//------------------------------schedule_local---------------------------------
// Topological sort within a block. Someday become a real scheduler.
bool PhaseCFG::schedule_local(Block* block, 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 (trace_opto_pipelining()) {
tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
for (uint i = 0;i < block->number_of_nodes(); i++) {
tty->print("# ");
block->get_node(i)->fast_dump();
}
tty->print_cr("#");
}
#endif
// RootNode is already sorted
if (block->number_of_nodes() == 1) {
return true;
}
// Move PhiNodes and ParmNodes from 1 to cnt up to the start
uint node_cnt = block->end_idx();
uint phi_cnt = 1;
uint i;
for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
Node *n = block->get_node(i);
if( n->is_Phi() || // Found a PhiNode or ParmNode
(n->is_Proj() && n->in(0) == block->head()) ) {
// Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
block->map_node(block->get_node(phi_cnt), i);
block->map_node(n, phi_cnt++); // 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 && get_block_for_node(m) == block && !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(get_block_for_node(oop_store)->_dom_depth <= block->_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< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
ready_cnt.at_put(block->get_node(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 = block->get_node(i3); // Get pre-scheduled
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* m = n->fast_out(j);
if (get_block_for_node(m) == block) { // 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 = block->get_node(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(block, block->head(), next_call);
#ifndef PRODUCT
if (trace_opto_pipelining()) {
for (uint j=0; j< block->number_of_nodes(); j++) {
Node *n = block->get_node(j);
int idx = n->_idx;
tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
tty->print("latency:%3d ", get_latency_for_node(n));
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 (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(block, worklist, ready_cnt, next_call, phi_cnt);
block->map_node(n, phi_cnt++); // Schedule him next
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# select %d: %s", n->_idx, n->Name());
tty->print(", latency:%d", get_latency_for_node(n));
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(block, 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 (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
map_node_to_block(proj, block);
block->insert_node(proj, phi_cnt++);
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 (get_block_for_node(m) != block) {
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 != block->end_idx() ) {
// did not schedule all. Retry, Bailout, or Die
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 (trace_opto_pipelining()) {
tty->print_cr("#");
tty->print_cr("# after schedule_local");
for (uint i = 0;i < block->number_of_nodes();i++) {
tty->print("# ");
block->get_node(i)->fast_dump();
}
tty->cr();
}
#endif
return true;
}
//------------------------------call_catch_cleanup-----------------------------
// If we inserted any instructions between a Call and his CatchNode,
// clone the instructions on all paths below the Catch.
void Block::call_catch_cleanup(Block_Array &bbs) {
// End of region to clone
uint end = end_idx();
if( !_nodes[end]->is_Catch() ) return;
// Start of region to clone
uint beg = end;
while( _nodes[beg-1]->Opcode() != Op_MachProj ||
!_nodes[beg-1]->in(0)->is_Call() ) {
beg--;
assert(beg > 0,"Catch cleanup walking beyond block boundry");
}
if( beg == end ) return;
// Clone along all Catch output paths. Clone area between the 'beg' and
// 'end' indices.
for( uint i = 0; i < _num_succs; i++ ) {
Block *sb = _succs[i];
// Clone the entire area; ignoring the edge fixup for now.
for( uint j = end; j > beg; j-- ) {
Node *clone = _nodes[j-1]->clone();
sb->_nodes.insert( 1, clone );
bbs.map(clone->_idx,sb);
}
}
// Fixup edges. Check the def-use info per cloned Node
for(uint i2 = beg; i2 < end; i2++ ) {
uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
Node *n = _nodes[i2]; // Node that got cloned
// Need DU safe iterator because of edge manipulation in calls.
Node_List *out = new Node_List(Thread::current()->resource_area());
for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
out->push(n->fast_out(j1));
}
uint max = out->size();
for (uint j = 0; j < max; j++) {// For all users
Node *use = out->pop();
Block *buse = bbs[use->_idx];
if( use->is_Phi() ) {
for( uint k = 1; k < use->req(); k++ )
if( use->in(k) == n )
catch_cleanup_one_use( use, bbs[buse->pred(k)->_idx], n, this, bbs, beg, n_clone_idx, k );
} else {
catch_cleanup_one_use( use, buse, n, this, bbs, beg, n_clone_idx, -1 );
}
} // End for all users
} // End of for all Nodes in cloned area
// Remove the now-dead cloned ops
for(uint i3 = beg; i3 < end; i3++ ) {
_nodes[beg]->disconnect_inputs(NULL);
_nodes.remove(beg);
}
// If the successor blocks have a CreateEx node, move it back to the top
for(uint i4 = 0; i4 < _num_succs; i4++ ) {
Block *sb = _succs[i4];
MachNode *cex = sb->_nodes[1+end-beg]->is_Mach();
if( cex && cex->ideal_Opcode() == Op_CreateEx ) {
sb->_nodes.remove(1+end-beg);
sb->_nodes.insert(1,cex);
}
}
}
//------------------------------do_unswitching-----------------------------
// Clone loop with an invariant test (that does not exit) and
// insert a clone of the test that selects which version to
// execute.
void PhaseIdealLoop::do_unswitching (IdealLoopTree *loop, Node_List &old_new) {
// Find first invariant test that doesn't exit the loop
LoopNode *head = loop->_head->as_Loop();
IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop);
assert(unswitch_iff != NULL, "should be at least one");
// Need to revert back to normal loop
if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
head->as_CountedLoop()->set_normal_loop();
}
ProjNode* proj_true = create_slow_version_of_loop(loop, old_new);
assert(proj_true->is_IfTrue() && proj_true->unique_ctrl_out() == head, "by construction");
// Increment unswitch count
LoopNode* head_clone = old_new[head->_idx]->as_Loop();
int nct = head->unswitch_count() + 1;
head->set_unswitch_count(nct);
head_clone->set_unswitch_count(nct);
// Add test to new "if" outside of loop
IfNode* invar_iff = proj_true->in(0)->as_If();
Node* invar_iff_c = invar_iff->in(0);
BoolNode* bol = unswitch_iff->in(1)->as_Bool();
invar_iff->set_req(1, bol);
invar_iff->_prob = unswitch_iff->_prob;
ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
// Hoist invariant casts out of each loop to the appropriate
// control projection.
Node_List worklist;
for (DUIterator_Fast imax, i = unswitch_iff->fast_outs(imax); i < imax; i++) {
ProjNode* proj= unswitch_iff->fast_out(i)->as_Proj();
// Copy to a worklist for easier manipulation
for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
Node* use = proj->fast_out(j);
if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
worklist.push(use);
}
}
ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
while (worklist.size() > 0) {
Node* use = worklist.pop();
Node* nuse = use->clone();
nuse->set_req(0, invar_proj);
_igvn.hash_delete(use);
use->set_req(1, nuse);
_igvn._worklist.push(use);
register_new_node(nuse, invar_proj);
// Same for the clone
Node* use_clone = old_new[use->_idx];
_igvn.hash_delete(use_clone);
use_clone->set_req(1, nuse);
_igvn._worklist.push(use_clone);
}
}
// Hardwire the control paths in the loops into if(true) and if(false)
_igvn.hash_delete(unswitch_iff);
short_circuit_if(unswitch_iff, proj_true);
_igvn._worklist.push(unswitch_iff);
IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
_igvn.hash_delete(unswitch_iff_clone);
short_circuit_if(unswitch_iff_clone, proj_false);
_igvn._worklist.push(unswitch_iff_clone);
// Reoptimize loops
loop->record_for_igvn();
for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
Node *n = loop->_body[i];
Node *n_clone = old_new[n->_idx];
_igvn._worklist.push(n_clone);
}
#ifndef PRODUCT
if (TraceLoopUnswitching) {
tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d",
head->_idx, unswitch_iff->_idx,
old_new[head->_idx]->_idx, unswitch_iff_clone->_idx);
}
#endif
C->set_major_progress();
}
//------------------------------schedule_local---------------------------------
// Topological sort within a block. Someday become a real scheduler.
bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, 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[n->_idx] = local; // Count em up
// A few node types require changing a required edge to a precedence edge
// before allocation.
if( UseConcMarkSweepGC || UseG1GC ) {
if( n->is_Mach() && n->as_Mach()->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(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
}
}
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[_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
ready_cnt[m->_idx]--; // 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[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[idx]);
tty->print("latency:%3d ", cfg->_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
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;
}
// 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( !--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 && !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;
}