duke@435: /* never@3499: * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "classfile/systemDictionary.hpp" stefank@2314: #include "code/debugInfoRec.hpp" stefank@2314: #include "code/nmethod.hpp" stefank@2314: #include "code/pcDesc.hpp" stefank@2314: #include "code/scopeDesc.hpp" stefank@2314: #include "interpreter/bytecode.hpp" stefank@2314: #include "interpreter/interpreter.hpp" stefank@2314: #include "interpreter/oopMapCache.hpp" stefank@2314: #include "memory/allocation.inline.hpp" stefank@2314: #include "memory/oopFactory.hpp" stefank@2314: #include "memory/resourceArea.hpp" stefank@2314: #include "oops/methodOop.hpp" stefank@2314: #include "oops/oop.inline.hpp" stefank@2314: #include "prims/jvmtiThreadState.hpp" stefank@2314: #include "runtime/biasedLocking.hpp" stefank@2314: #include "runtime/compilationPolicy.hpp" stefank@2314: #include "runtime/deoptimization.hpp" stefank@2314: #include "runtime/interfaceSupport.hpp" stefank@2314: #include "runtime/sharedRuntime.hpp" stefank@2314: #include "runtime/signature.hpp" stefank@2314: #include "runtime/stubRoutines.hpp" stefank@2314: #include "runtime/thread.hpp" stefank@2314: #include "runtime/vframe.hpp" stefank@2314: #include "runtime/vframeArray.hpp" stefank@2314: #include "runtime/vframe_hp.hpp" stefank@2314: #include "utilities/events.hpp" stefank@2314: #include "utilities/xmlstream.hpp" stefank@2314: #ifdef TARGET_ARCH_x86 stefank@2314: # include "vmreg_x86.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_sparc stefank@2314: # include "vmreg_sparc.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_zero stefank@2314: # include "vmreg_zero.inline.hpp" stefank@2314: #endif bobv@2508: #ifdef TARGET_ARCH_arm bobv@2508: # include "vmreg_arm.inline.hpp" bobv@2508: #endif bobv@2508: #ifdef TARGET_ARCH_ppc bobv@2508: # include "vmreg_ppc.inline.hpp" bobv@2508: #endif stefank@2314: #ifdef COMPILER2 stefank@2314: #ifdef TARGET_ARCH_MODEL_x86_32 stefank@2314: # include "adfiles/ad_x86_32.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_x86_64 stefank@2314: # include "adfiles/ad_x86_64.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_sparc stefank@2314: # include "adfiles/ad_sparc.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_zero stefank@2314: # include "adfiles/ad_zero.hpp" stefank@2314: #endif bobv@2508: #ifdef TARGET_ARCH_MODEL_arm bobv@2508: # include "adfiles/ad_arm.hpp" bobv@2508: #endif bobv@2508: #ifdef TARGET_ARCH_MODEL_ppc bobv@2508: # include "adfiles/ad_ppc.hpp" bobv@2508: #endif stefank@2314: #endif duke@435: duke@435: bool DeoptimizationMarker::_is_active = false; duke@435: duke@435: Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, duke@435: int caller_adjustment, never@2901: int caller_actual_parameters, duke@435: int number_of_frames, duke@435: intptr_t* frame_sizes, duke@435: address* frame_pcs, duke@435: BasicType return_type) { duke@435: _size_of_deoptimized_frame = size_of_deoptimized_frame; duke@435: _caller_adjustment = caller_adjustment; never@2901: _caller_actual_parameters = caller_actual_parameters; duke@435: _number_of_frames = number_of_frames; duke@435: _frame_sizes = frame_sizes; duke@435: _frame_pcs = frame_pcs; zgu@3900: _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); duke@435: _return_type = return_type; bdelsart@3130: _initial_info = 0; duke@435: // PD (x86 only) duke@435: _counter_temp = 0; duke@435: _unpack_kind = 0; duke@435: _sender_sp_temp = 0; duke@435: duke@435: _total_frame_sizes = size_of_frames(); duke@435: } duke@435: duke@435: duke@435: Deoptimization::UnrollBlock::~UnrollBlock() { zgu@3900: FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); zgu@3900: FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); zgu@3900: FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); duke@435: } duke@435: duke@435: duke@435: intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { duke@435: assert(register_number < RegisterMap::reg_count, "checking register number"); duke@435: return &_register_block[register_number * 2]; duke@435: } duke@435: duke@435: duke@435: duke@435: int Deoptimization::UnrollBlock::size_of_frames() const { duke@435: // Acount first for the adjustment of the initial frame duke@435: int result = _caller_adjustment; duke@435: for (int index = 0; index < number_of_frames(); index++) { duke@435: result += frame_sizes()[index]; duke@435: } duke@435: return result; duke@435: } duke@435: duke@435: duke@435: void Deoptimization::UnrollBlock::print() { duke@435: ttyLocker ttyl; duke@435: tty->print_cr("UnrollBlock"); duke@435: tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); duke@435: tty->print( " frame_sizes: "); duke@435: for (int index = 0; index < number_of_frames(); index++) { duke@435: tty->print("%d ", frame_sizes()[index]); duke@435: } duke@435: tty->cr(); duke@435: } duke@435: duke@435: duke@435: // In order to make fetch_unroll_info work properly with escape duke@435: // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and duke@435: // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation duke@435: // of previously eliminated objects occurs in realloc_objects, which is duke@435: // called from the method fetch_unroll_info_helper below. duke@435: JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) duke@435: // It is actually ok to allocate handles in a leaf method. It causes no safepoints, duke@435: // but makes the entry a little slower. There is however a little dance we have to duke@435: // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro duke@435: duke@435: // fetch_unroll_info() is called at the beginning of the deoptimization duke@435: // handler. Note this fact before we start generating temporary frames duke@435: // that can confuse an asynchronous stack walker. This counter is duke@435: // decremented at the end of unpack_frames(). duke@435: thread->inc_in_deopt_handler(); duke@435: duke@435: return fetch_unroll_info_helper(thread); duke@435: JRT_END duke@435: duke@435: duke@435: // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) duke@435: Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { duke@435: duke@435: // Note: there is a safepoint safety issue here. No matter whether we enter duke@435: // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once duke@435: // the vframeArray is created. duke@435: // duke@435: duke@435: // Allocate our special deoptimization ResourceMark duke@435: DeoptResourceMark* dmark = new DeoptResourceMark(thread); duke@435: assert(thread->deopt_mark() == NULL, "Pending deopt!"); duke@435: thread->set_deopt_mark(dmark); duke@435: duke@435: frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect duke@435: RegisterMap map(thread, true); duke@435: RegisterMap dummy_map(thread, false); duke@435: // Now get the deoptee with a valid map duke@435: frame deoptee = stub_frame.sender(&map); iveresov@2169: // Set the deoptee nmethod iveresov@2169: assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); iveresov@2169: thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); duke@435: never@2868: if (VerifyStack) { never@2868: thread->validate_frame_layout(); never@2868: } never@2868: duke@435: // Create a growable array of VFrames where each VFrame represents an inlined duke@435: // Java frame. This storage is allocated with the usual system arena. duke@435: assert(deoptee.is_compiled_frame(), "Wrong frame type"); duke@435: GrowableArray* chunk = new GrowableArray(10); duke@435: vframe* vf = vframe::new_vframe(&deoptee, &map, thread); duke@435: while (!vf->is_top()) { duke@435: assert(vf->is_compiled_frame(), "Wrong frame type"); duke@435: chunk->push(compiledVFrame::cast(vf)); duke@435: vf = vf->sender(); duke@435: } duke@435: assert(vf->is_compiled_frame(), "Wrong frame type"); duke@435: chunk->push(compiledVFrame::cast(vf)); duke@435: duke@435: #ifdef COMPILER2 duke@435: // Reallocate the non-escaping objects and restore their fields. Then duke@435: // relock objects if synchronization on them was eliminated. kvn@3406: if (DoEscapeAnalysis || EliminateNestedLocks) { kvn@479: if (EliminateAllocations) { kvn@518: assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); kvn@479: GrowableArray* objects = chunk->at(0)->scope()->objects(); kvn@1688: kvn@1688: // The flag return_oop() indicates call sites which return oop kvn@1688: // in compiled code. Such sites include java method calls, kvn@1688: // runtime calls (for example, used to allocate new objects/arrays kvn@1688: // on slow code path) and any other calls generated in compiled code. kvn@1688: // It is not guaranteed that we can get such information here only kvn@1688: // by analyzing bytecode in deoptimized frames. This is why this flag kvn@1688: // is set during method compilation (see Compile::Process_OopMap_Node()). kvn@1688: bool save_oop_result = chunk->at(0)->scope()->return_oop(); kvn@1688: Handle return_value; kvn@1688: if (save_oop_result) { kvn@1688: // Reallocation may trigger GC. If deoptimization happened on return from kvn@1688: // call which returns oop we need to save it since it is not in oopmap. kvn@1688: oop result = deoptee.saved_oop_result(&map); kvn@1688: assert(result == NULL || result->is_oop(), "must be oop"); kvn@1688: return_value = Handle(thread, result); kvn@1688: assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); kvn@1688: if (TraceDeoptimization) { kvn@1688: tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); kvn@1688: } kvn@1688: } kvn@479: bool reallocated = false; kvn@479: if (objects != NULL) { kvn@479: JRT_BLOCK kvn@479: reallocated = realloc_objects(thread, &deoptee, objects, THREAD); kvn@479: JRT_END duke@435: } kvn@479: if (reallocated) { kvn@479: reassign_fields(&deoptee, &map, objects); duke@435: #ifndef PRODUCT duke@435: if (TraceDeoptimization) { duke@435: ttyLocker ttyl; kvn@479: tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); kvn@479: print_objects(objects); kvn@1688: } kvn@1688: #endif kvn@479: } kvn@1688: if (save_oop_result) { kvn@1688: // Restore result. kvn@1688: deoptee.set_saved_oop_result(&map, return_value()); kvn@479: } kvn@479: } kvn@479: if (EliminateLocks) { kvn@518: #ifndef PRODUCT kvn@518: bool first = true; kvn@518: #endif kvn@479: for (int i = 0; i < chunk->length(); i++) { kvn@518: compiledVFrame* cvf = chunk->at(i); kvn@518: assert (cvf->scope() != NULL,"expect only compiled java frames"); kvn@518: GrowableArray* monitors = cvf->monitors(); kvn@518: if (monitors->is_nonempty()) { kvn@518: relock_objects(monitors, thread); kvn@479: #ifndef PRODUCT kvn@479: if (TraceDeoptimization) { kvn@479: ttyLocker ttyl; kvn@479: for (int j = 0; j < monitors->length(); j++) { kvn@518: MonitorInfo* mi = monitors->at(j); kvn@518: if (mi->eliminated()) { kvn@518: if (first) { kvn@518: first = false; kvn@518: tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); kvn@518: } kvn@518: tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); kvn@479: } duke@435: } duke@435: } kvn@479: #endif duke@435: } duke@435: } duke@435: } duke@435: } duke@435: #endif // COMPILER2 duke@435: // Ensure that no safepoint is taken after pointers have been stored duke@435: // in fields of rematerialized objects. If a safepoint occurs from here on duke@435: // out the java state residing in the vframeArray will be missed. duke@435: No_Safepoint_Verifier no_safepoint; duke@435: duke@435: vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); duke@435: duke@435: assert(thread->vframe_array_head() == NULL, "Pending deopt!");; duke@435: thread->set_vframe_array_head(array); duke@435: duke@435: // Now that the vframeArray has been created if we have any deferred local writes duke@435: // added by jvmti then we can free up that structure as the data is now in the duke@435: // vframeArray duke@435: duke@435: if (thread->deferred_locals() != NULL) { duke@435: GrowableArray* list = thread->deferred_locals(); duke@435: int i = 0; duke@435: do { duke@435: // Because of inlining we could have multiple vframes for a single frame duke@435: // and several of the vframes could have deferred writes. Find them all. duke@435: if (list->at(i)->id() == array->original().id()) { duke@435: jvmtiDeferredLocalVariableSet* dlv = list->at(i); duke@435: list->remove_at(i); duke@435: // individual jvmtiDeferredLocalVariableSet are CHeapObj's duke@435: delete dlv; duke@435: } else { duke@435: i++; duke@435: } duke@435: } while ( i < list->length() ); duke@435: if (list->length() == 0) { duke@435: thread->set_deferred_locals(NULL); duke@435: // free the list and elements back to C heap. duke@435: delete list; duke@435: } duke@435: duke@435: } duke@435: twisti@2047: #ifndef SHARK duke@435: // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. duke@435: CodeBlob* cb = stub_frame.cb(); duke@435: // Verify we have the right vframeArray duke@435: assert(cb->frame_size() >= 0, "Unexpected frame size"); duke@435: intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); duke@435: twisti@1639: // If the deopt call site is a MethodHandle invoke call site we have twisti@1639: // to adjust the unpack_sp. twisti@1639: nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); twisti@1639: if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) twisti@1639: unpack_sp = deoptee.unextended_sp(); twisti@1639: duke@435: #ifdef ASSERT duke@435: assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); duke@435: #endif twisti@2047: #else twisti@2047: intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); twisti@2047: #endif // !SHARK twisti@2047: duke@435: // This is a guarantee instead of an assert because if vframe doesn't match duke@435: // we will unpack the wrong deoptimized frame and wind up in strange places duke@435: // where it will be very difficult to figure out what went wrong. Better duke@435: // to die an early death here than some very obscure death later when the duke@435: // trail is cold. duke@435: // Note: on ia64 this guarantee can be fooled by frames with no memory stack duke@435: // in that it will fail to detect a problem when there is one. This needs duke@435: // more work in tiger timeframe. duke@435: guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); duke@435: duke@435: int number_of_frames = array->frames(); duke@435: duke@435: // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost duke@435: // virtual activation, which is the reverse of the elements in the vframes array. zgu@3900: intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); duke@435: // +1 because we always have an interpreter return address for the final slot. zgu@3900: address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); duke@435: int popframe_extra_args = 0; duke@435: // Create an interpreter return address for the stub to use as its return duke@435: // address so the skeletal frames are perfectly walkable duke@435: frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); duke@435: duke@435: // PopFrame requires that the preserved incoming arguments from the recently-popped topmost duke@435: // activation be put back on the expression stack of the caller for reexecution duke@435: if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { duke@435: popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); duke@435: } duke@435: never@2901: // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized never@2901: // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather never@2901: // than simply use array->sender.pc(). This requires us to walk the current set of frames never@2901: // never@2901: frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame never@2901: deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller never@2901: never@2901: // It's possible that the number of paramters at the call site is never@2901: // different than number of arguments in the callee when method never@2901: // handles are used. If the caller is interpreted get the real never@2901: // value so that the proper amount of space can be added to it's never@2901: // frame. twisti@3238: bool caller_was_method_handle = false; never@2901: if (deopt_sender.is_interpreted_frame()) { never@2901: methodHandle method = deopt_sender.interpreter_frame_method(); twisti@3251: Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); twisti@3251: if (cur.is_method_handle_invoke()) { twisti@3238: // Method handle invokes may involve fairly arbitrary chains of twisti@3238: // calls so it's impossible to know how much actual space the twisti@3238: // caller has for locals. twisti@3238: caller_was_method_handle = true; twisti@3238: } never@2901: } never@2901: duke@435: // duke@435: // frame_sizes/frame_pcs[0] oldest frame (int or c2i) duke@435: // frame_sizes/frame_pcs[1] next oldest frame (int) duke@435: // frame_sizes/frame_pcs[n] youngest frame (int) duke@435: // duke@435: // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame duke@435: // owns the space for the return address to it's caller). Confusing ain't it. duke@435: // duke@435: // The vframe array can address vframes with indices running from duke@435: // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. duke@435: // When we create the skeletal frames we need the oldest frame to be in the zero slot duke@435: // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. duke@435: // so things look a little strange in this loop. duke@435: // twisti@3238: int callee_parameters = 0; twisti@3238: int callee_locals = 0; duke@435: for (int index = 0; index < array->frames(); index++ ) { duke@435: // frame[number_of_frames - 1 ] = on_stack_size(youngest) duke@435: // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) duke@435: // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) never@2901: int caller_parms = callee_parameters; twisti@3238: if ((index == array->frames() - 1) && caller_was_method_handle) { twisti@3238: caller_parms = 0; never@2901: } never@2901: frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms, never@2901: callee_parameters, duke@435: callee_locals, duke@435: index == 0, duke@435: popframe_extra_args); duke@435: // This pc doesn't have to be perfect just good enough to identify the frame duke@435: // as interpreted so the skeleton frame will be walkable duke@435: // The correct pc will be set when the skeleton frame is completely filled out duke@435: // The final pc we store in the loop is wrong and will be overwritten below duke@435: frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; duke@435: duke@435: callee_parameters = array->element(index)->method()->size_of_parameters(); duke@435: callee_locals = array->element(index)->method()->max_locals(); duke@435: popframe_extra_args = 0; duke@435: } duke@435: duke@435: // Compute whether the root vframe returns a float or double value. duke@435: BasicType return_type; duke@435: { duke@435: HandleMark hm; duke@435: methodHandle method(thread, array->element(0)->method()); never@2462: Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); coleenp@2497: return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; duke@435: } duke@435: duke@435: // Compute information for handling adapters and adjusting the frame size of the caller. duke@435: int caller_adjustment = 0; duke@435: duke@435: // Compute the amount the oldest interpreter frame will have to adjust duke@435: // its caller's stack by. If the caller is a compiled frame then duke@435: // we pretend that the callee has no parameters so that the duke@435: // extension counts for the full amount of locals and not just duke@435: // locals-parms. This is because without a c2i adapter the parm duke@435: // area as created by the compiled frame will not be usable by duke@435: // the interpreter. (Depending on the calling convention there duke@435: // may not even be enough space). duke@435: duke@435: // QQQ I'd rather see this pushed down into last_frame_adjust duke@435: // and have it take the sender (aka caller). duke@435: twisti@3238: if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { duke@435: caller_adjustment = last_frame_adjust(0, callee_locals); twisti@3238: } else if (callee_locals > callee_parameters) { duke@435: // The caller frame may need extending to accommodate duke@435: // non-parameter locals of the first unpacked interpreted frame. duke@435: // Compute that adjustment. twisti@3238: caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); duke@435: } duke@435: duke@435: // If the sender is deoptimized the we must retrieve the address of the handler duke@435: // since the frame will "magically" show the original pc before the deopt duke@435: // and we'd undo the deopt. duke@435: duke@435: frame_pcs[0] = deopt_sender.raw_pc(); duke@435: twisti@2047: #ifndef SHARK duke@435: assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); twisti@2047: #endif // SHARK duke@435: duke@435: UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, duke@435: caller_adjustment * BytesPerWord, twisti@3238: caller_was_method_handle ? 0 : callee_parameters, duke@435: number_of_frames, duke@435: frame_sizes, duke@435: frame_pcs, duke@435: return_type); bdelsart@3130: // On some platforms, we need a way to pass some platform dependent bdelsart@3130: // information to the unpacking code so the skeletal frames come out bdelsart@3130: // correct (initial fp value, unextended sp, ...) bdelsart@3130: info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); duke@435: duke@435: if (array->frames() > 1) { duke@435: if (VerifyStack && TraceDeoptimization) { duke@435: tty->print_cr("Deoptimizing method containing inlining"); duke@435: } duke@435: } duke@435: duke@435: array->set_unroll_block(info); duke@435: return info; duke@435: } duke@435: duke@435: // Called to cleanup deoptimization data structures in normal case duke@435: // after unpacking to stack and when stack overflow error occurs duke@435: void Deoptimization::cleanup_deopt_info(JavaThread *thread, duke@435: vframeArray *array) { duke@435: duke@435: // Get array if coming from exception duke@435: if (array == NULL) { duke@435: array = thread->vframe_array_head(); duke@435: } duke@435: thread->set_vframe_array_head(NULL); duke@435: duke@435: // Free the previous UnrollBlock duke@435: vframeArray* old_array = thread->vframe_array_last(); duke@435: thread->set_vframe_array_last(array); duke@435: duke@435: if (old_array != NULL) { duke@435: UnrollBlock* old_info = old_array->unroll_block(); duke@435: old_array->set_unroll_block(NULL); duke@435: delete old_info; duke@435: delete old_array; duke@435: } duke@435: duke@435: // Deallocate any resource creating in this routine and any ResourceObjs allocated duke@435: // inside the vframeArray (StackValueCollections) duke@435: duke@435: delete thread->deopt_mark(); duke@435: thread->set_deopt_mark(NULL); iveresov@2169: thread->set_deopt_nmethod(NULL); duke@435: duke@435: duke@435: if (JvmtiExport::can_pop_frame()) { duke@435: #ifndef CC_INTERP duke@435: // Regardless of whether we entered this routine with the pending duke@435: // popframe condition bit set, we should always clear it now duke@435: thread->clear_popframe_condition(); duke@435: #else duke@435: // C++ interpeter will clear has_pending_popframe when it enters duke@435: // with method_resume. For deopt_resume2 we clear it now. duke@435: if (thread->popframe_forcing_deopt_reexecution()) duke@435: thread->clear_popframe_condition(); duke@435: #endif /* CC_INTERP */ duke@435: } duke@435: duke@435: // unpack_frames() is called at the end of the deoptimization handler duke@435: // and (in C2) at the end of the uncommon trap handler. Note this fact duke@435: // so that an asynchronous stack walker can work again. This counter is duke@435: // incremented at the beginning of fetch_unroll_info() and (in C2) at duke@435: // the beginning of uncommon_trap(). duke@435: thread->dec_in_deopt_handler(); duke@435: } duke@435: duke@435: duke@435: // Return BasicType of value being returned duke@435: JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) duke@435: duke@435: // We are already active int he special DeoptResourceMark any ResourceObj's we duke@435: // allocate will be freed at the end of the routine. duke@435: duke@435: // It is actually ok to allocate handles in a leaf method. It causes no safepoints, duke@435: // but makes the entry a little slower. There is however a little dance we have to duke@435: // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro duke@435: ResetNoHandleMark rnhm; // No-op in release/product versions duke@435: HandleMark hm; duke@435: duke@435: frame stub_frame = thread->last_frame(); duke@435: duke@435: // Since the frame to unpack is the top frame of this thread, the vframe_array_head duke@435: // must point to the vframeArray for the unpack frame. duke@435: vframeArray* array = thread->vframe_array_head(); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceDeoptimization) { duke@435: tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); duke@435: } duke@435: #endif never@3499: Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", never@3499: stub_frame.pc(), stub_frame.sp(), exec_mode); duke@435: duke@435: UnrollBlock* info = array->unroll_block(); duke@435: duke@435: // Unpack the interpreter frames and any adapter frame (c2 only) we might create. never@2901: array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); duke@435: duke@435: BasicType bt = info->return_type(); duke@435: duke@435: // If we have an exception pending, claim that the return type is an oop duke@435: // so the deopt_blob does not overwrite the exception_oop. duke@435: duke@435: if (exec_mode == Unpack_exception) duke@435: bt = T_OBJECT; duke@435: duke@435: // Cleanup thread deopt data duke@435: cleanup_deopt_info(thread, array); duke@435: duke@435: #ifndef PRODUCT duke@435: if (VerifyStack) { duke@435: ResourceMark res_mark; duke@435: never@2868: thread->validate_frame_layout(); never@2868: duke@435: // Verify that the just-unpacked frames match the interpreter's duke@435: // notions of expression stack and locals duke@435: vframeArray* cur_array = thread->vframe_array_last(); duke@435: RegisterMap rm(thread, false); duke@435: rm.set_include_argument_oops(false); duke@435: bool is_top_frame = true; duke@435: int callee_size_of_parameters = 0; duke@435: int callee_max_locals = 0; duke@435: for (int i = 0; i < cur_array->frames(); i++) { duke@435: vframeArrayElement* el = cur_array->element(i); duke@435: frame* iframe = el->iframe(); duke@435: guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); duke@435: duke@435: // Get the oop map for this bci duke@435: InterpreterOopMap mask; duke@435: int cur_invoke_parameter_size = 0; duke@435: bool try_next_mask = false; duke@435: int next_mask_expression_stack_size = -1; duke@435: int top_frame_expression_stack_adjustment = 0; duke@435: methodHandle mh(thread, iframe->interpreter_frame_method()); duke@435: OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); duke@435: BytecodeStream str(mh); duke@435: str.set_start(iframe->interpreter_frame_bci()); duke@435: int max_bci = mh->code_size(); duke@435: // Get to the next bytecode if possible duke@435: assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); duke@435: // Check to see if we can grab the number of outgoing arguments duke@435: // at an uncommon trap for an invoke (where the compiler duke@435: // generates debug info before the invoke has executed) duke@435: Bytecodes::Code cur_code = str.next(); duke@435: if (cur_code == Bytecodes::_invokevirtual || duke@435: cur_code == Bytecodes::_invokespecial || duke@435: cur_code == Bytecodes::_invokestatic || duke@435: cur_code == Bytecodes::_invokeinterface) { never@2462: Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); coleenp@2497: Symbol* signature = invoke.signature(); duke@435: ArgumentSizeComputer asc(signature); duke@435: cur_invoke_parameter_size = asc.size(); duke@435: if (cur_code != Bytecodes::_invokestatic) { duke@435: // Add in receiver duke@435: ++cur_invoke_parameter_size; duke@435: } duke@435: } duke@435: if (str.bci() < max_bci) { duke@435: Bytecodes::Code bc = str.next(); duke@435: if (bc >= 0) { duke@435: // The interpreter oop map generator reports results before duke@435: // the current bytecode has executed except in the case of duke@435: // calls. It seems to be hard to tell whether the compiler duke@435: // has emitted debug information matching the "state before" duke@435: // a given bytecode or the state after, so we try both duke@435: switch (cur_code) { duke@435: case Bytecodes::_invokevirtual: duke@435: case Bytecodes::_invokespecial: duke@435: case Bytecodes::_invokestatic: duke@435: case Bytecodes::_invokeinterface: duke@435: case Bytecodes::_athrow: duke@435: break; duke@435: default: { duke@435: InterpreterOopMap next_mask; duke@435: OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); duke@435: next_mask_expression_stack_size = next_mask.expression_stack_size(); duke@435: // Need to subtract off the size of the result type of duke@435: // the bytecode because this is not described in the duke@435: // debug info but returned to the interpreter in the TOS duke@435: // caching register duke@435: BasicType bytecode_result_type = Bytecodes::result_type(cur_code); duke@435: if (bytecode_result_type != T_ILLEGAL) { duke@435: top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; duke@435: } duke@435: assert(top_frame_expression_stack_adjustment >= 0, ""); duke@435: try_next_mask = true; duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Verify stack depth and oops in frame duke@435: // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) duke@435: if (!( duke@435: /* SPARC */ duke@435: (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || duke@435: /* x86 */ duke@435: (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || duke@435: (try_next_mask && duke@435: (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - duke@435: top_frame_expression_stack_adjustment))) || duke@435: (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || duke@435: (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && duke@435: (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) duke@435: )) { duke@435: ttyLocker ttyl; duke@435: duke@435: // Print out some information that will help us debug the problem duke@435: tty->print_cr("Wrong number of expression stack elements during deoptimization"); duke@435: tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); duke@435: tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", duke@435: iframe->interpreter_frame_expression_stack_size()); duke@435: tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); duke@435: tty->print_cr(" try_next_mask = %d", try_next_mask); duke@435: tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); duke@435: tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); duke@435: tty->print_cr(" callee_max_locals = %d", callee_max_locals); duke@435: tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); duke@435: tty->print_cr(" exec_mode = %d", exec_mode); duke@435: tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); duke@435: tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); duke@435: tty->print_cr(" Interpreted frames:"); duke@435: for (int k = 0; k < cur_array->frames(); k++) { duke@435: vframeArrayElement* el = cur_array->element(k); duke@435: tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); duke@435: } duke@435: cur_array->print_on_2(tty); duke@435: guarantee(false, "wrong number of expression stack elements during deopt"); duke@435: } duke@435: VerifyOopClosure verify; duke@435: iframe->oops_interpreted_do(&verify, &rm, false); duke@435: callee_size_of_parameters = mh->size_of_parameters(); duke@435: callee_max_locals = mh->max_locals(); duke@435: is_top_frame = false; duke@435: } duke@435: } duke@435: #endif /* !PRODUCT */ duke@435: duke@435: duke@435: return bt; duke@435: JRT_END duke@435: duke@435: duke@435: int Deoptimization::deoptimize_dependents() { duke@435: Threads::deoptimized_wrt_marked_nmethods(); duke@435: return 0; duke@435: } duke@435: duke@435: duke@435: #ifdef COMPILER2 duke@435: bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray* objects, TRAPS) { duke@435: Handle pending_exception(thread->pending_exception()); duke@435: const char* exception_file = thread->exception_file(); duke@435: int exception_line = thread->exception_line(); duke@435: thread->clear_pending_exception(); duke@435: duke@435: for (int i = 0; i < objects->length(); i++) { duke@435: assert(objects->at(i)->is_object(), "invalid debug information"); duke@435: ObjectValue* sv = (ObjectValue*) objects->at(i); duke@435: duke@435: KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); duke@435: oop obj = NULL; duke@435: duke@435: if (k->oop_is_instance()) { duke@435: instanceKlass* ik = instanceKlass::cast(k()); duke@435: obj = ik->allocate_instance(CHECK_(false)); duke@435: } else if (k->oop_is_typeArray()) { duke@435: typeArrayKlass* ak = typeArrayKlass::cast(k()); duke@435: assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); duke@435: int len = sv->field_size() / type2size[ak->element_type()]; duke@435: obj = ak->allocate(len, CHECK_(false)); duke@435: } else if (k->oop_is_objArray()) { duke@435: objArrayKlass* ak = objArrayKlass::cast(k()); duke@435: obj = ak->allocate(sv->field_size(), CHECK_(false)); duke@435: } duke@435: duke@435: assert(obj != NULL, "allocation failed"); duke@435: assert(sv->value().is_null(), "redundant reallocation"); duke@435: sv->set_value(obj); duke@435: } duke@435: duke@435: if (pending_exception.not_null()) { duke@435: thread->set_pending_exception(pending_exception(), exception_file, exception_line); duke@435: } duke@435: duke@435: return true; duke@435: } duke@435: duke@435: // This assumes that the fields are stored in ObjectValue in the same order duke@435: // they are yielded by do_nonstatic_fields. duke@435: class FieldReassigner: public FieldClosure { duke@435: frame* _fr; duke@435: RegisterMap* _reg_map; duke@435: ObjectValue* _sv; duke@435: instanceKlass* _ik; duke@435: oop _obj; duke@435: duke@435: int _i; duke@435: public: duke@435: FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : duke@435: _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} duke@435: duke@435: int i() const { return _i; } duke@435: duke@435: duke@435: void do_field(fieldDescriptor* fd) { kvn@479: intptr_t val; duke@435: StackValue* value = duke@435: StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); duke@435: int offset = fd->offset(); duke@435: switch (fd->field_type()) { duke@435: case T_OBJECT: case T_ARRAY: duke@435: assert(value->type() == T_OBJECT, "Agreement."); duke@435: _obj->obj_field_put(offset, value->get_obj()()); duke@435: break; duke@435: duke@435: case T_LONG: case T_DOUBLE: { duke@435: assert(value->type() == T_INT, "Agreement."); duke@435: StackValue* low = duke@435: StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); kvn@479: #ifdef _LP64 kvn@479: jlong res = (jlong)low->get_int(); kvn@479: #else kvn@479: #ifdef SPARC kvn@479: // For SPARC we have to swap high and low words. kvn@479: jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); kvn@479: #else duke@435: jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); kvn@479: #endif //SPARC kvn@479: #endif duke@435: _obj->long_field_put(offset, res); duke@435: break; duke@435: } kvn@479: // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. duke@435: case T_INT: case T_FLOAT: // 4 bytes. duke@435: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: _obj->int_field_put(offset, (jint)*((jint*)&val)); duke@435: break; duke@435: duke@435: case T_SHORT: case T_CHAR: // 2 bytes duke@435: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: _obj->short_field_put(offset, (jshort)*((jint*)&val)); duke@435: break; duke@435: kvn@479: case T_BOOLEAN: case T_BYTE: // 1 byte duke@435: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); duke@435: break; duke@435: duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: _i++; duke@435: } duke@435: }; duke@435: duke@435: // restore elements of an eliminated type array duke@435: void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { duke@435: int index = 0; kvn@479: intptr_t val; duke@435: duke@435: for (int i = 0; i < sv->field_size(); i++) { duke@435: StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); duke@435: switch(type) { kvn@479: case T_LONG: case T_DOUBLE: { kvn@479: assert(value->type() == T_INT, "Agreement."); kvn@479: StackValue* low = kvn@479: StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); kvn@479: #ifdef _LP64 kvn@479: jlong res = (jlong)low->get_int(); kvn@479: #else kvn@479: #ifdef SPARC kvn@479: // For SPARC we have to swap high and low words. kvn@479: jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); kvn@479: #else kvn@479: jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); kvn@479: #endif //SPARC kvn@479: #endif kvn@479: obj->long_at_put(index, res); kvn@479: break; kvn@479: } kvn@479: kvn@479: // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. kvn@479: case T_INT: case T_FLOAT: // 4 bytes. kvn@479: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: obj->int_at_put(index, (jint)*((jint*)&val)); kvn@479: break; kvn@479: kvn@479: case T_SHORT: case T_CHAR: // 2 bytes kvn@479: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: obj->short_at_put(index, (jshort)*((jint*)&val)); kvn@479: break; kvn@479: kvn@479: case T_BOOLEAN: case T_BYTE: // 1 byte kvn@479: assert(value->type() == T_INT, "Agreement."); kvn@479: val = value->get_int(); kvn@479: obj->bool_at_put(index, (jboolean)*((jint*)&val)); kvn@479: break; kvn@479: duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: index++; duke@435: } duke@435: } duke@435: duke@435: duke@435: // restore fields of an eliminated object array duke@435: void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { duke@435: for (int i = 0; i < sv->field_size(); i++) { duke@435: StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); duke@435: assert(value->type() == T_OBJECT, "object element expected"); duke@435: obj->obj_at_put(i, value->get_obj()()); duke@435: } duke@435: } duke@435: duke@435: duke@435: // restore fields of all eliminated objects and arrays duke@435: void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray* objects) { duke@435: for (int i = 0; i < objects->length(); i++) { duke@435: ObjectValue* sv = (ObjectValue*) objects->at(i); duke@435: KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); duke@435: Handle obj = sv->value(); duke@435: assert(obj.not_null(), "reallocation was missed"); duke@435: duke@435: if (k->oop_is_instance()) { duke@435: instanceKlass* ik = instanceKlass::cast(k()); duke@435: FieldReassigner reassign(fr, reg_map, sv, obj()); duke@435: ik->do_nonstatic_fields(&reassign); duke@435: } else if (k->oop_is_typeArray()) { duke@435: typeArrayKlass* ak = typeArrayKlass::cast(k()); duke@435: reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); duke@435: } else if (k->oop_is_objArray()) { duke@435: reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: // relock objects for which synchronization was eliminated kvn@518: void Deoptimization::relock_objects(GrowableArray* monitors, JavaThread* thread) { duke@435: for (int i = 0; i < monitors->length(); i++) { kvn@518: MonitorInfo* mon_info = monitors->at(i); kvn@518: if (mon_info->eliminated()) { kvn@518: assert(mon_info->owner() != NULL, "reallocation was missed"); kvn@518: Handle obj = Handle(mon_info->owner()); kvn@518: markOop mark = obj->mark(); kvn@518: if (UseBiasedLocking && mark->has_bias_pattern()) { kvn@518: // New allocated objects may have the mark set to anonymously biased. kvn@518: // Also the deoptimized method may called methods with synchronization kvn@518: // where the thread-local object is bias locked to the current thread. kvn@518: assert(mark->is_biased_anonymously() || kvn@518: mark->biased_locker() == thread, "should be locked to current thread"); kvn@518: // Reset mark word to unbiased prototype. kvn@518: markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); kvn@518: obj->set_mark(unbiased_prototype); kvn@518: } kvn@518: BasicLock* lock = mon_info->lock(); kvn@518: ObjectSynchronizer::slow_enter(obj, lock, thread); duke@435: } kvn@518: assert(mon_info->owner()->is_locked(), "object must be locked now"); duke@435: } duke@435: } duke@435: duke@435: duke@435: #ifndef PRODUCT duke@435: // print information about reallocated objects duke@435: void Deoptimization::print_objects(GrowableArray* objects) { duke@435: fieldDescriptor fd; duke@435: duke@435: for (int i = 0; i < objects->length(); i++) { duke@435: ObjectValue* sv = (ObjectValue*) objects->at(i); duke@435: KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); duke@435: Handle obj = sv->value(); duke@435: duke@435: tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); duke@435: k->as_klassOop()->print_value(); duke@435: tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); duke@435: tty->cr(); duke@435: duke@435: if (Verbose) { duke@435: k->oop_print_on(obj(), tty); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: #endif // COMPILER2 duke@435: duke@435: vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray* chunk) { never@3499: Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceDeoptimization) { duke@435: ttyLocker ttyl; duke@435: tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); duke@435: fr.print_on(tty); duke@435: tty->print_cr(" Virtual frames (innermost first):"); duke@435: for (int index = 0; index < chunk->length(); index++) { duke@435: compiledVFrame* vf = chunk->at(index); duke@435: tty->print(" %2d - ", index); duke@435: vf->print_value(); duke@435: int bci = chunk->at(index)->raw_bci(); duke@435: const char* code_name; duke@435: if (bci == SynchronizationEntryBCI) { duke@435: code_name = "sync entry"; duke@435: } else { never@2462: Bytecodes::Code code = vf->method()->code_at(bci); duke@435: code_name = Bytecodes::name(code); duke@435: } duke@435: tty->print(" - %s", code_name); duke@435: tty->print_cr(" @ bci %d ", bci); duke@435: if (Verbose) { duke@435: vf->print(); duke@435: tty->cr(); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: // Register map for next frame (used for stack crawl). We capture duke@435: // the state of the deopt'ing frame's caller. Thus if we need to duke@435: // stuff a C2I adapter we can properly fill in the callee-save duke@435: // register locations. duke@435: frame caller = fr.sender(reg_map); duke@435: int frame_size = caller.sp() - fr.sp(); duke@435: duke@435: frame sender = caller; duke@435: duke@435: // Since the Java thread being deoptimized will eventually adjust it's own stack, duke@435: // the vframeArray containing the unpacking information is allocated in the C heap. duke@435: // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). duke@435: vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); duke@435: duke@435: // Compare the vframeArray to the collected vframes duke@435: assert(array->structural_compare(thread, chunk), "just checking"); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceDeoptimization) { duke@435: ttyLocker ttyl; duke@435: tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); duke@435: } duke@435: #endif // PRODUCT duke@435: duke@435: return array; duke@435: } duke@435: duke@435: duke@435: static void collect_monitors(compiledVFrame* cvf, GrowableArray* objects_to_revoke) { duke@435: GrowableArray* monitors = cvf->monitors(); duke@435: for (int i = 0; i < monitors->length(); i++) { duke@435: MonitorInfo* mon_info = monitors->at(i); kvn@1253: if (!mon_info->eliminated() && mon_info->owner() != NULL) { duke@435: objects_to_revoke->append(Handle(mon_info->owner())); duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { duke@435: if (!UseBiasedLocking) { duke@435: return; duke@435: } duke@435: duke@435: GrowableArray* objects_to_revoke = new GrowableArray(); duke@435: duke@435: // Unfortunately we don't have a RegisterMap available in most of duke@435: // the places we want to call this routine so we need to walk the duke@435: // stack again to update the register map. duke@435: if (map == NULL || !map->update_map()) { duke@435: StackFrameStream sfs(thread, true); duke@435: bool found = false; duke@435: while (!found && !sfs.is_done()) { duke@435: frame* cur = sfs.current(); duke@435: sfs.next(); duke@435: found = cur->id() == fr.id(); duke@435: } duke@435: assert(found, "frame to be deoptimized not found on target thread's stack"); duke@435: map = sfs.register_map(); duke@435: } duke@435: duke@435: vframe* vf = vframe::new_vframe(&fr, map, thread); duke@435: compiledVFrame* cvf = compiledVFrame::cast(vf); duke@435: // Revoke monitors' biases in all scopes duke@435: while (!cvf->is_top()) { duke@435: collect_monitors(cvf, objects_to_revoke); duke@435: cvf = compiledVFrame::cast(cvf->sender()); duke@435: } duke@435: collect_monitors(cvf, objects_to_revoke); duke@435: duke@435: if (SafepointSynchronize::is_at_safepoint()) { duke@435: BiasedLocking::revoke_at_safepoint(objects_to_revoke); duke@435: } else { duke@435: BiasedLocking::revoke(objects_to_revoke); duke@435: } duke@435: } duke@435: duke@435: duke@435: void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { duke@435: if (!UseBiasedLocking) { duke@435: return; duke@435: } duke@435: duke@435: assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); duke@435: GrowableArray* objects_to_revoke = new GrowableArray(); duke@435: for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { duke@435: if (jt->has_last_Java_frame()) { duke@435: StackFrameStream sfs(jt, true); duke@435: while (!sfs.is_done()) { duke@435: frame* cur = sfs.current(); duke@435: if (cb->contains(cur->pc())) { duke@435: vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); duke@435: compiledVFrame* cvf = compiledVFrame::cast(vf); duke@435: // Revoke monitors' biases in all scopes duke@435: while (!cvf->is_top()) { duke@435: collect_monitors(cvf, objects_to_revoke); duke@435: cvf = compiledVFrame::cast(cvf->sender()); duke@435: } duke@435: collect_monitors(cvf, objects_to_revoke); duke@435: } duke@435: sfs.next(); duke@435: } duke@435: } duke@435: } duke@435: BiasedLocking::revoke_at_safepoint(objects_to_revoke); duke@435: } duke@435: duke@435: duke@435: void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { duke@435: assert(fr.can_be_deoptimized(), "checking frame type"); duke@435: duke@435: gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); duke@435: duke@435: // Patch the nmethod so that when execution returns to it we will duke@435: // deopt the execution state and return to the interpreter. duke@435: fr.deoptimize(thread); duke@435: } duke@435: duke@435: void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { duke@435: // Deoptimize only if the frame comes from compile code. duke@435: // Do not deoptimize the frame which is already patched duke@435: // during the execution of the loops below. duke@435: if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { duke@435: return; duke@435: } duke@435: ResourceMark rm; duke@435: DeoptimizationMarker dm; duke@435: if (UseBiasedLocking) { duke@435: revoke_biases_of_monitors(thread, fr, map); duke@435: } duke@435: deoptimize_single_frame(thread, fr); duke@435: duke@435: } duke@435: duke@435: never@2260: void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { never@2260: assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), never@2260: "can only deoptimize other thread at a safepoint"); duke@435: // Compute frame and register map based on thread and sp. duke@435: RegisterMap reg_map(thread, UseBiasedLocking); duke@435: frame fr = thread->last_frame(); duke@435: while (fr.id() != id) { duke@435: fr = fr.sender(®_map); duke@435: } duke@435: deoptimize(thread, fr, ®_map); duke@435: } duke@435: duke@435: never@2260: void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { never@2260: if (thread == Thread::current()) { never@2260: Deoptimization::deoptimize_frame_internal(thread, id); never@2260: } else { never@2260: VM_DeoptimizeFrame deopt(thread, id); never@2260: VMThread::execute(&deopt); never@2260: } never@2260: } never@2260: never@2260: duke@435: // JVMTI PopFrame support duke@435: JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) duke@435: { duke@435: thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); duke@435: } duke@435: JRT_END duke@435: duke@435: twisti@2047: #if defined(COMPILER2) || defined(SHARK) duke@435: void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { duke@435: // in case of an unresolved klass entry, load the class. duke@435: if (constant_pool->tag_at(index).is_unresolved_klass()) { duke@435: klassOop tk = constant_pool->klass_at(index, CHECK); duke@435: return; duke@435: } duke@435: duke@435: if (!constant_pool->tag_at(index).is_symbol()) return; duke@435: duke@435: Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); coleenp@2497: Symbol* symbol = constant_pool->symbol_at(index); duke@435: duke@435: // class name? duke@435: if (symbol->byte_at(0) != '(') { duke@435: Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); duke@435: SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); duke@435: return; duke@435: } duke@435: duke@435: // then it must be a signature! coleenp@2497: ResourceMark rm(THREAD); duke@435: for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { duke@435: if (ss.is_object()) { coleenp@2497: Symbol* class_name = ss.as_symbol(CHECK); duke@435: Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); duke@435: SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { duke@435: EXCEPTION_MARK; duke@435: load_class_by_index(constant_pool, index, THREAD); duke@435: if (HAS_PENDING_EXCEPTION) { duke@435: // Exception happened during classloading. We ignore the exception here, since it duke@435: // is going to be rethrown since the current activation is going to be deoptimzied and duke@435: // the interpreter will re-execute the bytecode. duke@435: CLEAR_PENDING_EXCEPTION; duke@435: } duke@435: } duke@435: duke@435: JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { duke@435: HandleMark hm; duke@435: duke@435: // uncommon_trap() is called at the beginning of the uncommon trap duke@435: // handler. Note this fact before we start generating temporary frames duke@435: // that can confuse an asynchronous stack walker. This counter is duke@435: // decremented at the end of unpack_frames(). duke@435: thread->inc_in_deopt_handler(); duke@435: duke@435: // We need to update the map if we have biased locking. duke@435: RegisterMap reg_map(thread, UseBiasedLocking); duke@435: frame stub_frame = thread->last_frame(); duke@435: frame fr = stub_frame.sender(®_map); duke@435: // Make sure the calling nmethod is not getting deoptimized and removed duke@435: // before we are done with it. duke@435: nmethodLocker nl(fr.pc()); duke@435: never@3499: // Log a message never@3499: Events::log_deopt_message(thread, "Uncommon trap %d fr.pc " INTPTR_FORMAT, never@3499: trap_request, fr.pc()); never@3499: duke@435: { duke@435: ResourceMark rm; duke@435: duke@435: // Revoke biases of any monitors in the frame to ensure we can migrate them duke@435: revoke_biases_of_monitors(thread, fr, ®_map); duke@435: duke@435: DeoptReason reason = trap_request_reason(trap_request); duke@435: DeoptAction action = trap_request_action(trap_request); duke@435: jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 duke@435: duke@435: vframe* vf = vframe::new_vframe(&fr, ®_map, thread); duke@435: compiledVFrame* cvf = compiledVFrame::cast(vf); duke@435: duke@435: nmethod* nm = cvf->code(); duke@435: duke@435: ScopeDesc* trap_scope = cvf->scope(); duke@435: methodHandle trap_method = trap_scope->method(); duke@435: int trap_bci = trap_scope->bci(); never@2462: Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); duke@435: duke@435: // Record this event in the histogram. duke@435: gather_statistics(reason, action, trap_bc); duke@435: duke@435: // Ensure that we can record deopt. history: duke@435: bool create_if_missing = ProfileTraps; duke@435: duke@435: methodDataHandle trap_mdo duke@435: (THREAD, get_method_data(thread, trap_method, create_if_missing)); duke@435: duke@435: // Print a bunch of diagnostics, if requested. duke@435: if (TraceDeoptimization || LogCompilation) { duke@435: ResourceMark rm; duke@435: ttyLocker ttyl; duke@435: char buf[100]; duke@435: if (xtty != NULL) { duke@435: xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", duke@435: os::current_thread_id(), duke@435: format_trap_request(buf, sizeof(buf), trap_request)); duke@435: nm->log_identity(xtty); duke@435: } coleenp@2497: Symbol* class_name = NULL; duke@435: bool unresolved = false; duke@435: if (unloaded_class_index >= 0) { duke@435: constantPoolHandle constants (THREAD, trap_method->constants()); duke@435: if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { coleenp@2497: class_name = constants->klass_name_at(unloaded_class_index); duke@435: unresolved = true; duke@435: if (xtty != NULL) duke@435: xtty->print(" unresolved='1'"); duke@435: } else if (constants->tag_at(unloaded_class_index).is_symbol()) { coleenp@2497: class_name = constants->symbol_at(unloaded_class_index); duke@435: } duke@435: if (xtty != NULL) duke@435: xtty->name(class_name); duke@435: } duke@435: if (xtty != NULL && trap_mdo.not_null()) { duke@435: // Dump the relevant MDO state. duke@435: // This is the deopt count for the current reason, any previous duke@435: // reasons or recompiles seen at this point. duke@435: int dcnt = trap_mdo->trap_count(reason); duke@435: if (dcnt != 0) duke@435: xtty->print(" count='%d'", dcnt); duke@435: ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); duke@435: int dos = (pdata == NULL)? 0: pdata->trap_state(); duke@435: if (dos != 0) { duke@435: xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); duke@435: if (trap_state_is_recompiled(dos)) { duke@435: int recnt2 = trap_mdo->overflow_recompile_count(); duke@435: if (recnt2 != 0) duke@435: xtty->print(" recompiles2='%d'", recnt2); duke@435: } duke@435: } duke@435: } duke@435: if (xtty != NULL) { duke@435: xtty->stamp(); duke@435: xtty->end_head(); duke@435: } duke@435: if (TraceDeoptimization) { // make noise on the tty duke@435: tty->print("Uncommon trap occurred in"); duke@435: nm->method()->print_short_name(tty); duke@435: tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", duke@435: fr.pc(), duke@435: (int) os::current_thread_id(), duke@435: trap_reason_name(reason), duke@435: trap_action_name(action), duke@435: unloaded_class_index); coleenp@2497: if (class_name != NULL) { duke@435: tty->print(unresolved ? " unresolved class: " : " symbol: "); duke@435: class_name->print_symbol_on(tty); duke@435: } duke@435: tty->cr(); duke@435: } duke@435: if (xtty != NULL) { duke@435: // Log the precise location of the trap. duke@435: for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { duke@435: xtty->begin_elem("jvms bci='%d'", sd->bci()); duke@435: xtty->method(sd->method()); duke@435: xtty->end_elem(); duke@435: if (sd->is_top()) break; duke@435: } duke@435: xtty->tail("uncommon_trap"); duke@435: } duke@435: } duke@435: // (End diagnostic printout.) duke@435: duke@435: // Load class if necessary duke@435: if (unloaded_class_index >= 0) { duke@435: constantPoolHandle constants(THREAD, trap_method->constants()); duke@435: load_class_by_index(constants, unloaded_class_index); duke@435: } duke@435: duke@435: // Flush the nmethod if necessary and desirable. duke@435: // duke@435: // We need to avoid situations where we are re-flushing the nmethod duke@435: // because of a hot deoptimization site. Repeated flushes at the same duke@435: // point need to be detected by the compiler and avoided. If the compiler duke@435: // cannot avoid them (or has a bug and "refuses" to avoid them), this duke@435: // module must take measures to avoid an infinite cycle of recompilation duke@435: // and deoptimization. There are several such measures: duke@435: // duke@435: // 1. If a recompilation is ordered a second time at some site X duke@435: // and for the same reason R, the action is adjusted to 'reinterpret', duke@435: // to give the interpreter time to exercise the method more thoroughly. duke@435: // If this happens, the method's overflow_recompile_count is incremented. duke@435: // duke@435: // 2. If the compiler fails to reduce the deoptimization rate, then duke@435: // the method's overflow_recompile_count will begin to exceed the set duke@435: // limit PerBytecodeRecompilationCutoff. If this happens, the action duke@435: // is adjusted to 'make_not_compilable', and the method is abandoned duke@435: // to the interpreter. This is a performance hit for hot methods, duke@435: // but is better than a disastrous infinite cycle of recompilations. duke@435: // (Actually, only the method containing the site X is abandoned.) duke@435: // duke@435: // 3. In parallel with the previous measures, if the total number of duke@435: // recompilations of a method exceeds the much larger set limit duke@435: // PerMethodRecompilationCutoff, the method is abandoned. duke@435: // This should only happen if the method is very large and has duke@435: // many "lukewarm" deoptimizations. The code which enforces this duke@435: // limit is elsewhere (class nmethod, class methodOopDesc). duke@435: // duke@435: // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance duke@435: // to recompile at each bytecode independently of the per-BCI cutoff. duke@435: // duke@435: // The decision to update code is up to the compiler, and is encoded duke@435: // in the Action_xxx code. If the compiler requests Action_none duke@435: // no trap state is changed, no compiled code is changed, and the duke@435: // computation suffers along in the interpreter. duke@435: // duke@435: // The other action codes specify various tactics for decompilation duke@435: // and recompilation. Action_maybe_recompile is the loosest, and duke@435: // allows the compiled code to stay around until enough traps are seen, duke@435: // and until the compiler gets around to recompiling the trapping method. duke@435: // duke@435: // The other actions cause immediate removal of the present code. duke@435: duke@435: bool update_trap_state = true; duke@435: bool make_not_entrant = false; duke@435: bool make_not_compilable = false; iveresov@2138: bool reprofile = false; duke@435: switch (action) { duke@435: case Action_none: duke@435: // Keep the old code. duke@435: update_trap_state = false; duke@435: break; duke@435: case Action_maybe_recompile: duke@435: // Do not need to invalidate the present code, but we can duke@435: // initiate another duke@435: // Start compiler without (necessarily) invalidating the nmethod. duke@435: // The system will tolerate the old code, but new code should be duke@435: // generated when possible. duke@435: break; duke@435: case Action_reinterpret: duke@435: // Go back into the interpreter for a while, and then consider duke@435: // recompiling form scratch. duke@435: make_not_entrant = true; duke@435: // Reset invocation counter for outer most method. duke@435: // This will allow the interpreter to exercise the bytecodes duke@435: // for a while before recompiling. duke@435: // By contrast, Action_make_not_entrant is immediate. duke@435: // duke@435: // Note that the compiler will track null_check, null_assert, duke@435: // range_check, and class_check events and log them as if they duke@435: // had been traps taken from compiled code. This will update duke@435: // the MDO trap history so that the next compilation will duke@435: // properly detect hot trap sites. iveresov@2138: reprofile = true; duke@435: break; duke@435: case Action_make_not_entrant: duke@435: // Request immediate recompilation, and get rid of the old code. duke@435: // Make them not entrant, so next time they are called they get duke@435: // recompiled. Unloaded classes are loaded now so recompile before next duke@435: // time they are called. Same for uninitialized. The interpreter will duke@435: // link the missing class, if any. duke@435: make_not_entrant = true; duke@435: break; duke@435: case Action_make_not_compilable: duke@435: // Give up on compiling this method at all. duke@435: make_not_entrant = true; duke@435: make_not_compilable = true; duke@435: break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: duke@435: // Setting +ProfileTraps fixes the following, on all platforms: duke@435: // 4852688: ProfileInterpreter is off by default for ia64. The result is duke@435: // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the duke@435: // recompile relies on a methodDataOop to record heroic opt failures. duke@435: duke@435: // Whether the interpreter is producing MDO data or not, we also need duke@435: // to use the MDO to detect hot deoptimization points and control duke@435: // aggressive optimization. kvn@1641: bool inc_recompile_count = false; kvn@1641: ProfileData* pdata = NULL; duke@435: if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { duke@435: assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); duke@435: uint this_trap_count = 0; duke@435: bool maybe_prior_trap = false; duke@435: bool maybe_prior_recompile = false; kvn@1641: pdata = query_update_method_data(trap_mdo, trap_bci, reason, duke@435: //outputs: duke@435: this_trap_count, duke@435: maybe_prior_trap, duke@435: maybe_prior_recompile); duke@435: // Because the interpreter also counts null, div0, range, and class duke@435: // checks, these traps from compiled code are double-counted. duke@435: // This is harmless; it just means that the PerXTrapLimit values duke@435: // are in effect a little smaller than they look. duke@435: duke@435: DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); duke@435: if (per_bc_reason != Reason_none) { duke@435: // Now take action based on the partially known per-BCI history. duke@435: if (maybe_prior_trap duke@435: && this_trap_count >= (uint)PerBytecodeTrapLimit) { duke@435: // If there are too many traps at this BCI, force a recompile. duke@435: // This will allow the compiler to see the limit overflow, and duke@435: // take corrective action, if possible. The compiler generally duke@435: // does not use the exact PerBytecodeTrapLimit value, but instead duke@435: // changes its tactics if it sees any traps at all. This provides duke@435: // a little hysteresis, delaying a recompile until a trap happens duke@435: // several times. duke@435: // duke@435: // Actually, since there is only one bit of counter per BCI, duke@435: // the possible per-BCI counts are {0,1,(per-method count)}. duke@435: // This produces accurate results if in fact there is only duke@435: // one hot trap site, but begins to get fuzzy if there are duke@435: // many sites. For example, if there are ten sites each duke@435: // trapping two or more times, they each get the blame for duke@435: // all of their traps. duke@435: make_not_entrant = true; duke@435: } duke@435: duke@435: // Detect repeated recompilation at the same BCI, and enforce a limit. duke@435: if (make_not_entrant && maybe_prior_recompile) { duke@435: // More than one recompile at this point. kvn@1641: inc_recompile_count = maybe_prior_trap; duke@435: } duke@435: } else { duke@435: // For reasons which are not recorded per-bytecode, we simply duke@435: // force recompiles unconditionally. duke@435: // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) duke@435: make_not_entrant = true; duke@435: } duke@435: duke@435: // Go back to the compiler if there are too many traps in this method. duke@435: if (this_trap_count >= (uint)PerMethodTrapLimit) { duke@435: // If there are too many traps in this method, force a recompile. duke@435: // This will allow the compiler to see the limit overflow, and duke@435: // take corrective action, if possible. duke@435: // (This condition is an unlikely backstop only, because the duke@435: // PerBytecodeTrapLimit is more likely to take effect first, duke@435: // if it is applicable.) duke@435: make_not_entrant = true; duke@435: } duke@435: duke@435: // Here's more hysteresis: If there has been a recompile at duke@435: // this trap point already, run the method in the interpreter duke@435: // for a while to exercise it more thoroughly. duke@435: if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { iveresov@2138: reprofile = true; duke@435: } duke@435: kvn@1641: } kvn@1641: kvn@1641: // Take requested actions on the method: kvn@1641: kvn@1641: // Recompile kvn@1641: if (make_not_entrant) { kvn@1641: if (!nm->make_not_entrant()) { kvn@1641: return; // the call did not change nmethod's state kvn@1641: } kvn@1641: kvn@1641: if (pdata != NULL) { duke@435: // Record the recompilation event, if any. duke@435: int tstate0 = pdata->trap_state(); duke@435: int tstate1 = trap_state_set_recompiled(tstate0, true); duke@435: if (tstate1 != tstate0) duke@435: pdata->set_trap_state(tstate1); duke@435: } duke@435: } duke@435: kvn@1641: if (inc_recompile_count) { kvn@1641: trap_mdo->inc_overflow_recompile_count(); kvn@1641: if ((uint)trap_mdo->overflow_recompile_count() > kvn@1641: (uint)PerBytecodeRecompilationCutoff) { kvn@1641: // Give up on the method containing the bad BCI. kvn@1641: if (trap_method() == nm->method()) { kvn@1641: make_not_compilable = true; kvn@1641: } else { iveresov@2138: trap_method->set_not_compilable(CompLevel_full_optimization); kvn@1641: // But give grace to the enclosing nm->method(). kvn@1641: } kvn@1641: } kvn@1641: } duke@435: iveresov@2138: // Reprofile iveresov@2138: if (reprofile) { iveresov@2138: CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); duke@435: } duke@435: duke@435: // Give up compiling iveresov@2138: if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { duke@435: assert(make_not_entrant, "consistent"); iveresov@2138: nm->method()->set_not_compilable(CompLevel_full_optimization); duke@435: } duke@435: duke@435: } // Free marked resources duke@435: duke@435: } duke@435: JRT_END duke@435: duke@435: methodDataOop duke@435: Deoptimization::get_method_data(JavaThread* thread, methodHandle m, duke@435: bool create_if_missing) { duke@435: Thread* THREAD = thread; duke@435: methodDataOop mdo = m()->method_data(); duke@435: if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { duke@435: // Build an MDO. Ignore errors like OutOfMemory; duke@435: // that simply means we won't have an MDO to update. duke@435: methodOopDesc::build_interpreter_method_data(m, THREAD); duke@435: if (HAS_PENDING_EXCEPTION) { duke@435: assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); duke@435: CLEAR_PENDING_EXCEPTION; duke@435: } duke@435: mdo = m()->method_data(); duke@435: } duke@435: return mdo; duke@435: } duke@435: duke@435: ProfileData* duke@435: Deoptimization::query_update_method_data(methodDataHandle trap_mdo, duke@435: int trap_bci, duke@435: Deoptimization::DeoptReason reason, duke@435: //outputs: duke@435: uint& ret_this_trap_count, duke@435: bool& ret_maybe_prior_trap, duke@435: bool& ret_maybe_prior_recompile) { duke@435: uint prior_trap_count = trap_mdo->trap_count(reason); duke@435: uint this_trap_count = trap_mdo->inc_trap_count(reason); duke@435: duke@435: // If the runtime cannot find a place to store trap history, duke@435: // it is estimated based on the general condition of the method. duke@435: // If the method has ever been recompiled, or has ever incurred duke@435: // a trap with the present reason , then this BCI is assumed duke@435: // (pessimistically) to be the culprit. duke@435: bool maybe_prior_trap = (prior_trap_count != 0); duke@435: bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); duke@435: ProfileData* pdata = NULL; duke@435: duke@435: duke@435: // For reasons which are recorded per bytecode, we check per-BCI data. duke@435: DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); duke@435: if (per_bc_reason != Reason_none) { duke@435: // Find the profile data for this BCI. If there isn't one, duke@435: // try to allocate one from the MDO's set of spares. duke@435: // This will let us detect a repeated trap at this point. duke@435: pdata = trap_mdo->allocate_bci_to_data(trap_bci); duke@435: duke@435: if (pdata != NULL) { duke@435: // Query the trap state of this profile datum. duke@435: int tstate0 = pdata->trap_state(); duke@435: if (!trap_state_has_reason(tstate0, per_bc_reason)) duke@435: maybe_prior_trap = false; duke@435: if (!trap_state_is_recompiled(tstate0)) duke@435: maybe_prior_recompile = false; duke@435: duke@435: // Update the trap state of this profile datum. duke@435: int tstate1 = tstate0; duke@435: // Record the reason. duke@435: tstate1 = trap_state_add_reason(tstate1, per_bc_reason); duke@435: // Store the updated state on the MDO, for next time. duke@435: if (tstate1 != tstate0) duke@435: pdata->set_trap_state(tstate1); duke@435: } else { kvn@1641: if (LogCompilation && xtty != NULL) { kvn@1641: ttyLocker ttyl; duke@435: // Missing MDP? Leave a small complaint in the log. duke@435: xtty->elem("missing_mdp bci='%d'", trap_bci); kvn@1641: } duke@435: } duke@435: } duke@435: duke@435: // Return results: duke@435: ret_this_trap_count = this_trap_count; duke@435: ret_maybe_prior_trap = maybe_prior_trap; duke@435: ret_maybe_prior_recompile = maybe_prior_recompile; duke@435: return pdata; duke@435: } duke@435: duke@435: void duke@435: Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { duke@435: ResourceMark rm; duke@435: // Ignored outputs: duke@435: uint ignore_this_trap_count; duke@435: bool ignore_maybe_prior_trap; duke@435: bool ignore_maybe_prior_recompile; duke@435: query_update_method_data(trap_mdo, trap_bci, duke@435: (DeoptReason)reason, duke@435: ignore_this_trap_count, duke@435: ignore_maybe_prior_trap, duke@435: ignore_maybe_prior_recompile); duke@435: } duke@435: duke@435: Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { duke@435: duke@435: // Still in Java no safepoints duke@435: { duke@435: // This enters VM and may safepoint duke@435: uncommon_trap_inner(thread, trap_request); duke@435: } duke@435: return fetch_unroll_info_helper(thread); duke@435: } duke@435: duke@435: // Local derived constants. duke@435: // Further breakdown of DataLayout::trap_state, as promised by DataLayout. duke@435: const int DS_REASON_MASK = DataLayout::trap_mask >> 1; duke@435: const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; duke@435: duke@435: //---------------------------trap_state_reason--------------------------------- duke@435: Deoptimization::DeoptReason duke@435: Deoptimization::trap_state_reason(int trap_state) { duke@435: // This assert provides the link between the width of DataLayout::trap_bits duke@435: // and the encoding of "recorded" reasons. It ensures there are enough duke@435: // bits to store all needed reasons in the per-BCI MDO profile. duke@435: assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); duke@435: int recompile_bit = (trap_state & DS_RECOMPILE_BIT); duke@435: trap_state -= recompile_bit; duke@435: if (trap_state == DS_REASON_MASK) { duke@435: return Reason_many; duke@435: } else { duke@435: assert((int)Reason_none == 0, "state=0 => Reason_none"); duke@435: return (DeoptReason)trap_state; duke@435: } duke@435: } duke@435: //-------------------------trap_state_has_reason------------------------------- duke@435: int Deoptimization::trap_state_has_reason(int trap_state, int reason) { duke@435: assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); duke@435: assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); duke@435: int recompile_bit = (trap_state & DS_RECOMPILE_BIT); duke@435: trap_state -= recompile_bit; duke@435: if (trap_state == DS_REASON_MASK) { duke@435: return -1; // true, unspecifically (bottom of state lattice) duke@435: } else if (trap_state == reason) { duke@435: return 1; // true, definitely duke@435: } else if (trap_state == 0) { duke@435: return 0; // false, definitely (top of state lattice) duke@435: } else { duke@435: return 0; // false, definitely duke@435: } duke@435: } duke@435: //-------------------------trap_state_add_reason------------------------------- duke@435: int Deoptimization::trap_state_add_reason(int trap_state, int reason) { duke@435: assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); duke@435: int recompile_bit = (trap_state & DS_RECOMPILE_BIT); duke@435: trap_state -= recompile_bit; duke@435: if (trap_state == DS_REASON_MASK) { duke@435: return trap_state + recompile_bit; // already at state lattice bottom duke@435: } else if (trap_state == reason) { duke@435: return trap_state + recompile_bit; // the condition is already true duke@435: } else if (trap_state == 0) { duke@435: return reason + recompile_bit; // no condition has yet been true duke@435: } else { duke@435: return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom duke@435: } duke@435: } duke@435: //-----------------------trap_state_is_recompiled------------------------------ duke@435: bool Deoptimization::trap_state_is_recompiled(int trap_state) { duke@435: return (trap_state & DS_RECOMPILE_BIT) != 0; duke@435: } duke@435: //-----------------------trap_state_set_recompiled----------------------------- duke@435: int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { duke@435: if (z) return trap_state | DS_RECOMPILE_BIT; duke@435: else return trap_state & ~DS_RECOMPILE_BIT; duke@435: } duke@435: //---------------------------format_trap_state--------------------------------- duke@435: // This is used for debugging and diagnostics, including hotspot.log output. duke@435: const char* Deoptimization::format_trap_state(char* buf, size_t buflen, duke@435: int trap_state) { duke@435: DeoptReason reason = trap_state_reason(trap_state); duke@435: bool recomp_flag = trap_state_is_recompiled(trap_state); duke@435: // Re-encode the state from its decoded components. duke@435: int decoded_state = 0; duke@435: if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) duke@435: decoded_state = trap_state_add_reason(decoded_state, reason); duke@435: if (recomp_flag) duke@435: decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); duke@435: // If the state re-encodes properly, format it symbolically. duke@435: // Because this routine is used for debugging and diagnostics, duke@435: // be robust even if the state is a strange value. duke@435: size_t len; duke@435: if (decoded_state != trap_state) { duke@435: // Random buggy state that doesn't decode?? duke@435: len = jio_snprintf(buf, buflen, "#%d", trap_state); duke@435: } else { duke@435: len = jio_snprintf(buf, buflen, "%s%s", duke@435: trap_reason_name(reason), duke@435: recomp_flag ? " recompiled" : ""); duke@435: } duke@435: if (len >= buflen) duke@435: buf[buflen-1] = '\0'; duke@435: return buf; duke@435: } duke@435: duke@435: duke@435: //--------------------------------statics-------------------------------------- duke@435: Deoptimization::DeoptAction Deoptimization::_unloaded_action duke@435: = Deoptimization::Action_reinterpret; duke@435: const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { duke@435: // Note: Keep this in sync. with enum DeoptReason. duke@435: "none", duke@435: "null_check", duke@435: "null_assert", duke@435: "range_check", duke@435: "class_check", duke@435: "array_check", duke@435: "intrinsic", kvn@1641: "bimorphic", duke@435: "unloaded", duke@435: "uninitialized", duke@435: "unreached", duke@435: "unhandled", duke@435: "constraint", duke@435: "div0_check", cfang@1607: "age", kvn@2877: "predicate", kvn@2877: "loop_limit_check" duke@435: }; duke@435: const char* Deoptimization::_trap_action_name[Action_LIMIT] = { duke@435: // Note: Keep this in sync. with enum DeoptAction. duke@435: "none", duke@435: "maybe_recompile", duke@435: "reinterpret", duke@435: "make_not_entrant", duke@435: "make_not_compilable" duke@435: }; duke@435: duke@435: const char* Deoptimization::trap_reason_name(int reason) { duke@435: if (reason == Reason_many) return "many"; duke@435: if ((uint)reason < Reason_LIMIT) duke@435: return _trap_reason_name[reason]; duke@435: static char buf[20]; duke@435: sprintf(buf, "reason%d", reason); duke@435: return buf; duke@435: } duke@435: const char* Deoptimization::trap_action_name(int action) { duke@435: if ((uint)action < Action_LIMIT) duke@435: return _trap_action_name[action]; duke@435: static char buf[20]; duke@435: sprintf(buf, "action%d", action); duke@435: return buf; duke@435: } duke@435: duke@435: // This is used for debugging and diagnostics, including hotspot.log output. duke@435: const char* Deoptimization::format_trap_request(char* buf, size_t buflen, duke@435: int trap_request) { duke@435: jint unloaded_class_index = trap_request_index(trap_request); duke@435: const char* reason = trap_reason_name(trap_request_reason(trap_request)); duke@435: const char* action = trap_action_name(trap_request_action(trap_request)); duke@435: size_t len; duke@435: if (unloaded_class_index < 0) { duke@435: len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", duke@435: reason, action); duke@435: } else { duke@435: len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", duke@435: reason, action, unloaded_class_index); duke@435: } duke@435: if (len >= buflen) duke@435: buf[buflen-1] = '\0'; duke@435: return buf; duke@435: } duke@435: duke@435: juint Deoptimization::_deoptimization_hist duke@435: [Deoptimization::Reason_LIMIT] duke@435: [1 + Deoptimization::Action_LIMIT] duke@435: [Deoptimization::BC_CASE_LIMIT] duke@435: = {0}; duke@435: duke@435: enum { duke@435: LSB_BITS = 8, duke@435: LSB_MASK = right_n_bits(LSB_BITS) duke@435: }; duke@435: duke@435: void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, duke@435: Bytecodes::Code bc) { duke@435: assert(reason >= 0 && reason < Reason_LIMIT, "oob"); duke@435: assert(action >= 0 && action < Action_LIMIT, "oob"); duke@435: _deoptimization_hist[Reason_none][0][0] += 1; // total duke@435: _deoptimization_hist[reason][0][0] += 1; // per-reason total duke@435: juint* cases = _deoptimization_hist[reason][1+action]; duke@435: juint* bc_counter_addr = NULL; duke@435: juint bc_counter = 0; duke@435: // Look for an unused counter, or an exact match to this BC. duke@435: if (bc != Bytecodes::_illegal) { duke@435: for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { duke@435: juint* counter_addr = &cases[bc_case]; duke@435: juint counter = *counter_addr; duke@435: if ((counter == 0 && bc_counter_addr == NULL) duke@435: || (Bytecodes::Code)(counter & LSB_MASK) == bc) { duke@435: // this counter is either free or is already devoted to this BC duke@435: bc_counter_addr = counter_addr; duke@435: bc_counter = counter | bc; duke@435: } duke@435: } duke@435: } duke@435: if (bc_counter_addr == NULL) { duke@435: // Overflow, or no given bytecode. duke@435: bc_counter_addr = &cases[BC_CASE_LIMIT-1]; duke@435: bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB duke@435: } duke@435: *bc_counter_addr = bc_counter + (1 << LSB_BITS); duke@435: } duke@435: duke@435: jint Deoptimization::total_deoptimization_count() { duke@435: return _deoptimization_hist[Reason_none][0][0]; duke@435: } duke@435: duke@435: jint Deoptimization::deoptimization_count(DeoptReason reason) { duke@435: assert(reason >= 0 && reason < Reason_LIMIT, "oob"); duke@435: return _deoptimization_hist[reason][0][0]; duke@435: } duke@435: duke@435: void Deoptimization::print_statistics() { duke@435: juint total = total_deoptimization_count(); duke@435: juint account = total; duke@435: if (total != 0) { duke@435: ttyLocker ttyl; duke@435: if (xtty != NULL) xtty->head("statistics type='deoptimization'"); duke@435: tty->print_cr("Deoptimization traps recorded:"); duke@435: #define PRINT_STAT_LINE(name, r) \ duke@435: tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); duke@435: PRINT_STAT_LINE("total", total); duke@435: // For each non-zero entry in the histogram, print the reason, duke@435: // the action, and (if specifically known) the type of bytecode. duke@435: for (int reason = 0; reason < Reason_LIMIT; reason++) { duke@435: for (int action = 0; action < Action_LIMIT; action++) { duke@435: juint* cases = _deoptimization_hist[reason][1+action]; duke@435: for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { duke@435: juint counter = cases[bc_case]; duke@435: if (counter != 0) { duke@435: char name[1*K]; duke@435: Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); duke@435: if (bc_case == BC_CASE_LIMIT && (int)bc == 0) duke@435: bc = Bytecodes::_illegal; duke@435: sprintf(name, "%s/%s/%s", duke@435: trap_reason_name(reason), duke@435: trap_action_name(action), duke@435: Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); duke@435: juint r = counter >> LSB_BITS; duke@435: tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); duke@435: account -= r; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: if (account != 0) { duke@435: PRINT_STAT_LINE("unaccounted", account); duke@435: } duke@435: #undef PRINT_STAT_LINE duke@435: if (xtty != NULL) xtty->tail("statistics"); duke@435: } duke@435: } twisti@2047: #else // COMPILER2 || SHARK duke@435: duke@435: duke@435: // Stubs for C1 only system. duke@435: bool Deoptimization::trap_state_is_recompiled(int trap_state) { duke@435: return false; duke@435: } duke@435: duke@435: const char* Deoptimization::trap_reason_name(int reason) { duke@435: return "unknown"; duke@435: } duke@435: duke@435: void Deoptimization::print_statistics() { duke@435: // no output duke@435: } duke@435: duke@435: void duke@435: Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { duke@435: // no udpate duke@435: } duke@435: duke@435: int Deoptimization::trap_state_has_reason(int trap_state, int reason) { duke@435: return 0; duke@435: } duke@435: duke@435: void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, duke@435: Bytecodes::Code bc) { duke@435: // no update duke@435: } duke@435: duke@435: const char* Deoptimization::format_trap_state(char* buf, size_t buflen, duke@435: int trap_state) { duke@435: jio_snprintf(buf, buflen, "#%d", trap_state); duke@435: return buf; duke@435: } duke@435: twisti@2047: #endif // COMPILER2 || SHARK