JVMState* ParseGenerator::generate(JVMState* jvms, Parse* parent_parser) {
Compile* C = Compile::current();
if (is_osr()) {
// The JVMS for a OSR has a single argument (see its TypeFunc).
assert(jvms->depth() == 1, "no inline OSR");
}
if (C->failing()) {
return NULL; // bailing out of the compile; do not try to parse
}
Parse parser(jvms, method(), _expected_uses, parent_parser);
// Grab signature for matching/allocation
#ifdef ASSERT
if (parser.tf() != (parser.depth() == 1 ? C->tf() : tf())) {
MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
assert(C->env()->system_dictionary_modification_counter_changed(),
"Must invalidate if TypeFuncs differ");
}
#endif
GraphKit& exits = parser.exits();
if (C->failing()) {
while (exits.pop_exception_state() != NULL) ;
return NULL;
}
assert(exits.jvms()->same_calls_as(jvms), "sanity");
// Simply return the exit state of the parser,
// augmented by any exceptional states.
return exits.transfer_exceptions_into_jvms();
}
void LateInlineCallGenerator::do_late_inline() {
// Can't inline it
if (call_node() == NULL || call_node()->outcnt() == 0 ||
call_node()->in(0) == NULL || call_node()->in(0)->is_top())
return;
CallStaticJavaNode* call = call_node();
// Make a clone of the JVMState that appropriate to use for driving a parse
Compile* C = Compile::current();
JVMState* jvms = call->jvms()->clone_shallow(C);
uint size = call->req();
SafePointNode* map = new (C, size) SafePointNode(size, jvms);
for (uint i1 = 0; i1 < size; i1++) {
map->init_req(i1, call->in(i1));
}
// Make sure the state is a MergeMem for parsing.
if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
map->set_req(TypeFunc::Memory, MergeMemNode::make(C, map->in(TypeFunc::Memory)));
}
// Make enough space for the expression stack and transfer the incoming arguments
int nargs = method()->arg_size();
jvms->set_map(map);
map->ensure_stack(jvms, jvms->method()->max_stack());
if (nargs > 0) {
for (int i1 = 0; i1 < nargs; i1++) {
map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1));
}
}
CompileLog* log = C->log();
if (log != NULL) {
log->head("late_inline method='%d'", log->identify(method()));
JVMState* p = jvms;
while (p != NULL) {
log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
p = p->caller();
}
log->tail("late_inline");
}
// Setup default node notes to be picked up by the inlining
Node_Notes* old_nn = C->default_node_notes();
if (old_nn != NULL) {
Node_Notes* entry_nn = old_nn->clone(C);
entry_nn->set_jvms(jvms);
C->set_default_node_notes(entry_nn);
}
// Now perform the inling using the synthesized JVMState
JVMState* new_jvms = _inline_cg->generate(jvms);
if (new_jvms == NULL) return; // no change
if (C->failing()) return;
// Capture any exceptional control flow
GraphKit kit(new_jvms);
// Find the result object
Node* result = C->top();
int result_size = method()->return_type()->size();
if (result_size != 0 && !kit.stopped()) {
result = (result_size == 1) ? kit.pop() : kit.pop_pair();
}
kit.replace_call(call, result);
}
void LateInlineCallGenerator::do_late_inline() {
// Can't inline it
CallStaticJavaNode* call = call_node();
if (call == NULL || call->outcnt() == 0 ||
call->in(0) == NULL || call->in(0)->is_top()) {
return;
}
const TypeTuple *r = call->tf()->domain();
for (int i1 = 0; i1 < method()->arg_size(); i1++) {
if (call->in(TypeFunc::Parms + i1)->is_top() && r->field_at(TypeFunc::Parms + i1) != Type::HALF) {
assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
return;
}
}
if (call->in(TypeFunc::Memory)->is_top()) {
assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
return;
}
Compile* C = Compile::current();
// Remove inlined methods from Compiler's lists.
if (call->is_macro()) {
C->remove_macro_node(call);
}
// Make a clone of the JVMState that appropriate to use for driving a parse
JVMState* old_jvms = call->jvms();
JVMState* jvms = old_jvms->clone_shallow(C);
uint size = call->req();
SafePointNode* map = new (C) SafePointNode(size, jvms);
for (uint i1 = 0; i1 < size; i1++) {
map->init_req(i1, call->in(i1));
}
// Make sure the state is a MergeMem for parsing.
if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory));
C->initial_gvn()->set_type_bottom(mem);
map->set_req(TypeFunc::Memory, mem);
}
uint nargs = method()->arg_size();
// blow away old call arguments
Node* top = C->top();
for (uint i1 = 0; i1 < nargs; i1++) {
map->set_req(TypeFunc::Parms + i1, top);
}
jvms->set_map(map);
// Make enough space in the expression stack to transfer
// the incoming arguments and return value.
map->ensure_stack(jvms, jvms->method()->max_stack());
for (uint i1 = 0; i1 < nargs; i1++) {
map->set_argument(jvms, i1, call->in(TypeFunc::Parms + i1));
}
// This check is done here because for_method_handle_inline() method
// needs jvms for inlined state.
if (!do_late_inline_check(jvms)) {
map->disconnect_inputs(NULL, C);
return;
}
C->print_inlining_insert(this);
CompileLog* log = C->log();
if (log != NULL) {
log->head("late_inline method='%d'", log->identify(method()));
JVMState* p = jvms;
while (p != NULL) {
log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
p = p->caller();
}
log->tail("late_inline");
}
// Setup default node notes to be picked up by the inlining
Node_Notes* old_nn = C->default_node_notes();
if (old_nn != NULL) {
Node_Notes* entry_nn = old_nn->clone(C);
entry_nn->set_jvms(jvms);
C->set_default_node_notes(entry_nn);
}
// Now perform the inling using the synthesized JVMState
JVMState* new_jvms = _inline_cg->generate(jvms, NULL);
if (new_jvms == NULL) return; // no change
if (C->failing()) return;
// Capture any exceptional control flow
GraphKit kit(new_jvms);
// Find the result object
Node* result = C->top();
int result_size = method()->return_type()->size();
if (result_size != 0 && !kit.stopped()) {
result = (result_size == 1) ? kit.pop() : kit.pop_pair();
}
C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops());
C->env()->notice_inlined_method(_inline_cg->method());
C->set_inlining_progress(true);
kit.replace_call(call, result);
}
//------------------------------schedule_local---------------------------------
// Topological sort within a block. Someday become a real scheduler.
bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
// Already "sorted" are the block start Node (as the first entry), and
// the block-ending Node and any trailing control projections. We leave
// these alone. PhiNodes and ParmNodes are made to follow the block start
// Node. Everything else gets topo-sorted.
#ifndef PRODUCT
if (cfg->trace_opto_pipelining()) {
tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
for (uint i = 0;i < _nodes.size();i++) {
tty->print("# ");
_nodes[i]->fast_dump();
}
tty->print_cr("#");
}
#endif
// RootNode is already sorted
if( _nodes.size() == 1 ) return true;
// Move PhiNodes and ParmNodes from 1 to cnt up to the start
uint node_cnt = end_idx();
uint phi_cnt = 1;
uint i;
for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
Node *n = _nodes[i];
if( n->is_Phi() || // Found a PhiNode or ParmNode
(n->is_Proj() && n->in(0) == head()) ) {
// Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
_nodes.map(i,_nodes[phi_cnt]);
_nodes.map(phi_cnt++,n); // swap Phi/Parm up front
} else { // All others
// Count block-local inputs to 'n'
uint cnt = n->len(); // Input count
uint local = 0;
for( uint j=0; j<cnt; j++ ) {
Node *m = n->in(j);
if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
local++; // One more block-local input
}
ready_cnt.at_put(n->_idx, local); // Count em up
#ifdef ASSERT
if( UseConcMarkSweepGC || UseG1GC ) {
if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
// Check the precedence edges
for (uint prec = n->req(); prec < n->len(); prec++) {
Node* oop_store = n->in(prec);
if (oop_store != NULL) {
assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
}
}
}
}
#endif
// A few node types require changing a required edge to a precedence edge
// before allocation.
if( n->is_Mach() && n->req() > TypeFunc::Parms &&
(n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
// MemBarAcquire could be created without Precedent edge.
// del_req() replaces the specified edge with the last input edge
// and then removes the last edge. If the specified edge > number of
// edges the last edge will be moved outside of the input edges array
// and the edge will be lost. This is why this code should be
// executed only when Precedent (== TypeFunc::Parms) edge is present.
Node *x = n->in(TypeFunc::Parms);
n->del_req(TypeFunc::Parms);
n->add_prec(x);
}
}
}
for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
ready_cnt.at_put(_nodes[i2]->_idx, 0);
// All the prescheduled guys do not hold back internal nodes
uint i3;
for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
Node *n = _nodes[i3]; // Get pre-scheduled
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* m = n->fast_out(j);
if( cfg->_bbs[m->_idx] ==this ) { // Local-block user
int m_cnt = ready_cnt.at(m->_idx)-1;
ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
}
}
}
Node_List delay;
// Make a worklist
Node_List worklist;
for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
Node *m = _nodes[i4];
if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
if (m->is_iteratively_computed()) {
// Push induction variable increments last to allow other uses
// of the phi to be scheduled first. The select() method breaks
// ties in scheduling by worklist order.
delay.push(m);
} else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
// Force the CreateEx to the top of the list so it's processed
// first and ends up at the start of the block.
worklist.insert(0, m);
} else {
worklist.push(m); // Then on to worklist!
}
}
}
while (delay.size()) {
Node* d = delay.pop();
worklist.push(d);
}
// Warm up the 'next_call' heuristic bits
needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
#ifndef PRODUCT
if (cfg->trace_opto_pipelining()) {
for (uint j=0; j<_nodes.size(); j++) {
Node *n = _nodes[j];
int idx = n->_idx;
tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx));
tty->print("%4d: %s\n", idx, n->Name());
}
}
#endif
uint max_idx = (uint)ready_cnt.length();
// Pull from worklist and schedule
while( worklist.size() ) { // Worklist is not ready
#ifndef PRODUCT
if (cfg->trace_opto_pipelining()) {
tty->print("# ready list:");
for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
Node *n = worklist[i]; // Get Node on worklist
tty->print(" %d", n->_idx);
}
tty->cr();
}
#endif
// Select and pop a ready guy from worklist
Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
_nodes.map(phi_cnt++,n); // Schedule him next
#ifndef PRODUCT
if (cfg->trace_opto_pipelining()) {
tty->print("# select %d: %s", n->_idx, n->Name());
tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
n->dump();
if (Verbose) {
tty->print("# ready list:");
for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
Node *n = worklist[i]; // Get Node on worklist
tty->print(" %d", n->_idx);
}
tty->cr();
}
}
#endif
if( n->is_MachCall() ) {
MachCallNode *mcall = n->as_MachCall();
phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
continue;
}
if (n->is_Mach() && n->as_Mach()->has_call()) {
RegMask regs;
regs.Insert(matcher.c_frame_pointer());
regs.OR(n->out_RegMask());
MachProjNode *proj = new (matcher.C, 1) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
cfg->_bbs.map(proj->_idx,this);
_nodes.insert(phi_cnt++, proj);
add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
}
// Children are now all ready
for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
Node* m = n->fast_out(i5); // Get user
if( cfg->_bbs[m->_idx] != this ) continue;
if( m->is_Phi() ) continue;
if (m->_idx >= max_idx) { // new node, skip it
assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
continue;
}
int m_cnt = ready_cnt.at(m->_idx)-1;
ready_cnt.at_put(m->_idx, m_cnt);
if( m_cnt == 0 )
worklist.push(m);
}
}
if( phi_cnt != end_idx() ) {
// did not schedule all. Retry, Bailout, or Die
Compile* C = matcher.C;
if (C->subsume_loads() == true && !C->failing()) {
// Retry with subsume_loads == false
// If this is the first failure, the sentinel string will "stick"
// to the Compile object, and the C2Compiler will see it and retry.
C->record_failure(C2Compiler::retry_no_subsuming_loads());
}
// assert( phi_cnt == end_idx(), "did not schedule all" );
return false;
}
#ifndef PRODUCT
if (cfg->trace_opto_pipelining()) {
tty->print_cr("#");
tty->print_cr("# after schedule_local");
for (uint i = 0;i < _nodes.size();i++) {
tty->print("# ");
_nodes[i]->fast_dump();
}
tty->cr();
}
#endif
return true;
}
