization::print_statistics() { |
duke@435 | 1772 | juint total = total_deoptimization_count(); |
duke@435 | 1773 | juint account = total; |
duke@435 | 1774 | if (total != 0) { |
duke@435 | 1775 | ttyLocker ttyl; |
duke@435 | 1776 | if (xtty != NULL) xtty->head("statistics type='deoptimization'"); |
duke@435 | 1777 | tty->print_cr("Deoptimization traps recorded:"); |
duke@435 | 1778 | #define PRINT_STAT_LINE(name, r) \ |
duke@435 | 1779 | tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); |
duke@435 | 1780 | PRINT_STAT_LINE("total", total); |
duke@435 | 1781 | // For each non-zero entry in the histogram, print the reason, |
duke@435 | 1782 | // the action, and (if specifically known) the type of bytecode. |
duke@435 | 1783 | for (int reason = 0; reason < Reason_LIMIT; reason++) { |
duke@435 | 1784 | for (int action = 0; action < Action_LIMIT; action++) { |
duke@435 | 1785 | juint* cases = _deoptimization_hist[reason][1+action]; |
duke@435 | 1786 | for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { |
duke@435 | 1787 | juint counter = cases[bc_case]; |
duke@435 | 1788 | if (counter != 0) { |
duke@435 | 1789 | char name[1*K]; |
duke@435 | 1790 | Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); |
duke@435 | 1791 | if (bc_case == BC_CASE_LIMIT && (int)bc == 0) |
duke@435 | 1792 | bc = Bytecodes::_illegal; |
duke@435 | 1793 | sprintf(name, "%s/%s/%s", |
duke@435 | 1794 | trap_reason_name(reason), |
duke@435 | 1795 | trap_action_name(action), |
duke@435 | 1796 | Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); |
duke@435 | 1797 | juint r = counter >> LSB_BITS; |
duke@435 | 1798 | tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); |
duke@435 | 1799 | account -= r; |
duke@435 | 1800 | } |
duke@435 | 1801 | } |
duke@435 | 1802 | } |
duke@435 | 1803 | } |
duke@435 | 1804 | if (account != 0) { |
duke@435 | 1805 | PRINT_STAT_LINE("unaccounted", account); |
duke@435 | 1806 | } |
duke@435 | 1807 | #undef PRINT_STAT_LINE |
duke@435 | 1808 | if (xtty != NULL) xtty->tail("statistics"); |
duke@435 | 1809 | } |
duke@435 | 1810 | } |
duke@435 | 1811 | #else // COMPILER2 |
duke@435 | 1812 | |
duke@435 | 1813 | |
duke@435 | 1814 | // Stubs for C1 only system. |
duke@435 | 1815 | bool Deoptimization::trap_state_is_recompiled(int trap_state) { |
duke@435 | 1816 | return false; |
duke@435 | 1817 | } |
duke@435 | 1818 | |
duke@435 | 1819 | const char* Deoptimization::trap_reason_name(int reason) { |
duke@435 | 1820 | return "unknown"; |
duke@435 | 1821 | } |
duke@435 | 1822 | |
duke@435 | 1823 | void Deoptimization::print_statistics() { |
duke@435 | 1824 | // no output |
duke@435 | 1825 | } |
duke@435 | 1826 | |
duke@435 | 1827 | void |
duke@435 | 1828 | Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { |
duke@435 | 1829 | // no udpate |
duke@435 | 1830 | } |
duke@435 | 1831 | |
duke@435 | 1832 | int Deoptimization::trap_state_has_reason(int trap_state, int reason) { |
duke@435 | 1833 | return 0; |
duke@435 | 1834 | } |
duke@435 | 1835 | |
duke@435 | 1836 | void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, |
duke@435 | 1837 | Bytecodes::Code bc) { |
duke@435 | 1838 | // no update |
duke@435 | 1839 | } |
duke@435 | 1840 | |
duke@435 | 1841 | const char* Deoptimization::format_trap_state(char* buf, size_t buflen, |
duke@435 | 1842 | int trap_state) { |
duke@435 | 1843 | jio_snprintf(buf, buflen, "#%d", trap_state); |
duke@435 | 1844 | return buf; |
duke@435 | 1845 | } |
duke@435 | 1846 | |
duke@435 | 1847 | #endif // COMPILER2 |