6953144: Tiered compilation
Summary: Infrastructure for tiered compilation support (interpreter + c1 + c2) for 32 and 64 bit. Simple tiered policy implementation.
Reviewed-by: kvn, never, phh, twisti
/*
* Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "incls/_precompiled.incl"
#include "incls/_deoptimization.cpp.incl"
bool DeoptimizationMarker::_is_active = false;
Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
int caller_adjustment,
int number_of_frames,
intptr_t* frame_sizes,
address* frame_pcs,
BasicType return_type) {
_size_of_deoptimized_frame = size_of_deoptimized_frame;
_caller_adjustment = caller_adjustment;
_number_of_frames = number_of_frames;
_frame_sizes = frame_sizes;
_frame_pcs = frame_pcs;
_register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2);
_return_type = return_type;
// PD (x86 only)
_counter_temp = 0;
_initial_fp = 0;
_unpack_kind = 0;
_sender_sp_temp = 0;
_total_frame_sizes = size_of_frames();
}
Deoptimization::UnrollBlock::~UnrollBlock() {
FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
FREE_C_HEAP_ARRAY(intptr_t, _register_block);
}
intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
assert(register_number < RegisterMap::reg_count, "checking register number");
return &_register_block[register_number * 2];
}
int Deoptimization::UnrollBlock::size_of_frames() const {
// Acount first for the adjustment of the initial frame
int result = _caller_adjustment;
for (int index = 0; index < number_of_frames(); index++) {
result += frame_sizes()[index];
}
return result;
}
void Deoptimization::UnrollBlock::print() {
ttyLocker ttyl;
tty->print_cr("UnrollBlock");
tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
tty->print( " frame_sizes: ");
for (int index = 0; index < number_of_frames(); index++) {
tty->print("%d ", frame_sizes()[index]);
}
tty->cr();
}
// In order to make fetch_unroll_info work properly with escape
// analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
// ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
// of previously eliminated objects occurs in realloc_objects, which is
// called from the method fetch_unroll_info_helper below.
JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
// It is actually ok to allocate handles in a leaf method. It causes no safepoints,
// but makes the entry a little slower. There is however a little dance we have to
// do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
// fetch_unroll_info() is called at the beginning of the deoptimization
// handler. Note this fact before we start generating temporary frames
// that can confuse an asynchronous stack walker. This counter is
// decremented at the end of unpack_frames().
thread->inc_in_deopt_handler();
return fetch_unroll_info_helper(thread);
JRT_END
// This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
// Note: there is a safepoint safety issue here. No matter whether we enter
// via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
// the vframeArray is created.
//
// Allocate our special deoptimization ResourceMark
DeoptResourceMark* dmark = new DeoptResourceMark(thread);
assert(thread->deopt_mark() == NULL, "Pending deopt!");
thread->set_deopt_mark(dmark);
frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
RegisterMap map(thread, true);
RegisterMap dummy_map(thread, false);
// Now get the deoptee with a valid map
frame deoptee = stub_frame.sender(&map);
// Create a growable array of VFrames where each VFrame represents an inlined
// Java frame. This storage is allocated with the usual system arena.
assert(deoptee.is_compiled_frame(), "Wrong frame type");
GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
while (!vf->is_top()) {
assert(vf->is_compiled_frame(), "Wrong frame type");
chunk->push(compiledVFrame::cast(vf));
vf = vf->sender();
}
assert(vf->is_compiled_frame(), "Wrong frame type");
chunk->push(compiledVFrame::cast(vf));
#ifdef COMPILER2
// Reallocate the non-escaping objects and restore their fields. Then
// relock objects if synchronization on them was eliminated.
if (DoEscapeAnalysis) {
if (EliminateAllocations) {
assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
// The flag return_oop() indicates call sites which return oop
// in compiled code. Such sites include java method calls,
// runtime calls (for example, used to allocate new objects/arrays
// on slow code path) and any other calls generated in compiled code.
// It is not guaranteed that we can get such information here only
// by analyzing bytecode in deoptimized frames. This is why this flag
// is set during method compilation (see Compile::Process_OopMap_Node()).
bool save_oop_result = chunk->at(0)->scope()->return_oop();
Handle return_value;
if (save_oop_result) {
// Reallocation may trigger GC. If deoptimization happened on return from
// call which returns oop we need to save it since it is not in oopmap.
oop result = deoptee.saved_oop_result(&map);
assert(result == NULL || result->is_oop(), "must be oop");
return_value = Handle(thread, result);
assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
if (TraceDeoptimization) {
tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
}
}
bool reallocated = false;
if (objects != NULL) {
JRT_BLOCK
reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
JRT_END
}
if (reallocated) {
reassign_fields(&deoptee, &map, objects);
#ifndef PRODUCT
if (TraceDeoptimization) {
ttyLocker ttyl;
tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
print_objects(objects);
}
#endif
}
if (save_oop_result) {
// Restore result.
deoptee.set_saved_oop_result(&map, return_value());
}
}
if (EliminateLocks) {
#ifndef PRODUCT
bool first = true;
#endif
for (int i = 0; i < chunk->length(); i++) {
compiledVFrame* cvf = chunk->at(i);
assert (cvf->scope() != NULL,"expect only compiled java frames");
GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
if (monitors->is_nonempty()) {
relock_objects(monitors, thread);
#ifndef PRODUCT
if (TraceDeoptimization) {
ttyLocker ttyl;
for (int j = 0; j < monitors->length(); j++) {
MonitorInfo* mi = monitors->at(j);
if (mi->eliminated()) {
if (first) {
first = false;
tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
}
tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
}
}
}
#endif
}
}
}
}
#endif // COMPILER2
// Ensure that no safepoint is taken after pointers have been stored
// in fields of rematerialized objects. If a safepoint occurs from here on
// out the java state residing in the vframeArray will be missed.
No_Safepoint_Verifier no_safepoint;
vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
thread->set_vframe_array_head(array);
// Now that the vframeArray has been created if we have any deferred local writes
// added by jvmti then we can free up that structure as the data is now in the
// vframeArray
if (thread->deferred_locals() != NULL) {
GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
int i = 0;
do {
// Because of inlining we could have multiple vframes for a single frame
// and several of the vframes could have deferred writes. Find them all.
if (list->at(i)->id() == array->original().id()) {
jvmtiDeferredLocalVariableSet* dlv = list->at(i);
list->remove_at(i);
// individual jvmtiDeferredLocalVariableSet are CHeapObj's
delete dlv;
} else {
i++;
}
} while ( i < list->length() );
if (list->length() == 0) {
thread->set_deferred_locals(NULL);
// free the list and elements back to C heap.
delete list;
}
}
#ifndef SHARK
// Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
CodeBlob* cb = stub_frame.cb();
// Verify we have the right vframeArray
assert(cb->frame_size() >= 0, "Unexpected frame size");
intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
// If the deopt call site is a MethodHandle invoke call site we have
// to adjust the unpack_sp.
nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
unpack_sp = deoptee.unextended_sp();
#ifdef ASSERT
assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp);
#endif
#else
intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
#endif // !SHARK
// This is a guarantee instead of an assert because if vframe doesn't match
// we will unpack the wrong deoptimized frame and wind up in strange places
// where it will be very difficult to figure out what went wrong. Better
// to die an early death here than some very obscure death later when the
// trail is cold.
// Note: on ia64 this guarantee can be fooled by frames with no memory stack
// in that it will fail to detect a problem when there is one. This needs
// more work in tiger timeframe.
guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
int number_of_frames = array->frames();
// Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
// virtual activation, which is the reverse of the elements in the vframes array.
intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames);
// +1 because we always have an interpreter return address for the final slot.
address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1);
int callee_parameters = 0;
int callee_locals = 0;
int popframe_extra_args = 0;
// Create an interpreter return address for the stub to use as its return
// address so the skeletal frames are perfectly walkable
frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
// PopFrame requires that the preserved incoming arguments from the recently-popped topmost
// activation be put back on the expression stack of the caller for reexecution
if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
}
//
// frame_sizes/frame_pcs[0] oldest frame (int or c2i)
// frame_sizes/frame_pcs[1] next oldest frame (int)
// frame_sizes/frame_pcs[n] youngest frame (int)
//
// Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
// owns the space for the return address to it's caller). Confusing ain't it.
//
// The vframe array can address vframes with indices running from
// 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
// When we create the skeletal frames we need the oldest frame to be in the zero slot
// in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
// so things look a little strange in this loop.
//
for (int index = 0; index < array->frames(); index++ ) {
// frame[number_of_frames - 1 ] = on_stack_size(youngest)
// frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
// frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
callee_locals,
index == 0,
popframe_extra_args);
// This pc doesn't have to be perfect just good enough to identify the frame
// as interpreted so the skeleton frame will be walkable
// The correct pc will be set when the skeleton frame is completely filled out
// The final pc we store in the loop is wrong and will be overwritten below
frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
callee_parameters = array->element(index)->method()->size_of_parameters();
callee_locals = array->element(index)->method()->max_locals();
popframe_extra_args = 0;
}
// Compute whether the root vframe returns a float or double value.
BasicType return_type;
{
HandleMark hm;
methodHandle method(thread, array->element(0)->method());
Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci());
return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL;
}
// Compute information for handling adapters and adjusting the frame size of the caller.
int caller_adjustment = 0;
// Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
// itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
// than simply use array->sender.pc(). This requires us to walk the current set of frames
//
frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
// Compute the amount the oldest interpreter frame will have to adjust
// its caller's stack by. If the caller is a compiled frame then
// we pretend that the callee has no parameters so that the
// extension counts for the full amount of locals and not just
// locals-parms. This is because without a c2i adapter the parm
// area as created by the compiled frame will not be usable by
// the interpreter. (Depending on the calling convention there
// may not even be enough space).
// QQQ I'd rather see this pushed down into last_frame_adjust
// and have it take the sender (aka caller).
if (deopt_sender.is_compiled_frame()) {
caller_adjustment = last_frame_adjust(0, callee_locals);
} else if (callee_locals > callee_parameters) {
// The caller frame may need extending to accommodate
// non-parameter locals of the first unpacked interpreted frame.
// Compute that adjustment.
caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
}
// If the sender is deoptimized the we must retrieve the address of the handler
// since the frame will "magically" show the original pc before the deopt
// and we'd undo the deopt.
frame_pcs[0] = deopt_sender.raw_pc();
#ifndef SHARK
assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
#endif // SHARK
UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
caller_adjustment * BytesPerWord,
number_of_frames,
frame_sizes,
frame_pcs,
return_type);
#if defined(IA32) || defined(AMD64)
// We need a way to pass fp to the unpacking code so the skeletal frames
// come out correct. This is only needed for x86 because of c2 using ebp
// as an allocatable register. So this update is useless (and harmless)
// on the other platforms. It would be nice to do this in a different
// way but even the old style deoptimization had a problem with deriving
// this value. NEEDS_CLEANUP
// Note: now that c1 is using c2's deopt blob we must do this on all
// x86 based platforms
intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes());
*fp_addr = array->sender().fp(); // was adapter_caller
#endif /* IA32 || AMD64 */
if (array->frames() > 1) {
if (VerifyStack && TraceDeoptimization) {
tty->print_cr("Deoptimizing method containing inlining");
}
}
array->set_unroll_block(info);
return info;
}
// Called to cleanup deoptimization data structures in normal case
// after unpacking to stack and when stack overflow error occurs
void Deoptimization::cleanup_deopt_info(JavaThread *thread,
vframeArray *array) {
// Get array if coming from exception
if (array == NULL) {
array = thread->vframe_array_head();
}
thread->set_vframe_array_head(NULL);
// Free the previous UnrollBlock
vframeArray* old_array = thread->vframe_array_last();
thread->set_vframe_array_last(array);
if (old_array != NULL) {
UnrollBlock* old_info = old_array->unroll_block();
old_array->set_unroll_block(NULL);
delete old_info;
delete old_array;
}
// Deallocate any resource creating in this routine and any ResourceObjs allocated
// inside the vframeArray (StackValueCollections)
delete thread->deopt_mark();
thread->set_deopt_mark(NULL);
if (JvmtiExport::can_pop_frame()) {
#ifndef CC_INTERP
// Regardless of whether we entered this routine with the pending
// popframe condition bit set, we should always clear it now
thread->clear_popframe_condition();
#else
// C++ interpeter will clear has_pending_popframe when it enters
// with method_resume. For deopt_resume2 we clear it now.
if (thread->popframe_forcing_deopt_reexecution())
thread->clear_popframe_condition();
#endif /* CC_INTERP */
}
// unpack_frames() is called at the end of the deoptimization handler
// and (in C2) at the end of the uncommon trap handler. Note this fact
// so that an asynchronous stack walker can work again. This counter is
// incremented at the beginning of fetch_unroll_info() and (in C2) at
// the beginning of uncommon_trap().
thread->dec_in_deopt_handler();
}
// Return BasicType of value being returned
JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
// We are already active int he special DeoptResourceMark any ResourceObj's we
// allocate will be freed at the end of the routine.
// It is actually ok to allocate handles in a leaf method. It causes no safepoints,
// but makes the entry a little slower. There is however a little dance we have to
// do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
ResetNoHandleMark rnhm; // No-op in release/product versions
HandleMark hm;
frame stub_frame = thread->last_frame();
// Since the frame to unpack is the top frame of this thread, the vframe_array_head
// must point to the vframeArray for the unpack frame.
vframeArray* array = thread->vframe_array_head();
#ifndef PRODUCT
if (TraceDeoptimization) {
tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
}
#endif
UnrollBlock* info = array->unroll_block();
// Unpack the interpreter frames and any adapter frame (c2 only) we might create.
array->unpack_to_stack(stub_frame, exec_mode);
BasicType bt = info->return_type();
// If we have an exception pending, claim that the return type is an oop
// so the deopt_blob does not overwrite the exception_oop.
if (exec_mode == Unpack_exception)
bt = T_OBJECT;
// Cleanup thread deopt data
cleanup_deopt_info(thread, array);
#ifndef PRODUCT
if (VerifyStack) {
ResourceMark res_mark;
// Verify that the just-unpacked frames match the interpreter's
// notions of expression stack and locals
vframeArray* cur_array = thread->vframe_array_last();
RegisterMap rm(thread, false);
rm.set_include_argument_oops(false);
bool is_top_frame = true;
int callee_size_of_parameters = 0;
int callee_max_locals = 0;
for (int i = 0; i < cur_array->frames(); i++) {
vframeArrayElement* el = cur_array->element(i);
frame* iframe = el->iframe();
guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
// Get the oop map for this bci
InterpreterOopMap mask;
int cur_invoke_parameter_size = 0;
bool try_next_mask = false;
int next_mask_expression_stack_size = -1;
int top_frame_expression_stack_adjustment = 0;
methodHandle mh(thread, iframe->interpreter_frame_method());
OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
BytecodeStream str(mh);
str.set_start(iframe->interpreter_frame_bci());
int max_bci = mh->code_size();
// Get to the next bytecode if possible
assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
// Check to see if we can grab the number of outgoing arguments
// at an uncommon trap for an invoke (where the compiler
// generates debug info before the invoke has executed)
Bytecodes::Code cur_code = str.next();
if (cur_code == Bytecodes::_invokevirtual ||
cur_code == Bytecodes::_invokespecial ||
cur_code == Bytecodes::_invokestatic ||
cur_code == Bytecodes::_invokeinterface) {
Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci());
symbolHandle signature(thread, invoke->signature());
ArgumentSizeComputer asc(signature);
cur_invoke_parameter_size = asc.size();
if (cur_code != Bytecodes::_invokestatic) {
// Add in receiver
++cur_invoke_parameter_size;
}
}
if (str.bci() < max_bci) {
Bytecodes::Code bc = str.next();
if (bc >= 0) {
// The interpreter oop map generator reports results before
// the current bytecode has executed except in the case of
// calls. It seems to be hard to tell whether the compiler
// has emitted debug information matching the "state before"
// a given bytecode or the state after, so we try both
switch (cur_code) {
case Bytecodes::_invokevirtual:
case Bytecodes::_invokespecial:
case Bytecodes::_invokestatic:
case Bytecodes::_invokeinterface:
case Bytecodes::_athrow:
break;
default: {
InterpreterOopMap next_mask;
OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
next_mask_expression_stack_size = next_mask.expression_stack_size();
// Need to subtract off the size of the result type of
// the bytecode because this is not described in the
// debug info but returned to the interpreter in the TOS
// caching register
BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
if (bytecode_result_type != T_ILLEGAL) {
top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
}
assert(top_frame_expression_stack_adjustment >= 0, "");
try_next_mask = true;
break;
}
}
}
}
// Verify stack depth and oops in frame
// This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
if (!(
/* SPARC */
(iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
/* x86 */
(iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
(try_next_mask &&
(iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
top_frame_expression_stack_adjustment))) ||
(is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
(is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
(iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
)) {
ttyLocker ttyl;
// Print out some information that will help us debug the problem
tty->print_cr("Wrong number of expression stack elements during deoptimization");
tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
iframe->interpreter_frame_expression_stack_size());
tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
tty->print_cr(" try_next_mask = %d", try_next_mask);
tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
tty->print_cr(" callee_max_locals = %d", callee_max_locals);
tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
tty->print_cr(" exec_mode = %d", exec_mode);
tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
tty->print_cr(" Interpreted frames:");
for (int k = 0; k < cur_array->frames(); k++) {
vframeArrayElement* el = cur_array->element(k);
tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
}
cur_array->print_on_2(tty);
guarantee(false, "wrong number of expression stack elements during deopt");
}
VerifyOopClosure verify;
iframe->oops_interpreted_do(&verify, &rm, false);
callee_size_of_parameters = mh->size_of_parameters();
callee_max_locals = mh->max_locals();
is_top_frame = false;
}
}
#endif /* !PRODUCT */
return bt;
JRT_END
int Deoptimization::deoptimize_dependents() {
Threads::deoptimized_wrt_marked_nmethods();
return 0;
}
#ifdef COMPILER2
bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
Handle pending_exception(thread->pending_exception());
const char* exception_file = thread->exception_file();
int exception_line = thread->exception_line();
thread->clear_pending_exception();
for (int i = 0; i < objects->length(); i++) {
assert(objects->at(i)->is_object(), "invalid debug information");
ObjectValue* sv = (ObjectValue*) objects->at(i);
KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
oop obj = NULL;
if (k->oop_is_instance()) {
instanceKlass* ik = instanceKlass::cast(k());
obj = ik->allocate_instance(CHECK_(false));
} else if (k->oop_is_typeArray()) {
typeArrayKlass* ak = typeArrayKlass::cast(k());
assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
int len = sv->field_size() / type2size[ak->element_type()];
obj = ak->allocate(len, CHECK_(false));
} else if (k->oop_is_objArray()) {
objArrayKlass* ak = objArrayKlass::cast(k());
obj = ak->allocate(sv->field_size(), CHECK_(false));
}
assert(obj != NULL, "allocation failed");
assert(sv->value().is_null(), "redundant reallocation");
sv->set_value(obj);
}
if (pending_exception.not_null()) {
thread->set_pending_exception(pending_exception(), exception_file, exception_line);
}
return true;
}
// This assumes that the fields are stored in ObjectValue in the same order
// they are yielded by do_nonstatic_fields.
class FieldReassigner: public FieldClosure {
frame* _fr;
RegisterMap* _reg_map;
ObjectValue* _sv;
instanceKlass* _ik;
oop _obj;
int _i;
public:
FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
_fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
int i() const { return _i; }
void do_field(fieldDescriptor* fd) {
intptr_t val;
StackValue* value =
StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
int offset = fd->offset();
switch (fd->field_type()) {
case T_OBJECT: case T_ARRAY:
assert(value->type() == T_OBJECT, "Agreement.");
_obj->obj_field_put(offset, value->get_obj()());
break;
case T_LONG: case T_DOUBLE: {
assert(value->type() == T_INT, "Agreement.");
StackValue* low =
StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
#ifdef _LP64
jlong res = (jlong)low->get_int();
#else
#ifdef SPARC
// For SPARC we have to swap high and low words.
jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
#else
jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
#endif //SPARC
#endif
_obj->long_field_put(offset, res);
break;
}
// Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
case T_INT: case T_FLOAT: // 4 bytes.
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
_obj->int_field_put(offset, (jint)*((jint*)&val));
break;
case T_SHORT: case T_CHAR: // 2 bytes
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
_obj->short_field_put(offset, (jshort)*((jint*)&val));
break;
case T_BOOLEAN: case T_BYTE: // 1 byte
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
_obj->bool_field_put(offset, (jboolean)*((jint*)&val));
break;
default:
ShouldNotReachHere();
}
_i++;
}
};
// restore elements of an eliminated type array
void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
int index = 0;
intptr_t val;
for (int i = 0; i < sv->field_size(); i++) {
StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
switch(type) {
case T_LONG: case T_DOUBLE: {
assert(value->type() == T_INT, "Agreement.");
StackValue* low =
StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
#ifdef _LP64
jlong res = (jlong)low->get_int();
#else
#ifdef SPARC
// For SPARC we have to swap high and low words.
jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
#else
jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
#endif //SPARC
#endif
obj->long_at_put(index, res);
break;
}
// Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
case T_INT: case T_FLOAT: // 4 bytes.
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
obj->int_at_put(index, (jint)*((jint*)&val));
break;
case T_SHORT: case T_CHAR: // 2 bytes
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
obj->short_at_put(index, (jshort)*((jint*)&val));
break;
case T_BOOLEAN: case T_BYTE: // 1 byte
assert(value->type() == T_INT, "Agreement.");
val = value->get_int();
obj->bool_at_put(index, (jboolean)*((jint*)&val));
break;
default:
ShouldNotReachHere();
}
index++;
}
}
// restore fields of an eliminated object array
void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
for (int i = 0; i < sv->field_size(); i++) {
StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
assert(value->type() == T_OBJECT, "object element expected");
obj->obj_at_put(i, value->get_obj()());
}
}
// restore fields of all eliminated objects and arrays
void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
for (int i = 0; i < objects->length(); i++) {
ObjectValue* sv = (ObjectValue*) objects->at(i);
KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
Handle obj = sv->value();
assert(obj.not_null(), "reallocation was missed");
if (k->oop_is_instance()) {
instanceKlass* ik = instanceKlass::cast(k());
FieldReassigner reassign(fr, reg_map, sv, obj());
ik->do_nonstatic_fields(&reassign);
} else if (k->oop_is_typeArray()) {
typeArrayKlass* ak = typeArrayKlass::cast(k());
reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
} else if (k->oop_is_objArray()) {
reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
}
}
}
// relock objects for which synchronization was eliminated
void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
for (int i = 0; i < monitors->length(); i++) {
MonitorInfo* mon_info = monitors->at(i);
if (mon_info->eliminated()) {
assert(mon_info->owner() != NULL, "reallocation was missed");
Handle obj = Handle(mon_info->owner());
markOop mark = obj->mark();
if (UseBiasedLocking && mark->has_bias_pattern()) {
// New allocated objects may have the mark set to anonymously biased.
// Also the deoptimized method may called methods with synchronization
// where the thread-local object is bias locked to the current thread.
assert(mark->is_biased_anonymously() ||
mark->biased_locker() == thread, "should be locked to current thread");
// Reset mark word to unbiased prototype.
markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
obj->set_mark(unbiased_prototype);
}
BasicLock* lock = mon_info->lock();
ObjectSynchronizer::slow_enter(obj, lock, thread);
}
assert(mon_info->owner()->is_locked(), "object must be locked now");
}
}
#ifndef PRODUCT
// print information about reallocated objects
void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
fieldDescriptor fd;
for (int i = 0; i < objects->length(); i++) {
ObjectValue* sv = (ObjectValue*) objects->at(i);
KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
Handle obj = sv->value();
tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
k->as_klassOop()->print_value();
tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
tty->cr();
if (Verbose) {
k->oop_print_on(obj(), tty);
}
}
}
#endif
#endif // COMPILER2
vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
#ifndef PRODUCT
if (TraceDeoptimization) {
ttyLocker ttyl;
tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
fr.print_on(tty);
tty->print_cr(" Virtual frames (innermost first):");
for (int index = 0; index < chunk->length(); index++) {
compiledVFrame* vf = chunk->at(index);
tty->print(" %2d - ", index);
vf->print_value();
int bci = chunk->at(index)->raw_bci();
const char* code_name;
if (bci == SynchronizationEntryBCI) {
code_name = "sync entry";
} else {
Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci);
code_name = Bytecodes::name(code);
}
tty->print(" - %s", code_name);
tty->print_cr(" @ bci %d ", bci);
if (Verbose) {
vf->print();
tty->cr();
}
}
}
#endif
// Register map for next frame (used for stack crawl). We capture
// the state of the deopt'ing frame's caller. Thus if we need to
// stuff a C2I adapter we can properly fill in the callee-save
// register locations.
frame caller = fr.sender(reg_map);
int frame_size = caller.sp() - fr.sp();
frame sender = caller;
// Since the Java thread being deoptimized will eventually adjust it's own stack,
// the vframeArray containing the unpacking information is allocated in the C heap.
// For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
// Compare the vframeArray to the collected vframes
assert(array->structural_compare(thread, chunk), "just checking");
Events::log("# vframes = %d", (intptr_t)chunk->length());
#ifndef PRODUCT
if (TraceDeoptimization) {
ttyLocker ttyl;
tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
}
#endif // PRODUCT
return array;
}
static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
for (int i = 0; i < monitors->length(); i++) {
MonitorInfo* mon_info = monitors->at(i);
if (!mon_info->eliminated() && mon_info->owner() != NULL) {
objects_to_revoke->append(Handle(mon_info->owner()));
}
}
}
void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
if (!UseBiasedLocking) {
return;
}
GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
// Unfortunately we don't have a RegisterMap available in most of
// the places we want to call this routine so we need to walk the
// stack again to update the register map.
if (map == NULL || !map->update_map()) {
StackFrameStream sfs(thread, true);
bool found = false;
while (!found && !sfs.is_done()) {
frame* cur = sfs.current();
sfs.next();
found = cur->id() == fr.id();
}
assert(found, "frame to be deoptimized not found on target thread's stack");
map = sfs.register_map();
}
vframe* vf = vframe::new_vframe(&fr, map, thread);
compiledVFrame* cvf = compiledVFrame::cast(vf);
// Revoke monitors' biases in all scopes
while (!cvf->is_top()) {
collect_monitors(cvf, objects_to_revoke);
cvf = compiledVFrame::cast(cvf->sender());
}
collect_monitors(cvf, objects_to_revoke);
if (SafepointSynchronize::is_at_safepoint()) {
BiasedLocking::revoke_at_safepoint(objects_to_revoke);
} else {
BiasedLocking::revoke(objects_to_revoke);
}
}
void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
if (!UseBiasedLocking) {
return;
}
assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
if (jt->has_last_Java_frame()) {
StackFrameStream sfs(jt, true);
while (!sfs.is_done()) {
frame* cur = sfs.current();
if (cb->contains(cur->pc())) {
vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
compiledVFrame* cvf = compiledVFrame::cast(vf);
// Revoke monitors' biases in all scopes
while (!cvf->is_top()) {
collect_monitors(cvf, objects_to_revoke);
cvf = compiledVFrame::cast(cvf->sender());
}
collect_monitors(cvf, objects_to_revoke);
}
sfs.next();
}
}
}
BiasedLocking::revoke_at_safepoint(objects_to_revoke);
}
void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
assert(fr.can_be_deoptimized(), "checking frame type");
gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id());
// Patch the nmethod so that when execution returns to it we will
// deopt the execution state and return to the interpreter.
fr.deoptimize(thread);
}
void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
// Deoptimize only if the frame comes from compile code.
// Do not deoptimize the frame which is already patched
// during the execution of the loops below.
if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
return;
}
ResourceMark rm;
DeoptimizationMarker dm;
if (UseBiasedLocking) {
revoke_biases_of_monitors(thread, fr, map);
}
deoptimize_single_frame(thread, fr);
}
void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
// Compute frame and register map based on thread and sp.
RegisterMap reg_map(thread, UseBiasedLocking);
frame fr = thread->last_frame();
while (fr.id() != id) {
fr = fr.sender(®_map);
}
deoptimize(thread, fr, ®_map);
}
// JVMTI PopFrame support
JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
{
thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
}
JRT_END
#if defined(COMPILER2) || defined(SHARK)
void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
// in case of an unresolved klass entry, load the class.
if (constant_pool->tag_at(index).is_unresolved_klass()) {
klassOop tk = constant_pool->klass_at(index, CHECK);
return;
}
if (!constant_pool->tag_at(index).is_symbol()) return;
Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
symbolHandle symbol (THREAD, constant_pool->symbol_at(index));
// class name?
if (symbol->byte_at(0) != '(') {
Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
return;
}
// then it must be a signature!
for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
if (ss.is_object()) {
symbolOop s = ss.as_symbol(CHECK);
symbolHandle class_name (THREAD, s);
Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
}
}
}
void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
EXCEPTION_MARK;
load_class_by_index(constant_pool, index, THREAD);
if (HAS_PENDING_EXCEPTION) {
// Exception happened during classloading. We ignore the exception here, since it
// is going to be rethrown since the current activation is going to be deoptimzied and
// the interpreter will re-execute the bytecode.
CLEAR_PENDING_EXCEPTION;
}
}
JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
HandleMark hm;
// uncommon_trap() is called at the beginning of the uncommon trap
// handler. Note this fact before we start generating temporary frames
// that can confuse an asynchronous stack walker. This counter is
// decremented at the end of unpack_frames().
thread->inc_in_deopt_handler();
// We need to update the map if we have biased locking.
RegisterMap reg_map(thread, UseBiasedLocking);
frame stub_frame = thread->last_frame();
frame fr = stub_frame.sender(®_map);
// Make sure the calling nmethod is not getting deoptimized and removed
// before we are done with it.
nmethodLocker nl(fr.pc());
{
ResourceMark rm;
// Revoke biases of any monitors in the frame to ensure we can migrate them
revoke_biases_of_monitors(thread, fr, ®_map);
DeoptReason reason = trap_request_reason(trap_request);
DeoptAction action = trap_request_action(trap_request);
jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request);
vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
compiledVFrame* cvf = compiledVFrame::cast(vf);
nmethod* nm = cvf->code();
ScopeDesc* trap_scope = cvf->scope();
methodHandle trap_method = trap_scope->method();
int trap_bci = trap_scope->bci();
Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code();
// Record this event in the histogram.
gather_statistics(reason, action, trap_bc);
// Ensure that we can record deopt. history:
bool create_if_missing = ProfileTraps;
methodDataHandle trap_mdo
(THREAD, get_method_data(thread, trap_method, create_if_missing));
// Print a bunch of diagnostics, if requested.
if (TraceDeoptimization || LogCompilation) {
ResourceMark rm;
ttyLocker ttyl;
char buf[100];
if (xtty != NULL) {
xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
os::current_thread_id(),
format_trap_request(buf, sizeof(buf), trap_request));
nm->log_identity(xtty);
}
symbolHandle class_name;
bool unresolved = false;
if (unloaded_class_index >= 0) {
constantPoolHandle constants (THREAD, trap_method->constants());
if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
class_name = symbolHandle(THREAD,
constants->klass_name_at(unloaded_class_index));
unresolved = true;
if (xtty != NULL)
xtty->print(" unresolved='1'");
} else if (constants->tag_at(unloaded_class_index).is_symbol()) {
class_name = symbolHandle(THREAD,
constants->symbol_at(unloaded_class_index));
}
if (xtty != NULL)
xtty->name(class_name);
}
if (xtty != NULL && trap_mdo.not_null()) {
// Dump the relevant MDO state.
// This is the deopt count for the current reason, any previous
// reasons or recompiles seen at this point.
int dcnt = trap_mdo->trap_count(reason);
if (dcnt != 0)
xtty->print(" count='%d'", dcnt);
ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
int dos = (pdata == NULL)? 0: pdata->trap_state();
if (dos != 0) {
xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
if (trap_state_is_recompiled(dos)) {
int recnt2 = trap_mdo->overflow_recompile_count();
if (recnt2 != 0)
xtty->print(" recompiles2='%d'", recnt2);
}
}
}
if (xtty != NULL) {
xtty->stamp();
xtty->end_head();
}
if (TraceDeoptimization) { // make noise on the tty
tty->print("Uncommon trap occurred in");
nm->method()->print_short_name(tty);
tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
fr.pc(),
(int) os::current_thread_id(),
trap_reason_name(reason),
trap_action_name(action),
unloaded_class_index);
if (class_name.not_null()) {
tty->print(unresolved ? " unresolved class: " : " symbol: ");
class_name->print_symbol_on(tty);
}
tty->cr();
}
if (xtty != NULL) {
// Log the precise location of the trap.
for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
xtty->begin_elem("jvms bci='%d'", sd->bci());
xtty->method(sd->method());
xtty->end_elem();
if (sd->is_top()) break;
}
xtty->tail("uncommon_trap");
}
}
// (End diagnostic printout.)
// Load class if necessary
if (unloaded_class_index >= 0) {
constantPoolHandle constants(THREAD, trap_method->constants());
load_class_by_index(constants, unloaded_class_index);
}
// Flush the nmethod if necessary and desirable.
//
// We need to avoid situations where we are re-flushing the nmethod
// because of a hot deoptimization site. Repeated flushes at the same
// point need to be detected by the compiler and avoided. If the compiler
// cannot avoid them (or has a bug and "refuses" to avoid them), this
// module must take measures to avoid an infinite cycle of recompilation
// and deoptimization. There are several such measures:
//
// 1. If a recompilation is ordered a second time at some site X
// and for the same reason R, the action is adjusted to 'reinterpret',
// to give the interpreter time to exercise the method more thoroughly.
// If this happens, the method's overflow_recompile_count is incremented.
//
// 2. If the compiler fails to reduce the deoptimization rate, then
// the method's overflow_recompile_count will begin to exceed the set
// limit PerBytecodeRecompilationCutoff. If this happens, the action
// is adjusted to 'make_not_compilable', and the method is abandoned
// to the interpreter. This is a performance hit for hot methods,
// but is better than a disastrous infinite cycle of recompilations.
// (Actually, only the method containing the site X is abandoned.)
//
// 3. In parallel with the previous measures, if the total number of
// recompilations of a method exceeds the much larger set limit
// PerMethodRecompilationCutoff, the method is abandoned.
// This should only happen if the method is very large and has
// many "lukewarm" deoptimizations. The code which enforces this
// limit is elsewhere (class nmethod, class methodOopDesc).
//
// Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
// to recompile at each bytecode independently of the per-BCI cutoff.
//
// The decision to update code is up to the compiler, and is encoded
// in the Action_xxx code. If the compiler requests Action_none
// no trap state is changed, no compiled code is changed, and the
// computation suffers along in the interpreter.
//
// The other action codes specify various tactics for decompilation
// and recompilation. Action_maybe_recompile is the loosest, and
// allows the compiled code to stay around until enough traps are seen,
// and until the compiler gets around to recompiling the trapping method.
//
// The other actions cause immediate removal of the present code.
bool update_trap_state = true;
bool make_not_entrant = false;
bool make_not_compilable = false;
bool reprofile = false;
switch (action) {
case Action_none:
// Keep the old code.
update_trap_state = false;
break;
case Action_maybe_recompile:
// Do not need to invalidate the present code, but we can
// initiate another
// Start compiler without (necessarily) invalidating the nmethod.
// The system will tolerate the old code, but new code should be
// generated when possible.
break;
case Action_reinterpret:
// Go back into the interpreter for a while, and then consider
// recompiling form scratch.
make_not_entrant = true;
// Reset invocation counter for outer most method.
// This will allow the interpreter to exercise the bytecodes
// for a while before recompiling.
// By contrast, Action_make_not_entrant is immediate.
//
// Note that the compiler will track null_check, null_assert,
// range_check, and class_check events and log them as if they
// had been traps taken from compiled code. This will update
// the MDO trap history so that the next compilation will
// properly detect hot trap sites.
reprofile = true;
break;
case Action_make_not_entrant:
// Request immediate recompilation, and get rid of the old code.
// Make them not entrant, so next time they are called they get
// recompiled. Unloaded classes are loaded now so recompile before next
// time they are called. Same for uninitialized. The interpreter will
// link the missing class, if any.
make_not_entrant = true;
break;
case Action_make_not_compilable:
// Give up on compiling this method at all.
make_not_entrant = true;
make_not_compilable = true;
break;
default:
ShouldNotReachHere();
}
// Setting +ProfileTraps fixes the following, on all platforms:
// 4852688: ProfileInterpreter is off by default for ia64. The result is
// infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
// recompile relies on a methodDataOop to record heroic opt failures.
// Whether the interpreter is producing MDO data or not, we also need
// to use the MDO to detect hot deoptimization points and control
// aggressive optimization.
bool inc_recompile_count = false;
ProfileData* pdata = NULL;
if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
uint this_trap_count = 0;
bool maybe_prior_trap = false;
bool maybe_prior_recompile = false;
pdata = query_update_method_data(trap_mdo, trap_bci, reason,
//outputs:
this_trap_count,
maybe_prior_trap,
maybe_prior_recompile);
// Because the interpreter also counts null, div0, range, and class
// checks, these traps from compiled code are double-counted.
// This is harmless; it just means that the PerXTrapLimit values
// are in effect a little smaller than they look.
DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
if (per_bc_reason != Reason_none) {
// Now take action based on the partially known per-BCI history.
if (maybe_prior_trap
&& this_trap_count >= (uint)PerBytecodeTrapLimit) {
// If there are too many traps at this BCI, force a recompile.
// This will allow the compiler to see the limit overflow, and
// take corrective action, if possible. The compiler generally
// does not use the exact PerBytecodeTrapLimit value, but instead
// changes its tactics if it sees any traps at all. This provides
// a little hysteresis, delaying a recompile until a trap happens
// several times.
//
// Actually, since there is only one bit of counter per BCI,
// the possible per-BCI counts are {0,1,(per-method count)}.
// This produces accurate results if in fact there is only
// one hot trap site, but begins to get fuzzy if there are
// many sites. For example, if there are ten sites each
// trapping two or more times, they each get the blame for
// all of their traps.
make_not_entrant = true;
}
// Detect repeated recompilation at the same BCI, and enforce a limit.
if (make_not_entrant && maybe_prior_recompile) {
// More than one recompile at this point.
inc_recompile_count = maybe_prior_trap;
}
} else {
// For reasons which are not recorded per-bytecode, we simply
// force recompiles unconditionally.
// (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
make_not_entrant = true;
}
// Go back to the compiler if there are too many traps in this method.
if (this_trap_count >= (uint)PerMethodTrapLimit) {
// If there are too many traps in this method, force a recompile.
// This will allow the compiler to see the limit overflow, and
// take corrective action, if possible.
// (This condition is an unlikely backstop only, because the
// PerBytecodeTrapLimit is more likely to take effect first,
// if it is applicable.)
make_not_entrant = true;
}
// Here's more hysteresis: If there has been a recompile at
// this trap point already, run the method in the interpreter
// for a while to exercise it more thoroughly.
if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
reprofile = true;
}
}
// Take requested actions on the method:
// Recompile
if (make_not_entrant) {
if (!nm->make_not_entrant()) {
return; // the call did not change nmethod's state
}
if (pdata != NULL) {
// Record the recompilation event, if any.
int tstate0 = pdata->trap_state();
int tstate1 = trap_state_set_recompiled(tstate0, true);
if (tstate1 != tstate0)
pdata->set_trap_state(tstate1);
}
}
if (inc_recompile_count) {
trap_mdo->inc_overflow_recompile_count();
if ((uint)trap_mdo->overflow_recompile_count() >
(uint)PerBytecodeRecompilationCutoff) {
// Give up on the method containing the bad BCI.
if (trap_method() == nm->method()) {
make_not_compilable = true;
} else {
trap_method->set_not_compilable(CompLevel_full_optimization);
// But give grace to the enclosing nm->method().
}
}
}
// Reprofile
if (reprofile) {
CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
}
// Give up compiling
if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
assert(make_not_entrant, "consistent");
nm->method()->set_not_compilable(CompLevel_full_optimization);
}
} // Free marked resources
}
JRT_END
methodDataOop
Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
bool create_if_missing) {
Thread* THREAD = thread;
methodDataOop mdo = m()->method_data();
if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
// Build an MDO. Ignore errors like OutOfMemory;
// that simply means we won't have an MDO to update.
methodOopDesc::build_interpreter_method_data(m, THREAD);
if (HAS_PENDING_EXCEPTION) {
assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
CLEAR_PENDING_EXCEPTION;
}
mdo = m()->method_data();
}
return mdo;
}
ProfileData*
Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
int trap_bci,
Deoptimization::DeoptReason reason,
//outputs:
uint& ret_this_trap_count,
bool& ret_maybe_prior_trap,
bool& ret_maybe_prior_recompile) {
uint prior_trap_count = trap_mdo->trap_count(reason);
uint this_trap_count = trap_mdo->inc_trap_count(reason);
// If the runtime cannot find a place to store trap history,
// it is estimated based on the general condition of the method.
// If the method has ever been recompiled, or has ever incurred
// a trap with the present reason , then this BCI is assumed
// (pessimistically) to be the culprit.
bool maybe_prior_trap = (prior_trap_count != 0);
bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
ProfileData* pdata = NULL;
// For reasons which are recorded per bytecode, we check per-BCI data.
DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
if (per_bc_reason != Reason_none) {
// Find the profile data for this BCI. If there isn't one,
// try to allocate one from the MDO's set of spares.
// This will let us detect a repeated trap at this point.
pdata = trap_mdo->allocate_bci_to_data(trap_bci);
if (pdata != NULL) {
// Query the trap state of this profile datum.
int tstate0 = pdata->trap_state();
if (!trap_state_has_reason(tstate0, per_bc_reason))
maybe_prior_trap = false;
if (!trap_state_is_recompiled(tstate0))
maybe_prior_recompile = false;
// Update the trap state of this profile datum.
int tstate1 = tstate0;
// Record the reason.
tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
// Store the updated state on the MDO, for next time.
if (tstate1 != tstate0)
pdata->set_trap_state(tstate1);
} else {
if (LogCompilation && xtty != NULL) {
ttyLocker ttyl;
// Missing MDP? Leave a small complaint in the log.
xtty->elem("missing_mdp bci='%d'", trap_bci);
}
}
}
// Return results:
ret_this_trap_count = this_trap_count;
ret_maybe_prior_trap = maybe_prior_trap;
ret_maybe_prior_recompile = maybe_prior_recompile;
return pdata;
}
void
Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
ResourceMark rm;
// Ignored outputs:
uint ignore_this_trap_count;
bool ignore_maybe_prior_trap;
bool ignore_maybe_prior_recompile;
query_update_method_data(trap_mdo, trap_bci,
(DeoptReason)reason,
ignore_this_trap_count,
ignore_maybe_prior_trap,
ignore_maybe_prior_recompile);
}
Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
// Still in Java no safepoints
{
// This enters VM and may safepoint
uncommon_trap_inner(thread, trap_request);
}
return fetch_unroll_info_helper(thread);
}
// Local derived constants.
// Further breakdown of DataLayout::trap_state, as promised by DataLayout.
const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
//---------------------------trap_state_reason---------------------------------
Deoptimization::DeoptReason
Deoptimization::trap_state_reason(int trap_state) {
// This assert provides the link between the width of DataLayout::trap_bits
// and the encoding of "recorded" reasons. It ensures there are enough
// bits to store all needed reasons in the per-BCI MDO profile.
assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
trap_state -= recompile_bit;
if (trap_state == DS_REASON_MASK) {
return Reason_many;
} else {
assert((int)Reason_none == 0, "state=0 => Reason_none");
return (DeoptReason)trap_state;
}
}
//-------------------------trap_state_has_reason-------------------------------
int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
trap_state -= recompile_bit;
if (trap_state == DS_REASON_MASK) {
return -1; // true, unspecifically (bottom of state lattice)
} else if (trap_state == reason) {
return 1; // true, definitely
} else if (trap_state == 0) {
return 0; // false, definitely (top of state lattice)
} else {
return 0; // false, definitely
}
}
//-------------------------trap_state_add_reason-------------------------------
int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
trap_state -= recompile_bit;
if (trap_state == DS_REASON_MASK) {
return trap_state + recompile_bit; // already at state lattice bottom
} else if (trap_state == reason) {
return trap_state + recompile_bit; // the condition is already true
} else if (trap_state == 0) {
return reason + recompile_bit; // no condition has yet been true
} else {
return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
}
}
//-----------------------trap_state_is_recompiled------------------------------
bool Deoptimization::trap_state_is_recompiled(int trap_state) {
return (trap_state & DS_RECOMPILE_BIT) != 0;
}
//-----------------------trap_state_set_recompiled-----------------------------
int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
if (z) return trap_state | DS_RECOMPILE_BIT;
else return trap_state & ~DS_RECOMPILE_BIT;
}
//---------------------------format_trap_state---------------------------------
// This is used for debugging and diagnostics, including hotspot.log output.
const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
int trap_state) {
DeoptReason reason = trap_state_reason(trap_state);
bool recomp_flag = trap_state_is_recompiled(trap_state);
// Re-encode the state from its decoded components.
int decoded_state = 0;
if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
decoded_state = trap_state_add_reason(decoded_state, reason);
if (recomp_flag)
decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
// If the state re-encodes properly, format it symbolically.
// Because this routine is used for debugging and diagnostics,
// be robust even if the state is a strange value.
size_t len;
if (decoded_state != trap_state) {
// Random buggy state that doesn't decode??
len = jio_snprintf(buf, buflen, "#%d", trap_state);
} else {
len = jio_snprintf(buf, buflen, "%s%s",
trap_reason_name(reason),
recomp_flag ? " recompiled" : "");
}
if (len >= buflen)
buf[buflen-1] = '\0';
return buf;
}
//--------------------------------statics--------------------------------------
Deoptimization::DeoptAction Deoptimization::_unloaded_action
= Deoptimization::Action_reinterpret;
const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
// Note: Keep this in sync. with enum DeoptReason.
"none",
"null_check",
"null_assert",
"range_check",
"class_check",
"array_check",
"intrinsic",
"bimorphic",
"unloaded",
"uninitialized",
"unreached",
"unhandled",
"constraint",
"div0_check",
"age",
"predicate"
};
const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
// Note: Keep this in sync. with enum DeoptAction.
"none",
"maybe_recompile",
"reinterpret",
"make_not_entrant",
"make_not_compilable"
};
const char* Deoptimization::trap_reason_name(int reason) {
if (reason == Reason_many) return "many";
if ((uint)reason < Reason_LIMIT)
return _trap_reason_name[reason];
static char buf[20];
sprintf(buf, "reason%d", reason);
return buf;
}
const char* Deoptimization::trap_action_name(int action) {
if ((uint)action < Action_LIMIT)
return _trap_action_name[action];
static char buf[20];
sprintf(buf, "action%d", action);
return buf;
}
// This is used for debugging and diagnostics, including hotspot.log output.
const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
int trap_request) {
jint unloaded_class_index = trap_request_index(trap_request);
const char* reason = trap_reason_name(trap_request_reason(trap_request));
const char* action = trap_action_name(trap_request_action(trap_request));
size_t len;
if (unloaded_class_index < 0) {
len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
reason, action);
} else {
len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
reason, action, unloaded_class_index);
}
if (len >= buflen)
buf[buflen-1] = '\0';
return buf;
}
juint Deoptimization::_deoptimization_hist
[Deoptimization::Reason_LIMIT]
[1 + Deoptimization::Action_LIMIT]
[Deoptimization::BC_CASE_LIMIT]
= {0};
enum {
LSB_BITS = 8,
LSB_MASK = right_n_bits(LSB_BITS)
};
void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
Bytecodes::Code bc) {
assert(reason >= 0 && reason < Reason_LIMIT, "oob");
assert(action >= 0 && action < Action_LIMIT, "oob");
_deoptimization_hist[Reason_none][0][0] += 1; // total
_deoptimization_hist[reason][0][0] += 1; // per-reason total
juint* cases = _deoptimization_hist[reason][1+action];
juint* bc_counter_addr = NULL;
juint bc_counter = 0;
// Look for an unused counter, or an exact match to this BC.
if (bc != Bytecodes::_illegal) {
for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
juint* counter_addr = &cases[bc_case];
juint counter = *counter_addr;
if ((counter == 0 && bc_counter_addr == NULL)
|| (Bytecodes::Code)(counter & LSB_MASK) == bc) {
// this counter is either free or is already devoted to this BC
bc_counter_addr = counter_addr;
bc_counter = counter | bc;
}
}
}
if (bc_counter_addr == NULL) {
// Overflow, or no given bytecode.
bc_counter_addr = &cases[BC_CASE_LIMIT-1];
bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
}
*bc_counter_addr = bc_counter + (1 << LSB_BITS);
}
jint Deoptimization::total_deoptimization_count() {
return _deoptimization_hist[Reason_none][0][0];
}
jint Deoptimization::deoptimization_count(DeoptReason reason) {
assert(reason >= 0 && reason < Reason_LIMIT, "oob");
return _deoptimization_hist[reason][0][0];
}
void Deoptimization::print_statistics() {
juint total = total_deoptimization_count();
juint account = total;
if (total != 0) {
ttyLocker ttyl;
if (xtty != NULL) xtty->head("statistics type='deoptimization'");
tty->print_cr("Deoptimization traps recorded:");
#define PRINT_STAT_LINE(name, r) \
tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
PRINT_STAT_LINE("total", total);
// For each non-zero entry in the histogram, print the reason,
// the action, and (if specifically known) the type of bytecode.
for (int reason = 0; reason < Reason_LIMIT; reason++) {
for (int action = 0; action < Action_LIMIT; action++) {
juint* cases = _deoptimization_hist[reason][1+action];
for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
juint counter = cases[bc_case];
if (counter != 0) {
char name[1*K];
Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
bc = Bytecodes::_illegal;
sprintf(name, "%s/%s/%s",
trap_reason_name(reason),
trap_action_name(action),
Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
juint r = counter >> LSB_BITS;
tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
account -= r;
}
}
}
}
if (account != 0) {
PRINT_STAT_LINE("unaccounted", account);
}
#undef PRINT_STAT_LINE
if (xtty != NULL) xtty->tail("statistics");
}
}
#else // COMPILER2 || SHARK
// Stubs for C1 only system.
bool Deoptimization::trap_state_is_recompiled(int trap_state) {
return false;
}
const char* Deoptimization::trap_reason_name(int reason) {
return "unknown";
}
void Deoptimization::print_statistics() {
// no output
}
void
Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
// no udpate
}
int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
return 0;
}
void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
Bytecodes::Code bc) {
// no update
}
const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
int trap_state) {
jio_snprintf(buf, buflen, "#%d", trap_state);
return buf;
}
#endif // COMPILER2 || SHARK