#include "precompiled.hpp"
#include "compiler/compileLog.hpp"
#include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
#include "gc_implementation/g1/heapRegion.hpp"
#include "gc_interface/collectedHeap.hpp"
#include "memory/barrierSet.hpp"
#include "memory/cardTableModRefBS.hpp"
#include "opto/addnode.hpp"
#include "opto/graphKit.hpp"
#include "opto/idealKit.hpp"
#include "opto/locknode.hpp"
#include "opto/machnode.hpp"
#include "opto/parse.hpp"
#include "opto/rootnode.hpp"
#include "opto/runtime.hpp"
#include "runtime/deoptimization.hpp"
#include "runtime/sharedRuntime.hpp"
//----------------------------GraphKit-----------------------------------------
// Main utility constructor.
GraphKit::GraphKit(JVMState* jvms)
: Phase(Phase::Parser),
_env(C->env()),
_gvn(*C->initial_gvn())
{
_exceptions = jvms->map()->next_exception();
if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
set_jvms(jvms);
}
// Private constructor for parser.
GraphKit::GraphKit()
: Phase(Phase::Parser),
_env(C->env()),
_gvn(*C->initial_gvn())
{
_exceptions = NULL;
set_map(NULL);
debug_only(_sp = -99);
debug_only(set_bci(-99));
}
//---------------------------clean_stack---------------------------------------
// Clear away rubbish from the stack area of the JVM state.
// This destroys any arguments that may be waiting on the stack.
void GraphKit::clean_stack(int from_sp) {
SafePointNode* map = this->map();
JVMState* jvms = this->jvms();
int stk_size = jvms->stk_size();
int stkoff = jvms->stkoff();
Node* top = this->top();
for (int i = from_sp; i < stk_size; i++) {
if (map->in(stkoff + i) != top) {
map->set_req(stkoff + i, top);
}
}
}
//--------------------------------sync_jvms-----------------------------------
// Make sure our current jvms agrees with our parse state.
JVMState* GraphKit::sync_jvms() const {
JVMState* jvms = this->jvms();
jvms->set_bci(bci()); // Record the new bci in the JVMState
jvms->set_sp(sp()); // Record the new sp in the JVMState
assert(jvms_in_sync(), "jvms is now in sync");
return jvms;
}
//--------------------------------sync_jvms_for_reexecute---------------------
// Make sure our current jvms agrees with our parse state. This version
// uses the reexecute_sp for reexecuting bytecodes.
JVMState* GraphKit::sync_jvms_for_reexecute() {
JVMState* jvms = this->jvms();
jvms->set_bci(bci()); // Record the new bci in the JVMState
jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
return jvms;
}
#ifdef ASSERT
bool GraphKit::jvms_in_sync() const {
Parse* parse = is_Parse();
if (parse == NULL) {
if (bci() != jvms()->bci()) return false;
if (sp() != (int)jvms()->sp()) return false;
return true;
}
if (jvms()->method() != parse->method()) return false;
if (jvms()->bci() != parse->bci()) return false;
int jvms_sp = jvms()->sp();
if (jvms_sp != parse->sp()) return false;
int jvms_depth = jvms()->depth();
if (jvms_depth != parse->depth()) return false;
return true;
}
// Local helper checks for special internal merge points
// used to accumulate and merge exception states.
// They are marked by the region's in(0) edge being the map itself.
// Such merge points must never "escape" into the parser at large,
// until they have been handed to gvn.transform.
static bool is_hidden_merge(Node* reg) {
if (reg == NULL) return false;
if (reg->is_Phi()) {
reg = reg->in(0);
if (reg == NULL) return false;
}
return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
}
void GraphKit::verify_map() const {
if (map() == NULL) return; // null map is OK
assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
}
void GraphKit::verify_exception_state(SafePointNode* ex_map) {
assert(ex_map->next_exception() == NULL, "not already part of a chain");
assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
}
#endif
//---------------------------stop_and_kill_map---------------------------------
// Set _map to NULL, signalling a stop to further bytecode execution.
// First smash the current map's control to a constant, to mark it dead.
void GraphKit::stop_and_kill_map() {
SafePointNode* dead_map = stop();
if (dead_map != NULL) {
dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
assert(dead_map->is_killed(), "must be so marked");
}
}
//--------------------------------stopped--------------------------------------
// Tell if _map is NULL, or control is top.
bool GraphKit::stopped() {
if (map() == NULL) return true;
else if (control() == top()) return true;
else return false;
}
//-----------------------------has_ex_handler----------------------------------
// Tell if this method or any caller method has exception handlers.
bool GraphKit::has_ex_handler() {
for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
return true;
}
}
return false;
}
//------------------------------save_ex_oop------------------------------------
// Save an exception without blowing stack contents or other JVM state.
void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
ex_map->add_req(ex_oop);
debug_only(verify_exception_state(ex_map));
}
inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
Node* ex_oop = ex_map->in(ex_map->req()-1);
if (clear_it) ex_map->del_req(ex_map->req()-1);
return ex_oop;
}
//-----------------------------saved_ex_oop------------------------------------
// Recover a saved exception from its map.
Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
return common_saved_ex_oop(ex_map, false);
}
//--------------------------clear_saved_ex_oop---------------------------------
// Erase a previously saved exception from its map.
Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
return common_saved_ex_oop(ex_map, true);
}
#ifdef ASSERT
//---------------------------has_saved_ex_oop----------------------------------
// Erase a previously saved exception from its map.
bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
return ex_map->req() == ex_map->jvms()->endoff()+1;
}
#endif
//-------------------------make_exception_state--------------------------------
// Turn the current JVM state into an exception state, appending the ex_oop.
SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
sync_jvms();
SafePointNode* ex_map = stop(); // do not manipulate this map any more
set_saved_ex_oop(ex_map, ex_oop);
return ex_map;
}
//--------------------------add_exception_state--------------------------------
// Add an exception to my list of exceptions.
void GraphKit::add_exception_state(SafePointNode* ex_map) {
if (ex_map == NULL || ex_map->control() == top()) {
return;
}
#ifdef ASSERT
verify_exception_state(ex_map);
if (has_exceptions()) {
assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
}
#endif
// If there is already an exception of exactly this type, merge with it.
// In particular, null-checks and other low-level exceptions common up here.
Node* ex_oop = saved_ex_oop(ex_map);
const Type* ex_type = _gvn.type(ex_oop);
if (ex_oop == top()) {
// No action needed.
return;
}
assert(ex_type->isa_instptr(), "exception must be an instance");
for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
// We check sp also because call bytecodes can generate exceptions
// both before and after arguments are popped!
if (ex_type2 == ex_type
&& e2->_jvms->sp() == ex_map->_jvms->sp()) {
combine_exception_states(ex_map, e2);
return;
}
}
// No pre-existing exception of the same type. Chain it on the list.
push_exception_state(ex_map);
}
//-----------------------add_exception_states_from-----------------------------
void GraphKit::add_exception_states_from(JVMState* jvms) {
SafePointNode* ex_map = jvms->map()->next_exception();
if (ex_map != NULL) {
jvms->map()->set_next_exception(NULL);
for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
next_map = ex_map->next_exception();
ex_map->set_next_exception(NULL);
add_exception_state(ex_map);
}
}
}
//-----------------------transfer_exceptions_into_jvms-------------------------
JVMState* GraphKit::transfer_exceptions_into_jvms() {
if (map() == NULL) {
// We need a JVMS to carry the exceptions, but the map has gone away.
// Create a scratch JVMS, cloned from any of the exception states...
if (has_exceptions()) {
_map = _exceptions;
_map = clone_map();
_map->set_next_exception(NULL);
clear_saved_ex_oop(_map);
debug_only(verify_map());
} else {
// ...or created from scratch
JVMState* jvms = new (C) JVMState(_method, NULL);
jvms->set_bci(_bci);
jvms->set_sp(_sp);
jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
set_jvms(jvms);
for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
set_all_memory(top());
while (map()->req() < jvms->endoff()) map()->add_req(top());
}
// (This is a kludge, in case you didn't notice.)
set_control(top());
}
JVMState* jvms = sync_jvms();
assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
jvms->map()->set_next_exception(_exceptions);
_exceptions = NULL; // done with this set of exceptions
return jvms;
}
static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
assert(is_hidden_merge(dstphi), "must be a special merge node");
assert(is_hidden_merge(srcphi), "must be a special merge node");
uint limit = srcphi->req();
for (uint i = PhiNode::Input; i < limit; i++) {
dstphi->add_req(srcphi->in(i));
}
}
static inline void add_one_req(Node* dstphi, Node* src) {
assert(is_hidden_merge(dstphi), "must be a special merge node");
assert(!is_hidden_merge(src), "must not be a special merge node");
dstphi->add_req(src);
}
//-----------------------combine_exception_states------------------------------
// This helper function combines exception states by building phis on a
// specially marked state-merging region. These regions and phis are
// untransformed, and can build up gradually. The region is marked by
// having a control input of its exception map, rather than NULL. Such
// regions do not appear except in this function, and in use_exception_state.
void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
if (failing()) return; // dying anyway...
JVMState* ex_jvms = ex_map->_jvms;
assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
assert(ex_map->req() == phi_map->req(), "matching maps");
uint tos = ex_jvms->stkoff() + ex_jvms->sp();
Node* hidden_merge_mark = root();
Node* region = phi_map->control();
MergeMemNode* phi_mem = phi_map->merged_memory();
MergeMemNode* ex_mem = ex_map->merged_memory();
if (region->in(0) != hidden_merge_mark) {
// The control input is not (yet) a specially-marked region in phi_map.
// Make it so, and build some phis.
region = new (C) RegionNode(2);
_gvn.set_type(region, Type::CONTROL);
region->set_req(0, hidden_merge_mark); // marks an internal ex-state
region->init_req(1, phi_map->control());
phi_map->set_control(region);
Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
record_for_igvn(io_phi);
_gvn.set_type(io_phi, Type::ABIO);
phi_map->set_i_o(io_phi);
for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
Node* m = mms.memory();
Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
record_for_igvn(m_phi);
_gvn.set_type(m_phi, Type::MEMORY);
mms.set_memory(m_phi);
}
}
// Either or both of phi_map and ex_map might already be converted into phis.
Node* ex_control = ex_map->control();
// if there is special marking on ex_map also, we add multiple edges from src
bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
// how wide was the destination phi_map, originally?
uint orig_width = region->req();
if (add_multiple) {
add_n_reqs(region, ex_control);
add_n_reqs(phi_map->i_o(), ex_map->i_o());
} else {
// ex_map has no merges, so we just add single edges everywhere
add_one_req(region, ex_control);
add_one_req(phi_map->i_o(), ex_map->i_o());
}
for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
if (mms.is_empty()) {
// get a copy of the base memory, and patch some inputs into it
const TypePtr* adr_type = mms.adr_type(C);
Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
mms.set_memory(phi);
// Prepare to append interesting stuff onto the newly sliced phi:
while (phi->req() > orig_width) phi->del_req(phi->req()-1);
}
// Append stuff from ex_map:
if (add_multiple) {
add_n_reqs(mms.memory(), mms.memory2());
} else {
add_one_req(mms.memory(), mms.memory2());
}
}
uint limit = ex_map->req();
for (uint i = TypeFunc::Parms; i < limit; i++) {
// Skip everything in the JVMS after tos. (The ex_oop follows.)
if (i == tos) i = ex_jvms->monoff();
Node* src = ex_map->in(i);
Node* dst = phi_map->in(i);
if (src != dst) {
PhiNode* phi;
if (dst->in(0) != region) {
dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
record_for_igvn(phi);
_gvn.set_type(phi, phi->type());
phi_map->set_req(i, dst);
// Prepare to append interesting stuff onto the new phi:
while (dst->req() > orig_width) dst->del_req(dst->req()-1);
} else {
assert(dst->is_Phi(), "nobody else uses a hidden region");
phi = dst->as_Phi();
}
if (add_multiple && src->in(0) == ex_control) {
// Both are phis.
add_n_reqs(dst, src);
} else {
while (dst->req() < region->req()) add_one_req(dst, src);
}
const Type* srctype = _gvn.type(src);
if (phi->type() != srctype) {
const Type* dsttype = phi->type()->meet_speculative(srctype);
if (phi->type() != dsttype) {
phi->set_type(dsttype);
_gvn.set_type(phi, dsttype);
}
}
}
}
phi_map->merge_replaced_nodes_with(ex_map);
}
//--------------------------use_exception_state--------------------------------
Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
if (failing()) { stop(); return top(); }
Node* region = phi_map->control();
Node* hidden_merge_mark = root();
assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
Node* ex_oop = clear_saved_ex_oop(phi_map);
if (region->in(0) == hidden_merge_mark) {
// Special marking for internal ex-states. Process the phis now.
region->set_req(0, region); // now it's an ordinary region
set_jvms(phi_map->jvms()); // ...so now we can use it as a map
// Note: Setting the jvms also sets the bci and sp.
set_control(_gvn.transform(region));
uint tos = jvms()->stkoff() + sp();
for (uint i = 1; i < tos; i++) {
Node* x = phi_map->in(i);
if (x->in(0) == region) {
assert(x->is_Phi(), "expected a special phi");
phi_map->set_req(i, _gvn.transform(x));
}
}
for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
Node* x = mms.memory();
if (x->in(0) == region) {
assert(x->is_Phi(), "nobody else uses a hidden region");
mms.set_memory(_gvn.transform(x));
}
}
if (ex_oop->in(0) == region) {
assert(ex_oop->is_Phi(), "expected a special phi");
ex_oop = _gvn.transform(ex_oop);
}
} else {
set_jvms(phi_map->jvms());
}
assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
return ex_oop;
}
//---------------------------------java_bc-------------------------------------
Bytecodes::Code GraphKit::java_bc() const {
ciMethod* method = this->method();
int bci = this->bci();
if (method != NULL && bci != InvocationEntryBci)
return method->java_code_at_bci(bci);
else
return Bytecodes::_illegal;
}
void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
bool must_throw) {
// if the exception capability is set, then we will generate code
// to check the JavaThread.should_post_on_exceptions flag to see
// if we actually need to report exception events (for this
// thread). If we don't need to report exception events, we will
// take the normal fast path provided by add_exception_events. If
// exception event reporting is enabled for this thread, we will
// take the uncommon_trap in the BuildCutout below.
// first must access the should_post_on_exceptions_flag in this thread's JavaThread
Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
// Test the should_post_on_exceptions_flag vs. 0
Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
// Branch to slow_path if should_post_on_exceptions_flag was true
{ BuildCutout unless(this, tst, PROB_MAX);
// Do not try anything fancy if we're notifying the VM on every throw.
// Cf. case Bytecodes::_athrow in parse2.cpp.
uncommon_trap(reason, Deoptimization::Action_none,
(ciKlass*)NULL, (char*)NULL, must_throw);
}
}
//------------------------------builtin_throw----------------------------------
void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
bool must_throw = true;
if (env()->jvmti_can_post_on_exceptions()) {
// check if we must post exception events, take uncommon trap if so
uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
// here if should_post_on_exceptions is false
// continue on with the normal codegen
}
// If this particular condition has not yet happened at this
// bytecode, then use the uncommon trap mechanism, and allow for
// a future recompilation if several traps occur here.
// If the throw is hot, try to use a more complicated inline mechanism
// which keeps execution inside the compiled code.
bool treat_throw_as_hot = false;
ciMethodData* md = method()->method_data();
if (ProfileTraps) {
if (too_many_traps(reason)) {
treat_throw_as_hot = true;
}
// (If there is no MDO at all, assume it is early in
// execution, and that any deopts are part of the
// startup transient, and don't need to be remembered.)
// Also, if there is a local exception handler, treat all throws
// as hot if there has been at least one in this method.
if (C->trap_count(reason) != 0
&& method()->method_data()->trap_count(reason) != 0
&& has_ex_handler()) {
treat_throw_as_hot = true;
}
}
// If this throw happens frequently, an uncommon trap might cause
// a performance pothole. If there is a local exception handler,
// and if this particular bytecode appears to be deoptimizing often,
// let us handle the throw inline, with a preconstructed instance.
// Note: If the deopt count has blown up, the uncommon trap
// runtime is going to flush this nmethod, not matter what.
if (treat_throw_as_hot
&& (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
// If the throw is local, we use a pre-existing instance and
// punt on the backtrace. This would lead to a missing backtrace
// (a repeat of 4292742) if the backtrace object is ever asked
// for its backtrace.
// Fixing this remaining case of 4292742 requires some flavor of
// escape analysis. Leave that for the future.
ciInstance* ex_obj = NULL;
switch (reason) {
case Deoptimization::Reason_null_check:
ex_obj = env()->NullPointerException_instance();
break;
case Deoptimization::Reason_div0_check:
ex_obj = env()->ArithmeticException_instance();
break;
case Deoptimization::Reason_range_check:
ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
break;
case Deoptimization::Reason_class_check:
if (java_bc() == Bytecodes::_aastore) {
ex_obj = env()->ArrayStoreException_instance();
} else {
ex_obj = env()->ClassCastException_instance();
}
break;
}
if (failing()) { stop(); return; } // exception allocation might fail
if (ex_obj != NULL) {
// Cheat with a preallocated exception object.
if (C->log() != NULL)
C->log()->elem("hot_throw preallocated='1' reason='%s'",
Deoptimization::trap_reason_name(reason));
const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
// Clear the detail message of the preallocated exception object.
// Weblogic sometimes mutates the detail message of exceptions
// using reflection.
int offset = java_lang_Throwable::get_detailMessage_offset();
const TypePtr* adr_typ = ex_con->add_offset(offset);
Node *adr = basic_plus_adr(ex_node, ex_node, offset);
const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
// Conservatively release stores of object references.
Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
add_exception_state(make_exception_state(ex_node));
return;
}
}
// %%% Maybe add entry to OptoRuntime which directly throws the exc.?
// It won't be much cheaper than bailing to the interp., since we'll
// have to pass up all the debug-info, and the runtime will have to
// create the stack trace.
// Usual case: Bail to interpreter.
// Reserve the right to recompile if we haven't seen anything yet.
assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
if (treat_throw_as_hot
&& (method()->method_data()->trap_recompiled_at(bci(), NULL)
|| C->too_many_traps(reason))) {
// We cannot afford to take more traps here. Suffer in the interpreter.
if (C->log() != NULL)
C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
Deoptimization::trap_reason_name(reason),
C->trap_count(reason));
action = Deoptimization::Action_none;
}
// "must_throw" prunes the JVM state to include only the stack, if there
// are no local exception handlers. This should cut down on register
// allocation time and code size, by drastically reducing the number
// of in-edges on the call to the uncommon trap.
uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
}
//----------------------------PreserveJVMState---------------------------------
PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
debug_only(kit->verify_map());
_kit = kit;
_map = kit->map(); // preserve the map
_sp = kit->sp();
kit->set_map(clone_map ? kit->clone_map() : NULL);
#ifdef ASSERT
_bci = kit->bci();
Parse* parser = kit->is_Parse();
int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
_block = block;
#endif
}
PreserveJVMState::~PreserveJVMState() {
GraphKit* kit = _kit;
#ifdef ASSERT
assert(kit->bci() == _bci, "bci must not shift");
Parse* parser = kit->is_Parse();
int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
assert(block == _block, "block must not shift");
#endif
kit->set_map(_map);
kit->set_sp(_sp);
}
//-----------------------------BuildCutout-------------------------------------
BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
: PreserveJVMState(kit)
{
assert(p->is_Con() || p->is_Bool(), "test must be a bool");
SafePointNode* outer_map = _map; // preserved map is caller's
SafePointNode* inner_map = kit->map();
IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
}
BuildCutout::~BuildCutout() {
GraphKit* kit = _kit;
assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
}
//---------------------------PreserveReexecuteState----------------------------
PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
assert(!kit->stopped(), "must call stopped() before");
_kit = kit;
_sp = kit->sp();
_reexecute = kit->jvms()->_reexecute;
}
PreserveReexecuteState::~PreserveReexecuteState() {
if (_kit->stopped()) return;
_kit->jvms()->_reexecute = _reexecute;
_kit->set_sp(_sp);
}
//------------------------------clone_map--------------------------------------
// Implementation of PreserveJVMState
//
// Only clone_map(...) here. If this function is only used in the
// PreserveJVMState class we may want to get rid of this extra
// function eventually and do it all there.
SafePointNode* GraphKit::clone_map() {
if (map() == NULL) return NULL;
// Clone the memory edge first
Node* mem = MergeMemNode::make(C, map()->memory());
gvn().set_type_bottom(mem);
SafePointNode *clonemap = (SafePointNode*)map()->clone();
JVMState* jvms = this->jvms();
JVMState* clonejvms = jvms->clone_shallow(C);
clonemap->set_memory(mem);
clonemap->set_jvms(clonejvms);
clonejvms->set_map(clonemap);
record_for_igvn(clonemap);
gvn().set_type_bottom(clonemap);
return clonemap;
}
//-----------------------------set_map_clone-----------------------------------
void GraphKit::set_map_clone(SafePointNode* m) {
_map = m;
_map = clone_map();
_map->set_next_exception(NULL);
debug_only(verify_map());
}
//----------------------------kill_dead_locals---------------------------------
// Detect any locals which are known to be dead, and force them to top.
void GraphKit::kill_dead_locals() {
// Consult the liveness information for the locals. If any
// of them are unused, then they can be replaced by top(). This
// should help register allocation time and cut down on the size
// of the deoptimization information.
// This call is made from many of the bytecode handling
// subroutines called from the Big Switch in do_one_bytecode.
// Every bytecode which might include a slow path is responsible
// for killing its dead locals. The more consistent we
// are about killing deads, the fewer useless phis will be
// constructed for them at various merge points.
// bci can be -1 (InvocationEntryBci). We return the entry
// liveness for the method.
if (method() == NULL || method()->code_size() == 0) {
// We are building a graph for a call to a native method.
// All locals are live.
return;
}
ResourceMark rm;
// Consult the liveness information for the locals. If any
// of them are unused, then they can be replaced by top(). This
// should help register allocation time and cut down on the size
// of the deoptimization information.
MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
int len = (int)live_locals.size();
assert(len <= jvms()->loc_size(), "too many live locals");
for (int local = 0; local < len; local++) {
if (!live_locals.at(local)) {
set_local(local, top());
}
}
}
#ifdef ASSERT
//-------------------------dead_locals_are_killed------------------------------
// Return true if all dead locals are set to top in the map.
// Used to assert "clean" debug info at various points.
bool GraphKit::dead_locals_are_killed() {
if (method() == NULL || method()->code_size() == 0) {
// No locals need to be dead, so all is as it should be.
return true;
}
// Make sure somebody called kill_dead_locals upstream.
ResourceMark rm;
for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
if (jvms->loc_size() == 0) continue; // no locals to consult
SafePointNode* map = jvms->map();
ciMethod* method = jvms->method();
int bci = jvms->bci();
if (jvms == this->jvms()) {
bci = this->bci(); // it might not yet be synched
}
MethodLivenessResult live_locals = method->liveness_at_bci(bci);
int len = (int)live_locals.size();
if (!live_locals.is_valid() || len == 0)
// This method is trivial, or is poisoned by a breakpoint.
return true;
assert(len == jvms->loc_size(), "live map consistent with locals map");
for (int local = 0; local < len; local++) {
if (!live_locals.at(local) && map->local(jvms, local) != top()) {
if (PrintMiscellaneous && (Verbose || WizardMode)) {
tty->print_cr("Zombie local %d: ", local);
jvms->dump();
}
return false;
}
}
}
return true;
}
#endif //ASSERT
// Helper function for enforcing certain bytecodes to reexecute if
// deoptimization happens
static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
ciMethod* cur_method = jvms->method();
int cur_bci = jvms->bci();
if (cur_method != NULL && cur_bci != InvocationEntryBci) {
Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
return Interpreter::bytecode_should_reexecute(code) ||
is_anewarray && code == Bytecodes::_multianewarray;
// Reexecute _multianewarray bytecode which was replaced with
// sequence of [a]newarray. See Parse::do_multianewarray().
//
// Note: interpreter should not have it set since this optimization
// is limited by dimensions and guarded by flag so in some cases
// multianewarray() runtime calls will be generated and
// the bytecode should not be reexecutes (stack will not be reset).
} else
return false;
}
// Helper function for adding JVMState and debug information to node
void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
// Add the safepoint edges to the call (or other safepoint).
// Make sure dead locals are set to top. This
// should help register allocation time and cut down on the size
// of the deoptimization information.
assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
// Walk the inline list to fill in the correct set of JVMState's
// Also fill in the associated edges for each JVMState.
// If the bytecode needs to be reexecuted we need to put
// the arguments back on the stack.
const bool should_reexecute = jvms()->should_reexecute();
JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
// NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
// undefined if the bci is different. This is normal for Parse but it
// should not happen for LibraryCallKit because only one bci is processed.
assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
"in LibraryCallKit the reexecute bit should not change");
// If we are guaranteed to throw, we can prune everything but the
// input to the current bytecode.
bool can_prune_locals = false;
uint stack_slots_not_pruned = 0;
int inputs = 0, depth = 0;
if (must_throw) {
assert(method() == youngest_jvms->method(), "sanity");
if (compute_stack_effects(inputs, depth)) {
can_prune_locals = true;
stack_slots_not_pruned = inputs;
}
}
if (env()->should_retain_local_variables()) {
// At any safepoint, this method can get breakpointed, which would
// then require an immediate deoptimization.
can_prune_locals = false; // do not prune locals
stack_slots_not_pruned = 0;
}
// do not scribble on the input jvms
JVMState* out_jvms = youngest_jvms->clone_deep(C);
call->set_jvms(out_jvms); // Start jvms list for call node
// For a known set of bytecodes, the interpreter should reexecute them if
// deoptimization happens. We set the reexecute state for them here
if (out_jvms->is_reexecute_undefined() && //don't change if already specified
should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
}
// Presize the call:
DEBUG_ONLY(uint non_debug_edges = call->req());
call->add_req_batch(top(), youngest_jvms->debug_depth());
assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
// Set up edges so that the call looks like this:
// Call [state:] ctl io mem fptr retadr
// [parms:] parm0 ... parmN
// [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
// [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
// [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
// Note that caller debug info precedes callee debug info.
// Fill pointer walks backwards from "young:" to "root:" in the diagram above:
uint debug_ptr = call->req();
// Loop over the map input edges associated with jvms, add them
// to the call node, & reset all offsets to match call node array.
for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
uint debug_end = debug_ptr;
uint debug_start = debug_ptr - in_jvms->debug_size();
debug_ptr = debug_start; // back up the ptr
uint p = debug_start; // walks forward in [debug_start, debug_end)
uint j, k, l;
SafePointNode* in_map = in_jvms->map();
out_jvms->set_map(call);
if (can_prune_locals) {
assert(in_jvms->method() == out_jvms->method(), "sanity");
// If the current throw can reach an exception handler in this JVMS,
// then we must keep everything live that can reach that handler.
// As a quick and dirty approximation, we look for any handlers at all.
if (in_jvms->method()->has_exception_handlers()) {
can_prune_locals = false;
}
}
// Add the Locals
k = in_jvms->locoff();
l = in_jvms->loc_size();
out_jvms->set_locoff(p);
if (!can_prune_locals) {
for (j = 0; j < l; j++)
call->set_req(p++, in_map->in(k+j));
} else {
p += l; // already set to top above by add_req_batch
}
// Add the Expression Stack
k = in_jvms->stkoff();
l = in_jvms->sp();
out_jvms->set_stkoff(p);
if (!can_prune_locals) {
for (j = 0; j < l; j++)
call->set_req(p++, in_map->in(k+j));
} else if (can_prune_locals && stack_slots_not_pruned != 0) {
// Divide stack into {S0,...,S1}, where S0 is set to top.
uint s1 = stack_slots_not_pruned;
stack_slots_not_pruned = 0; // for next iteration
if (s1 > l) s1 = l;
uint s0 = l - s1;
p += s0; // skip the tops preinstalled by add_req_batch
for (j = s0; j < l; j++)
call->set_req(p++, in_map->in(k+j));
} else {
p += l; // already set to top above by add_req_batch
}
// Add the Monitors
k = in_jvms->monoff();
l = in_jvms->mon_size();
out_jvms->set_monoff(p);
for (j = 0; j < l; j++)
call->set_req(p++, in_map->in(k+j));
// Copy any scalar object fields.
k = in_jvms->scloff();
l = in_jvms->scl_size();
out_jvms->set_scloff(p);
for (j = 0; j < l; j++)
call->set_req(p++, in_map->in(k+j));
// Finish the new jvms.
out_jvms->set_endoff(p);
assert(out_jvms->endoff() == debug_end, "fill ptr must match");
assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
// Update the two tail pointers in parallel.
out_jvms = out_jvms->caller();
in_jvms = in_jvms->caller();
}
assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
// Test the correctness of JVMState::debug_xxx accessors:
assert(call->jvms()->debug_start() == non_debug_edges, "");
assert(call->jvms()->debug_end() == call->req(), "");
assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
}
bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
Bytecodes::Code code = java_bc();
if (code == Bytecodes::_wide) {
code = method()->java_code_at_bci(bci() + 1);
}
BasicType rtype = T_ILLEGAL;
int rsize = 0;
if (code != Bytecodes::_illegal) {
depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
if (rtype < T_CONFLICT)
rsize = type2size[rtype];
}
switch (code) {
case Bytecodes::_illegal:
return false;
case Bytecodes::_ldc:
case Bytecodes::_ldc_w:
case Bytecodes::_ldc2_w:
inputs = 0;
break;
case Bytecodes::_dup: inputs = 1; break;
case Bytecodes::_dup_x1: inputs = 2; break;
case Bytecodes::_dup_x2: inputs = 3; break;
case Bytecodes::_dup2: inputs = 2; break;
case Bytecodes::_dup2_x1: inputs = 3; break;
case Bytecodes::_dup2_x2: inputs = 4; break;
case Bytecodes::_swap: inputs = 2; break;
case Bytecodes::_arraylength: inputs = 1; break;
case Bytecodes::_getstatic:
case Bytecodes::_putstatic:
case Bytecodes::_getfield:
case Bytecodes::_putfield:
{
bool ignored_will_link;
ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
int size = field->type()->size();
bool is_get = (depth >= 0), is_static = (depth & 1);
inputs = (is_static ? 0 : 1);
if (is_get) {
depth = size - inputs;
} else {
inputs += size; // putxxx pops the value from the stack
depth = - inputs;
}
}
break;
case Bytecodes::_invokevirtual:
case Bytecodes::_invokespecial:
case Bytecodes::_invokestatic:
case Bytecodes::_invokedynamic:
case Bytecodes::_invokeinterface:
{
bool ignored_will_link;
ciSignature* declared_signature = NULL;
ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
assert(declared_signature != NULL, "cannot be null");
inputs = declared_signature->arg_size_for_bc(code);
int size = declared_signature->return_type()->size();
depth = size - inputs;
}
break;
case Bytecodes::_multianewarray:
{
ciBytecodeStream iter(method());
iter.reset_to_bci(bci());
iter.next();
inputs = iter.get_dimensions();
assert(rsize == 1, "");
depth = rsize - inputs;
}
break;
case Bytecodes::_ireturn:
case Bytecodes::_lreturn:
case Bytecodes::_freturn:
case Bytecodes::_dreturn:
case Bytecodes::_areturn:
assert(rsize = -depth, "");
inputs = rsize;
break;
case Bytecodes::_jsr:
case Bytecodes::_jsr_w:
inputs = 0;
depth = 1; // S.B. depth=1, not zero
break;
default:
// bytecode produces a typed result
inputs = rsize - depth;
assert(inputs >= 0, "");
break;
}
#ifdef ASSERT
// spot check
int outputs = depth + inputs;
assert(outputs >= 0, "sanity");
switch (code) {
case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
}
#endif //ASSERT
return true;
}
//------------------------------basic_plus_adr---------------------------------
Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
// short-circuit a common case
if (offset == intcon(0)) return ptr;
return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
}
Node* GraphKit::ConvI2L(Node* offset) {
// short-circuit a common case
jint offset_con = find_int_con(offset, Type::OffsetBot);
if (offset_con != Type::OffsetBot) {
return longcon((jlong) offset_con);
}
return _gvn.transform( new (C) ConvI2LNode(offset));
}
Node* GraphKit::ConvI2UL(Node* offset) {
juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
if (offset_con != (juint) Type::OffsetBot) {
return longcon((julong) offset_con);
}
Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
return _gvn.transform( new (C) AndLNode(conv, mask) );
}
Node* GraphKit::ConvL2I(Node* offset) {
// short-circuit a common case
jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
if (offset_con != (jlong)Type::OffsetBot) {
return intcon((int) offset_con);
}
return _gvn.transform( new (C) ConvL2INode(offset));
}
//-------------------------load_object_klass-----------------------------------
Node* GraphKit::load_object_klass(Node* obj) {
// Special-case a fresh allocation to avoid building nodes:
Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
if (akls != NULL) return akls;
Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
}
//-------------------------load_array_length-----------------------------------
Node* GraphKit::load_array_length(Node* array) {
// Special-case a fresh allocation to avoid building nodes:
AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
Node *alen;
if (alloc == NULL) {
Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
} else {
alen = alloc->Ideal_length();
Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
if (ccast != alen) {
alen = _gvn.transform(ccast);
}
}
return alen;
}
//------------------------------do_null_check----------------------------------
// Helper function to do a NULL pointer check. Returned value is
// the incoming address with NULL casted away. You are allowed to use the
// not-null value only if you are control dependent on the test.
extern int explicit_null_checks_inserted,
explicit_null_checks_elided;
Node* GraphKit::null_check_common(Node* value, BasicType type,
// optional arguments for variations:
bool assert_null,
Node* *null_control) {
assert(!assert_null || null_control == NULL, "not both at once");
if (stopped()) return top();
if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
// For some performance testing, we may wish to suppress null checking.
value = cast_not_null(value); // Make it appear to be non-null (4962416).
return value;
}
explicit_null_checks_inserted++;
// Construct NULL check
Node *chk = NULL;
switch(type) {
case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
case T_ARRAY : // fall through
type = T_OBJECT; // simplify further tests
case T_OBJECT : {
const Type *t = _gvn.type( value );
const TypeOopPtr* tp = t->isa_oopptr();
if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
// Only for do_null_check, not any of its siblings:
&& !assert_null && null_control == NULL) {
// Usually, any field access or invocation on an unloaded oop type
// will simply fail to link, since the statically linked class is
// likely also to be unloaded. However, in -Xcomp mode, sometimes
// the static class is loaded but the sharper oop type is not.
// Rather than checking for this obscure case in lots of places,
// we simply observe that a null check on an unloaded class
// will always be followed by a nonsense operation, so we
// can just issue the uncommon trap here.
// Our access to the unloaded class will only be correct
// after it has been loaded and initialized, which requires
// a trip through the interpreter.
#ifndef PRODUCT
if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
#endif
uncommon_trap(Deoptimization::Reason_unloaded,
Deoptimization::Action_reinterpret,
tp->klass(), "!loaded");
return top();
}
if (assert_null) {
// See if the type is contained in NULL_PTR.
// If so, then the value is already null.
if (t->higher_equal(TypePtr::NULL_PTR)) {
explicit_null_checks_elided++;
return value; // Elided null assert quickly!
}
} else {
// See if mixing in the NULL pointer changes type.
// If so, then the NULL pointer was not allowed in the original
// type. In other words, "value" was not-null.
if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
// same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
explicit_null_checks_elided++;
return value; // Elided null check quickly!
}
}
chk = new (C) CmpPNode( value, null() );
break;
}
default:
fatal(err_msg_res("unexpected type: %s", type2name(type)));
}
assert(chk != NULL, "sanity check");
chk = _gvn.transform(chk);
BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
BoolNode *btst = new (C) BoolNode( chk, btest);
Node *tst = _gvn.transform( btst );
//-----------
// if peephole optimizations occurred, a prior test existed.
// If a prior test existed, maybe it dominates as we can avoid this test.
if (tst != btst && type == T_OBJECT) {
// At this point we want to scan up the CFG to see if we can
// find an identical test (and so avoid this test altogether).
Node *cfg = control();
int depth = 0;
while( depth < 16 ) { // Limit search depth for speed
if( cfg->Opcode() == Op_IfTrue &&
cfg->in(0)->in(1) == tst ) {
// Found prior test. Use "cast_not_null" to construct an identical
// CastPP (and hence hash to) as already exists for the prior test.
// Return that casted value.
if (assert_null) {
replace_in_map(value, null());
return null(); // do not issue the redundant test
}
Node *oldcontrol = control();
set_control(cfg);
Node *res = cast_not_null(value);
set_control(oldcontrol);
explicit_null_checks_elided++;
return res;
}
cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
if (cfg == NULL) break; // Quit at region nodes
depth++;
}
}
//-----------
// Branch to failure if null
float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
Deoptimization::DeoptReason reason;
if (assert_null)
reason = Deoptimization::Reason_null_assert;
else if (type == T_OBJECT)
reason = Deoptimization::Reason_null_check;
else
reason = Deoptimization::Reason_div0_check;
// %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
// ciMethodData::has_trap_at will return a conservative -1 if any
// must-be-null assertion has failed. This could cause performance
// problems for a method after its first do_null_assert failure.
// Consider using 'Reason_class_check' instead?
// To cause an implicit null check, we set the not-null probability
// to the maximum (PROB_MAX). For an explicit check the probability
// is set to a smaller value.
if (null_control != NULL || too_many_traps(reason)) {
// probability is less likely
ok_prob = PROB_LIKELY_MAG(3);
} else if (!assert_null &&
(ImplicitNullCheckThreshold > 0) &&
method() != NULL &&
(method()->method_data()->trap_count(reason)
>= (uint)ImplicitNullCheckThreshold)) {
ok_prob = PROB_LIKELY_MAG(3);
}
if (null_control != NULL) {
IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
set_control( _gvn.transform( new (C) IfTrueNode(iff)));
if (null_true == top())
explicit_null_checks_elided++;
(*null_control) = null_true;
} else {
BuildCutout unless(this, tst, ok_prob);
// Check for optimizer eliding test at parse time
if (stopped()) {
// Failure not possible; do not bother making uncommon trap.
explicit_null_checks_elided++;
} else if (assert_null) {
uncommon_trap(reason,
Deoptimization::Action_make_not_entrant,
NULL, "assert_null");
} else {
replace_in_map(value, zerocon(type));
builtin_throw(reason);
}
}
// Must throw exception, fall-thru not possible?
if (stopped()) {
return top(); // No result
}
if (assert_null) {
// Cast obj to null on this path.
replace_in_map(value, zerocon(type));
return zerocon(type);
}
// Cast obj to not-null on this path, if there is no null_control.
// (If there is a null_control, a non-null value may come back to haunt us.)
if (type == T_OBJECT) {
Node* cast = cast_not_null(value, false);
if (null_control == NULL || (*null_control) == top())
replace_in_map(value, cast);
value = cast;
}
return value;
}
//------------------------------cast_not_null----------------------------------
// Cast obj to not-null on this path
Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
const Type *t = _gvn.type(obj);
const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
// Object is already not-null?
if( t == t_not_null ) return obj;
Node *cast = new (C) CastPPNode(obj,t_not_null);
cast->init_req(0, control());
cast = _gvn.transform( cast );
// Scan for instances of 'obj' in the current JVM mapping.
// These instances are known to be not-null after the test.
if (do_replace_in_map)
replace_in_map(obj, cast);
return cast; // Return casted value
}
//--------------------------replace_in_map-------------------------------------
void GraphKit::replace_in_map(Node* old, Node* neww) {
if (old == neww) {
return;
}
map()->replace_edge(old, neww);
// Note: This operation potentially replaces any edge
// on the map. This includes locals, stack, and monitors
// of the current (innermost) JVM state.
// don't let inconsistent types from profiling escape this
// method
const Type* told = _gvn.type(old);
const Type* tnew = _gvn.type(neww);
if (!tnew->higher_equal(told)) {
return;
}
map()->record_replaced_node(old, neww);
}
//=============================================================================
//--------------------------------memory---------------------------------------
Node* GraphKit::memory(uint alias_idx) {
MergeMemNode* mem = merged_memory();
Node* p = mem->memory_at(alias_idx);
_gvn.set_type(p, Type::MEMORY); // must be mapped
return p;
}
//-----------------------------reset_memory------------------------------------
Node* GraphKit::reset_memory() {
Node* mem = map()->memory();
// do not use this node for any more parsing!
debug_only( map()->set_memory((Node*)NULL) );
return _gvn.transform( mem );
}
//------------------------------set_all_memory---------------------------------
void GraphKit::set_all_memory(Node* newmem) {
Node* mergemem = MergeMemNode::make(C, newmem);
gvn().set_type_bottom(mergemem);
map()->set_memory(mergemem);
}
//------------------------------set_all_memory_call----------------------------
void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
set_all_memory(newmem);
}
//=============================================================================
//
// parser factory methods for MemNodes
//
// These are layered on top of the factory methods in LoadNode and StoreNode,
// and integrate with the parser's memory state and _gvn engine.
//
// factory methods in "int adr_idx"
Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
int adr_idx,
MemNode::MemOrd mo,
LoadNode::ControlDependency control_dependency,
bool require_atomic_access,
bool unaligned,
bool mismatched) {
assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
const TypePtr* adr_type = NULL; // debug-mode-only argument
debug_only(adr_type = C->get_adr_type(adr_idx));
Node* mem = memory(adr_idx);
Node* ld;
if (require_atomic_access && bt == T_LONG) {
ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
} else if (require_atomic_access && bt == T_DOUBLE) {
ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
} else {
ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency);
}
if (unaligned) {
ld->as_Load()->set_unaligned_access();
}
if (mismatched) {
ld->as_Load()->set_mismatched_access();
}
ld = _gvn.transform(ld);
if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
// Improve graph before escape analysis and boxing elimination.
record_for_igvn(ld);
}
return ld;
}
Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
int adr_idx,
MemNode::MemOrd mo,
bool require_atomic_access,
bool unaligned,
bool mismatched) {
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
const TypePtr* adr_type = NULL;
debug_only(adr_type = C->get_adr_type(adr_idx));
Node *mem = memory(adr_idx);
Node* st;
if (require_atomic_access && bt == T_LONG) {
st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
} else if (require_atomic_access && bt == T_DOUBLE) {
st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
} else {
st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
}
if (unaligned) {
st->as_Store()->set_unaligned_access();
}
if (mismatched) {
st->as_Store()->set_mismatched_access();
}
st = _gvn.transform(st);
set_memory(st, adr_idx);
// Back-to-back stores can only remove intermediate store with DU info
// so push on worklist for optimizer.
if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
record_for_igvn(st);
return st;
}
void GraphKit::pre_barrier(bool do_load,
Node* ctl,
Node* obj,
Node* adr,
uint adr_idx,
Node* val,
const TypeOopPtr* val_type,
Node* pre_val,
BasicType bt) {
BarrierSet* bs = Universe::heap()->barrier_set();
set_control(ctl);
switch (bs->kind()) {
case BarrierSet::G1SATBCT:
case BarrierSet::G1SATBCTLogging:
g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
break;
case BarrierSet::CardTableModRef:
case BarrierSet::CardTableExtension:
case BarrierSet::ModRef:
break;
case BarrierSet::Other:
default :
ShouldNotReachHere();
}
}
bool GraphKit::can_move_pre_barrier() const {
BarrierSet* bs = Universe::heap()->barrier_set();
switch (bs->kind()) {
case BarrierSet::G1SATBCT:
case BarrierSet::G1SATBCTLogging:
return true; // Can move it if no safepoint
case BarrierSet::CardTableModRef:
case BarrierSet::CardTableExtension:
case BarrierSet::ModRef:
return true; // There is no pre-barrier
case BarrierSet::Other:
default :
ShouldNotReachHere();
}
return false;
}
void GraphKit::post_barrier(Node* ctl,
Node* store,
Node* obj,
Node* adr,
uint adr_idx,
Node* val,
BasicType bt,
bool use_precise) {
BarrierSet* bs = Universe::heap()->barrier_set();
set_control(ctl);
switch (bs->kind()) {
case BarrierSet::G1SATBCT:
case BarrierSet::G1SATBCTLogging:
g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
break;
case BarrierSet::CardTableModRef:
case BarrierSet::CardTableExtension:
write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
break;
case BarrierSet::ModRef:
break;
case BarrierSet::Other:
default :
ShouldNotReachHere();
}
}
Node* GraphKit::store_oop(Node* ctl,
Node* obj,
Node* adr,
const TypePtr* adr_type,
Node* val,
const TypeOopPtr* val_type,
BasicType bt,
bool use_precise,
MemNode::MemOrd mo,
bool mismatched) {
// Transformation of a value which could be NULL pointer (CastPP #NULL)
// could be delayed during Parse (for example, in adjust_map_after_if()).
// Execute transformation here to avoid barrier generation in such case.
if (_gvn.type(val) == TypePtr::NULL_PTR)
val = _gvn.makecon(TypePtr::NULL_PTR);
set_control(ctl);
if (stopped()) return top(); // Dead path ?
assert(bt == T_OBJECT, "sanity");
assert(val != NULL, "not dead path");
uint adr_idx = C->get_alias_index(adr_type);
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
pre_barrier(true /* do_load */,
control(), obj, adr, adr_idx, val, val_type,
NULL /* pre_val */,
bt);
Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
return store;
}
// Could be an array or object we don't know at compile time (unsafe ref.)
Node* GraphKit::store_oop_to_unknown(Node* ctl,
Node* obj, // containing obj
Node* adr, // actual adress to store val at
const TypePtr* adr_type,
Node* val,
BasicType bt,
MemNode::MemOrd mo,
bool mismatched) {
Compile::AliasType* at = C->alias_type(adr_type);
const TypeOopPtr* val_type = NULL;
if (adr_type->isa_instptr()) {
if (at->field() != NULL) {
// known field. This code is a copy of the do_put_xxx logic.
ciField* field = at->field();
if (!field->type()->is_loaded()) {
val_type = TypeInstPtr::BOTTOM;
} else {
val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
}
}
} else if (adr_type->isa_aryptr()) {
val_type = adr_type->is_aryptr()->elem()->make_oopptr();
}
if (val_type == NULL) {
val_type = TypeInstPtr::BOTTOM;
}
return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
}
//-------------------------array_element_address-------------------------
Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
const TypeInt* sizetype, Node* ctrl) {
uint shift = exact_log2(type2aelembytes(elembt));
uint header = arrayOopDesc::base_offset_in_bytes(elembt);
// short-circuit a common case (saves lots of confusing waste motion)
jint idx_con = find_int_con(idx, -1);
if (idx_con >= 0) {
intptr_t offset = header + ((intptr_t)idx_con << shift);
return basic_plus_adr(ary, offset);
}
// must be correct type for alignment purposes
Node* base = basic_plus_adr(ary, header);
#ifdef _LP64
// The scaled index operand to AddP must be a clean 64-bit value.
// Java allows a 32-bit int to be incremented to a negative
// value, which appears in a 64-bit register as a large
// positive number. Using that large positive number as an
// operand in pointer arithmetic has bad consequences.
// On the other hand, 32-bit overflow is rare, and the possibility
// can often be excluded, if we annotate the ConvI2L node with
// a type assertion that its value is known to be a small positive
// number. (The prior range check has ensured this.)
// This assertion is used by ConvI2LNode::Ideal.
int index_max = max_jint - 1; // array size is max_jint, index is one less
if (sizetype != NULL) index_max = sizetype->_hi - 1;
const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl);
#endif
Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
return basic_plus_adr(ary, base, scale);
}
//-------------------------load_array_element-------------------------
Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
const Type* elemtype = arytype->elem();
BasicType elembt = elemtype->array_element_basic_type();
Node* adr = array_element_address(ary, idx, elembt, arytype->size());
if (elembt == T_NARROWOOP) {
elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
}
Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
return ld;
}
//-------------------------set_arguments_for_java_call-------------------------
// Arguments (pre-popped from the stack) are taken from the JVMS.
void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
// Add the call arguments:
uint nargs = call->method()->arg_size();
for (uint i = 0; i < nargs; i++) {
Node* arg = argument(i);
call->init_req(i + TypeFunc::Parms, arg);
}
}
//---------------------------set_edges_for_java_call---------------------------
// Connect a newly created call into the current JVMS.
// A return value node (if any) is returned from set_edges_for_java_call.
void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
// Add the predefined inputs:
call->init_req( TypeFunc::Control, control() );
call->init_req( TypeFunc::I_O , i_o() );
call->init_req( TypeFunc::Memory , reset_memory() );
call->init_req( TypeFunc::FramePtr, frameptr() );
call->init_req( TypeFunc::ReturnAdr, top() );
add_safepoint_edges(call, must_throw);
Node* xcall = _gvn.transform(call);
if (xcall == top()) {
set_control(top());
return;
}
assert(xcall == call, "call identity is stable");
// Re-use the current map to produce the result.
set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
set_all_memory_call(xcall, separate_io_proj);
//return xcall; // no need, caller already has it
}
Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
if (stopped()) return top(); // maybe the call folded up?
// Capture the return value, if any.
Node* ret;
if (call->method() == NULL ||
call->method()->return_type()->basic_type() == T_VOID)
ret = top();
else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
// Note: Since any out-of-line call can produce an exception,
// we always insert an I_O projection from the call into the result.
make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
if (separate_io_proj) {
// The caller requested separate projections be used by the fall
// through and exceptional paths, so replace the projections for
// the fall through path.
set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
}
return ret;
}
//--------------------set_predefined_input_for_runtime_call--------------------
// Reading and setting the memory state is way conservative here.
// The real problem is that I am not doing real Type analysis on memory,
// so I cannot distinguish card mark stores from other stores. Across a GC
// point the Store Barrier and the card mark memory has to agree. I cannot
// have a card mark store and its barrier split across the GC point from
// either above or below. Here I get that to happen by reading ALL of memory.
// A better answer would be to separate out card marks from other memory.
// For now, return the input memory state, so that it can be reused
// after the call, if this call has restricted memory effects.
Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
// Set fixed predefined input arguments
Node* memory = reset_memory();
Node* m = narrow_mem == NULL ? memory : narrow_mem;
call->init_req( TypeFunc::Control, control() );
call->init_req( TypeFunc::I_O, top() ); // does no i/o
call->init_req( TypeFunc::Memory, m ); // may gc ptrs
call->init_req( TypeFunc::FramePtr, frameptr() );
call->init_req( TypeFunc::ReturnAdr, top() );
return memory;
}
//-------------------set_predefined_output_for_runtime_call--------------------
// Set control and memory (not i_o) from the call.
// If keep_mem is not NULL, use it for the output state,
// except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
// If hook_mem is NULL, this call produces no memory effects at all.
// If hook_mem is a Java-visible memory slice (such as arraycopy operands),
// then only that memory slice is taken from the call.
// In the last case, we must put an appropriate memory barrier before
// the call, so as to create the correct anti-dependencies on loads
// preceding the call.
void GraphKit::set_predefined_output_for_runtime_call(Node* call,
Node* keep_mem,
const TypePtr* hook_mem) {
// no i/o
set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
if (keep_mem) {
// First clone the existing memory state
set_all_memory(keep_mem);
if (hook_mem != NULL) {
// Make memory for the call
Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
// Set the RawPtr memory state only. This covers all the heap top/GC stuff
// We also use hook_mem to extract specific effects from arraycopy stubs.
set_memory(mem, hook_mem);
}
// ...else the call has NO memory effects.
// Make sure the call advertises its memory effects precisely.
// This lets us build accurate anti-dependences in gcm.cpp.
assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
"call node must be constructed correctly");
} else {
assert(hook_mem == NULL, "");
// This is not a "slow path" call; all memory comes from the call.
set_all_memory_call(call);
}
}
// Replace the call with the current state of the kit.
void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
JVMState* ejvms = NULL;
if (has_exceptions()) {
ejvms = transfer_exceptions_into_jvms();
}
ReplacedNodes replaced_nodes = map()->replaced_nodes();
ReplacedNodes replaced_nodes_exception;
Node* ex_ctl = top();
SafePointNode* final_state = stop();
// Find all the needed outputs of this call
CallProjections callprojs;
call->extract_projections(&callprojs, true);
Node* init_mem = call->in(TypeFunc::Memory);
Node* final_mem = final_state->in(TypeFunc::Memory);
Node* final_ctl = final_state->in(TypeFunc::Control);
Node* final_io = final_state->in(TypeFunc::I_O);
// Replace all the old call edges with the edges from the inlining result
if (callprojs.fallthrough_catchproj != NULL) {
C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
}
if (callprojs.fallthrough_memproj != NULL) {
if (final_mem->is_MergeMem()) {
// Parser's exits MergeMem was not transformed but may be optimized
final_mem = _gvn.transform(final_mem);
}
C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
}
if (callprojs.fallthrough_ioproj != NULL) {
C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
}
// Replace the result with the new result if it exists and is used
if (callprojs.resproj != NULL && result != NULL) {
C->gvn_replace_by(callprojs.resproj, result);
}
if (ejvms == NULL) {
// No exception edges to simply kill off those paths
if (callprojs.catchall_catchproj != NULL) {
C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
}
if (callprojs.catchall_memproj != NULL) {
C->gvn_replace_by(callprojs.catchall_memproj, C->top());
}
if (callprojs.catchall_ioproj != NULL) {
C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
}
// Replace the old exception object with top
if (callprojs.exobj != NULL) {
C->gvn_replace_by(callprojs.exobj, C->top());
}
} else {
GraphKit ekit(ejvms);
// Load my combined exception state into the kit, with all phis transformed:
SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
replaced_nodes_exception = ex_map->replaced_nodes();
Node* ex_oop = ekit.use_exception_state(ex_map);
if (callprojs.catchall_catchproj != NULL) {
C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
ex_ctl = ekit.control();
}
if (callprojs.catchall_memproj != NULL) {
C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
}
if (callprojs.catchall_ioproj != NULL) {
C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
}
// Replace the old exception object with the newly created one
if (callprojs.exobj != NULL) {
C->gvn_replace_by(callprojs.exobj, ex_oop);
}
}
// Disconnect the call from the graph
call->disconnect_inputs(NULL, C);
C->gvn_replace_by(call, C->top());
// Clean up any MergeMems that feed other MergeMems since the
// optimizer doesn't like that.
if (final_mem->is_MergeMem()) {
Node_List wl;
for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
Node* m = i.get();
if (m->is_MergeMem() && !wl.contains(m)) {
wl.push(m);
}
}
while (wl.size() > 0) {
_gvn.transform(wl.pop());
}
}
if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
replaced_nodes.apply(C, final_ctl);
}
if (!ex_ctl->is_top() && do_replaced_nodes) {
replaced_nodes_exception.apply(C, ex_ctl);
}
}
//------------------------------increment_counter------------------------------
// for statistics: increment a VM counter by 1
void GraphKit::increment_counter(address counter_addr) {
Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
increment_counter(adr1);
}
void GraphKit::increment_counter(Node* counter_addr) {
int adr_type = Compile::AliasIdxRaw;
Node* ctrl = control();
Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
}
//------------------------------uncommon_trap----------------------------------
// Bail out to the interpreter in mid-method. Implemented by calling the
// uncommon_trap blob. This helper function inserts a runtime call with the
// right debug info.
void GraphKit::uncommon_trap(int trap_request,
ciKlass* klass, const char* comment,
bool must_throw,
bool keep_exact_action) {
if (failing()) stop();
if (stopped()) return; // trap reachable?
// Note: If ProfileTraps is true, and if a deopt. actually
// occurs here, the runtime will make sure an MDO exists. There is
// no need to call method()->ensure_method_data() at this point.
// Set the stack pointer to the right value for reexecution:
set_sp(reexecute_sp());
#ifdef ASSERT
if (!must_throw) {
// Make sure the stack has at least enough depth to execute
// the current bytecode.
int inputs, ignored_depth;
if (compute_stack_effects(inputs, ignored_depth)) {
assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
Bytecodes::name(java_bc()), sp(), inputs));
}
}
#endif
Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
switch (action) {
case Deoptimization::Action_maybe_recompile:
case Deoptimization::Action_reinterpret:
// Temporary fix for 6529811 to allow virtual calls to be sure they
// get the chance to go from mono->bi->mega
if (!keep_exact_action &&
Deoptimization::trap_request_index(trap_request) < 0 &&
too_many_recompiles(reason)) {
// This BCI is causing too many recompilations.
if (C->log() != NULL) {
C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
Deoptimization::trap_reason_name(reason),
Deoptimization::trap_action_name(action));
}
action = Deoptimization::Action_none;
trap_request = Deoptimization::make_trap_request(reason, action);
} else {
C->set_trap_can_recompile(true);
}
break;
case Deoptimization::Action_make_not_entrant:
C->set_trap_can_recompile(true);
break;
#ifdef ASSERT
case Deoptimization::Action_none:
case Deoptimization::Action_make_not_compilable:
break;
default:
fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
break;
#endif
}
if (TraceOptoParse) {
char buf[100];
tty->print_cr("Uncommon trap %s at bci:%d",
Deoptimization::format_trap_request(buf, sizeof(buf),
trap_request), bci());
}
CompileLog* log = C->log();
if (log != NULL) {
int kid = (klass == NULL)? -1: log->identify(klass);
log->begin_elem("uncommon_trap bci='%d'", bci());
char buf[100];
log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
trap_request));
if (kid >= 0) log->print(" klass='%d'", kid);
if (comment != NULL) log->print(" comment='%s'", comment);
log->end_elem();
}
// Make sure any guarding test views this path as very unlikely
Node *i0 = control()->in(0);
if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
IfNode *iff = i0->as_If();
float f = iff->_prob; // Get prob
if (control()->Opcode() == Op_IfTrue) {
if (f > PROB_UNLIKELY_MAG(4))
iff->_prob = PROB_MIN;
} else {
if (f < PROB_LIKELY_MAG(4))
iff->_prob = PROB_MAX;
}
}
// Clear out dead values from the debug info.
kill_dead_locals();
// Now insert the uncommon trap subroutine call
address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
const TypePtr* no_memory_effects = NULL;
// Pass the index of the class to be loaded
Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
(must_throw ? RC_MUST_THROW : 0),
OptoRuntime::uncommon_trap_Type(),
call_addr, "uncommon_trap", no_memory_effects,
intcon(trap_request));
assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
"must extract request correctly from the graph");
assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
call->set_req(TypeFunc::ReturnAdr, returnadr());
// The debug info is the only real input to this call.
// Halt-and-catch fire here. The above call should never return!
HaltNode* halt = new(C) HaltNode(control(), frameptr());
_gvn.set_type_bottom(halt);
root()->add_req(halt);
stop_and_kill_map();
}
//--------------------------just_allocated_object------------------------------
// Report the object that was just allocated.
// It must be the case that there are no intervening safepoints.
// We use this to determine if an object is so "fresh" that
// it does not require card marks.
Node* GraphKit::just_allocated_object(Node* current_control) {
if (C->recent_alloc_ctl() == current_control)
return C->recent_alloc_obj();
return NULL;
}
void GraphKit::round_double_arguments(ciMethod* dest_method) {
// (Note: TypeFunc::make has a cache that makes this fast.)
const TypeFunc* tf = TypeFunc::make(dest_method);
int nargs = tf->_domain->_cnt - TypeFunc::Parms;
for (int j = 0; j < nargs; j++) {
const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
if( targ->basic_type() == T_DOUBLE ) {
// If any parameters are doubles, they must be rounded before
// the call, dstore_rounding does gvn.transform
Node *arg = argument(j);
arg = dstore_rounding(arg);
set_argument(j, arg);
}
}
}
/**
* Record profiling data exact_kls for Node n with the type system so
* that it can propagate it (speculation)
*
* @param n node that the type applies to
* @param exact_kls type from profiling
*
* @return node with improved type
*/
Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
const Type* current_type = _gvn.type(n);
assert(UseTypeSpeculation, "type speculation must be on");
const TypeOopPtr* speculative = current_type->speculative();
if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
const TypeOopPtr* xtype = tklass->as_instance_type();
assert(xtype->klass_is_exact(), "Should be exact");
// record the new speculative type's depth
speculative = xtype->with_inline_depth(jvms()->depth());
}
if (speculative != current_type->speculative()) {
// Build a type with a speculative type (what we think we know
// about the type but will need a guard when we use it)
const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
// We're changing the type, we need a new CheckCast node to carry
// the new type. The new type depends on the control: what
// profiling tells us is only valid from here as far as we can
// tell.
Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
cast = _gvn.transform(cast);
replace_in_map(n, cast);
n = cast;
}
return n;
}
/**
* Record profiling data from receiver profiling at an invoke with the
* type system so that it can propagate it (speculation)
*
* @param n receiver node
*
* @return node with improved type
*/
Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
if (!UseTypeSpeculation) {
return n;
}
ciKlass* exact_kls = profile_has_unique_klass();
return record_profile_for_speculation(n, exact_kls);
}
/**
* Record profiling data from argument profiling at an invoke with the
* type system so that it can propagate it (speculation)
*
* @param dest_method target method for the call
* @param bc what invoke bytecode is this?
*/
void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
if (!UseTypeSpeculation) {
return;
}
const TypeFunc* tf = TypeFunc::make(dest_method);
int nargs = tf->_domain->_cnt - TypeFunc::Parms;
int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
ciKlass* better_type = method()->argument_profiled_type(bci(), i);
if (better_type != NULL) {
record_profile_for_speculation(argument(j), better_type);
}
i++;
}
}
}
/**
* Record profiling data from parameter profiling at an invoke with
* the type system so that it can propagate it (speculation)
*/
void GraphKit::record_profiled_parameters_for_speculation() {
if (!UseTypeSpeculation) {
return;
}
for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
if (_gvn.type(local(i))->isa_oopptr()) {
ciKlass* better_type = method()->parameter_profiled_type(j);
if (better_type != NULL) {
record_profile_for_speculation(local(i), better_type);
}
j++;
}
}
}
void GraphKit::round_double_result(ciMethod* dest_method) {
// A non-strict method may return a double value which has an extended
// exponent, but this must not be visible in a caller which is 'strict'
// If a strict caller invokes a non-strict callee, round a double result
BasicType result_type = dest_method->return_type()->basic_type();
assert( method() != NULL, "must have caller context");
if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
// Destination method's return value is on top of stack
// dstore_rounding() does gvn.transform
Node *result = pop_pair();
result = dstore_rounding(result);
push_pair(result);
}
}
// rounding for strict float precision conformance
Node* GraphKit::precision_rounding(Node* n) {
return UseStrictFP && _method->flags().is_strict()
&& UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
? _gvn.transform( new (C) RoundFloatNode(0, n) )
: n;
}
// rounding for strict double precision conformance
Node* GraphKit::dprecision_rounding(Node *n) {
return UseStrictFP && _method->flags().is_strict()
&& UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
? _gvn.transform( new (C) RoundDoubleNode(0, n) )
: n;
}
// rounding for non-strict double stores
Node* GraphKit::dstore_rounding(Node* n) {
return Matcher::strict_fp_requires_explicit_rounding
&& UseSSE <= 1
? _gvn.transform( new (C) RoundDoubleNode(0, n) )
: n;
}
//=============================================================================
// Generate a fast path/slow path idiom. Graph looks like:
// [foo] indicates that 'foo' is a parameter
//
// [in] NULL
// \ /
// CmpP
// Bool ne
// If
// / \
// True False-<2>
// / |
// / cast_not_null
// Load | | ^
// [fast_test] | |
// gvn to opt_test | |
// / \ | <1>
// True False |
// | \\ |
// [slow_call] \[fast_result]
// Ctl Val \ \
// | \ \
// Catch <1> \ \
// / \ ^ \ \
// Ex No_Ex | \ \
// | \ \ | \ <2> \
// ... \ [slow_res] | | \ [null_result]
// \ \--+--+--- | |
// \ | / \ | /
// --------Region Phi
//
//=============================================================================
// Code is structured as a series of driver functions all called 'do_XXX' that
// call a set of helper functions. Helper functions first, then drivers.
//------------------------------null_check_oop---------------------------------
// Null check oop. Set null-path control into Region in slot 3.
// Make a cast-not-nullness use the other not-null control. Return cast.
Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
bool never_see_null, bool safe_for_replace) {
// Initial NULL check taken path
(*null_control) = top();
Node* cast = null_check_common(value, T_OBJECT, false, null_control);
// Generate uncommon_trap:
if (never_see_null && (*null_control) != top()) {
// If we see an unexpected null at a check-cast we record it and force a
// recompile; the offending check-cast will be compiled to handle NULLs.
// If we see more than one offending BCI, then all checkcasts in the
// method will be compiled to handle NULLs.
PreserveJVMState pjvms(this);
set_control(*null_control);
replace_in_map(value, null());
uncommon_trap(Deoptimization::Reason_null_check,
Deoptimization::Action_make_not_entrant);
(*null_control) = top(); // NULL path is dead
}
if ((*null_control) == top() && safe_for_replace) {
replace_in_map(value, cast);
}
// Cast away null-ness on the result
return cast;
}
//------------------------------opt_iff----------------------------------------
// Optimize the fast-check IfNode. Set the fast-path region slot 2.
// Return slow-path control.
Node* GraphKit::opt_iff(Node* region, Node* iff) {
IfNode *opt_iff = _gvn.transform(iff)->as_If();
// Fast path taken; set region slot 2
Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
region->init_req(2,fast_taken); // Capture fast-control
// Fast path not-taken, i.e. slow path
Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
return slow_taken;
}
//-----------------------------make_runtime_call-------------------------------
Node* GraphKit::make_runtime_call(int flags,
const TypeFunc* call_type, address call_addr,
const char* call_name,
const TypePtr* adr_type,
// The following parms are all optional.
// The first NULL ends the list.
Node* parm0, Node* parm1,
Node* parm2, Node* parm3,
Node* parm4, Node* parm5,
Node* parm6, Node* parm7) {
// Slow-path call
bool is_leaf = !(flags & RC_NO_LEAF);
bool has_io = (!is_leaf && !(flags & RC_NO_IO));
if (call_name == NULL) {
assert(!is_leaf, "must supply name for leaf");
call_name = OptoRuntime::stub_name(call_addr);
}
CallNode* call;
if (!is_leaf) {
call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
bci(), adr_type);
} else if (flags & RC_NO_FP) {
call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
} else {
call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
}
// The following is similar to set_edges_for_java_call,
// except that the memory effects of the call are restricted to AliasIdxRaw.
// Slow path call has no side-effects, uses few values
bool wide_in = !(flags & RC_NARROW_MEM);
bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
Node* prev_mem = NULL;
if (wide_in) {
prev_mem = set_predefined_input_for_runtime_call(call);
} else {
assert(!wide_out, "narrow in => narrow out");
Node* narrow_mem = memory(adr_type);
prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
}
// Hook each parm in order. Stop looking at the first NULL.
if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
/* close each nested if ===> */ } } } } } } } }
assert(call->in(call->req()-1) != NULL, "must initialize all parms");
if (!is_leaf) {
// Non-leaves can block and take safepoints:
add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
}
// Non-leaves can throw exceptions:
if (has_io) {
call->set_req(TypeFunc::I_O, i_o());
}
if (flags & RC_UNCOMMON) {
// Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
// (An "if" probability corresponds roughly to an unconditional count.
// Sort of.)
call->set_cnt(PROB_UNLIKELY_MAG(4));
}
Node* c = _gvn.transform(call);
assert(c == call, "cannot disappear");
if (wide_out) {
// Slow path call has full side-effects.
set_predefined_output_for_runtime_call(call);
} else {
// Slow path call has few side-effects, and/or sets few values.
set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
}
if (has_io) {
set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
}
return call;
}
//------------------------------merge_memory-----------------------------------
// Merge memory from one path into the current memory state.
void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
Node* old_slice = mms.force_memory();
Node* new_slice = mms.memory2();
if (old_slice != new_slice) {
PhiNode* phi;
if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
if (mms.is_empty()) {
// clone base memory Phi's inputs for this memory slice
assert(old_slice == mms.base_memory(), "sanity");
phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
_gvn.set_type(phi, Type::MEMORY);
for (uint i = 1; i < phi->req(); i++) {
phi->init_req(i, old_slice->in(i));
}
} else {
phi = old_slice->as_Phi(); // Phi was generated already
}
} else {
phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
_gvn.set_type(phi, Type::MEMORY);
}
phi->set_req(new_path, new_slice);
mms.set_memory(phi);
}
}
}
//------------------------------make_slow_call_ex------------------------------
// Make the exception handler hookups for the slow call
void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
if (stopped()) return;
// Make a catch node with just two handlers: fall-through and catch-all
Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
{ PreserveJVMState pjvms(this);
set_control(excp);
set_i_o(i_o);
if (excp != top()) {
if (deoptimize) {
// Deoptimize if an exception is caught. Don't construct exception state in this case.
uncommon_trap(Deoptimization::Reason_unhandled,
Deoptimization::Action_none);
} else {
// Create an exception state also.
// Use an exact type if the caller has specified a specific exception.
const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
}
}
}
// Get the no-exception control from the CatchNode.
set_control(norm);
}
//-------------------------------gen_subtype_check-----------------------------
// Generate a subtyping check. Takes as input the subtype and supertype.
// Returns 2 values: sets the default control() to the true path and returns
// the false path. Only reads invariant memory; sets no (visible) memory.
// The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
// but that's not exposed to the optimizer. This call also doesn't take in an
// Object; if you wish to check an Object you need to load the Object's class
// prior to coming here.
Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
// Fast check for identical types, perhaps identical constants.
// The types can even be identical non-constants, in cases
// involving Array.newInstance, Object.clone, etc.
if (subklass == superklass)
return top(); // false path is dead; no test needed.
if (_gvn.type(superklass)->singleton()) {
ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
// In the common case of an exact superklass, try to fold up the
// test before generating code. You may ask, why not just generate
// the code and then let it fold up? The answer is that the generated
// code will necessarily include null checks, which do not always
// completely fold away. If they are also needless, then they turn
// into a performance loss. Example:
// Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
// Here, the type of 'fa' is often exact, so the store check
// of fa[1]=x will fold up, without testing the nullness of x.
switch (static_subtype_check(superk, subk)) {
case SSC_always_false:
{
Node* always_fail = control();
set_control(top());
return always_fail;
}
case SSC_always_true:
return top();
case SSC_easy_test:
{
// Just do a direct pointer compare and be done.
Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
return _gvn.transform( new(C) IfFalseNode(iff) );
}
case SSC_full_test:
break;
default:
ShouldNotReachHere();
}
}
// %%% Possible further optimization: Even if the superklass is not exact,
// if the subklass is the unique subtype of the superklass, the check
// will always succeed. We could leave a dependency behind to ensure this.
// First load the super-klass's check-offset
Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
TypeInt::INT, MemNode::unordered));
int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
// Load from the sub-klass's super-class display list, or a 1-word cache of
// the secondary superclass list, or a failing value with a sentinel offset
// if the super-klass is an interface or exceptionally deep in the Java
// hierarchy and we have to scan the secondary superclass list the hard way.
// Worst-case type is a little odd: NULL is allowed as a result (usually
// klass loads can never produce a NULL).
Node *chk_off_X = ConvI2X(chk_off);
Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
// For some types like interfaces the following loadKlass is from a 1-word
// cache which is mutable so can't use immutable memory. Other
// types load from the super-class display table which is immutable.
Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
// Compile speed common case: ARE a subtype and we canNOT fail
if( superklass == nkls )
return top(); // false path is dead; no test needed.
// See if we get an immediate positive hit. Happens roughly 83% of the
// time. Test to see if the value loaded just previously from the subklass
// is exactly the superklass.
Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
// Compile speed common case: Check for being deterministic right now. If
// chk_off is a constant and not equal to cacheoff then we are NOT a
// subklass. In this case we need exactly the 1 test above and we can
// return those results immediately.
if (!might_be_cache) {
Node* not_subtype_ctrl = control();
set_control(iftrue1); // We need exactly the 1 test above
return not_subtype_ctrl;
}
// Gather the various success & failures here
RegionNode *r_ok_subtype = new (C) RegionNode(4);
record_for_igvn(r_ok_subtype);
RegionNode *r_not_subtype = new (C) RegionNode(3);
record_for_igvn(r_not_subtype);
r_ok_subtype->init_req(1, iftrue1);
// Check for immediate negative hit. Happens roughly 11% of the time (which
// is roughly 63% of the remaining cases). Test to see if the loaded
// check-offset points into the subklass display list or the 1-element
// cache. If it points to the display (and NOT the cache) and the display
// missed then it's not a subtype.
Node *cacheoff = _gvn.intcon(cacheoff_con);
Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
// Check for self. Very rare to get here, but it is taken 1/3 the time.
// No performance impact (too rare) but allows sharing of secondary arrays
// which has some footprint reduction.
Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
// -- Roads not taken here: --
// We could also have chosen to perform the self-check at the beginning
// of this code sequence, as the assembler does. This would not pay off
// the same way, since the optimizer, unlike the assembler, can perform
// static type analysis to fold away many successful self-checks.
// Non-foldable self checks work better here in second position, because
// the initial primary superclass check subsumes a self-check for most
// types. An exception would be a secondary type like array-of-interface,
// which does not appear in its own primary supertype display.
// Finally, we could have chosen to move the self-check into the
// PartialSubtypeCheckNode, and from there out-of-line in a platform
// dependent manner. But it is worthwhile to have the check here,
// where it can be perhaps be optimized. The cost in code space is
// small (register compare, branch).
// Now do a linear scan of the secondary super-klass array. Again, no real
// performance impact (too rare) but it's gotta be done.
// Since the code is rarely used, there is no penalty for moving it
// out of line, and it can only improve I-cache density.
// The decision to inline or out-of-line this final check is platform
// dependent, and is found in the AD file definition of PartialSubtypeCheck.
Node* psc = _gvn.transform(
new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
// Return false path; set default control to true path.
set_control( _gvn.transform(r_ok_subtype) );
return _gvn.transform(r_not_subtype);
}
//----------------------------static_subtype_check-----------------------------
// Shortcut important common cases when superklass is exact:
// (0) superklass is java.lang.Object (can occur in reflective code)
// (1) subklass is already limited to a subtype of superklass => always ok
// (2) subklass does not overlap with superklass => always fail
// (3) superklass has NO subtypes and we can check with a simple compare.
int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
if (StressReflectiveCode) {
return SSC_full_test; // Let caller generate the general case.
}
if (superk == env()->Object_klass()) {
return SSC_always_true; // (0) this test cannot fail
}
ciType* superelem = superk;
if (superelem->is_array_klass())
superelem = superelem->as_array_klass()->base_element_type();
if (!subk->is_interface()) { // cannot trust static interface types yet
if (subk->is_subtype_of(superk)) {
return SSC_always_true; // (1) false path dead; no dynamic test needed
}
if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
!superk->is_subtype_of(subk)) {
return SSC_always_false;
}
}
// If casting to an instance klass, it must have no subtypes
if (superk->is_interface()) {
// Cannot trust interfaces yet.
// %%% S.B. superk->nof_implementors() == 1
} else if (superelem->is_instance_klass()) {
ciInstanceKlass* ik = superelem->as_instance_klass();
if (!ik->has_subklass() && !ik->is_interface()) {
if (!ik->is_final()) {
// Add a dependency if there is a chance of a later subclass.
C->dependencies()->assert_leaf_type(ik);
}
return SSC_easy_test; // (3) caller can do a simple ptr comparison
}
} else {
// A primitive array type has no subtypes.
return SSC_easy_test; // (3) caller can do a simple ptr comparison
}
return SSC_full_test;
}
// Profile-driven exact type check:
Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
float prob,
Node* *casted_receiver) {
const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
Node* recv_klass = load_object_klass(receiver);
Node* want_klass = makecon(tklass);
Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
const TypeOopPtr* recv_xtype = tklass->as_instance_type();
assert(recv_xtype->klass_is_exact(), "");
// Subsume downstream occurrences of receiver with a cast to
// recv_xtype, since now we know what the type will be.
Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
(*casted_receiver) = _gvn.transform(cast);
// (User must make the replace_in_map call.)
return fail;
}
//------------------------------seems_never_null-------------------------------
// Use null_seen information if it is available from the profile.
// If we see an unexpected null at a type check we record it and force a
// recompile; the offending check will be recompiled to handle NULLs.
// If we see several offending BCIs, then all checks in the
// method will be recompiled.
bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
if (UncommonNullCast // Cutout for this technique
&& obj != null() // And not the -Xcomp stupid case?
&& !too_many_traps(Deoptimization::Reason_null_check)
) {
if (data == NULL)
// Edge case: no mature data. Be optimistic here.
return true;
// If the profile has not seen a null, assume it won't happen.
assert(java_bc() == Bytecodes::_checkcast ||
java_bc() == Bytecodes::_instanceof ||
java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
return !data->as_BitData()->null_seen();
}
return false;
}
//------------------------maybe_cast_profiled_receiver-------------------------
// If the profile has seen exactly one type, narrow to exactly that type.
// Subsequent type checks will always fold up.
Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
ciKlass* require_klass,
ciKlass* spec_klass,
bool safe_for_replace) {
if (!UseTypeProfile || !TypeProfileCasts) return NULL;
Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
// Make sure we haven't already deoptimized from this tactic.
if (too_many_traps(reason) || too_many_recompiles(reason))
return NULL;
// (No, this isn't a call, but it's enough like a virtual call
// to use the same ciMethod accessor to get the profile info...)
// If we have a speculative type use it instead of profiling (which
// may not help us)
ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
if (exact_kls != NULL) {// no cast failures here
if (require_klass == NULL ||
static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
// If we narrow the type to match what the type profile sees or
// the speculative type, we can then remove the rest of the
// cast.
// This is a win, even if the exact_kls is very specific,
// because downstream operations, such as method calls,
// will often benefit from the sharper type.
Node* exact_obj = not_null_obj; // will get updated in place...
Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
&exact_obj);
{ PreserveJVMState pjvms(this);
set_control(slow_ctl);
uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
}
if (safe_for_replace) {
replace_in_map(not_null_obj, exact_obj);
}
return exact_obj;
}
// assert(ssc == SSC_always_true)... except maybe the profile lied to us.
}
return NULL;
}
/**
* Cast obj to type and emit guard unless we had too many traps here
* already
*
* @param obj node being casted
* @param type type to cast the node to
* @param not_null true if we know node cannot be null
*/
Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
ciKlass* type,
bool not_null) {
// type == NULL if profiling tells us this object is always null
if (type != NULL) {
Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
!too_many_traps(class_reason) && !too_many_recompiles(class_reason)) {
Node* not_null_obj = NULL;
// not_null is true if we know the object is not null and
// there's no need for a null check
if (!not_null) {
Node* null_ctl = top();
not_null_obj = null_check_oop(obj, &null_ctl, true, true);
assert(null_ctl->is_top(), "no null control here");
} else {
not_null_obj = obj;
}
Node* exact_obj = not_null_obj;
ciKlass* exact_kls = type;
Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
&exact_obj);
{
PreserveJVMState pjvms(this);
set_control(slow_ctl);
uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
}
replace_in_map(not_null_obj, exact_obj);
obj = exact_obj;
}
} else {
if (!too_many_traps(Deoptimization::Reason_null_assert) &&
!too_many_recompiles(Deoptimization::Reason_null_assert)) {
Node* exact_obj = null_assert(obj);
replace_in_map(obj, exact_obj);
obj = exact_obj;
}
}
return obj;
}
//-------------------------------gen_instanceof--------------------------------
// Generate an instance-of idiom. Used by both the instance-of bytecode
// and the reflective instance-of call.
Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
kill_dead_locals(); // Benefit all the uncommon traps
assert( !stopped(), "dead parse path should be checked in callers" );
assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
"must check for not-null not-dead klass in callers");
// Make the merge point
enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
RegionNode* region = new(C) RegionNode(PATH_LIMIT);
Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
C->set_has_split_ifs(true); // Has chance for split-if optimization
ciProfileData* data = NULL;
if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
data = method()->method_data()->bci_to_data(bci());
}
bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
&& seems_never_null(obj, data));
// Null check; get casted pointer; set region slot 3
Node* null_ctl = top();
Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
// If not_null_obj is dead, only null-path is taken
if (stopped()) { // Doing instance-of on a NULL?
set_control(null_ctl);
return intcon(0);
}
region->init_req(_null_path, null_ctl);
phi ->init_req(_null_path, intcon(0)); // Set null path value
if (null_ctl == top()) {
// Do this eagerly, so that pattern matches like is_diamond_phi
// will work even during parsing.
assert(_null_path == PATH_LIMIT-1, "delete last");
region->del_req(_null_path);
phi ->del_req(_null_path);
}
// Do we know the type check always succeed?
bool known_statically = false;
if (_gvn.type(superklass)->singleton()) {
ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
if (subk != NULL && subk->is_loaded()) {
int static_res = static_subtype_check(superk, subk);
known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
}
}
if (known_statically && UseTypeSpeculation) {
// If we know the type check always succeeds then we don't use the
// profiling data at this bytecode. Don't lose it, feed it to the
// type system as a speculative type.
not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
} else {
const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
// We may not have profiling here or it may not help us. If we
// have a speculative type use it to perform an exact cast.
ciKlass* spec_obj_type = obj_type->speculative_type();
if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
if (stopped()) { // Profile disagrees with this path.
set_control(null_ctl); // Null is the only remaining possibility.
return intcon(0);
}
if (cast_obj != NULL) {
not_null_obj = cast_obj;
}
}
}
// Load the object's klass
Node* obj_klass = load_object_klass(not_null_obj);
// Generate the subtype check
Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
// Plug in the success path to the general merge in slot 1.
region->init_req(_obj_path, control());
phi ->init_req(_obj_path, intcon(1));
// Plug in the failing path to the general merge in slot 2.
region->init_req(_fail_path, not_subtype_ctrl);
phi ->init_req(_fail_path, intcon(0));
// Return final merged results
set_control( _gvn.transform(region) );
record_for_igvn(region);
return _gvn.transform(phi);
}
//-------------------------------gen_checkcast---------------------------------
// Generate a checkcast idiom. Used by both the checkcast bytecode and the
// array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
// uncommon-trap paths work. Adjust stack after this call.
// If failure_control is supplied and not null, it is filled in with
// the control edge for the cast failure. Otherwise, an appropriate
// uncommon trap or exception is thrown.
Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
Node* *failure_control) {
kill_dead_locals(); // Benefit all the uncommon traps
const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
// Fast cutout: Check the case that the cast is vacuously true.
// This detects the common cases where the test will short-circuit
// away completely. We do this before we perform the null check,
// because if the test is going to turn into zero code, we don't
// want a residual null check left around. (Causes a slowdown,
// for example, in some objArray manipulations, such as a[i]=a[j].)
if (tk->singleton()) {
const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
if (objtp != NULL && objtp->klass() != NULL) {
switch (static_subtype_check(tk->klass(), objtp->klass())) {
case SSC_always_true:
// If we know the type check always succeed then we don't use
// the profiling data at this bytecode. Don't lose it, feed it
// to the type system as a speculative type.
return record_profiled_receiver_for_speculation(obj);
case SSC_always_false:
// It needs a null check because a null will *pass* the cast check.
// A non-null value will always produce an exception.
return null_assert(obj);
}
}
}
ciProfileData* data = NULL;
bool safe_for_replace = false;
if (failure_control == NULL) { // use MDO in regular case only
assert(java_bc() == Bytecodes::_aastore ||
java_bc() == Bytecodes::_checkcast,
"interpreter profiles type checks only for these BCs");
data = method()->method_data()->bci_to_data(bci());
safe_for_replace = true;
}
// Make the merge point
enum { _obj_path = 1, _null_path, PATH_LIMIT };
RegionNode* region = new (C) RegionNode(PATH_LIMIT);
Node* phi = new (C) PhiNode(region, toop);
C->set_has_split_ifs(true); // Has chance for split-if optimization
// Use null-cast information if it is available
bool never_see_null = ((failure_control == NULL) // regular case only
&& seems_never_null(obj, data));
// Null check; get casted pointer; set region slot 3
Node* null_ctl = top();
Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
// If not_null_obj is dead, only null-path is taken
if (stopped()) { // Doing instance-of on a NULL?
set_control(null_ctl);
return null();
}
region->init_req(_null_path, null_ctl);
phi ->init_req(_null_path, null()); // Set null path value
if (null_ctl == top()) {
// Do this eagerly, so that pattern matches like is_diamond_phi
// will work even during parsing.
assert(_null_path == PATH_LIMIT-1, "delete last");
region->del_req(_null_path);
phi ->del_req(_null_path);
}
Node* cast_obj = NULL;
if (tk->klass_is_exact()) {
// The following optimization tries to statically cast the speculative type of the object
// (for example obtained during profiling) to the type of the superklass and then do a
// dynamic check that the type of the object is what we expect. To work correctly
// for checkcast and aastore the type of superklass should be exact.
const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
// We may not have profiling here or it may not help us. If we have
// a speculative type use it to perform an exact cast.
ciKlass* spec_obj_type = obj_type->speculative_type();
if (spec_obj_type != NULL ||
(data != NULL &&
// Counter has never been decremented (due to cast failure).
// ...This is a reasonable thing to expect. It is true of
// all casts inserted by javac to implement generic types.
data->as_CounterData()->count() >= 0)) {
cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
if (cast_obj != NULL) {
if (failure_control != NULL) // failure is now impossible
(*failure_control) = top();
// adjust the type of the phi to the exact klass:
phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
}
}
}
if (cast_obj == NULL) {
// Load the object's klass
Node* obj_klass = load_object_klass(not_null_obj);
// Generate the subtype check
Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
// Plug in success path into the merge
cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
not_null_obj, toop));
// Failure path ends in uncommon trap (or may be dead - failure impossible)
if (failure_control == NULL) {
if (not_subtype_ctrl != top()) { // If failure is possible
PreserveJVMState pjvms(this);
set_control(not_subtype_ctrl);
builtin_throw(Deoptimization::Reason_class_check, obj_klass);
}
} else {
(*failure_control) = not_subtype_ctrl;
}
}
region->init_req(_obj_path, control());
phi ->init_req(_obj_path, cast_obj);
// A merge of NULL or Casted-NotNull obj
Node* res = _gvn.transform(phi);
// Note I do NOT always 'replace_in_map(obj,result)' here.
// if( tk->klass()->can_be_primary_super() )
// This means that if I successfully store an Object into an array-of-String
// I 'forget' that the Object is really now known to be a String. I have to
// do this because we don't have true union types for interfaces - if I store
// a Baz into an array-of-Interface and then tell the optimizer it's an
// Interface, I forget that it's also a Baz and cannot do Baz-like field
// references to it. FIX THIS WHEN UNION TYPES APPEAR!
// replace_in_map( obj, res );
// Return final merged results
set_control( _gvn.transform(region) );
record_for_igvn(region);
return res;
}
//------------------------------next_monitor-----------------------------------
// What number should be given to the next monitor?
int GraphKit::next_monitor() {
int current = jvms()->monitor_depth()* C->sync_stack_slots();
int next = current + C->sync_stack_slots();
// Keep the toplevel high water mark current:
if (C->fixed_slots() < next) C->set_fixed_slots(next);
return current;
}
//------------------------------insert_mem_bar---------------------------------
// Memory barrier to avoid floating things around
// The membar serves as a pinch point between both control and all memory slices.
Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
mb->init_req(TypeFunc::Control, control());
mb->init_req(TypeFunc::Memory, reset_memory());
Node* membar = _gvn.transform(mb);
set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
set_all_memory_call(membar);
return membar;
}
//-------------------------insert_mem_bar_volatile----------------------------
// Memory barrier to avoid floating things around
// The membar serves as a pinch point between both control and memory(alias_idx).
// If you want to make a pinch point on all memory slices, do not use this
// function (even with AliasIdxBot); use insert_mem_bar() instead.
Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
// When Parse::do_put_xxx updates a volatile field, it appends a series
// of MemBarVolatile nodes, one for *each* volatile field alias category.
// The first membar is on the same memory slice as the field store opcode.
// This forces the membar to follow the store. (Bug 6500685 broke this.)
// All the other membars (for other volatile slices, including AliasIdxBot,
// which stands for all unknown volatile slices) are control-dependent
// on the first membar. This prevents later volatile loads or stores
// from sliding up past the just-emitted store.
MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
mb->set_req(TypeFunc::Control,control());
if (alias_idx == Compile::AliasIdxBot) {
mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
} else {
assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
mb->set_req(TypeFunc::Memory, memory(alias_idx));
}
Node* membar = _gvn.transform(mb);
set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
if (alias_idx == Compile::AliasIdxBot) {
merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
} else {
set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
}
return membar;
}
//------------------------------shared_lock------------------------------------
// Emit locking code.
FastLockNode* GraphKit::shared_lock(Node* obj) {
// bci is either a monitorenter bc or InvocationEntryBci
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
if( !GenerateSynchronizationCode )
return NULL; // Not locking things?
if (stopped()) // Dead monitor?
return NULL;
assert(dead_locals_are_killed(), "should kill locals before sync. point");
// Box the stack location
Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
Node* mem = reset_memory();
FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
// Create the counters for this fast lock.
flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
}
// Create the rtm counters for this fast lock if needed.
flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
// Add monitor to debug info for the slow path. If we block inside the
// slow path and de-opt, we need the monitor hanging around
map()->push_monitor( flock );
const TypeFunc *tf = LockNode::lock_type();
LockNode *lock = new (C) LockNode(C, tf);
lock->init_req( TypeFunc::Control, control() );
lock->init_req( TypeFunc::Memory , mem );
lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
lock->init_req( TypeFunc::FramePtr, frameptr() );
lock->init_req( TypeFunc::ReturnAdr, top() );
lock->init_req(TypeFunc::Parms + 0, obj);
lock->init_req(TypeFunc::Parms + 1, box);
lock->init_req(TypeFunc::Parms + 2, flock);
add_safepoint_edges(lock);
lock = _gvn.transform( lock )->as_Lock();
// lock has no side-effects, sets few values
set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
insert_mem_bar(Op_MemBarAcquireLock);
// Add this to the worklist so that the lock can be eliminated
record_for_igvn(lock);
#ifndef PRODUCT
if (PrintLockStatistics) {
// Update the counter for this lock. Don't bother using an atomic
// operation since we don't require absolute accuracy.
lock->create_lock_counter(map()->jvms());
increment_counter(lock->counter()->addr());
}
#endif
return flock;
}
//------------------------------shared_unlock----------------------------------
// Emit unlocking code.
void GraphKit::shared_unlock(Node* box, Node* obj) {
// bci is either a monitorenter bc or InvocationEntryBci
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
if( !GenerateSynchronizationCode )
return;
if (stopped()) { // Dead monitor?
map()->pop_monitor(); // Kill monitor from debug info
return;
}
// Memory barrier to avoid floating things down past the locked region
insert_mem_bar(Op_MemBarReleaseLock);
const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
UnlockNode *unlock = new (C) UnlockNode(C, tf);
#ifdef ASSERT
unlock->set_dbg_jvms(sync_jvms());
#endif
uint raw_idx = Compile::AliasIdxRaw;
unlock->init_req( TypeFunc::Control, control() );
unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
unlock->init_req( TypeFunc::FramePtr, frameptr() );
unlock->init_req( TypeFunc::ReturnAdr, top() );
unlock->init_req(TypeFunc::Parms + 0, obj);
unlock->init_req(TypeFunc::Parms + 1, box);
unlock = _gvn.transform(unlock)->as_Unlock();
Node* mem = reset_memory();
// unlock has no side-effects, sets few values
set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
// Kill monitor from debug info
map()->pop_monitor( );
}
//-------------------------------get_layout_helper-----------------------------
// If the given klass is a constant or known to be an array,
// fetch the constant layout helper value into constant_value
// and return (Node*)NULL. Otherwise, load the non-constant
// layout helper value, and return the node which represents it.
// This two-faced routine is useful because allocation sites
// almost always feature constant types.
Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
if (!StressReflectiveCode && inst_klass != NULL) {
ciKlass* klass = inst_klass->klass();
bool xklass = inst_klass->klass_is_exact();
if (xklass || klass->is_array_klass()) {
jint lhelper = klass->layout_helper();
if (lhelper != Klass::_lh_neutral_value) {
constant_value = lhelper;
return (Node*) NULL;
}
}
}
constant_value = Klass::_lh_neutral_value; // put in a known value
Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
}
// We just put in an allocate/initialize with a big raw-memory effect.
// Hook selected additional alias categories on the initialization.
static void hook_memory_on_init(GraphKit& kit, int alias_idx,
MergeMemNode* init_in_merge,
Node* init_out_raw) {
DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
Node* prevmem = kit.memory(alias_idx);
init_in_merge->set_memory_at(alias_idx, prevmem);
kit.set_memory(init_out_raw, alias_idx);
}
//---------------------------set_output_for_allocation-------------------------
Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
const TypeOopPtr* oop_type,
bool deoptimize_on_exception) {
int rawidx = Compile::AliasIdxRaw;
alloc->set_req( TypeFunc::FramePtr, frameptr() );
add_safepoint_edges(alloc);
Node* allocx = _gvn.transform(alloc);
set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
// create memory projection for i_o
set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
// create a memory projection as for the normal control path
Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
set_memory(malloc, rawidx);
// a normal slow-call doesn't change i_o, but an allocation does
// we create a separate i_o projection for the normal control path
set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
// put in an initialization barrier
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
rawoop)->as_Initialize();
assert(alloc->initialization() == init, "2-way macro link must work");
assert(init ->allocation() == alloc, "2-way macro link must work");
{
// Extract memory strands which may participate in the new object's
// initialization, and source them from the new InitializeNode.
// This will allow us to observe initializations when they occur,
// and link them properly (as a group) to the InitializeNode.
assert(init->in(InitializeNode::Memory) == malloc, "");
MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
init->set_req(InitializeNode::Memory, minit_in);
record_for_igvn(minit_in); // fold it up later, if possible
Node* minit_out = memory(rawidx);
assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
if (oop_type->isa_aryptr()) {
const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
int elemidx = C->get_alias_index(telemref);
hook_memory_on_init(*this, elemidx, minit_in, minit_out);
} else if (oop_type->isa_instptr()) {
ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
ciField* field = ik->nonstatic_field_at(i);
if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
continue; // do not bother to track really large numbers of fields
// Find (or create) the alias category for this field:
int fieldidx = C->alias_type(field)->index();
hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
}
}
}
// Cast raw oop to the real thing...
Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
javaoop = _gvn.transform(javaoop);
C->set_recent_alloc(control(), javaoop);
assert(just_allocated_object(control()) == javaoop, "just allocated");
#ifdef ASSERT
{ // Verify that the AllocateNode::Ideal_allocation recognizers work:
assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
"Ideal_allocation works");
assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
"Ideal_allocation works");
if (alloc->is_AllocateArray()) {
assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
"Ideal_allocation works");
assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
"Ideal_allocation works");
} else {
assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
}
}
#endif //ASSERT
return javaoop;
}
//---------------------------new_instance--------------------------------------
// This routine takes a klass_node which may be constant (for a static type)
// or may be non-constant (for reflective code). It will work equally well
// for either, and the graph will fold nicely if the optimizer later reduces
// the type to a constant.
// The optional arguments are for specialized use by intrinsics:
// - If 'extra_slow_test' if not null is an extra condition for the slow-path.
// - If 'return_size_val', report the the total object size to the caller.
// - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
Node* GraphKit::new_instance(Node* klass_node,
Node* extra_slow_test,
Node* *return_size_val,
bool deoptimize_on_exception) {
// Compute size in doublewords
// The size is always an integral number of doublewords, represented
// as a positive bytewise size stored in the klass's layout_helper.
// The layout_helper also encodes (in a low bit) the need for a slow path.
jint layout_con = Klass::_lh_neutral_value;
Node* layout_val = get_layout_helper(klass_node, layout_con);
int layout_is_con = (layout_val == NULL);
if (extra_slow_test == NULL) extra_slow_test = intcon(0);
// Generate the initial go-slow test. It's either ALWAYS (return a
// Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
// case) a computed value derived from the layout_helper.
Node* initial_slow_test = NULL;
if (layout_is_con) {
assert(!StressReflectiveCode, "stress mode does not use these paths");
bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
} else { // reflective case
// This reflective path is used by Unsafe.allocateInstance.
// (It may be stress-tested by specifying StressReflectiveCode.)
// Basically, we want to get into the VM is there's an illegal argument.
Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
if (extra_slow_test != intcon(0)) {
initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
}
// (Macro-expander will further convert this to a Bool, if necessary.)
}
// Find the size in bytes. This is easy; it's the layout_helper.
// The size value must be valid even if the slow path is taken.
Node* size = NULL;
if (layout_is_con) {
size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
} else { // reflective case
// This reflective path is used by clone and Unsafe.allocateInstance.
size = ConvI2X(layout_val);
// Clear the low bits to extract layout_helper_size_in_bytes:
assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
size = _gvn.transform( new (C) AndXNode(size, mask) );
}
if (return_size_val != NULL) {
(*return_size_val) = size;
}
// This is a precise notnull oop of the klass.
// (Actually, it need not be precise if this is a reflective allocation.)
// It's what we cast the result to.
const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
if (!tklass) tklass = TypeKlassPtr::OBJECT;
const TypeOopPtr* oop_type = tklass->as_instance_type();
// Now generate allocation code
// The entire memory state is needed for slow path of the allocation
// since GC and deoptimization can happened.
Node *mem = reset_memory();
set_all_memory(mem); // Create new memory state
AllocateNode* alloc
= new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
control(), mem, i_o(),
size, klass_node,
initial_slow_test);
return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
}
//-------------------------------new_array-------------------------------------
// helper for both newarray and anewarray
// The 'length' parameter is (obviously) the length of the array.
// See comments on new_instance for the meaning of the other arguments.
Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
Node* length, // number of array elements
int nargs, // number of arguments to push back for uncommon trap
Node* *return_size_val,
bool deoptimize_on_exception) {
jint layout_con = Klass::_lh_neutral_value;
Node* layout_val = get_layout_helper(klass_node, layout_con);
int layout_is_con = (layout_val == NULL);
if (!layout_is_con && !StressReflectiveCode &&
!too_many_traps(Deoptimization::Reason_class_check)) {
// This is a reflective array creation site.
// Optimistically assume that it is a subtype of Object[],
// so that we can fold up all the address arithmetic.
layout_con = Klass::array_layout_helper(T_OBJECT);
Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
{ BuildCutout unless(this, bol_lh, PROB_MAX);
inc_sp(nargs);
uncommon_trap(Deoptimization::Reason_class_check,
Deoptimization::Action_maybe_recompile);
}
layout_val = NULL;
layout_is_con = true;
}
// Generate the initial go-slow test. Make sure we do not overflow
// if length is huge (near 2Gig) or negative! We do not need
// exact double-words here, just a close approximation of needed
// double-words. We can't add any offset or rounding bits, lest we
// take a size -1 of bytes and make it positive. Use an unsigned
// compare, so negative sizes look hugely positive.
int fast_size_limit = FastAllocateSizeLimit;
if (layout_is_con) {
assert(!StressReflectiveCode, "stress mode does not use these paths");
// Increase the size limit if we have exact knowledge of array type.
int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
fast_size_limit <<= (LogBytesPerLong - log2_esize);
}
Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
// --- Size Computation ---
// array_size = round_to_heap(array_header + (length << elem_shift));
// where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
// and round_to(x, y) == ((x + y-1) & ~(y-1))
// The rounding mask is strength-reduced, if possible.
int round_mask = MinObjAlignmentInBytes - 1;
Node* header_size = NULL;
int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
// (T_BYTE has the weakest alignment and size restrictions...)
if (layout_is_con) {
int hsize = Klass::layout_helper_header_size(layout_con);
int eshift = Klass::layout_helper_log2_element_size(layout_con);
BasicType etype = Klass::layout_helper_element_type(layout_con);
if ((round_mask & ~right_n_bits(eshift)) == 0)
round_mask = 0; // strength-reduce it if it goes away completely
assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
assert(header_size_min <= hsize, "generic minimum is smallest");
header_size_min = hsize;
header_size = intcon(hsize + round_mask);
} else {
Node* hss = intcon(Klass::_lh_header_size_shift);
Node* hsm = intcon(Klass::_lh_header_size_mask);
Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
Node* mask = intcon(round_mask);
header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
}
Node* elem_shift = NULL;
if (layout_is_con) {
int eshift = Klass::layout_helper_log2_element_size(layout_con);
if (eshift != 0)
elem_shift = intcon(eshift);
} else {
// There is no need to mask or shift this value.
// The semantics of LShiftINode include an implicit mask to 0x1F.
assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
elem_shift = layout_val;
}
// Transition to native address size for all offset calculations:
Node* lengthx = ConvI2X(length);
Node* headerx = ConvI2X(header_size);
#ifdef _LP64
{ const TypeInt* tilen = _gvn.find_int_type(length);
if (tilen != NULL && tilen->_lo < 0) {
// Add a manual constraint to a positive range. Cf. array_element_address.
jlong size_max = fast_size_limit;
if (size_max > tilen->_hi) size_max = tilen->_hi;
const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
// Only do a narrow I2L conversion if the range check passed.
IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
_gvn.transform(iff);
RegionNode* region = new (C) RegionNode(3);
_gvn.set_type(region, Type::CONTROL);
lengthx = new (C) PhiNode(region, TypeLong::LONG);
_gvn.set_type(lengthx, TypeLong::LONG);
// Range check passed. Use ConvI2L node with narrow type.
Node* passed = IfFalse(iff);
region->init_req(1, passed);
// Make I2L conversion control dependent to prevent it from
// floating above the range check during loop optimizations.
lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
// Range check failed. Use ConvI2L with wide type because length may be invalid.
region->init_req(2, IfTrue(iff));
lengthx->init_req(2, ConvI2X(length));
set_control(region);
record_for_igvn(region);
record_for_igvn(lengthx);
}
}
#endif
// Combine header size (plus rounding) and body size. Then round down.
// This computation cannot overflow, because it is used only in two
// places, one where the length is sharply limited, and the other
// after a successful allocation.
Node* abody = lengthx;
if (elem_shift != NULL)
abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
if (round_mask != 0) {
Node* mask = MakeConX(~round_mask);
size = _gvn.transform( new(C) AndXNode(size, mask) );
}
// else if round_mask == 0, the size computation is self-rounding
if (return_size_val != NULL) {
// This is the size
(*return_size_val) = size;
}
// Now generate allocation code
// The entire memory state is needed for slow path of the allocation
// since GC and deoptimization can happened.
Node *mem = reset_memory();
set_all_memory(mem); // Create new memory state
if (initial_slow_test->is_Bool()) {
// Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
}
// Create the AllocateArrayNode and its result projections
AllocateArrayNode* alloc
= new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
control(), mem, i_o(),
size, klass_node,
initial_slow_test,
length);
// Cast to correct type. Note that the klass_node may be constant or not,
// and in the latter case the actual array type will be inexact also.
// (This happens via a non-constant argument to inline_native_newArray.)
// In any case, the value of klass_node provides the desired array type.
const TypeInt* length_type = _gvn.find_int_type(length);
const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
if (ary_type->isa_aryptr() && length_type != NULL) {
// Try to get a better type than POS for the size
ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
}
Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
// Cast length on remaining path to be as narrow as possible
if (map()->find_edge(length) >= 0) {
Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
if (ccast != length) {
_gvn.set_type_bottom(ccast);
record_for_igvn(ccast);
replace_in_map(length, ccast);
}
}
return javaoop;
}
// The following "Ideal_foo" functions are placed here because they recognize
// the graph shapes created by the functions immediately above.
//---------------------------Ideal_allocation----------------------------------
// Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
if (ptr == NULL) { // reduce dumb test in callers
return NULL;
}
if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
ptr = ptr->in(1);
if (ptr == NULL) return NULL;
}
// Return NULL for allocations with several casts:
// j.l.reflect.Array.newInstance(jobject, jint)
// Object.clone()
// to keep more precise type from last cast.
if (ptr->is_Proj()) {
Node* allo = ptr->in(0);
if (allo != NULL && allo->is_Allocate()) {
return allo->as_Allocate();
}
}
// Report failure to match.
return NULL;
}
// Fancy version which also strips off an offset (and reports it to caller).
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
intptr_t& offset) {
Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
if (base == NULL) return NULL;
return Ideal_allocation(base, phase);
}
// Trace Initialize <- Proj[Parm] <- Allocate
AllocateNode* InitializeNode::allocation() {
Node* rawoop = in(InitializeNode::RawAddress);
if (rawoop->is_Proj()) {
Node* alloc = rawoop->in(0);
if (alloc->is_Allocate()) {
return alloc->as_Allocate();
}
}
return NULL;
}
// Trace Allocate -> Proj[Parm] -> Initialize
InitializeNode* AllocateNode::initialization() {
ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
if (rawoop == NULL) return NULL;
for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
Node* init = rawoop->fast_out(i);
if (init->is_Initialize()) {
assert(init->as_Initialize()->allocation() == this, "2-way link");
return init->as_Initialize();
}
}
return NULL;
}
//----------------------------- loop predicates ---------------------------
//------------------------------add_predicate_impl----------------------------
void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
// Too many traps seen?
if (too_many_traps(reason)) {
#ifdef ASSERT
if (TraceLoopPredicate) {
int tc = C->trap_count(reason);
tty->print("too many traps=%s tcount=%d in ",
Deoptimization::trap_reason_name(reason), tc);
method()->print(); // which method has too many predicate traps
tty->cr();
}
#endif
// We cannot afford to take more traps here,
// do not generate predicate.
return;
}
Node *cont = _gvn.intcon(1);
Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
C->add_predicate_opaq(opq);
{
PreserveJVMState pjvms(this);
set_control(iffalse);
inc_sp(nargs);
uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
}
Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
set_control(iftrue);
}
//------------------------------add_predicate---------------------------------
void GraphKit::add_predicate(int nargs) {
if (UseLoopPredicate) {
add_predicate_impl(Deoptimization::Reason_predicate, nargs);
}
// loop's limit check predicate should be near the loop.
if (LoopLimitCheck) {
add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
}
}
//----------------------------- store barriers ----------------------------
#define __ ideal.
void GraphKit::sync_kit(IdealKit& ideal) {
set_all_memory(__ merged_memory());
set_i_o(__ i_o());
set_control(__ ctrl());
}
void GraphKit::final_sync(IdealKit& ideal) {
// Final sync IdealKit and graphKit.
sync_kit(ideal);
}
// vanilla/CMS post barrier
// Insert a write-barrier store. This is to let generational GC work; we have
// to flag all oop-stores before the next GC point.
void GraphKit::write_barrier_post(Node* oop_store,
Node* obj,
Node* adr,
uint adr_idx,
Node* val,
bool use_precise) {
// No store check needed if we're storing a NULL or an old object
// (latter case is probably a string constant). The concurrent
// mark sweep garbage collector, however, needs to have all nonNull
// oop updates flagged via card-marks.
if (val != NULL && val->is_Con()) {
// must be either an oop or NULL
const Type* t = val->bottom_type();
if (t == TypePtr::NULL_PTR || t == Type::TOP)
// stores of null never (?) need barriers
return;
}
if (use_ReduceInitialCardMarks()
&& obj == just_allocated_object(control())) {
// We can skip marks on a freshly-allocated object in Eden.
// Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
// That routine informs GC to take appropriate compensating steps,
// upon a slow-path allocation, so as to make this card-mark
// elision safe.
return;
}
if (!use_precise) {
// All card marks for a (non-array) instance are in one place:
adr = obj;
}
// (Else it's an array (or unknown), and we want more precise card marks.)
assert(adr != NULL, "");
IdealKit ideal(this, true);
// Convert the pointer to an int prior to doing math on it
Node* cast = __ CastPX(__ ctrl(), adr);
// Divide by card size
assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
"Only one we handle so far.");
Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
// Combine card table base and card offset
Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
// Get the alias_index for raw card-mark memory
int adr_type = Compile::AliasIdxRaw;
Node* zero = __ ConI(0); // Dirty card value
BasicType bt = T_BYTE;
if (UseCondCardMark) {
// The classic GC reference write barrier is typically implemented
// as a store into the global card mark table. Unfortunately
// unconditional stores can result in false sharing and excessive
// coherence traffic as well as false transactional aborts.
// UseCondCardMark enables MP "polite" conditional card mark
// stores. In theory we could relax the load from ctrl() to
// no_ctrl, but that doesn't buy much latitude.
Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
__ if_then(card_val, BoolTest::ne, zero);
}
// Smash zero into card
if( !UseConcMarkSweepGC ) {
#if defined(AARCH64)
__ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
#else
__ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
#endif
} else {
// Specialized path for CM store barrier
__ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
}
if (UseCondCardMark) {
__ end_if();
}
// Final sync IdealKit and GraphKit.
final_sync(ideal);
}
// G1 pre/post barriers
void GraphKit::g1_write_barrier_pre(bool do_load,
Node* obj,
Node* adr,
uint alias_idx,
Node* val,
const TypeOopPtr* val_type,
Node* pre_val,
BasicType bt) {
// Some sanity checks
// Note: val is unused in this routine.
if (do_load) {
// We need to generate the load of the previous value
assert(obj != NULL, "must have a base");
assert(adr != NULL, "where are loading from?");
assert(pre_val == NULL, "loaded already?");
assert(val_type != NULL, "need a type");
} else {
// In this case both val_type and alias_idx are unused.
assert(pre_val != NULL, "must be loaded already");
// Nothing to be done if pre_val is null.
if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
}
assert(bt == T_OBJECT, "or we shouldn't be here");
IdealKit ideal(this, true);
Node* tls = __ thread(); // ThreadLocalStorage
Node* no_ctrl = NULL;
Node* no_base = __ top();
Node* zero = __ ConI(0);
Node* zeroX = __ ConX(0);
float likely = PROB_LIKELY(0.999);
float unlikely = PROB_UNLIKELY(0.999);
BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
// Offsets into the thread
const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
PtrQueue::byte_offset_of_active());
const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
PtrQueue::byte_offset_of_index());
const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
PtrQueue::byte_offset_of_buf());
// Now the actual pointers into the thread
Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
// Now some of the values
Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
// if (!marking)
__ if_then(marking, BoolTest::ne, zero, unlikely); {
BasicType index_bt = TypeX_X->basic_type();
assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
if (do_load) {
// load original value
// alias_idx correct??
pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
}
// if (pre_val != NULL)
__ if_then(pre_val, BoolTest::ne, null()); {
Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
// is the queue for this thread full?
__ if_then(index, BoolTest::ne, zeroX, likely); {
// decrement the index
Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
// Now get the buffer location we will log the previous value into and store it
Node *log_addr = __ AddP(no_base, buffer, next_index);
__ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
// update the index
__ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
} __ else_(); {
// logging buffer is full, call the runtime
const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
__ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
} __ end_if(); // (!index)
} __ end_if(); // (pre_val != NULL)
} __ end_if(); // (!marking)
// Final sync IdealKit and GraphKit.
final_sync(ideal);
}
//
// Update the card table and add card address to the queue
//
void GraphKit::g1_mark_card(IdealKit& ideal,
Node* card_adr,
Node* oop_store,
uint oop_alias_idx,
Node* index,
Node* index_adr,
Node* buffer,
const TypeFunc* tf) {
Node* zero = __ ConI(0);
Node* zeroX = __ ConX(0);
Node* no_base = __ top();
BasicType card_bt = T_BYTE;
// Smash zero into card. MUST BE ORDERED WRT TO STORE
__ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
// Now do the queue work
__ if_then(index, BoolTest::ne, zeroX); {
Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
Node* log_addr = __ AddP(no_base, buffer, next_index);
// Order, see storeCM.
__ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
__ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
} __ else_(); {
__ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
} __ end_if();
}
void GraphKit::g1_write_barrier_post(Node* oop_store,
Node* obj,
Node* adr,
uint alias_idx,
Node* val,
BasicType bt,
bool use_precise) {
// If we are writing a NULL then we need no post barrier
if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
// Must be NULL
const Type* t = val->bottom_type();
assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
// No post barrier if writing NULLx
return;
}
if (!use_precise) {
// All card marks for a (non-array) instance are in one place:
adr = obj;
}
// (Else it's an array (or unknown), and we want more precise card marks.)
assert(adr != NULL, "");
IdealKit ideal(this, true);
Node* tls = __ thread(); // ThreadLocalStorage
Node* no_base = __ top();
float likely = PROB_LIKELY(0.999);
float unlikely = PROB_UNLIKELY(0.999);
Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
Node* zeroX = __ ConX(0);
// Get the alias_index for raw card-mark memory
const TypePtr* card_type = TypeRawPtr::BOTTOM;
const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
// Offsets into the thread
const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
PtrQueue::byte_offset_of_index());
const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
PtrQueue::byte_offset_of_buf());
// Pointers into the thread
Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
// Now some values
// Use ctrl to avoid hoisting these values past a safepoint, which could
// potentially reset these fields in the JavaThread.
Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
// Convert the store obj pointer to an int prior to doing math on it
// Must use ctrl to prevent "integerized oop" existing across safepoint
Node* cast = __ CastPX(__ ctrl(), adr);
// Divide pointer by card size
Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
// Combine card table base and card offset
Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
// If we know the value being stored does it cross regions?
if (val != NULL) {
// Does the store cause us to cross regions?
// Should be able to do an unsigned compare of region_size instead of
// and extra shift. Do we have an unsigned compare??
// Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
// if (xor_res == 0) same region so skip
__ if_then(xor_res, BoolTest::ne, zeroX); {
// No barrier if we are storing a NULL
__ if_then(val, BoolTest::ne, null(), unlikely); {
// Ok must mark the card if not already dirty
// load the original value of the card
Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
__ if_then(card_val, BoolTest::ne, young_card); {
sync_kit(ideal);
// Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
insert_mem_bar(Op_MemBarVolatile, oop_store);
__ sync_kit(this);
Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
__ if_then(card_val_reload, BoolTest::ne, dirty_card); {
g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
} __ end_if();
} __ end_if();
} __ end_if();
} __ end_if();
} else {
// Object.clone() instrinsic uses this path.
g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
}
// Final sync IdealKit and GraphKit.
final_sync(ideal);
}
#undef __
Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
if (java_lang_String::has_offset_field()) {
int offset_offset = java_lang_String::offset_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
int offset_field_idx = C->get_alias_index(offset_field_type);
return make_load(ctrl,
basic_plus_adr(str, str, offset_offset),
TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
} else {
return intcon(0);
}
}
Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
if (java_lang_String::has_count_field()) {
int count_offset = java_lang_String::count_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* count_field_type = string_type->add_offset(count_offset);
int count_field_idx = C->get_alias_index(count_field_type);
return make_load(ctrl,
basic_plus_adr(str, str, count_offset),
TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
} else {
return load_array_length(load_String_value(ctrl, str));
}
}
Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
int value_offset = java_lang_String::value_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* value_field_type = string_type->add_offset(value_offset);
const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
TypeAry::make(TypeInt::CHAR,TypeInt::POS),
ciTypeArrayKlass::make(T_CHAR), true, 0);
int value_field_idx = C->get_alias_index(value_field_type);
Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
value_type, T_OBJECT, value_field_idx, MemNode::unordered);
// String.value field is known to be @Stable.
if (UseImplicitStableValues) {
load = cast_array_to_stable(load, value_type);
}
return load;
}
void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
int offset_offset = java_lang_String::offset_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
int offset_field_idx = C->get_alias_index(offset_field_type);
store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
value, T_INT, offset_field_idx, MemNode::unordered);
}
void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
int value_offset = java_lang_String::value_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* value_field_type = string_type->add_offset(value_offset);
store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
}
void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
int count_offset = java_lang_String::count_offset_in_bytes();
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
false, NULL, 0);
const TypePtr* count_field_type = string_type->add_offset(count_offset);
int count_field_idx = C->get_alias_index(count_field_type);
store_to_memory(ctrl, basic_plus_adr(str, count_offset),
value, T_INT, count_field_idx, MemNode::unordered);
}
Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
// Reify the property as a CastPP node in Ideal graph to comply with monotonicity
// assumption of CCP analysis.
return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
}
C:\hotspot-69087d08d473\src\share\vm/opto/graphKit.hpp
/*
* Copyright (c) 2001, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_GRAPHKIT_HPP
#define SHARE_VM_OPTO_GRAPHKIT_HPP
#include "ci/ciEnv.hpp"
#include "ci/ciMethodData.hpp"
#include "opto/addnode.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/compile.hpp"
#include "opto/divnode.hpp"
#include "opto/mulnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
#include "opto/type.hpp"
#include "runtime/deoptimization.hpp"
class FastLockNode;
class FastUnlockNode;
class IdealKit;
class LibraryCallKit;
class Parse;
class RootNode;
//-----------------------------------------------------------------------------
//----------------------------GraphKit-----------------------------------------
// Toolkit for building the common sorts of subgraphs.
// Does not know about bytecode parsing or type-flow results.
// It is able to create graphs implementing the semantics of most
// or all bytecodes, so that it can expand intrinsics and calls.
// It may depend on JVMState structure, but it must not depend
// on specific bytecode streams.
class GraphKit : public Phase {
friend class PreserveJVMState;
protected:
ciEnv* _env; // Compilation environment
PhaseGVN &_gvn; // Some optimizations while parsing
SafePointNode* _map; // Parser map from JVM to Nodes
SafePointNode* _exceptions;// Parser map(s) for exception state(s)
int _bci; // JVM Bytecode Pointer
ciMethod* _method; // JVM Current Method
private:
int _sp; // JVM Expression Stack Pointer; don't modify directly!
private:
SafePointNode* map_not_null() const {
assert(_map != NULL, "must call stopped() to test for reset compiler map");
return _map;
}
public:
GraphKit(); // empty constructor
GraphKit(JVMState* jvms); // the JVM state on which to operate
#ifdef ASSERT
~GraphKit() {
assert(!has_exceptions(), "user must call transfer_exceptions_into_jvms");
}
#endif
virtual Parse* is_Parse() const { return NULL; }
virtual LibraryCallKit* is_LibraryCallKit() const { return NULL; }
ciEnv* env() const { return _env; }
PhaseGVN& gvn() const { return _gvn; }
void record_for_igvn(Node* n) const { C->record_for_igvn(n); } // delegate to Compile
// Handy well-known nodes:
Node* null() const { return zerocon(T_OBJECT); }
Node* top() const { return C->top(); }
RootNode* root() const { return C->root(); }
// Create or find a constant node
Node* intcon(jint con) const { return _gvn.intcon(con); }
Node* longcon(jlong con) const { return _gvn.longcon(con); }
Node* makecon(const Type *t) const { return _gvn.makecon(t); }
Node* zerocon(BasicType bt) const { return _gvn.zerocon(bt); }
// (See also macro MakeConX in type.hpp, which uses intcon or longcon.)
// Helper for byte_map_base
Node* byte_map_base_node() {
// Get base of card map
CardTableModRefBS* ct = (CardTableModRefBS*)(Universe::heap()->barrier_set());
assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
if (ct->byte_map_base != NULL) {
return makecon(TypeRawPtr::make((address)ct->byte_map_base));
} else {
return null();
}
}
jint find_int_con(Node* n, jint value_if_unknown) {
return _gvn.find_int_con(n, value_if_unknown);
}
jlong find_long_con(Node* n, jlong value_if_unknown) {
return _gvn.find_long_con(n, value_if_unknown);
}
// (See also macro find_intptr_t_con in type.hpp, which uses one of these.)
// JVM State accessors:
// Parser mapping from JVM indices into Nodes.
// Low slots are accessed by the StartNode::enum.
// Then come the locals at StartNode::Parms to StartNode::Parms+max_locals();
// Then come JVM stack slots.
// Finally come the monitors, if any.
// See layout accessors in class JVMState.
SafePointNode* map() const { return _map; }
bool has_exceptions() const { return _exceptions != NULL; }
JVMState* jvms() const { return map_not_null()->_jvms; }
int sp() const { return _sp; }
int bci() const { return _bci; }
Bytecodes::Code java_bc() const;
ciMethod* method() const { return _method; }
void set_jvms(JVMState* jvms) { set_map(jvms->map());
assert(jvms == this->jvms(), "sanity");
_sp = jvms->sp();
_bci = jvms->bci();
_method = jvms->has_method() ? jvms->method() : NULL; }
void set_map(SafePointNode* m) { _map = m; debug_only(verify_map()); }
void set_sp(int sp) { assert(sp >= 0, err_msg_res("sp must be non-negative: %d", sp)); _sp = sp; }
void clean_stack(int from_sp); // clear garbage beyond from_sp to top
void inc_sp(int i) { set_sp(sp() + i); }
void dec_sp(int i) { set_sp(sp() - i); }
void set_bci(int bci) { _bci = bci; }
// Make sure jvms has current bci & sp.
JVMState* sync_jvms() const;
JVMState* sync_jvms_for_reexecute();
#ifdef ASSERT
// Make sure JVMS has an updated copy of bci and sp.
// Also sanity-check method, depth, and monitor depth.
bool jvms_in_sync() const;
// Make sure the map looks OK.
void verify_map() const;
// Make sure a proposed exception state looks OK.
static void verify_exception_state(SafePointNode* ex_map);
#endif
// Clone the existing map state. (Implements PreserveJVMState.)
SafePointNode* clone_map();
// Set the map to a clone of the given one.
void set_map_clone(SafePointNode* m);
// Tell if the compilation is failing.
bool failing() const { return C->failing(); }
// Set _map to NULL, signalling a stop to further bytecode execution.
// Preserve the map intact for future use, and return it back to the caller.
SafePointNode* stop() { SafePointNode* m = map(); set_map(NULL); return m; }
// Stop, but first smash the map's inputs to NULL, to mark it dead.
void stop_and_kill_map();
// Tell if _map is NULL, or control is top.
bool stopped();
// Tell if this method or any caller method has exception handlers.
bool has_ex_handler();
// Save an exception without blowing stack contents or other JVM state.
// (The extra pointer is stuck with add_req on the map, beyond the JVMS.)
static void set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop);
// Recover a saved exception from its map.
static Node* saved_ex_oop(SafePointNode* ex_map);
// Recover a saved exception from its map, and remove it from the map.
static Node* clear_saved_ex_oop(SafePointNode* ex_map);
#ifdef ASSERT
// Recover a saved exception from its map, and remove it from the map.
static bool has_saved_ex_oop(SafePointNode* ex_map);
#endif
// Push an exception in the canonical position for handlers (stack(0)).
void push_ex_oop(Node* ex_oop) {
ensure_stack(1); // ensure room to push the exception
set_stack(0, ex_oop);
set_sp(1);
clean_stack(1);
}
// Detach and return an exception state.
SafePointNode* pop_exception_state() {
SafePointNode* ex_map = _exceptions;
if (ex_map != NULL) {
_exceptions = ex_map->next_exception();
ex_map->set_next_exception(NULL);
debug_only(verify_exception_state(ex_map));
}
return ex_map;
}
// Add an exception, using the given JVM state, without commoning.
void push_exception_state(SafePointNode* ex_map) {
debug_only(verify_exception_state(ex_map));
ex_map->set_next_exception(_exceptions);
_exceptions = ex_map;
}
// Turn the current JVM state into an exception state, appending the ex_oop.
SafePointNode* make_exception_state(Node* ex_oop);
// Add an exception, using the given JVM state.
// Combine all exceptions with a common exception type into a single state.
// (This is done via combine_exception_states.)
void add_exception_state(SafePointNode* ex_map);
// Combine all exceptions of any sort whatever into a single master state.
SafePointNode* combine_and_pop_all_exception_states() {
if (_exceptions == NULL) return NULL;
SafePointNode* phi_map = pop_exception_state();
SafePointNode* ex_map;
while ((ex_map = pop_exception_state()) != NULL) {
combine_exception_states(ex_map, phi_map);
}
return phi_map;
}
// Combine the two exception states, building phis as necessary.
// The second argument is updated to include contributions from the first.
void combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map);
// Reset the map to the given state. If there are any half-finished phis
// in it (created by combine_exception_states), transform them now.
// Returns the exception oop. (Caller must call push_ex_oop if required.)
Node* use_exception_state(SafePointNode* ex_map);
// Collect exceptions from a given JVM state into my exception list.
void add_exception_states_from(JVMState* jvms);
// Collect all raised exceptions into the current JVM state.
// Clear the current exception list and map, returns the combined states.
JVMState* transfer_exceptions_into_jvms();
// Helper to throw a built-in exception.
// Range checks take the offending index.
// Cast and array store checks take the offending class.
// Others do not take the optional argument.
// The JVMS must allow the bytecode to be re-executed
// via an uncommon trap.
void builtin_throw(Deoptimization::DeoptReason reason, Node* arg = NULL);
// Helper to check the JavaThread::_should_post_on_exceptions flag
// and branch to an uncommon_trap if it is true (with the specified reason and must_throw)
void uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
bool must_throw) ;
// Helper Functions for adding debug information
void kill_dead_locals();
#ifdef ASSERT
bool dead_locals_are_killed();
#endif
// The call may deoptimize. Supply required JVM state as debug info.
// If must_throw is true, the call is guaranteed not to return normally.
void add_safepoint_edges(SafePointNode* call,
bool must_throw = false);
// How many stack inputs does the current BC consume?
// And, how does the stack change after the bytecode?
// Returns false if unknown.
bool compute_stack_effects(int& inputs, int& depth);
// Add a fixed offset to a pointer
Node* basic_plus_adr(Node* base, Node* ptr, intptr_t offset) {
return basic_plus_adr(base, ptr, MakeConX(offset));
}
Node* basic_plus_adr(Node* base, intptr_t offset) {
return basic_plus_adr(base, base, MakeConX(offset));
}
// Add a variable offset to a pointer
Node* basic_plus_adr(Node* base, Node* offset) {
return basic_plus_adr(base, base, offset);
}
Node* basic_plus_adr(Node* base, Node* ptr, Node* offset);
// Some convenient shortcuts for common nodes
Node* IfTrue(IfNode* iff) { return _gvn.transform(new (C) IfTrueNode(iff)); }
Node* IfFalse(IfNode* iff) { return _gvn.transform(new (C) IfFalseNode(iff)); }
Node* AddI(Node* l, Node* r) { return _gvn.transform(new (C) AddINode(l, r)); }
Node* SubI(Node* l, Node* r) { return _gvn.transform(new (C) SubINode(l, r)); }
Node* MulI(Node* l, Node* r) { return _gvn.transform(new (C) MulINode(l, r)); }
Node* DivI(Node* ctl, Node* l, Node* r) { return _gvn.transform(new (C) DivINode(ctl, l, r)); }
Node* AndI(Node* l, Node* r) { return _gvn.transform(new (C) AndINode(l, r)); }
Node* OrI(Node* l, Node* r) { return _gvn.transform(new (C) OrINode(l, r)); }
Node* XorI(Node* l, Node* r) { return _gvn.transform(new (C) XorINode(l, r)); }
Node* MaxI(Node* l, Node* r) { return _gvn.transform(new (C) MaxINode(l, r)); }
Node* MinI(Node* l, Node* r) { return _gvn.transform(new (C) MinINode(l, r)); }
Node* LShiftI(Node* l, Node* r) { return _gvn.transform(new (C) LShiftINode(l, r)); }
Node* RShiftI(Node* l, Node* r) { return _gvn.transform(new (C) RShiftINode(l, r)); }
Node* URShiftI(Node* l, Node* r) { return _gvn.transform(new (C) URShiftINode(l, r)); }
Node* CmpI(Node* l, Node* r) { return _gvn.transform(new (C) CmpINode(l, r)); }
Node* CmpL(Node* l, Node* r) { return _gvn.transform(new (C) CmpLNode(l, r)); }
Node* CmpP(Node* l, Node* r) { return _gvn.transform(new (C) CmpPNode(l, r)); }
Node* Bool(Node* cmp, BoolTest::mask relop) { return _gvn.transform(new (C) BoolNode(cmp, relop)); }
Node* AddP(Node* b, Node* a, Node* o) { return _gvn.transform(new (C) AddPNode(b, a, o)); }
// Convert between int and long, and size_t.
// (See macros ConvI2X, etc., in type.hpp for ConvI2X, etc.)
Node* ConvI2L(Node* offset);
Node* ConvI2UL(Node* offset);
Node* ConvL2I(Node* offset);
// Find out the klass of an object.
Node* load_object_klass(Node* object);
// Find out the length of an array.
Node* load_array_length(Node* array);
// Helper function to do a NULL pointer check or ZERO check based on type.
// Throw an exception if a given value is null.
// Return the value cast to not-null.
// Be clever about equivalent dominating null checks.
Node* null_check_common(Node* value, BasicType type,
bool assert_null = false, Node* *null_control = NULL);
Node* null_check(Node* value, BasicType type = T_OBJECT) {
return null_check_common(value, type);
}
Node* null_check_receiver() {
assert(argument(0)->bottom_type()->isa_ptr(), "must be");
return null_check(argument(0));
}
Node* zero_check_int(Node* value) {
assert(value->bottom_type()->basic_type() == T_INT,
err_msg_res("wrong type: %s", type2name(value->bottom_type()->basic_type())));
return null_check_common(value, T_INT);
}
Node* zero_check_long(Node* value) {
assert(value->bottom_type()->basic_type() == T_LONG,
err_msg_res("wrong type: %s", type2name(value->bottom_type()->basic_type())));
return null_check_common(value, T_LONG);
}
// Throw an uncommon trap if a given value is __not__ null.
// Return the value cast to null, and be clever about dominating checks.
Node* null_assert(Node* value, BasicType type = T_OBJECT) {
return null_check_common(value, type, true);
}
// Null check oop. Return null-path control into (*null_control).
// Return a cast-not-null node which depends on the not-null control.
// If never_see_null, use an uncommon trap (*null_control sees a top).
// The cast is not valid along the null path; keep a copy of the original.
// If safe_for_replace, then we can replace the value with the cast
// in the parsing map (the cast is guaranteed to dominate the map)
Node* null_check_oop(Node* value, Node* *null_control,
bool never_see_null = false, bool safe_for_replace = false);
// Check the null_seen bit.
bool seems_never_null(Node* obj, ciProfileData* data);
// Check for unique class for receiver at call
ciKlass* profile_has_unique_klass() {
ciCallProfile profile = method()->call_profile_at_bci(bci());
if (profile.count() >= 0 && // no cast failures here
profile.has_receiver(0) &&
profile.morphism() == 1) {
return profile.receiver(0);
}
return NULL;
}
// record type from profiling with the type system
Node* record_profile_for_speculation(Node* n, ciKlass* exact_kls);
Node* record_profiled_receiver_for_speculation(Node* n);
void record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc);
void record_profiled_parameters_for_speculation();
// Use the type profile to narrow an object type.
Node* maybe_cast_profiled_receiver(Node* not_null_obj,
ciKlass* require_klass,
ciKlass* spec,
bool safe_for_replace);
// Cast obj to type and emit guard unless we had too many traps here already
Node* maybe_cast_profiled_obj(Node* obj,
ciKlass* type,
bool not_null = false);
// Cast obj to not-null on this path
Node* cast_not_null(Node* obj, bool do_replace_in_map = true);
// Replace all occurrences of one node by another.
void replace_in_map(Node* old, Node* neww);
void push(Node* n) { map_not_null(); _map->set_stack(_map->_jvms, _sp++ , n); }
Node* pop() { map_not_null(); return _map->stack( _map->_jvms, --_sp ); }
Node* peek(int off = 0) { map_not_null(); return _map->stack( _map->_jvms, _sp - off - 1 ); }
void push_pair(Node* ldval) {
push(ldval);
push(top()); // the halfword is merely a placeholder
}
void push_pair_local(int i) {
// longs are stored in locals in "push" order
push( local(i+0) ); // the real value
assert(local(i+1) == top(), "");
push(top()); // halfword placeholder
}
Node* pop_pair() {
// the second half is pushed last & popped first; it contains exactly nothing
Node* halfword = pop();
assert(halfword == top(), "");
// the long bits are pushed first & popped last:
return pop();
}
void set_pair_local(int i, Node* lval) {
// longs are stored in locals as a value/half pair (like doubles)
set_local(i+0, lval);
set_local(i+1, top());
}
// Push the node, which may be zero, one, or two words.
void push_node(BasicType n_type, Node* n) {
int n_size = type2size[n_type];
if (n_size == 1) push( n ); // T_INT, ...
else if (n_size == 2) push_pair( n ); // T_DOUBLE, T_LONG
else { assert(n_size == 0, "must be T_VOID"); }
}
Node* pop_node(BasicType n_type) {
int n_size = type2size[n_type];
if (n_size == 1) return pop();
else if (n_size == 2) return pop_pair();
else return NULL;
}
Node* control() const { return map_not_null()->control(); }
Node* i_o() const { return map_not_null()->i_o(); }
Node* returnadr() const { return map_not_null()->returnadr(); }
Node* frameptr() const { return map_not_null()->frameptr(); }
Node* local(uint idx) const { map_not_null(); return _map->local( _map->_jvms, idx); }
Node* stack(uint idx) const { map_not_null(); return _map->stack( _map->_jvms, idx); }
Node* argument(uint idx) const { map_not_null(); return _map->argument( _map->_jvms, idx); }
Node* monitor_box(uint idx) const { map_not_null(); return _map->monitor_box(_map->_jvms, idx); }
Node* monitor_obj(uint idx) const { map_not_null(); return _map->monitor_obj(_map->_jvms, idx); }
void set_control (Node* c) { map_not_null()->set_control(c); }
void set_i_o (Node* c) { map_not_null()->set_i_o(c); }
void set_local(uint idx, Node* c) { map_not_null(); _map->set_local( _map->_jvms, idx, c); }
void set_stack(uint idx, Node* c) { map_not_null(); _map->set_stack( _map->_jvms, idx, c); }
void set_argument(uint idx, Node* c){ map_not_null(); _map->set_argument(_map->_jvms, idx, c); }
void ensure_stack(uint stk_size) { map_not_null(); _map->ensure_stack(_map->_jvms, stk_size); }
// Access unaliased memory
Node* memory(uint alias_idx);
Node* memory(const TypePtr *tp) { return memory(C->get_alias_index(tp)); }
Node* memory(Node* adr) { return memory(_gvn.type(adr)->is_ptr()); }
// Access immutable memory
Node* immutable_memory() { return C->immutable_memory(); }
// Set unaliased memory
void set_memory(Node* c, uint alias_idx) { merged_memory()->set_memory_at(alias_idx, c); }
void set_memory(Node* c, const TypePtr *tp) { set_memory(c,C->get_alias_index(tp)); }
void set_memory(Node* c, Node* adr) { set_memory(c,_gvn.type(adr)->is_ptr()); }
// Get the entire memory state (probably a MergeMemNode), and reset it
// (The resetting prevents somebody from using the dangling Node pointer.)
Node* reset_memory();
// Get the entire memory state, asserted to be a MergeMemNode.
MergeMemNode* merged_memory() {
Node* mem = map_not_null()->memory();
assert(mem->is_MergeMem(), "parse memory is always pre-split");
return mem->as_MergeMem();
}
// Set the entire memory state; produce a new MergeMemNode.
void set_all_memory(Node* newmem);
// Create a memory projection from the call, then set_all_memory.
void set_all_memory_call(Node* call, bool separate_io_proj = false);
// Create a LoadNode, reading from the parser's memory state.
// (Note: require_atomic_access is useful only with T_LONG.)
//
// We choose the unordered semantics by default because we have
// adapted the `do_put_xxx' and `do_get_xxx' procedures for the case
// of volatile fields.
Node* make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency = LoadNode::DependsOnlyOnTest,
bool require_atomic_access = false, bool unaligned = false,
bool mismatched = false) {
// This version computes alias_index from bottom_type
return make_load(ctl, adr, t, bt, adr->bottom_type()->is_ptr(),
mo, control_dependency, require_atomic_access,
unaligned, mismatched);
}
Node* make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, const TypePtr* adr_type,
MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency = LoadNode::DependsOnlyOnTest,
bool require_atomic_access = false, bool unaligned = false,
bool mismatched = false) {
// This version computes alias_index from an address type
assert(adr_type != NULL, "use other make_load factory");
return make_load(ctl, adr, t, bt, C->get_alias_index(adr_type),
mo, control_dependency, require_atomic_access,
unaligned, mismatched);
}
// This is the base version which is given an alias index.
Node* make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, int adr_idx,
MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency = LoadNode::DependsOnlyOnTest,
bool require_atomic_access = false, bool unaligned = false,
bool mismatched = false);
// Create & transform a StoreNode and store the effect into the
// parser's memory state.
//
// We must ensure that stores of object references will be visible
// only after the object's initialization. So the clients of this
// procedure must indicate that the store requires `release'
// semantics, if the stored value is an object reference that might
// point to a new object and may become externally visible.
Node* store_to_memory(Node* ctl, Node* adr, Node* val, BasicType bt,
const TypePtr* adr_type,
MemNode::MemOrd mo,
bool require_atomic_access = false,
bool unaligned = false,
bool mismatched = false) {
// This version computes alias_index from an address type
assert(adr_type != NULL, "use other store_to_memory factory");
return store_to_memory(ctl, adr, val, bt,
C->get_alias_index(adr_type),
mo, require_atomic_access,
unaligned, mismatched);
}
// This is the base version which is given alias index
// Return the new StoreXNode
Node* store_to_memory(Node* ctl, Node* adr, Node* val, BasicType bt,
int adr_idx,
MemNode::MemOrd,
bool require_atomic_access = false,
bool unaligned = false,
bool mismatched = false);
// All in one pre-barrier, store, post_barrier
// Insert a write-barrier'd store. This is to let generational GC
// work; we have to flag all oop-stores before the next GC point.
//
// It comes in 3 flavors of store to an object, array, or unknown.
// We use precise card marks for arrays to avoid scanning the entire
// array. We use imprecise for object. We use precise for unknown
// since we don't know if we have an array or and object or even
// where the object starts.
//
// If val==NULL, it is taken to be a completely unknown value. QQQ
Node* store_oop(Node* ctl,
Node* obj, // containing obj
Node* adr, // actual adress to store val at
const TypePtr* adr_type,
Node* val,
const TypeOopPtr* val_type,
BasicType bt,
bool use_precise,
MemNode::MemOrd mo,
bool mismatched = false);
Node* store_oop_to_object(Node* ctl,
Node* obj, // containing obj
Node* adr, // actual adress to store val at
const TypePtr* adr_type,
Node* val,
const TypeOopPtr* val_type,
BasicType bt,
MemNode::MemOrd mo) {
return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, false, mo);
}
Node* store_oop_to_array(Node* ctl,
Node* obj, // containing obj
Node* adr, // actual adress to store val at
const TypePtr* adr_type,
Node* val,
const TypeOopPtr* val_type,
BasicType bt,
MemNode::MemOrd mo) {
return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
}
// Could be an array or object we don't know at compile time (unsafe ref.)
Node* store_oop_to_unknown(Node* ctl,
Node* obj, // containing obj
Node* adr, // actual adress to store val at
const TypePtr* adr_type,
Node* val,
BasicType bt,
MemNode::MemOrd mo,
bool mismatched = false);
// For the few case where the barriers need special help
void pre_barrier(bool do_load, Node* ctl,
Node* obj, Node* adr, uint adr_idx, Node* val, const TypeOopPtr* val_type,
Node* pre_val,
BasicType bt);
void post_barrier(Node* ctl, Node* store, Node* obj, Node* adr, uint adr_idx,
Node* val, BasicType bt, bool use_precise);
// Return addressing for an array element.
Node* array_element_address(Node* ary, Node* idx, BasicType elembt,
// Optional constraint on the array size:
const TypeInt* sizetype = NULL,
// Optional control dependency (for example, on range check)
Node* ctrl = NULL);
// Return a load of array element at idx.
Node* load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype);
//---------------- Dtrace support --------------------
void make_dtrace_method_entry_exit(ciMethod* method, bool is_entry);
void make_dtrace_method_entry(ciMethod* method) {
make_dtrace_method_entry_exit(method, true);
}
void make_dtrace_method_exit(ciMethod* method) {
make_dtrace_method_entry_exit(method, false);
}
//--------------- stub generation -------------------
public:
void gen_stub(address C_function,
const char *name,
int is_fancy_jump,
bool pass_tls,
bool return_pc);
//---------- help for generating calls --------------
// Do a null check on the receiver as it would happen before the call to
// callee (with all arguments still on the stack).
Node* null_check_receiver_before_call(ciMethod* callee) {
assert(!callee->is_static(), "must be a virtual method");
// Callsite signature can be different from actual method being called (i.e _linkTo* sites).
// Use callsite signature always.
ciMethod* declared_method = method()->get_method_at_bci(bci());
const int nargs = declared_method->arg_size();
inc_sp(nargs);
Node* n = null_check_receiver();
dec_sp(nargs);
return n;
}
// Fill in argument edges for the call from argument(0), argument(1), ...
// (The next step is to call set_edges_for_java_call.)
void set_arguments_for_java_call(CallJavaNode* call);
// Fill in non-argument edges for the call.
// Transform the call, and update the basics: control, i_o, memory.
// (The next step is usually to call set_results_for_java_call.)
void set_edges_for_java_call(CallJavaNode* call,
bool must_throw = false, bool separate_io_proj = false);
// Finish up a java call that was started by set_edges_for_java_call.
// Call add_exception on any throw arising from the call.
// Return the call result (transformed).
Node* set_results_for_java_call(CallJavaNode* call, bool separate_io_proj = false);
// Similar to set_edges_for_java_call, but simplified for runtime calls.
void set_predefined_output_for_runtime_call(Node* call) {
set_predefined_output_for_runtime_call(call, NULL, NULL);
}
void set_predefined_output_for_runtime_call(Node* call,
Node* keep_mem,
const TypePtr* hook_mem);
Node* set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem = NULL);
// Replace the call with the current state of the kit. Requires
// that the call was generated with separate io_projs so that
// exceptional control flow can be handled properly.
void replace_call(CallNode* call, Node* result, bool do_replaced_nodes = false);
// helper functions for statistics
void increment_counter(address counter_addr); // increment a debug counter
void increment_counter(Node* counter_addr); // increment a debug counter
// Bail out to the interpreter right now
// The optional klass is the one causing the trap.
// The optional reason is debug information written to the compile log.
// Optional must_throw is the same as with add_safepoint_edges.
void uncommon_trap(int trap_request,
ciKlass* klass = NULL, const char* reason_string = NULL,
bool must_throw = false, bool keep_exact_action = false);
// Shorthand, to avoid saying "Deoptimization::" so many times.
void uncommon_trap(Deoptimization::DeoptReason reason,
Deoptimization::DeoptAction action,
ciKlass* klass = NULL, const char* reason_string = NULL,
bool must_throw = false, bool keep_exact_action = false) {
uncommon_trap(Deoptimization::make_trap_request(reason, action),
klass, reason_string, must_throw, keep_exact_action);
}
// Bail out to the interpreter and keep exact action (avoid switching to Action_none).
void uncommon_trap_exact(Deoptimization::DeoptReason reason,
Deoptimization::DeoptAction action,
ciKlass* klass = NULL, const char* reason_string = NULL,
bool must_throw = false) {
uncommon_trap(Deoptimization::make_trap_request(reason, action),
klass, reason_string, must_throw, /*keep_exact_action=*/true);
}
// SP when bytecode needs to be reexecuted.
virtual int reexecute_sp() { return sp(); }
// Report if there were too many traps at the current method and bci.
// Report if a trap was recorded, and/or PerMethodTrapLimit was exceeded.
// If there is no MDO at all, report no trap unless told to assume it.
bool too_many_traps(Deoptimization::DeoptReason reason) {
return C->too_many_traps(method(), bci(), reason);
}
// Report if there were too many recompiles at the current method and bci.
bool too_many_recompiles(Deoptimization::DeoptReason reason) {
return C->too_many_recompiles(method(), bci(), reason);
}
// Returns the object (if any) which was created the moment before.
Node* just_allocated_object(Node* current_control);
static bool use_ReduceInitialCardMarks() {
return (ReduceInitialCardMarks
&& Universe::heap()->can_elide_tlab_store_barriers());
}
// Sync Ideal and Graph kits.
void sync_kit(IdealKit& ideal);
void final_sync(IdealKit& ideal);
// vanilla/CMS post barrier
void write_barrier_post(Node *store, Node* obj,
Node* adr, uint adr_idx, Node* val, bool use_precise);
// Allow reordering of pre-barrier with oop store and/or post-barrier.
// Used for load_store operations which loads old value.
bool can_move_pre_barrier() const;
// G1 pre/post barriers
void g1_write_barrier_pre(bool do_load,
Node* obj,
Node* adr,
uint alias_idx,
Node* val,
const TypeOopPtr* val_type,
Node* pre_val,
BasicType bt);
void g1_write_barrier_post(Node* store,
Node* obj,
Node* adr,
uint alias_idx,
Node* val,
BasicType bt,
bool use_precise);
// Helper function for g1
private:
void g1_mark_card(IdealKit& ideal, Node* card_adr, Node* store, uint oop_alias_idx,
Node* index, Node* index_adr,
Node* buffer, const TypeFunc* tf);
public:
// Helper function to round double arguments before a call
void round_double_arguments(ciMethod* dest_method);
void round_double_result(ciMethod* dest_method);
// rounding for strict float precision conformance
Node* precision_rounding(Node* n);
// rounding for strict double precision conformance
Node* dprecision_rounding(Node* n);
// rounding for non-strict double stores
Node* dstore_rounding(Node* n);
// Helper functions for fast/slow path codes
Node* opt_iff(Node* region, Node* iff);
Node* make_runtime_call(int flags,
const TypeFunc* call_type, address call_addr,
const char* call_name,
const TypePtr* adr_type, // NULL if no memory effects
Node* parm0 = NULL, Node* parm1 = NULL,
Node* parm2 = NULL, Node* parm3 = NULL,
Node* parm4 = NULL, Node* parm5 = NULL,
Node* parm6 = NULL, Node* parm7 = NULL);
enum { // flag values for make_runtime_call
RC_NO_FP = 1, // CallLeafNoFPNode
RC_NO_IO = 2, // do not hook IO edges
RC_NO_LEAF = 4, // CallStaticJavaNode
RC_MUST_THROW = 8, // flag passed to add_safepoint_edges
RC_NARROW_MEM = 16, // input memory is same as output
RC_UNCOMMON = 32, // freq. expected to be like uncommon trap
RC_LEAF = 0 // null value: no flags set
};
// merge in all memory slices from new_mem, along the given path
void merge_memory(Node* new_mem, Node* region, int new_path);
void make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize = false);
// Helper functions to build synchronizations
int next_monitor();
Node* insert_mem_bar(int opcode, Node* precedent = NULL);
Node* insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent = NULL);
// Optional 'precedent' is appended as an extra edge, to force ordering.
FastLockNode* shared_lock(Node* obj);
void shared_unlock(Node* box, Node* obj);
// helper functions for the fast path/slow path idioms
Node* fast_and_slow(Node* in, const Type *result_type, Node* null_result, IfNode* fast_test, Node* fast_result, address slow_call, const TypeFunc *slow_call_type, Node* slow_arg, Klass* ex_klass, Node* slow_result);
// Generate an instance-of idiom. Used by both the instance-of bytecode
// and the reflective instance-of call.
Node* gen_instanceof(Node *subobj, Node* superkls, bool safe_for_replace = false);
// Generate a check-cast idiom. Used by both the check-cast bytecode
// and the array-store bytecode
Node* gen_checkcast( Node *subobj, Node* superkls,
Node* *failure_control = NULL );
// Generate a subtyping check. Takes as input the subtype and supertype.
// Returns 2 values: sets the default control() to the true path and
// returns the false path. Only reads from constant memory taken from the
// default memory; does not write anything. It also doesn't take in an
// Object; if you wish to check an Object you need to load the Object's
// class prior to coming here.
Node* gen_subtype_check(Node* subklass, Node* superklass);
// Static parse-time type checking logic for gen_subtype_check:
enum { SSC_always_false, SSC_always_true, SSC_easy_test, SSC_full_test };
int static_subtype_check(ciKlass* superk, ciKlass* subk);
// Exact type check used for predicted calls and casts.
// Rewrites (*casted_receiver) to be casted to the stronger type.
// (Caller is responsible for doing replace_in_map.)
Node* type_check_receiver(Node* receiver, ciKlass* klass, float prob,
Node* *casted_receiver);
// implementation of object creation
Node* set_output_for_allocation(AllocateNode* alloc,
const TypeOopPtr* oop_type,
bool deoptimize_on_exception=false);
Node* get_layout_helper(Node* klass_node, jint& constant_value);
Node* new_instance(Node* klass_node,
Node* slow_test = NULL,
Node* *return_size_val = NULL,
bool deoptimize_on_exception = false);
Node* new_array(Node* klass_node, Node* count_val, int nargs,
Node* *return_size_val = NULL,
bool deoptimize_on_exception = false);
// java.lang.String helpers
Node* load_String_offset(Node* ctrl, Node* str);
Node* load_String_length(Node* ctrl, Node* str);
Node* load_String_value(Node* ctrl, Node* str);
void store_String_offset(Node* ctrl, Node* str, Node* value);
void store_String_length(Node* ctrl, Node* str, Node* value);
void store_String_value(Node* ctrl, Node* str, Node* value);
// Handy for making control flow
IfNode* create_and_map_if(Node* ctrl, Node* tst, float prob, float cnt) {
IfNode* iff = new (C) IfNode(ctrl, tst, prob, cnt);// New IfNode's
_gvn.set_type(iff, iff->Value(&_gvn)); // Value may be known at parse-time
// Place 'if' on worklist if it will be in graph
if (!tst->is_Con()) record_for_igvn(iff); // Range-check and Null-check removal is later
return iff;
}
IfNode* create_and_xform_if(Node* ctrl, Node* tst, float prob, float cnt) {
IfNode* iff = new (C) IfNode(ctrl, tst, prob, cnt);// New IfNode's
_gvn.transform(iff); // Value may be known at parse-time
// Place 'if' on worklist if it will be in graph
if (!tst->is_Con()) record_for_igvn(iff); // Range-check and Null-check removal is later
return iff;
}
// Insert a loop predicate into the graph
void add_predicate(int nargs = 0);
void add_predicate_impl(Deoptimization::DeoptReason reason, int nargs);
// Produce new array node of stable type
Node* cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type);
};
// Helper class to support building of control flow branches. Upon
// creation the map and sp at bci are cloned and restored upon de-
// struction. Typical use:
//
// { PreserveJVMState pjvms(this);
// // code of new branch
// }
// // here the JVM state at bci is established
class PreserveJVMState: public StackObj {
protected:
GraphKit* _kit;
#ifdef ASSERT
int _block; // PO of current block, if a Parse
int _bci;
#endif
SafePointNode* _map;
uint _sp;
public:
PreserveJVMState(GraphKit* kit, bool clone_map = true);
~PreserveJVMState();
};
// Helper class to build cutouts of the form if (p) ; else {x...}.
// The code {x...} must not fall through.
// The kit's main flow of control is set to the "then" continuation of if(p).
class BuildCutout: public PreserveJVMState {
public:
BuildCutout(GraphKit* kit, Node* p, float prob, float cnt = COUNT_UNKNOWN);
~BuildCutout();
};
// Helper class to preserve the original _reexecute bit and _sp and restore
// them back
class PreserveReexecuteState: public StackObj {
protected:
GraphKit* _kit;
uint _sp;
JVMState::ReexecuteState _reexecute;
public:
PreserveReexecuteState(GraphKit* kit);
~PreserveReexecuteState();
};
#endif // SHARE_VM_OPTO_GRAPHKIT_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/idealGraphPrinter.cpp
/*
* Copyright (c) 2007, 2018, 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 "precompiled.hpp"
#include "opto/chaitin.hpp"
#include "opto/idealGraphPrinter.hpp"
#include "opto/machnode.hpp"
#include "opto/parse.hpp"
#include "runtime/threadCritical.hpp"
#ifndef PRODUCT
// Constants
// Keep consistent with Java constants
const char *IdealGraphPrinter::INDENT = " ";
const char *IdealGraphPrinter::TOP_ELEMENT = "graphDocument";
const char *IdealGraphPrinter::GROUP_ELEMENT = "group";
const char *IdealGraphPrinter::GRAPH_ELEMENT = "graph";
const char *IdealGraphPrinter::PROPERTIES_ELEMENT = "properties";
const char *IdealGraphPrinter::EDGES_ELEMENT = "edges";
const char *IdealGraphPrinter::PROPERTY_ELEMENT = "p";
const char *IdealGraphPrinter::EDGE_ELEMENT = "edge";
const char *IdealGraphPrinter::NODE_ELEMENT = "node";
const char *IdealGraphPrinter::NODES_ELEMENT = "nodes";
const char *IdealGraphPrinter::REMOVE_EDGE_ELEMENT = "removeEdge";
const char *IdealGraphPrinter::REMOVE_NODE_ELEMENT = "removeNode";
const char *IdealGraphPrinter::METHOD_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::METHOD_IS_PUBLIC_PROPERTY = "public";
const char *IdealGraphPrinter::METHOD_IS_STATIC_PROPERTY = "static";
const char *IdealGraphPrinter::TRUE_VALUE = "true";
const char *IdealGraphPrinter::NODE_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::EDGE_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::NODE_ID_PROPERTY = "id";
const char *IdealGraphPrinter::FROM_PROPERTY = "from";
const char *IdealGraphPrinter::TO_PROPERTY = "to";
const char *IdealGraphPrinter::PROPERTY_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::GRAPH_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::INDEX_PROPERTY = "index";
const char *IdealGraphPrinter::METHOD_ELEMENT = "method";
const char *IdealGraphPrinter::INLINE_ELEMENT = "inline";
const char *IdealGraphPrinter::BYTECODES_ELEMENT = "bytecodes";
const char *IdealGraphPrinter::METHOD_BCI_PROPERTY = "bci";
const char *IdealGraphPrinter::METHOD_SHORT_NAME_PROPERTY = "shortName";
const char *IdealGraphPrinter::CONTROL_FLOW_ELEMENT = "controlFlow";
const char *IdealGraphPrinter::BLOCK_NAME_PROPERTY = "name";
const char *IdealGraphPrinter::BLOCK_DOMINATOR_PROPERTY = "dom";
const char *IdealGraphPrinter::BLOCK_ELEMENT = "block";
const char *IdealGraphPrinter::SUCCESSORS_ELEMENT = "successors";
const char *IdealGraphPrinter::SUCCESSOR_ELEMENT = "successor";
const char *IdealGraphPrinter::ASSEMBLY_ELEMENT = "assembly";
int IdealGraphPrinter::_file_count = 0;
IdealGraphPrinter *IdealGraphPrinter::printer() {
if (PrintIdealGraphLevel == 0) return NULL;
JavaThread *thread = JavaThread::current();
if (!thread->is_Compiler_thread()) return NULL;
CompilerThread *compiler_thread = (CompilerThread *)thread;
if (compiler_thread->ideal_graph_printer() == NULL) {
IdealGraphPrinter *printer = new IdealGraphPrinter();
compiler_thread->set_ideal_graph_printer(printer);
}
return compiler_thread->ideal_graph_printer();
}
void IdealGraphPrinter::clean_up() {
JavaThread *p;
for (p = Threads::first(); p; p = p->next()) {
if (p->is_Compiler_thread()) {
CompilerThread *c = (CompilerThread *)p;
IdealGraphPrinter *printer = c->ideal_graph_printer();
if (printer) {
delete printer;
}
c->set_ideal_graph_printer(NULL);
}
}
}
// Constructor, either file or network output
IdealGraphPrinter::IdealGraphPrinter() {
// By default dump both ins and outs since dead or unreachable code
// needs to appear in the graph. There are also some special cases
// in the mach where kill projections have no users but should
// appear in the dump.
_traverse_outs = true;
_should_send_method = true;
_output = NULL;
buffer[0] = 0;
_depth = 0;
_current_method = NULL;
assert(!_current_method, "current method must be initialized to NULL");
_stream = NULL;
if (PrintIdealGraphFile != NULL) {
ThreadCritical tc;
// User wants all output to go to files
if (_file_count != 0) {
ResourceMark rm;
stringStream st;
const char* dot = strrchr(PrintIdealGraphFile, '.');
if (dot) {
st.write(PrintIdealGraphFile, dot - PrintIdealGraphFile);
st.print("%d%s", _file_count, dot);
} else {
st.print("%s%d", PrintIdealGraphFile, _file_count);
}
fileStream *stream = new (ResourceObj::C_HEAP, mtCompiler) fileStream(st.as_string());
_output = stream;
} else {
fileStream *stream = new (ResourceObj::C_HEAP, mtCompiler) fileStream(PrintIdealGraphFile);
_output = stream;
}
_file_count++;
} else {
_stream = new (ResourceObj::C_HEAP, mtCompiler) networkStream();
// Try to connect to visualizer
if (_stream->connect(PrintIdealGraphAddress, PrintIdealGraphPort)) {
char c = 0;
_stream->read(&c, 1);
if (c != 'y') {
tty->print_cr("Client available, but does not want to receive data!");
_stream->close();
delete _stream;
_stream = NULL;
return;
}
_output = _stream;
} else {
// It would be nice if we could shut down cleanly but it should
// be an error if we can't connect to the visualizer.
fatal(err_msg_res("Couldn't connect to visualizer at %s:" INTX_FORMAT,
PrintIdealGraphAddress, PrintIdealGraphPort));
}
}
_xml = new (ResourceObj::C_HEAP, mtCompiler) xmlStream(_output);
head(TOP_ELEMENT);
}
// Destructor, close file or network stream
IdealGraphPrinter::~IdealGraphPrinter() {
tail(TOP_ELEMENT);
// tty->print_cr("Walk time: %d", (int)_walk_time.milliseconds());
// tty->print_cr("Output time: %d", (int)_output_time.milliseconds());
// tty->print_cr("Build blocks time: %d", (int)_build_blocks_time.milliseconds());
if(_xml) {
delete _xml;
_xml = NULL;
}
if (_stream) {
delete _stream;
if (_stream == _output) {
_output = NULL;
}
_stream = NULL;
}
if (_output) {
delete _output;
_output = NULL;
}
}
void IdealGraphPrinter::begin_elem(const char *s) {
_xml->begin_elem("%s", s);
}
void IdealGraphPrinter::end_elem() {
_xml->end_elem();
}
void IdealGraphPrinter::begin_head(const char *s) {
_xml->begin_head("%s", s);
}
void IdealGraphPrinter::end_head() {
_xml->end_head();
}
void IdealGraphPrinter::print_attr(const char *name, intptr_t val) {
stringStream stream;
stream.print(INTX_FORMAT, val);
print_attr(name, stream.as_string());
}
void IdealGraphPrinter::print_attr(const char *name, const char *val) {
_xml->print(" %s='", name);
text(val);
_xml->print("'");
}
void IdealGraphPrinter::head(const char *name) {
_xml->head("%s", name);
}
void IdealGraphPrinter::tail(const char *name) {
_xml->tail(name);
}
void IdealGraphPrinter::text(const char *s) {
_xml->text("%s", s);
}
void IdealGraphPrinter::print_prop(const char *name, int val) {
stringStream stream;
stream.print("%d", val);
print_prop(name, stream.as_string());
}
void IdealGraphPrinter::print_prop(const char *name, const char *val) {
begin_head(PROPERTY_ELEMENT);
print_attr(PROPERTY_NAME_PROPERTY, name);
end_head();
text(val);
tail(PROPERTY_ELEMENT);
}
void IdealGraphPrinter::print_method(ciMethod *method, int bci, InlineTree *tree) {
begin_head(METHOD_ELEMENT);
stringStream str;
method->print_name(&str);
stringStream shortStr;
method->print_short_name(&shortStr);
print_attr(METHOD_NAME_PROPERTY, str.as_string());
print_attr(METHOD_SHORT_NAME_PROPERTY, shortStr.as_string());
print_attr(METHOD_BCI_PROPERTY, bci);
end_head();
head(BYTECODES_ELEMENT);
output()->print_cr("<![CDATA[");
method->print_codes_on(output());
output()->print_cr("]]>");
tail(BYTECODES_ELEMENT);
head(INLINE_ELEMENT);
if (tree != NULL) {
GrowableArray<InlineTree *> subtrees = tree->subtrees();
for (int i = 0; i < subtrees.length(); i++) {
print_inline_tree(subtrees.at(i));
}
}
tail(INLINE_ELEMENT);
tail(METHOD_ELEMENT);
output()->flush();
}
void IdealGraphPrinter::print_inline_tree(InlineTree *tree) {
if (tree == NULL) return;
ciMethod *method = tree->method();
print_method(tree->method(), tree->caller_bci(), tree);
}
void IdealGraphPrinter::print_inlining(Compile* compile) {
// Print inline tree
if (_should_send_method) {
InlineTree *inlineTree = compile->ilt();
if (inlineTree != NULL) {
print_inline_tree(inlineTree);
} else {
// print this method only
}
}
}
// Has to be called whenever a method is compiled
void IdealGraphPrinter::begin_method(Compile* compile) {
ciMethod *method = compile->method();
assert(_output, "output stream must exist!");
assert(method, "null methods are not allowed!");
assert(!_current_method, "current method must be null!");
head(GROUP_ELEMENT);
head(PROPERTIES_ELEMENT);
// Print properties
// Add method name
stringStream strStream;
method->print_name(&strStream);
print_prop(METHOD_NAME_PROPERTY, strStream.as_string());
if (method->flags().is_public()) {
print_prop(METHOD_IS_PUBLIC_PROPERTY, TRUE_VALUE);
}
if (method->flags().is_static()) {
print_prop(METHOD_IS_STATIC_PROPERTY, TRUE_VALUE);
}
tail(PROPERTIES_ELEMENT);
if (_stream) {
char answer = 0;
_xml->flush();
int result = _stream->read(&answer, 1);
_should_send_method = (answer == 'y');
}
this->_current_method = method;
_xml->flush();
}
// Has to be called whenever a method has finished compilation
void IdealGraphPrinter::end_method() {
nmethod* method = (nmethod*)this->_current_method->code();
tail(GROUP_ELEMENT);
_current_method = NULL;
_xml->flush();
}
// Print indent
void IdealGraphPrinter::print_indent() {
tty->print_cr("printing ident %d", _depth);
for (int i = 0; i < _depth; i++) {
_xml->print("%s", INDENT);
}
}
bool IdealGraphPrinter::traverse_outs() {
return _traverse_outs;
}
void IdealGraphPrinter::set_traverse_outs(bool b) {
_traverse_outs = b;
}
intptr_t IdealGraphPrinter::get_node_id(Node *n) {
return (intptr_t)(n);
}
void IdealGraphPrinter::visit_node(Node *n, bool edges, VectorSet* temp_set) {
if (edges) {
// Output edge
intptr_t dest_id = get_node_id(n);
for ( uint i = 0; i < n->len(); i++ ) {
if ( n->in(i) ) {
Node *source = n->in(i);
begin_elem(EDGE_ELEMENT);
intptr_t source_id = get_node_id(source);
print_attr(FROM_PROPERTY, source_id);
print_attr(TO_PROPERTY, dest_id);
print_attr(INDEX_PROPERTY, i);
end_elem();
}
}
} else {
// Output node
begin_head(NODE_ELEMENT);
print_attr(NODE_ID_PROPERTY, get_node_id(n));
end_head();
head(PROPERTIES_ELEMENT);
Node *node = n;
#ifndef PRODUCT
Compile::current()->_in_dump_cnt++;
print_prop(NODE_NAME_PROPERTY, (const char *)node->Name());
const Type *t = node->bottom_type();
print_prop("type", t->msg());
print_prop("idx", node->_idx);
#ifdef ASSERT
print_prop("debug_idx", node->_debug_idx);
#endif
if (C->cfg() != NULL) {
Block* block = C->cfg()->get_block_for_node(node);
if (block == NULL) {
print_prop("block", C->cfg()->get_block(0)->_pre_order);
} else {
print_prop("block", block->_pre_order);
}
}
const jushort flags = node->flags();
if (flags & Node::Flag_is_Copy) {
print_prop("is_copy", "true");
}
if (flags & Node::Flag_rematerialize) {
print_prop("rematerialize", "true");
}
if (flags & Node::Flag_needs_anti_dependence_check) {
print_prop("needs_anti_dependence_check", "true");
}
if (flags & Node::Flag_is_macro) {
print_prop("is_macro", "true");
}
if (flags & Node::Flag_is_Con) {
print_prop("is_con", "true");
}
if (flags & Node::Flag_is_cisc_alternate) {
print_prop("is_cisc_alternate", "true");
}
if (flags & Node::Flag_is_dead_loop_safe) {
print_prop("is_dead_loop_safe", "true");
}
if (flags & Node::Flag_may_be_short_branch) {
print_prop("may_be_short_branch", "true");
}
if (flags & Node::Flag_has_call) {
print_prop("has_call", "true");
}
if (C->matcher() != NULL) {
if (C->matcher()->is_shared(node)) {
print_prop("is_shared", "true");
} else {
print_prop("is_shared", "false");
}
if (C->matcher()->is_dontcare(node)) {
print_prop("is_dontcare", "true");
} else {
print_prop("is_dontcare", "false");
}
#ifdef ASSERT
Node* old = C->matcher()->find_old_node(node);
if (old != NULL) {
print_prop("old_node_idx", old->_idx);
}
#endif
}
if (node->is_Proj()) {
print_prop("con", (int)node->as_Proj()->_con);
}
if (node->is_Mach()) {
print_prop("idealOpcode", (const char *)NodeClassNames[node->as_Mach()->ideal_Opcode()]);
}
buffer[0] = 0;
stringStream s2(buffer, sizeof(buffer) - 1);
node->dump_spec(&s2);
if (t != NULL && (t->isa_instptr() || t->isa_klassptr())) {
const TypeInstPtr *toop = t->isa_instptr();
const TypeKlassPtr *tkls = t->isa_klassptr();
ciKlass* klass = toop ? toop->klass() : (tkls ? tkls->klass() : NULL );
if( klass && klass->is_loaded() && klass->is_interface() ) {
s2.print(" Interface:");
} else if( toop ) {
s2.print(" Oop:");
} else if( tkls ) {
s2.print(" Klass:");
}
t->dump_on(&s2);
} else if( t == Type::MEMORY ) {
s2.print(" Memory:");
MemNode::dump_adr_type(node, node->adr_type(), &s2);
}
assert(s2.size() < sizeof(buffer), "size in range");
print_prop("dump_spec", buffer);
if (node->is_block_proj()) {
print_prop("is_block_proj", "true");
}
if (node->is_block_start()) {
print_prop("is_block_start", "true");
}
const char *short_name = "short_name";
if (strcmp(node->Name(), "Parm") == 0 && node->as_Proj()->_con >= TypeFunc::Parms) {
int index = node->as_Proj()->_con - TypeFunc::Parms;
if (index >= 10) {
print_prop(short_name, "PA");
} else {
sprintf(buffer, "P%d", index);
print_prop(short_name, buffer);
}
} else if (strcmp(node->Name(), "IfTrue") == 0) {
print_prop(short_name, "T");
} else if (strcmp(node->Name(), "IfFalse") == 0) {
print_prop(short_name, "F");
} else if ((node->is_Con() && node->is_Type()) || node->is_Proj()) {
if (t->base() == Type::Int && t->is_int()->is_con()) {
const TypeInt *typeInt = t->is_int();
assert(typeInt->is_con(), "must be constant");
jint value = typeInt->get_con();
// max. 2 chars allowed
if (value >= -9 && value <= 99) {
sprintf(buffer, "%d", value);
print_prop(short_name, buffer);
} else {
print_prop(short_name, "I");
}
} else if (t == Type::TOP) {
print_prop(short_name, "^");
} else if (t->base() == Type::Long && t->is_long()->is_con()) {
const TypeLong *typeLong = t->is_long();
assert(typeLong->is_con(), "must be constant");
jlong value = typeLong->get_con();
// max. 2 chars allowed
if (value >= -9 && value <= 99) {
sprintf(buffer, JLONG_FORMAT, value);
print_prop(short_name, buffer);
} else {
print_prop(short_name, "L");
}
} else if (t->base() == Type::KlassPtr) {
const TypeKlassPtr *typeKlass = t->is_klassptr();
print_prop(short_name, "CP");
} else if (t->base() == Type::Control) {
print_prop(short_name, "C");
} else if (t->base() == Type::Memory) {
print_prop(short_name, "M");
} else if (t->base() == Type::Abio) {
print_prop(short_name, "IO");
} else if (t->base() == Type::Return_Address) {
print_prop(short_name, "RA");
} else if (t->base() == Type::AnyPtr) {
print_prop(short_name, "P");
} else if (t->base() == Type::RawPtr) {
print_prop(short_name, "RP");
} else if (t->base() == Type::AryPtr) {
print_prop(short_name, "AP");
}
}
JVMState* caller = NULL;
if (node->is_SafePoint()) {
caller = node->as_SafePoint()->jvms();
} else {
Node_Notes* notes = C->node_notes_at(node->_idx);
if (notes != NULL) {
caller = notes->jvms();
}
}
if (caller != NULL) {
stringStream bciStream;
ciMethod* last = NULL;
int last_bci;
while(caller) {
if (caller->has_method()) {
last = caller->method();
last_bci = caller->bci();
}
bciStream.print("%d ", caller->bci());
caller = caller->caller();
}
print_prop("bci", bciStream.as_string());
if (last != NULL && last->has_linenumber_table() && last_bci >= 0) {
print_prop("line", last->line_number_from_bci(last_bci));
}
}
#ifdef ASSERT
if (node->debug_orig() != NULL) {
temp_set->Clear();
stringStream dorigStream;
Node* dorig = node->debug_orig();
while (dorig && temp_set->test_set(dorig->_idx)) {
dorigStream.print("%d ", dorig->_idx);
}
print_prop("debug_orig", dorigStream.as_string());
}
#endif
if (_chaitin && _chaitin != (PhaseChaitin *)((intptr_t)0xdeadbeef)) {
buffer[0] = 0;
_chaitin->dump_register(node, buffer);
print_prop("reg", buffer);
uint lrg_id = 0;
if (node->_idx < _chaitin->_lrg_map.size()) {
lrg_id = _chaitin->_lrg_map.live_range_id(node);
}
print_prop("lrg", lrg_id);
}
Compile::current()->_in_dump_cnt--;
#endif
tail(PROPERTIES_ELEMENT);
tail(NODE_ELEMENT);
}
}
void IdealGraphPrinter::walk_nodes(Node *start, bool edges, VectorSet* temp_set) {
VectorSet visited(Thread::current()->resource_area());
GrowableArray<Node *> nodeStack(Thread::current()->resource_area(), 0, 0, NULL);
nodeStack.push(start);
visited.test_set(start->_idx);
if (C->cfg() != NULL) {
// once we have a CFG there are some nodes that aren't really
// reachable but are in the CFG so add them here.
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
Block* block = C->cfg()->get_block(i);
for (uint s = 0; s < block->number_of_nodes(); s++) {
nodeStack.push(block->get_node(s));
}
}
}
while(nodeStack.length() > 0) {
Node *n = nodeStack.pop();
visit_node(n, edges, temp_set);
if (_traverse_outs) {
for (DUIterator i = n->outs(); n->has_out(i); i++) {
Node* p = n->out(i);
if (!visited.test_set(p->_idx)) {
nodeStack.push(p);
}
}
}
for ( uint i = 0; i < n->len(); i++ ) {
if ( n->in(i) ) {
if (!visited.test_set(n->in(i)->_idx)) {
nodeStack.push(n->in(i));
}
}
}
}
}
void IdealGraphPrinter::print_method(Compile* compile, const char *name, int level, bool clear_nodes) {
print(compile, name, (Node *)compile->root(), level, clear_nodes);
}
// Print current ideal graph
void IdealGraphPrinter::print(Compile* compile, const char *name, Node *node, int level, bool clear_nodes) {
if (!_current_method || !_should_send_method || level > PrintIdealGraphLevel) return;
this->C = compile;
// Warning, unsafe cast?
_chaitin = (PhaseChaitin *)C->regalloc();
begin_head(GRAPH_ELEMENT);
print_attr(GRAPH_NAME_PROPERTY, (const char *)name);
end_head();
VectorSet temp_set(Thread::current()->resource_area());
head(NODES_ELEMENT);
walk_nodes(node, false, &temp_set);
tail(NODES_ELEMENT);
head(EDGES_ELEMENT);
walk_nodes(node, true, &temp_set);
tail(EDGES_ELEMENT);
if (C->cfg() != NULL) {
head(CONTROL_FLOW_ELEMENT);
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
Block* block = C->cfg()->get_block(i);
begin_head(BLOCK_ELEMENT);
print_attr(BLOCK_NAME_PROPERTY, block->_pre_order);
end_head();
head(SUCCESSORS_ELEMENT);
for (uint s = 0; s < block->_num_succs; s++) {
begin_elem(SUCCESSOR_ELEMENT);
print_attr(BLOCK_NAME_PROPERTY, block->_succs[s]->_pre_order);
end_elem();
}
tail(SUCCESSORS_ELEMENT);
head(NODES_ELEMENT);
for (uint s = 0; s < block->number_of_nodes(); s++) {
begin_elem(NODE_ELEMENT);
print_attr(NODE_ID_PROPERTY, get_node_id(block->get_node(s)));
end_elem();
}
tail(NODES_ELEMENT);
tail(BLOCK_ELEMENT);
}
tail(CONTROL_FLOW_ELEMENT);
}
tail(GRAPH_ELEMENT);
output()->flush();
}
extern const char *NodeClassNames[];
outputStream *IdealGraphPrinter::output() {
return _xml;
}
#endif
C:\hotspot-69087d08d473\src\share\vm/opto/idealGraphPrinter.hpp
/*
* Copyright (c) 2007, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_IDEALGRAPHPRINTER_HPP
#define SHARE_VM_OPTO_IDEALGRAPHPRINTER_HPP
#include "libadt/dict.hpp"
#include "libadt/vectset.hpp"
#include "utilities/growableArray.hpp"
#include "utilities/ostream.hpp"
#include "utilities/xmlstream.hpp"
#ifndef PRODUCT
class Compile;
class PhaseIFG;
class PhaseChaitin;
class Matcher;
class Node;
class InlineTree;
class ciMethod;
class IdealGraphPrinter : public CHeapObj<mtCompiler> {
private:
static const char *INDENT;
static const char *TOP_ELEMENT;
static const char *GROUP_ELEMENT;
static const char *GRAPH_ELEMENT;
static const char *PROPERTIES_ELEMENT;
static const char *EDGES_ELEMENT;
static const char *PROPERTY_ELEMENT;
static const char *EDGE_ELEMENT;
static const char *NODE_ELEMENT;
static const char *NODES_ELEMENT;
static const char *CONTROL_FLOW_ELEMENT;
static const char *REMOVE_EDGE_ELEMENT;
static const char *REMOVE_NODE_ELEMENT;
static const char *METHOD_NAME_PROPERTY;
static const char *BLOCK_NAME_PROPERTY;
static const char *BLOCK_DOMINATOR_PROPERTY;
static const char *BLOCK_ELEMENT;
static const char *SUCCESSORS_ELEMENT;
static const char *SUCCESSOR_ELEMENT;
static const char *METHOD_IS_PUBLIC_PROPERTY;
static const char *METHOD_IS_STATIC_PROPERTY;
static const char *TRUE_VALUE;
static const char *NODE_NAME_PROPERTY;
static const char *EDGE_NAME_PROPERTY;
static const char *NODE_ID_PROPERTY;
static const char *FROM_PROPERTY;
static const char *TO_PROPERTY;
static const char *PROPERTY_NAME_PROPERTY;
static const char *GRAPH_NAME_PROPERTY;
static const char *INDEX_PROPERTY;
static const char *METHOD_ELEMENT;
static const char *INLINE_ELEMENT;
static const char *BYTECODES_ELEMENT;
static const char *METHOD_BCI_PROPERTY;
static const char *METHOD_SHORT_NAME_PROPERTY;
static const char *ASSEMBLY_ELEMENT;
elapsedTimer _walk_time;
elapsedTimer _output_time;
elapsedTimer _build_blocks_time;
static int _file_count;
networkStream *_stream;
xmlStream *_xml;
outputStream *_output;
ciMethod *_current_method;
int _depth;
char buffer[128];
bool _should_send_method;
PhaseChaitin* _chaitin;
bool _traverse_outs;
Compile *C;
static void pre_node(Node* node, void *env);
static void post_node(Node* node, void *env);
void print_indent();
void print_method(ciMethod *method, int bci, InlineTree *tree);
void print_inline_tree(InlineTree *tree);
void visit_node(Node *n, bool edges, VectorSet* temp_set);
void walk_nodes(Node *start, bool edges, VectorSet* temp_set);
void begin_elem(const char *s);
void end_elem();
void begin_head(const char *s);
void end_head();
void print_attr(const char *name, const char *val);
void print_attr(const char *name, intptr_t val);
void print_prop(const char *name, const char *val);
void print_prop(const char *name, int val);
void tail(const char *name);
void head(const char *name);
void text(const char *s);
intptr_t get_node_id(Node *n);
IdealGraphPrinter();
~IdealGraphPrinter();
public:
static void clean_up();
static IdealGraphPrinter *printer();
bool traverse_outs();
void set_traverse_outs(bool b);
outputStream *output();
void print_inlining(Compile* compile);
void begin_method(Compile* compile);
void end_method();
void print_method(Compile* compile, const char *name, int level=1, bool clear_nodes = false);
void print(Compile* compile, const char *name, Node *root, int level=1, bool clear_nodes = false);
void print_xml(const char *name);
};
#endif
#endif // SHARE_VM_OPTO_IDEALGRAPHPRINTER_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/idealKit.cpp
/*
* Copyright (c) 2005, 2013, 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 "precompiled.hpp"
#include "opto/addnode.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/idealKit.hpp"
#include "opto/runtime.hpp"
// Static initialization
// This declares the position where vars are kept in the cvstate
// For some degree of consistency we use the TypeFunc enum to
// soak up spots in the inputs even though we only use early Control
// and Memory slots. (So far.)
const uint IdealKit::first_var = TypeFunc::Parms + 1;
//----------------------------IdealKit-----------------------------------------
IdealKit::IdealKit(GraphKit* gkit, bool delay_all_transforms, bool has_declarations) :
_gvn(gkit->gvn()), C(gkit->C) {
_initial_ctrl = gkit->control();
_initial_memory = gkit->merged_memory();
_initial_i_o = gkit->i_o();
_delay_all_transforms = delay_all_transforms;
_var_ct = 0;
_cvstate = NULL;
// We can go memory state free or else we need the entire memory state
assert(_initial_memory == NULL || _initial_memory->Opcode() == Op_MergeMem, "memory must be pre-split");
assert(!_gvn.is_IterGVN(), "IdealKit can't be used during Optimize phase");
int init_size = 5;
_pending_cvstates = new (C->node_arena()) GrowableArray<Node*>(C->node_arena(), init_size, 0, 0);
DEBUG_ONLY(_state = new (C->node_arena()) GrowableArray<int>(C->node_arena(), init_size, 0, 0));
if (!has_declarations) {
declarations_done();
}
}
//----------------------------sync_kit-----------------------------------------
void IdealKit::sync_kit(GraphKit* gkit) {
set_all_memory(gkit->merged_memory());
set_i_o(gkit->i_o());
set_ctrl(gkit->control());
}
//-------------------------------if_then-------------------------------------
// Create: if(left relop right)
// / \
// iffalse iftrue
// Push the iffalse cvstate onto the stack. The iftrue becomes the current cvstate.
void IdealKit::if_then(Node* left, BoolTest::mask relop,
Node* right, float prob, float cnt, bool push_new_state) {
assert((state() & (BlockS|LoopS|IfThenS|ElseS)), "bad state for new If");
Node* bol;
if (left->bottom_type()->isa_ptr() == NULL) {
if (left->bottom_type()->isa_int() != NULL) {
bol = Bool(CmpI(left, right), relop);
} else {
assert(left->bottom_type()->isa_long() != NULL, "what else?");
bol = Bool(CmpL(left, right), relop);
}
} else {
bol = Bool(CmpP(left, right), relop);
}
// Delay gvn.tranform on if-nodes until construction is finished
// to prevent a constant bool input from discarding a control output.
IfNode* iff = delay_transform(new (C) IfNode(ctrl(), bol, prob, cnt))->as_If();
Node* then = IfTrue(iff);
Node* elsen = IfFalse(iff);
Node* else_cvstate = copy_cvstate();
else_cvstate->set_req(TypeFunc::Control, elsen);
_pending_cvstates->push(else_cvstate);
DEBUG_ONLY(if (push_new_state) _state->push(IfThenS));
set_ctrl(then);
}
//-------------------------------else_-------------------------------------
// Pop the else cvstate off the stack, and push the (current) then cvstate.
// The else cvstate becomes the current cvstate.
void IdealKit::else_() {
assert(state() == IfThenS, "bad state for new Else");
Node* else_cvstate = _pending_cvstates->pop();
DEBUG_ONLY(_state->pop());
// save current (then) cvstate for later use at endif
_pending_cvstates->push(_cvstate);
DEBUG_ONLY(_state->push(ElseS));
_cvstate = else_cvstate;
}
//-------------------------------end_if-------------------------------------
// Merge the "then" and "else" cvstates.
//
// The if_then() pushed a copy of the current state for later use
// as the initial state for a future "else" clause. The
// current state then became the initial state for the
// then clause. If an "else" clause was encountered, it will
// pop the top state and use it for it's initial state.
// It will also push the current state (the state at the end of
// the "then" clause) for latter use at the end_if.
//
// At the endif, the states are:
// 1) else exists a) current state is end of "else" clause
// b) top stack state is end of "then" clause
//
// 2) no else: a) current state is end of "then" clause
// b) top stack state is from the "if_then" which
// would have been the initial state of the else.
//
// Merging the states is accomplished by:
// 1) make a label for the merge
// 2) terminate the current state with a goto to the label
// 3) pop the top state from the stack and make it the
// current state
// 4) bind the label at the current state. Binding a label
// terminates the current state with a goto to the
// label and makes the label's state the current state.
//
void IdealKit::end_if() {
assert(state() & (IfThenS|ElseS), "bad state for new Endif");
Node* lab = make_label(1);
// Node* join_state = _pending_cvstates->pop();
/* merging, join */
goto_(lab);
_cvstate = _pending_cvstates->pop();
bind(lab);
DEBUG_ONLY(_state->pop());
}
//-------------------------------loop-------------------------------------
// Create the loop head portion (*) of:
// * iv = init
// * top: (region node)
// * if (iv relop limit) {
// loop body
// i = i + 1
// goto top
// * } else // exits loop
//
// Pushes the loop top cvstate first, then the else (loop exit) cvstate
// onto the stack.
void IdealKit::loop(GraphKit* gkit, int nargs, IdealVariable& iv, Node* init, BoolTest::mask relop, Node* limit, float prob, float cnt) {
assert((state() & (BlockS|LoopS|IfThenS|ElseS)), "bad state for new loop");
if (UseLoopPredicate) {
// Sync IdealKit and graphKit.
gkit->sync_kit(*this);
// Add loop predicate.
gkit->add_predicate(nargs);
// Update IdealKit memory.
sync_kit(gkit);
}
set(iv, init);
Node* head = make_label(1);
bind(head);
_pending_cvstates->push(head); // push for use at end_loop
_cvstate = copy_cvstate();
if_then(value(iv), relop, limit, prob, cnt, false /* no new state */);
DEBUG_ONLY(_state->push(LoopS));
assert(ctrl()->is_IfTrue(), "true branch stays in loop");
assert(_pending_cvstates->top()->in(TypeFunc::Control)->is_IfFalse(), "false branch exits loop");
}
//-------------------------------end_loop-------------------------------------
// Creates the goto top label.
// Expects the else (loop exit) cvstate to be on top of the
// stack, and the loop top cvstate to be 2nd.
void IdealKit::end_loop() {
assert((state() == LoopS), "bad state for new end_loop");
Node* exit = _pending_cvstates->pop();
Node* head = _pending_cvstates->pop();
goto_(head);
clear(head);
DEBUG_ONLY(_state->pop());
_cvstate = exit;
}
//-------------------------------make_label-------------------------------------
// Creates a label. The number of goto's
// must be specified (which should be 1 less than
// the number of precedessors.)
Node* IdealKit::make_label(int goto_ct) {
assert(_cvstate != NULL, "must declare variables before labels");
Node* lab = new_cvstate();
int sz = 1 + goto_ct + 1 /* fall thru */;
Node* reg = delay_transform(new (C) RegionNode(sz));
lab->init_req(TypeFunc::Control, reg);
return lab;
}
//-------------------------------bind-------------------------------------
// Bind a label at the current cvstate by simulating
// a goto to the label.
void IdealKit::bind(Node* lab) {
goto_(lab, true /* bind */);
_cvstate = lab;
}
//-------------------------------goto_-------------------------------------
// Make the current cvstate a predecessor of the label,
// creating phi's to merge values. If bind is true and
// this is not the last control edge, then ensure that
// all live values have phis created. Used to create phis
// at loop-top regions.
void IdealKit::goto_(Node* lab, bool bind) {
Node* reg = lab->in(TypeFunc::Control);
// find next empty slot in region
uint slot = 1;
while (slot < reg->req() && reg->in(slot) != NULL) slot++;
assert(slot < reg->req(), "too many gotos");
// If this is last predecessor, then don't force phi creation
if (slot == reg->req() - 1) bind = false;
reg->init_req(slot, ctrl());
assert(first_var + _var_ct == _cvstate->req(), "bad _cvstate size");
for (uint i = first_var; i < _cvstate->req(); i++) {
// l is the value of var reaching the label. Could be a single value
// reaching the label, or a phi that merges multiples values reaching
// the label. The latter is true if the label's input: in(..) is
// a phi whose control input is the region node for the label.
Node* l = lab->in(i);
// Get the current value of the var
Node* m = _cvstate->in(i);
// If the var went unused no need for a phi
if (m == NULL) {
continue;
} else if (l == NULL || m == l) {
// Only one unique value "m" is known to reach this label so a phi
// is not yet necessary unless:
// the label is being bound and all predecessors have not been seen,
// in which case "bind" will be true.
if (bind) {
m = promote_to_phi(m, reg);
}
// Record the phi/value used for this var in the label's cvstate
lab->set_req(i, m);
} else {
// More than one value for the variable reaches this label so
// a create a phi if one does not already exist.
if (!was_promoted_to_phi(l, reg)) {
l = promote_to_phi(l, reg);
lab->set_req(i, l);
}
// Record in the phi, the var's value from the current state
l->set_req(slot, m);
}
}
do_memory_merge(_cvstate, lab);
stop();
}
//-----------------------------promote_to_phi-----------------------------------
Node* IdealKit::promote_to_phi(Node* n, Node* reg) {
assert(!was_promoted_to_phi(n, reg), "n already promoted to phi on this region");
// Get a conservative type for the phi
const BasicType bt = n->bottom_type()->basic_type();
const Type* ct = Type::get_const_basic_type(bt);
return delay_transform(PhiNode::make(reg, n, ct));
}
//-----------------------------declarations_done-------------------------------
void IdealKit::declarations_done() {
_cvstate = new_cvstate(); // initialize current cvstate
set_ctrl(_initial_ctrl); // initialize control in current cvstate
set_all_memory(_initial_memory);// initialize memory in current cvstate
set_i_o(_initial_i_o); // initialize i_o in current cvstate
DEBUG_ONLY(_state->push(BlockS));
}
//-----------------------------transform-----------------------------------
Node* IdealKit::transform(Node* n) {
if (_delay_all_transforms) {
return delay_transform(n);
} else {
n = gvn().transform(n);
C->record_for_igvn(n);
return n;
}
}
//-----------------------------delay_transform-----------------------------------
Node* IdealKit::delay_transform(Node* n) {
// Delay transform until IterativeGVN
gvn().set_type(n, n->bottom_type());
C->record_for_igvn(n);
return n;
}
//-----------------------------new_cvstate-----------------------------------
Node* IdealKit::new_cvstate() {
uint sz = _var_ct + first_var;
return new (C) Node(sz);
}
//-----------------------------copy_cvstate-----------------------------------
Node* IdealKit::copy_cvstate() {
Node* ns = new_cvstate();
for (uint i = 0; i < ns->req(); i++) ns->init_req(i, _cvstate->in(i));
// We must clone memory since it will be updated as we do stores.
ns->set_req(TypeFunc::Memory, MergeMemNode::make(C, ns->in(TypeFunc::Memory)));
return ns;
}
//-----------------------------clear-----------------------------------
void IdealKit::clear(Node* m) {
for (uint i = 0; i < m->req(); i++) m->set_req(i, NULL);
}
//-----------------------------IdealVariable----------------------------
IdealVariable::IdealVariable(IdealKit &k) {
k.declare(this);
}
Node* IdealKit::memory(uint alias_idx) {
MergeMemNode* mem = merged_memory();
Node* p = mem->memory_at(alias_idx);
_gvn.set_type(p, Type::MEMORY); // must be mapped
return p;
}
void IdealKit::set_memory(Node* mem, uint alias_idx) {
merged_memory()->set_memory_at(alias_idx, mem);
}
//----------------------------- make_load ----------------------------
Node* IdealKit::load(Node* ctl,
Node* adr,
const Type* t,
BasicType bt,
int adr_idx,
bool require_atomic_access) {
assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
const TypePtr* adr_type = NULL; // debug-mode-only argument
debug_only(adr_type = C->get_adr_type(adr_idx));
Node* mem = memory(adr_idx);
Node* ld;
if (require_atomic_access && bt == T_LONG) {
ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, MemNode::unordered);
} else {
ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, MemNode::unordered);
}
return transform(ld);
}
Node* IdealKit::store(Node* ctl, Node* adr, Node *val, BasicType bt,
int adr_idx,
MemNode::MemOrd mo, bool require_atomic_access,
bool mismatched) {
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory");
const TypePtr* adr_type = NULL;
debug_only(adr_type = C->get_adr_type(adr_idx));
Node *mem = memory(adr_idx);
Node* st;
if (require_atomic_access && bt == T_LONG) {
st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
} else {
st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
}
if (mismatched) {
st->as_Store()->set_mismatched_access();
}
st = transform(st);
set_memory(st, adr_idx);
return st;
}
// Card mark store. Must be ordered so that it will come after the store of
// the oop.
Node* IdealKit::storeCM(Node* ctl, Node* adr, Node *val, Node* oop_store, int oop_adr_idx,
BasicType bt,
int adr_idx) {
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
const TypePtr* adr_type = NULL;
debug_only(adr_type = C->get_adr_type(adr_idx));
Node *mem = memory(adr_idx);
// Add required edge to oop_store, optimizer does not support precedence edges.
// Convert required edge to precedence edge before allocation.
Node* st = new (C) StoreCMNode(ctl, mem, adr, adr_type, val, oop_store, oop_adr_idx);
st = transform(st);
set_memory(st, adr_idx);
return st;
}
//---------------------------- do_memory_merge --------------------------------
// The memory from one merging cvstate needs to be merged with the memory for another
// join cvstate. If the join cvstate doesn't have a merged memory yet then we
// can just copy the state from the merging cvstate
// Merge one slow path into the rest of memory.
void IdealKit::do_memory_merge(Node* merging, Node* join) {
// Get the region for the join state
Node* join_region = join->in(TypeFunc::Control);
assert(join_region != NULL, "join region must exist");
if (join->in(TypeFunc::I_O) == NULL ) {
join->set_req(TypeFunc::I_O, merging->in(TypeFunc::I_O));
}
if (join->in(TypeFunc::Memory) == NULL ) {
join->set_req(TypeFunc::Memory, merging->in(TypeFunc::Memory));
return;
}
// The control flow for merging must have already been attached to the join region
// we need its index for the phis.
uint slot;
for (slot = 1; slot < join_region->req() ; slot ++ ) {
if (join_region->in(slot) == merging->in(TypeFunc::Control)) break;
}
assert(slot != join_region->req(), "edge must already exist");
MergeMemNode* join_m = join->in(TypeFunc::Memory)->as_MergeMem();
MergeMemNode* merging_m = merging->in(TypeFunc::Memory)->as_MergeMem();
// join_m should be an ancestor mergemem of merging
// Slow path memory comes from the current map (which is from a slow call)
// Fast path/null path memory comes from the call's input
// Merge the other fast-memory inputs with the new slow-default memory.
// for (MergeMemStream mms(merged_memory(), fast_mem->as_MergeMem()); mms.next_non_empty2(); ) {
for (MergeMemStream mms(join_m, merging_m); mms.next_non_empty2(); ) {
Node* join_slice = mms.force_memory();
Node* merging_slice = mms.memory2();
if (join_slice != merging_slice) {
PhiNode* phi;
// bool new_phi = false;
// Is the phi for this slice one that we created for this join region or simply
// one we copied? If it is ours then add
if (join_slice->is_Phi() && join_slice->as_Phi()->region() == join_region) {
phi = join_slice->as_Phi();
} else {
// create the phi with join_slice filling supplying memory for all of the
// control edges to the join region
phi = PhiNode::make(join_region, join_slice, Type::MEMORY, mms.adr_type(C));
phi = (PhiNode*) delay_transform(phi);
// gvn().set_type(phi, Type::MEMORY);
// new_phi = true;
}
// Now update the phi with the slice for the merging slice
phi->set_req(slot, merging_slice/* slow_path, slow_slice */);
// this updates join_m with the phi
mms.set_memory(phi);
}
}
Node* join_io = join->in(TypeFunc::I_O);
Node* merging_io = merging->in(TypeFunc::I_O);
if (join_io != merging_io) {
PhiNode* phi;
if (join_io->is_Phi() && join_io->as_Phi()->region() == join_region) {
phi = join_io->as_Phi();
} else {
phi = PhiNode::make(join_region, join_io, Type::ABIO);
phi = (PhiNode*) delay_transform(phi);
join->set_req(TypeFunc::I_O, phi);
}
phi->set_req(slot, merging_io);
}
}
//----------------------------- make_call ----------------------------
// Trivial runtime call
void IdealKit::make_leaf_call(const TypeFunc *slow_call_type,
address slow_call,
const char *leaf_name,
Node* parm0,
Node* parm1,
Node* parm2,
Node* parm3) {
// We only handle taking in RawMem and modifying RawMem
const TypePtr* adr_type = TypeRawPtr::BOTTOM;
uint adr_idx = C->get_alias_index(adr_type);
// Slow-path leaf call
CallNode *call = (CallNode*)new (C) CallLeafNode( slow_call_type, slow_call, leaf_name, adr_type);
// Set fixed predefined input arguments
call->init_req( TypeFunc::Control, ctrl() );
call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
// Narrow memory as only memory input
call->init_req( TypeFunc::Memory , memory(adr_idx));
call->init_req( TypeFunc::FramePtr, top() /* frameptr() */ );
call->init_req( TypeFunc::ReturnAdr, top() );
if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
if (parm3 != NULL) call->init_req(TypeFunc::Parms+3, parm3);
// Node *c = _gvn.transform(call);
call = (CallNode *) _gvn.transform(call);
Node *c = call; // dbx gets confused with call call->dump()
// Slow leaf call has no side-effects, sets few values
set_ctrl(transform( new (C) ProjNode(call,TypeFunc::Control) ));
// Make memory for the call
Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
// Set the RawPtr memory state only.
set_memory(mem, adr_idx);
assert(C->alias_type(call->adr_type()) == C->alias_type(adr_type),
"call node must be constructed correctly");
}
void IdealKit::make_leaf_call_no_fp(const TypeFunc *slow_call_type,
address slow_call,
const char *leaf_name,
const TypePtr* adr_type,
Node* parm0,
Node* parm1,
Node* parm2,
Node* parm3) {
// We only handle taking in RawMem and modifying RawMem
uint adr_idx = C->get_alias_index(adr_type);
// Slow-path leaf call
CallNode *call = (CallNode*)new (C) CallLeafNoFPNode( slow_call_type, slow_call, leaf_name, adr_type);
// Set fixed predefined input arguments
call->init_req( TypeFunc::Control, ctrl() );
call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
// Narrow memory as only memory input
call->init_req( TypeFunc::Memory , memory(adr_idx));
call->init_req( TypeFunc::FramePtr, top() /* frameptr() */ );
call->init_req( TypeFunc::ReturnAdr, top() );
if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
if (parm3 != NULL) call->init_req(TypeFunc::Parms+3, parm3);
// Node *c = _gvn.transform(call);
call = (CallNode *) _gvn.transform(call);
Node *c = call; // dbx gets confused with call call->dump()
// Slow leaf call has no side-effects, sets few values
set_ctrl(transform( new (C) ProjNode(call,TypeFunc::Control) ));
// Make memory for the call
Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
// Set the RawPtr memory state only.
set_memory(mem, adr_idx);
assert(C->alias_type(call->adr_type()) == C->alias_type(adr_type),
"call node must be constructed correctly");
}
C:\hotspot-69087d08d473\src\share\vm/opto/idealKit.hpp
/*
* Copyright (c) 2005, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_IDEALKIT_HPP
#define SHARE_VM_OPTO_IDEALKIT_HPP
#include "opto/addnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/connode.hpp"
#include "opto/divnode.hpp"
#include "opto/graphKit.hpp"
#include "opto/mulnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
#include "opto/type.hpp"
//-----------------------------------------------------------------------------
//----------------------------IdealKit-----------------------------------------
// Set of utilities for creating control flow and scalar SSA data flow.
// Control:
// if_then(left, relop, right)
// else_ (optional)
// end_if
// loop(iv variable, initial, relop, limit)
// - sets iv to initial for first trip
// - exits when relation on limit is true
// - the values of initial and limit should be loop invariant
// - no increment, must be explicitly coded
// - final value of iv is available after end_loop (until dead())
// end_loop
// make_label(number of gotos)
// goto_(label)
// bind(label)
// Data:
// ConI(integer constant) - create an integer constant
// set(variable, value) - assignment
// value(variable) - reference value
// dead(variable) - variable's value is no longer live
// increment(variable, value) - increment variable by value
// simple operations: AddI, SubI, AndI, LShiftI, etc.
// Example:
// Node* limit = ??
// IdealVariable i(kit), j(kit);
// declarations_done();
// Node* exit = make_label(1); // 1 goto
// set(j, ConI(0));
// loop(i, ConI(0), BoolTest::lt, limit); {
// if_then(value(i), BoolTest::gt, ConI(5)) {
// set(j, ConI(1));
// goto_(exit); dead(i);
// } end_if();
// increment(i, ConI(1));
// } end_loop(); dead(i);
// bind(exit);
//
// See string_indexOf for a more complete example.
class IdealKit;
// Variable definition for IdealKit
class IdealVariable: public StackObj {
friend class IdealKit;
private:
int _id;
void set_id(int id) { _id = id; }
public:
IdealVariable(IdealKit &k);
int id() { assert(has_id(),"uninitialized id"); return _id; }
bool has_id() { return _id >= 0; }
};
class IdealKit: public StackObj {
friend class IdealVariable;
// The main state (called a cvstate for Control and Variables)
// contains both the current values of the variables and the
// current set of predecessor control edges. The variable values
// are managed via a Node [in(1)..in(_var_ct)], and the predecessor
// control edges managed via a RegionNode. The in(0) of the Node
// for variables points to the RegionNode for the control edges.
protected:
Compile * const C;
PhaseGVN &_gvn;
GrowableArray<Node*>* _pending_cvstates; // stack of cvstates
Node* _cvstate; // current cvstate (control, memory and variables)
uint _var_ct; // number of variables
bool _delay_all_transforms; // flag forcing all transforms to be delayed
Node* _initial_ctrl; // saves initial control until variables declared
Node* _initial_memory; // saves initial memory until variables declared
Node* _initial_i_o; // saves initial i_o until variables declared
PhaseGVN& gvn() const { return _gvn; }
// Create a new cvstate filled with nulls
Node* new_cvstate(); // Create a new cvstate
Node* cvstate() { return _cvstate; } // current cvstate
Node* copy_cvstate(); // copy current cvstate
void set_memory(Node* mem, uint alias_idx );
void do_memory_merge(Node* merging, Node* join);
void clear(Node* m); // clear a cvstate
void stop() { clear(_cvstate); } // clear current cvstate
Node* delay_transform(Node* n);
Node* transform(Node* n); // gvn.transform or skip it
Node* promote_to_phi(Node* n, Node* reg);// Promote "n" to a phi on region "reg"
bool was_promoted_to_phi(Node* n, Node* reg) {
return (n->is_Phi() && n->in(0) == reg);
}
void declare(IdealVariable* v) { v->set_id(_var_ct++); }
// This declares the position where vars are kept in the cvstate
// For some degree of consistency we use the TypeFunc enum to
// soak up spots in the inputs even though we only use early Control
// and Memory slots. (So far.)
static const uint first_var; // = TypeFunc::Parms + 1;
#ifdef ASSERT
enum State { NullS=0, BlockS=1, LoopS=2, IfThenS=4, ElseS=8, EndifS= 16 };
GrowableArray<int>* _state;
State state() { return (State)(_state->top()); }
#endif
// Users should not care about slices only MergedMem so no access for them.
Node* memory(uint alias_idx);
public:
IdealKit(GraphKit* gkit, bool delay_all_transforms = false, bool has_declarations = false);
~IdealKit() {
stop();
}
void sync_kit(GraphKit* gkit);
// Control
Node* ctrl() { return _cvstate->in(TypeFunc::Control); }
void set_ctrl(Node* ctrl) { _cvstate->set_req(TypeFunc::Control, ctrl); }
Node* top() { return C->top(); }
MergeMemNode* merged_memory() { return _cvstate->in(TypeFunc::Memory)->as_MergeMem(); }
void set_all_memory(Node* mem) { _cvstate->set_req(TypeFunc::Memory, mem); }
Node* i_o() { return _cvstate->in(TypeFunc::I_O); }
void set_i_o(Node* c) { _cvstate->set_req(TypeFunc::I_O, c); }
void set(IdealVariable& v, Node* rhs) { _cvstate->set_req(first_var + v.id(), rhs); }
Node* value(IdealVariable& v) { return _cvstate->in(first_var + v.id()); }
void dead(IdealVariable& v) { set(v, (Node*)NULL); }
void if_then(Node* left, BoolTest::mask relop, Node* right,
float prob = PROB_FAIR, float cnt = COUNT_UNKNOWN,
bool push_new_state = true);
void else_();
void end_if();
void loop(GraphKit* gkit, int nargs, IdealVariable& iv, Node* init, BoolTest::mask cmp, Node* limit,
float prob = PROB_LIKELY(0.9), float cnt = COUNT_UNKNOWN);
void end_loop();
Node* make_label(int goto_ct);
void bind(Node* lab);
void goto_(Node* lab, bool bind = false);
void declarations_done();
Node* IfTrue(IfNode* iff) { return transform(new (C) IfTrueNode(iff)); }
Node* IfFalse(IfNode* iff) { return transform(new (C) IfFalseNode(iff)); }
// Data
Node* ConI(jint k) { return (Node*)gvn().intcon(k); }
Node* makecon(const Type *t) const { return _gvn.makecon(t); }
Node* AddI(Node* l, Node* r) { return transform(new (C) AddINode(l, r)); }
Node* SubI(Node* l, Node* r) { return transform(new (C) SubINode(l, r)); }
Node* AndI(Node* l, Node* r) { return transform(new (C) AndINode(l, r)); }
Node* MaxI(Node* l, Node* r) { return transform(new (C) MaxINode(l, r)); }
Node* LShiftI(Node* l, Node* r) { return transform(new (C) LShiftINode(l, r)); }
Node* CmpI(Node* l, Node* r) { return transform(new (C) CmpINode(l, r)); }
Node* Bool(Node* cmp, BoolTest::mask relop) { return transform(new (C) BoolNode(cmp, relop)); }
void increment(IdealVariable& v, Node* j) { set(v, AddI(value(v), j)); }
void decrement(IdealVariable& v, Node* j) { set(v, SubI(value(v), j)); }
Node* CmpL(Node* l, Node* r) { return transform(new (C) CmpLNode(l, r)); }
// TLS
Node* thread() { return gvn().transform(new (C) ThreadLocalNode()); }
// Pointers
// Raw address should be transformed regardless 'delay_transform' flag
// to produce canonical form CastX2P(offset).
Node* AddP(Node *base, Node *ptr, Node *off) { return _gvn.transform(new (C) AddPNode(base, ptr, off)); }
Node* CmpP(Node* l, Node* r) { return transform(new (C) CmpPNode(l, r)); }
#ifdef _LP64
Node* XorX(Node* l, Node* r) { return transform(new (C) XorLNode(l, r)); }
#else // _LP64
Node* XorX(Node* l, Node* r) { return transform(new (C) XorINode(l, r)); }
#endif // _LP64
Node* URShiftX(Node* l, Node* r) { return transform(new (C) URShiftXNode(l, r)); }
Node* ConX(jint k) { return (Node*)gvn().MakeConX(k); }
Node* CastPX(Node* ctl, Node* p) { return transform(new (C) CastP2XNode(ctl, p)); }
// Memory operations
// This is the base version which is given an alias index.
Node* load(Node* ctl,
Node* adr,
const Type* t,
BasicType bt,
int adr_idx,
bool require_atomic_access = false);
// Return the new StoreXNode
Node* store(Node* ctl,
Node* adr,
Node* val,
BasicType bt,
int adr_idx,
MemNode::MemOrd mo,
bool require_atomic_access = false,
bool mismatched = false
);
// Store a card mark ordered after store_oop
Node* storeCM(Node* ctl,
Node* adr,
Node* val,
Node* oop_store,
int oop_adr_idx,
BasicType bt,
int adr_idx);
// Trivial call
void make_leaf_call(const TypeFunc *slow_call_type,
address slow_call,
const char *leaf_name,
Node* parm0,
Node* parm1 = NULL,
Node* parm2 = NULL,
Node* parm3 = NULL);
void make_leaf_call_no_fp(const TypeFunc *slow_call_type,
address slow_call,
const char *leaf_name,
const TypePtr* adr_type,
Node* parm0,
Node* parm1,
Node* parm2,
Node* parm3);
};
#endif // SHARE_VM_OPTO_IDEALKIT_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/ifg.cpp
/*
* Copyright (c) 1998, 2014, 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 "precompiled.hpp"
#include "compiler/oopMap.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/block.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/chaitin.hpp"
#include "opto/coalesce.hpp"
#include "opto/connode.hpp"
#include "opto/indexSet.hpp"
#include "opto/machnode.hpp"
#include "opto/memnode.hpp"
#include "opto/opcodes.hpp"
PhaseIFG::PhaseIFG( Arena *arena ) : Phase(Interference_Graph), _arena(arena) {
}
void PhaseIFG::init( uint maxlrg ) {
_maxlrg = maxlrg;
_yanked = new (_arena) VectorSet(_arena);
_is_square = false;
// Make uninitialized adjacency lists
_adjs = (IndexSet*)_arena->Amalloc(sizeof(IndexSet)*maxlrg);
// Also make empty live range structures
_lrgs = (LRG *)_arena->Amalloc( maxlrg * sizeof(LRG) );
memset(_lrgs,0,sizeof(LRG)*maxlrg);
// Init all to empty
for( uint i = 0; i < maxlrg; i++ ) {
_adjs[i].initialize(maxlrg);
_lrgs[i].Set_All();
}
}
// Add edge between vertices a & b. These are sorted (triangular matrix),
// then the smaller number is inserted in the larger numbered array.
int PhaseIFG::add_edge( uint a, uint b ) {
lrgs(a).invalid_degree();
lrgs(b).invalid_degree();
// Sort a and b, so that a is bigger
assert( !_is_square, "only on triangular" );
if( a < b ) { uint tmp = a; a = b; b = tmp; }
return _adjs[a].insert( b );
}
// Add an edge between 'a' and everything in the vector.
void PhaseIFG::add_vector( uint a, IndexSet *vec ) {
// IFG is triangular, so do the inserts where 'a' < 'b'.
assert( !_is_square, "only on triangular" );
IndexSet *adjs_a = &_adjs[a];
if( !vec->count() ) return;
IndexSetIterator elements(vec);
uint neighbor;
while ((neighbor = elements.next()) != 0) {
add_edge( a, neighbor );
}
}
// Is there an edge between a and b?
int PhaseIFG::test_edge( uint a, uint b ) const {
// Sort a and b, so that a is larger
assert( !_is_square, "only on triangular" );
if( a < b ) { uint tmp = a; a = b; b = tmp; }
return _adjs[a].member(b);
}
// Convert triangular matrix to square matrix
void PhaseIFG::SquareUp() {
assert( !_is_square, "only on triangular" );
// Simple transpose
for( uint i = 0; i < _maxlrg; i++ ) {
IndexSetIterator elements(&_adjs[i]);
uint datum;
while ((datum = elements.next()) != 0) {
_adjs[datum].insert( i );
}
}
_is_square = true;
}
// Compute effective degree in bulk
void PhaseIFG::Compute_Effective_Degree() {
assert( _is_square, "only on square" );
for( uint i = 0; i < _maxlrg; i++ )
lrgs(i).set_degree(effective_degree(i));
}
int PhaseIFG::test_edge_sq( uint a, uint b ) const {
assert( _is_square, "only on square" );
// Swap, so that 'a' has the lesser count. Then binary search is on
// the smaller of a's list and b's list.
if( neighbor_cnt(a) > neighbor_cnt(b) ) { uint tmp = a; a = b; b = tmp; }
//return _adjs[a].unordered_member(b);
return _adjs[a].member(b);
}
// Union edges of B into A
void PhaseIFG::Union( uint a, uint b ) {
assert( _is_square, "only on square" );
IndexSet *A = &_adjs[a];
IndexSetIterator b_elements(&_adjs[b]);
uint datum;
while ((datum = b_elements.next()) != 0) {
if(A->insert(datum)) {
_adjs[datum].insert(a);
lrgs(a).invalid_degree();
lrgs(datum).invalid_degree();
}
}
}
// Yank a Node and all connected edges from the IFG. Return a
// list of neighbors (edges) yanked.
IndexSet *PhaseIFG::remove_node( uint a ) {
assert( _is_square, "only on square" );
assert( !_yanked->test(a), "" );
_yanked->set(a);
// I remove the LRG from all neighbors.
IndexSetIterator elements(&_adjs[a]);
LRG &lrg_a = lrgs(a);
uint datum;
while ((datum = elements.next()) != 0) {
_adjs[datum].remove(a);
lrgs(datum).inc_degree( -lrg_a.compute_degree(lrgs(datum)) );
}
return neighbors(a);
}
// Re-insert a yanked Node.
void PhaseIFG::re_insert( uint a ) {
assert( _is_square, "only on square" );
assert( _yanked->test(a), "" );
(*_yanked) >>= a;
IndexSetIterator elements(&_adjs[a]);
uint datum;
while ((datum = elements.next()) != 0) {
_adjs[datum].insert(a);
lrgs(datum).invalid_degree();
}
}
// Compute the degree between 2 live ranges. If both live ranges are
// aligned-adjacent powers-of-2 then we use the MAX size. If either is
// mis-aligned (or for Fat-Projections, not-adjacent) then we have to
// MULTIPLY the sizes. Inspect Brigg's thesis on register pairs to see why
// this is so.
int LRG::compute_degree( LRG &l ) const {
int tmp;
int num_regs = _num_regs;
int nregs = l.num_regs();
tmp = (_fat_proj || l._fat_proj) // either is a fat-proj?
? (num_regs * nregs) // then use product
: MAX2(num_regs,nregs); // else use max
return tmp;
}
// Compute effective degree for this live range. If both live ranges are
// aligned-adjacent powers-of-2 then we use the MAX size. If either is
// mis-aligned (or for Fat-Projections, not-adjacent) then we have to
// MULTIPLY the sizes. Inspect Brigg's thesis on register pairs to see why
// this is so.
int PhaseIFG::effective_degree( uint lidx ) const {
int eff = 0;
int num_regs = lrgs(lidx).num_regs();
int fat_proj = lrgs(lidx)._fat_proj;
IndexSet *s = neighbors(lidx);
IndexSetIterator elements(s);
uint nidx;
while((nidx = elements.next()) != 0) {
LRG &lrgn = lrgs(nidx);
int nregs = lrgn.num_regs();
eff += (fat_proj || lrgn._fat_proj) // either is a fat-proj?
? (num_regs * nregs) // then use product
: MAX2(num_regs,nregs); // else use max
}
return eff;
}
#ifndef PRODUCT
void PhaseIFG::dump() const {
tty->print_cr("-- Interference Graph --%s--",
_is_square ? "square" : "triangular" );
if( _is_square ) {
for( uint i = 0; i < _maxlrg; i++ ) {
tty->print( (*_yanked)[i] ? "XX " : " ");
tty->print("L%d: { ",i);
IndexSetIterator elements(&_adjs[i]);
uint datum;
while ((datum = elements.next()) != 0) {
tty->print("L%d ", datum);
}
tty->print_cr("}");
}
return;
}
// Triangular
for( uint i = 0; i < _maxlrg; i++ ) {
uint j;
tty->print( (*_yanked)[i] ? "XX " : " ");
tty->print("L%d: { ",i);
for( j = _maxlrg; j > i; j-- )
if( test_edge(j - 1,i) ) {
tty->print("L%d ",j - 1);
}
tty->print("| ");
IndexSetIterator elements(&_adjs[i]);
uint datum;
while ((datum = elements.next()) != 0) {
tty->print("L%d ", datum);
}
tty->print("}\n");
}
tty->print("\n");
}
void PhaseIFG::stats() const {
ResourceMark rm;
int *h_cnt = NEW_RESOURCE_ARRAY(int,_maxlrg*2);
memset( h_cnt, 0, sizeof(int)*_maxlrg*2 );
uint i;
for( i = 0; i < _maxlrg; i++ ) {
h_cnt[neighbor_cnt(i)]++;
}
tty->print_cr("--Histogram of counts--");
for( i = 0; i < _maxlrg*2; i++ )
if( h_cnt[i] )
tty->print("%d/%d ",i,h_cnt[i]);
tty->cr();
}
void PhaseIFG::verify( const PhaseChaitin *pc ) const {
// IFG is square, sorted and no need for Find
for( uint i = 0; i < _maxlrg; i++ ) {
assert(!((*_yanked)[i]) || !neighbor_cnt(i), "Is removed completely" );
IndexSet *set = &_adjs[i];
IndexSetIterator elements(set);
uint idx;
uint last = 0;
while ((idx = elements.next()) != 0) {
assert(idx != i, "Must have empty diagonal");
assert(pc->_lrg_map.find_const(idx) == idx, "Must not need Find");
assert(_adjs[idx].member(i), "IFG not square");
assert(!(*_yanked)[idx], "No yanked neighbors");
assert(last < idx, "not sorted increasing");
last = idx;
}
assert(!lrgs(i)._degree_valid || effective_degree(i) == lrgs(i).degree(), "degree is valid but wrong");
}
}
#endif
// Interfere this register with everything currently live. Use the RegMasks
// to trim the set of possible interferences. Return a count of register-only
// interferences as an estimate of register pressure.
void PhaseChaitin::interfere_with_live( uint r, IndexSet *liveout ) {
uint retval = 0;
// Interfere with everything live.
const RegMask &rm = lrgs(r).mask();
// Check for interference by checking overlap of regmasks.
// Only interfere if acceptable register masks overlap.
IndexSetIterator elements(liveout);
uint l;
while( (l = elements.next()) != 0 )
if( rm.overlap( lrgs(l).mask() ) )
_ifg->add_edge( r, l );
}
// Actually build the interference graph. Uses virtual registers only, no
// physical register masks. This allows me to be very aggressive when
// coalescing copies. Some of this aggressiveness will have to be undone
// later, but I'd rather get all the copies I can now (since unremoved copies
// at this point can end up in bad places). Copies I re-insert later I have
// more opportunity to insert them in low-frequency locations.
void PhaseChaitin::build_ifg_virtual( ) {
// For all blocks (in any order) do...
for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
Block* block = _cfg.get_block(i);
IndexSet* liveout = _live->live(block);
// The IFG is built by a single reverse pass over each basic block.
// Starting with the known live-out set, we remove things that get
// defined and add things that become live (essentially executing one
// pass of a standard LIVE analysis). Just before a Node defines a value
// (and removes it from the live-ness set) that value is certainly live.
// The defined value interferes with everything currently live. The
// value is then removed from the live-ness set and it's inputs are
// added to the live-ness set.
for (uint j = block->end_idx() + 1; j > 1; j--) {
Node* n = block->get_node(j - 1);
// Get value being defined
uint r = _lrg_map.live_range_id(n);
// Some special values do not allocate
if (r) {
// Remove from live-out set
liveout->remove(r);
// Copies do not define a new value and so do not interfere.
// Remove the copies source from the liveout set before interfering.
uint idx = n->is_Copy();
if (idx) {
liveout->remove(_lrg_map.live_range_id(n->in(idx)));
}
// Interfere with everything live
interfere_with_live(r, liveout);
}
// Make all inputs live
if (!n->is_Phi()) { // Phi function uses come from prior block
for(uint k = 1; k < n->req(); k++) {
liveout->insert(_lrg_map.live_range_id(n->in(k)));
}
}
// 2-address instructions always have the defined value live
// on entry to the instruction, even though it is being defined
// by the instruction. We pretend a virtual copy sits just prior
// to the instruction and kills the src-def'd register.
// In other words, for 2-address instructions the defined value
// interferes with all inputs.
uint idx;
if( n->is_Mach() && (idx = n->as_Mach()->two_adr()) ) {
const MachNode *mach = n->as_Mach();
// Sometimes my 2-address ADDs are commuted in a bad way.
// We generally want the USE-DEF register to refer to the
// loop-varying quantity, to avoid a copy.
uint op = mach->ideal_Opcode();
// Check that mach->num_opnds() == 3 to ensure instruction is
// not subsuming constants, effectively excludes addI_cin_imm
// Can NOT swap for instructions like addI_cin_imm since it
// is adding zero to yhi + carry and the second ideal-input
// points to the result of adding low-halves.
// Checking req() and num_opnds() does NOT distinguish addI_cout from addI_cout_imm
if( (op == Op_AddI && mach->req() == 3 && mach->num_opnds() == 3) &&
n->in(1)->bottom_type()->base() == Type::Int &&
// See if the ADD is involved in a tight data loop the wrong way
n->in(2)->is_Phi() &&
n->in(2)->in(2) == n ) {
Node *tmp = n->in(1);
n->set_req( 1, n->in(2) );
n->set_req( 2, tmp );
}
// Defined value interferes with all inputs
uint lidx = _lrg_map.live_range_id(n->in(idx));
for (uint k = 1; k < n->req(); k++) {
uint kidx = _lrg_map.live_range_id(n->in(k));
if (kidx != lidx) {
_ifg->add_edge(r, kidx);
}
}
}
} // End of forall instructions in block
} // End of forall blocks
}
uint PhaseChaitin::count_int_pressure( IndexSet *liveout ) {
IndexSetIterator elements(liveout);
uint lidx;
uint cnt = 0;
while ((lidx = elements.next()) != 0) {
if( lrgs(lidx).mask().is_UP() &&
lrgs(lidx).mask_size() &&
!lrgs(lidx)._is_float &&
!lrgs(lidx)._is_vector &&
lrgs(lidx).mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) )
cnt += lrgs(lidx).reg_pressure();
}
return cnt;
}
uint PhaseChaitin::count_float_pressure( IndexSet *liveout ) {
IndexSetIterator elements(liveout);
uint lidx;
uint cnt = 0;
while ((lidx = elements.next()) != 0) {
if( lrgs(lidx).mask().is_UP() &&
lrgs(lidx).mask_size() &&
(lrgs(lidx)._is_float || lrgs(lidx)._is_vector))
cnt += lrgs(lidx).reg_pressure();
}
return cnt;
}
// Adjust register pressure down by 1. Capture last hi-to-low transition,
static void lower_pressure( LRG *lrg, uint where, Block *b, uint *pressure, uint *hrp_index ) {
if (lrg->mask().is_UP() && lrg->mask_size()) {
if (lrg->_is_float || lrg->_is_vector) {
pressure[1] -= lrg->reg_pressure();
if( pressure[1] == (uint)FLOATPRESSURE ) {
hrp_index[1] = where;
if( pressure[1] > b->_freg_pressure )
b->_freg_pressure = pressure[1]+1;
}
} else if( lrg->mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) ) {
pressure[0] -= lrg->reg_pressure();
if( pressure[0] == (uint)INTPRESSURE ) {
hrp_index[0] = where;
if( pressure[0] > b->_reg_pressure )
b->_reg_pressure = pressure[0]+1;
}
}
}
}
// Build the interference graph using physical registers when available.
// That is, if 2 live ranges are simultaneously alive but in their acceptable
// register sets do not overlap, then they do not interfere.
uint PhaseChaitin::build_ifg_physical( ResourceArea *a ) {
NOT_PRODUCT( Compile::TracePhase t3("buildIFG", &_t_buildIFGphysical, TimeCompiler); )
uint must_spill = 0;
// For all blocks (in any order) do...
for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
Block* block = _cfg.get_block(i);
// Clone (rather than smash in place) the liveout info, so it is alive
// for the "collect_gc_info" phase later.
IndexSet liveout(_live->live(block));
uint last_inst = block->end_idx();
// Compute first nonphi node index
uint first_inst;
for (first_inst = 1; first_inst < last_inst; first_inst++) {
if (!block->get_node(first_inst)->is_Phi()) {
break;
}
}
// Spills could be inserted before CreateEx node which should be
// first instruction in block after Phis. Move CreateEx up.
for (uint insidx = first_inst; insidx < last_inst; insidx++) {
Node *ex = block->get_node(insidx);
if (ex->is_SpillCopy()) {
continue;
}
if (insidx > first_inst && ex->is_Mach() && ex->as_Mach()->ideal_Opcode() == Op_CreateEx) {
// If the CreateEx isn't above all the MachSpillCopies
// then move it to the top.
block->remove_node(insidx);
block->insert_node(ex, first_inst);
}
// Stop once a CreateEx or any other node is found
break;
}
// Reset block's register pressure values for each ifg construction
uint pressure[2], hrp_index[2];
pressure[0] = pressure[1] = 0;
hrp_index[0] = hrp_index[1] = last_inst+1;
block->_reg_pressure = block->_freg_pressure = 0;
// Liveout things are presumed live for the whole block. We accumulate
// 'area' accordingly. If they get killed in the block, we'll subtract
// the unused part of the block from the area.
int inst_count = last_inst - first_inst;
double cost = (inst_count <= 0) ? 0.0 : block->_freq * double(inst_count);
assert(!(cost < 0.0), "negative spill cost" );
IndexSetIterator elements(&liveout);
uint lidx;
while ((lidx = elements.next()) != 0) {
LRG &lrg = lrgs(lidx);
lrg._area += cost;
// Compute initial register pressure
if (lrg.mask().is_UP() && lrg.mask_size()) {
if (lrg._is_float || lrg._is_vector) { // Count float pressure
pressure[1] += lrg.reg_pressure();
if (pressure[1] > block->_freg_pressure) {
block->_freg_pressure = pressure[1];
}
// Count int pressure, but do not count the SP, flags
} else if(lrgs(lidx).mask().overlap(*Matcher::idealreg2regmask[Op_RegI])) {
pressure[0] += lrg.reg_pressure();
if (pressure[0] > block->_reg_pressure) {
block->_reg_pressure = pressure[0];
}
}
}
}
assert( pressure[0] == count_int_pressure (&liveout), "" );
assert( pressure[1] == count_float_pressure(&liveout), "" );
// The IFG is built by a single reverse pass over each basic block.
// Starting with the known live-out set, we remove things that get
// defined and add things that become live (essentially executing one
// pass of a standard LIVE analysis). Just before a Node defines a value
// (and removes it from the live-ness set) that value is certainly live.
// The defined value interferes with everything currently live. The
// value is then removed from the live-ness set and it's inputs are added
// to the live-ness set.
uint j;
for (j = last_inst + 1; j > 1; j--) {
Node* n = block->get_node(j - 1);
// Get value being defined
uint r = _lrg_map.live_range_id(n);
// Some special values do not allocate
if(r) {
// A DEF normally costs block frequency; rematerialized values are
// removed from the DEF sight, so LOWER costs here.
lrgs(r)._cost += n->rematerialize() ? 0 : block->_freq;
// If it is not live, then this instruction is dead. Probably caused
// by spilling and rematerialization. Who cares why, yank this baby.
if( !liveout.member(r) && n->Opcode() != Op_SafePoint ) {
Node *def = n->in(0);
if( !n->is_Proj() ||
// Could also be a flags-projection of a dead ADD or such.
(_lrg_map.live_range_id(def) && !liveout.member(_lrg_map.live_range_id(def)))) {
bool remove = true;
if (n->is_MachProj()) {
// Don't remove KILL projections if their "defining" nodes have
// memory effects (have SCMemProj projection node) -
// they are not dead even when their result is not used.
// For example, compareAndSwapL (and other CAS) and EncodeISOArray nodes.
// The method add_input_to_liveout() keeps such nodes alive (put them on liveout list)
// when it sees SCMemProj node in a block. Unfortunately SCMemProj node could be placed
// in block in such order that KILL MachProj nodes are processed first.
uint cnt = def->outcnt();
for (uint i = 0; i < cnt; i++) {
Node* proj = def->raw_out(i);
if (proj->Opcode() == Op_SCMemProj) {
remove = false;
break;
}
}
}
if (remove) {
block->remove_node(j - 1);
if (lrgs(r)._def == n) {
lrgs(r)._def = 0;
}
n->disconnect_inputs(NULL, C);
_cfg.unmap_node_from_block(n);
n->replace_by(C->top());
// Since yanking a Node from block, high pressure moves up one
hrp_index[0]--;
hrp_index[1]--;
continue;
}
}
// Fat-projections kill many registers which cannot be used to
// hold live ranges.
if (lrgs(r)._fat_proj) {
// Count the int-only registers
RegMask itmp = lrgs(r).mask();
itmp.AND(*Matcher::idealreg2regmask[Op_RegI]);
int iregs = itmp.Size();
if (pressure[0]+iregs > block->_reg_pressure) {
block->_reg_pressure = pressure[0] + iregs;
}
if (pressure[0] <= (uint)INTPRESSURE && pressure[0] + iregs > (uint)INTPRESSURE) {
hrp_index[0] = j - 1;
}
// Count the float-only registers
RegMask ftmp = lrgs(r).mask();
ftmp.AND(*Matcher::idealreg2regmask[Op_RegD]);
int fregs = ftmp.Size();
if (pressure[1] + fregs > block->_freg_pressure) {
block->_freg_pressure = pressure[1] + fregs;
}
if(pressure[1] <= (uint)FLOATPRESSURE && pressure[1]+fregs > (uint)FLOATPRESSURE) {
hrp_index[1] = j - 1;
}
}
} else { // Else it is live
// A DEF also ends 'area' partway through the block.
lrgs(r)._area -= cost;
assert(!(lrgs(r)._area < 0.0), "negative spill area" );
扫一扫
10-31
129
129
11-01
103
103
11-01
197
197
11-01
179
179
03-28
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
