//------------------------------schedule_local---------------------------------
// Topological sort within a block. Someday become a real scheduler.
bool Block::schedule_local(Matcher &matcher, Block_Array &bbs,int *ready_cnt, VectorSet &next_call, GrowableArray<uint> &node_latency) {
// 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 (TraceOptoPipelining) {
tty->print("# before schedule_local\n");
for (uint i = 0;i < _nodes.size();i++) {
tty->print("# ");
_nodes[i]->fast_dump();
}
tty->print("\n");
}
#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 && bbs[m->_idx] == this && !m->is_top() )
local++; // One more block-local input
}
ready_cnt[n->_idx] = local; // Count em up
// A few node types require changing a required edge to a precedence edge
// before allocation.
MachNode *m = n->is_Mach();
if( UseConcMarkSweepGC ) {
if( m && m->ideal_Opcode() == Op_StoreCM ) {
// Note: Required edges with an index greater than oper_input_base
// are not supported by the allocator.
// Note2: Can only depend on unmatched edge being last,
// can not depend on its absolute position.
Node *oop_store = n->in(n->req() - 1);
n->del_req(n->req() - 1);
n->add_prec(oop_store);
assert(bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
}
}
if( m && m->ideal_Opcode() == Op_MemBarAcquire ) {
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[_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( bbs[m->_idx] ==this ) // Local-block user
ready_cnt[m->_idx]--; // Fix ready count
}
}
// 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[m->_idx] ) // Zero ready count?
worklist.push(m); // Then on to worklist!
}
// Warm up the 'next_call' heuristic bits
needed_for_next_call(_nodes[0], next_call, bbs);
#ifndef PRODUCT
if (TraceOptoPipelining) {
for (uint j=0; j<_nodes.size(); j++) {
Node *n = _nodes[j];
int idx = n->_idx;
tty->print("# ready cnt:%3d ", ready_cnt[idx]);
tty->print("latency:%3d ", node_latency.at_grow(idx));
tty->print("%4d: %s\n", idx, n->Name());
}
}
#endif
// Pull from worklist and schedule
while( worklist.size() ) { // Worklist is not ready
#ifndef PRODUCT
uint before_size = worklist.size();
if (TraceOptoPipelining && before_size > 1) {
tty->print("# before select:");
for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
Node *n = worklist[i]; // Get Node on worklist
tty->print(" %3d", n->_idx);
}
tty->print("\n");
}
#endif
// Select and pop a ready guy from worklist
Node* n = select(worklist, bbs, ready_cnt, next_call, phi_cnt, node_latency);
_nodes.map(phi_cnt++,n); // Schedule him next
MachNode *m = n->is_Mach();
#ifndef PRODUCT
if (TraceOptoPipelining && before_size > 1) {
tty->print("# select %d: %s", n->_idx, n->Name());
tty->print(", latency:%d", node_latency.at_grow(n->_idx));
n->dump();
tty->print("# after select:");
for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
Node *n = worklist[i]; // Get Node on worklist
tty->print(" %4d", n->_idx);
}
tty->print("\n");
}
#endif
if( m ) {
MachCallNode *mcall = m->is_MachCall();
if( mcall ) {
phi_cnt = sched_call(matcher, bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
continue;
}
}
// 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( bbs[m->_idx] != this ) continue;
if( m->is_Phi() ) continue;
if( !--ready_cnt[m->_idx] )
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) {
// Retry with subsume_loads == false
C->set_result(Compile::Comp_subsumed_load_conflict);
} else {
// Bailout without retry
C->set_result(Compile::Comp_no_retry);
}
// assert( phi_cnt == end_idx(), "did not schedule all" );
return false;
}
#ifndef PRODUCT
if (TraceOptoPipelining) {
tty->print("# after schedule_local\n");
for (uint i = 0;i < _nodes.size();i++) {
tty->print("# ");
_nodes[i]->fast_dump();
}
tty->print("\n");
}
#endif
return true;
}
