//------------------------------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);
}
}
}
//------------------------------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;
}
//------------------------------implicit_null_check----------------------------
// Detect implicit-null-check opportunities. Basically, find NULL checks
// with suitable memory ops nearby. Use the memory op to do the NULL check.
// I can generate a memory op if there is not one nearby.
void Block::implicit_null_check(Block_Array &bbs, GrowableArray<uint> &latency, Node *proj, Node *val) {
// Assume if null check need for 0 offset then always needed
// Intel solaris doesn't support any null checks yet and no
// mechanism exists (yet) to set the switches at an os_cpu level
if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
// Make sure the ptr-is-null path appears to be uncommon!
float f = end()->is_Mach()->is_MachIf()->_prob;
if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
if( f > 0.0001 ) return;
uint bidx = 0; // Capture index of value into memop
bool was_store; // Memory op is a store op
// Search the successor block for a load or store who's base value is also
// the tested value. There may be several.
Node_List *out = new Node_List(Thread::current()->resource_area());
MachNode *best = NULL; // Best found so far
for (DUIterator i = val->outs(); val->has_out(i); i++) {
MachNode *mach = val->out(i)->is_Mach();
if( !mach ) continue;
was_store = false;
switch( mach->ideal_Opcode() ) {
case Op_LoadB:
case Op_LoadC:
case Op_LoadD:
case Op_LoadF:
case Op_LoadI:
case Op_LoadL:
case Op_LoadP:
case Op_LoadS:
case Op_LoadKlass:
case Op_LoadRange:
case Op_LoadD_unaligned:
case Op_LoadL_unaligned:
break;
case Op_StoreB:
case Op_StoreC:
case Op_StoreCM:
case Op_StoreD:
case Op_StoreF:
case Op_StoreI:
case Op_StoreL:
case Op_StoreP:
was_store = true; // Memory op is a store op
// Stores will have their address in slot 2 (memory in slot 1).
// If the value being nul-checked is in another slot, it means we
// are storing the checked value, which does NOT check the value!
if( mach->in(2) != val ) continue;
break; // Found a memory op?
case Op_StrComp:
// Not a legit memory op for implicit null check regardless of
// embedded loads
continue;
default: // Also check for embedded loads
if( !mach->check_for_anti_dependence() )
continue; // Not an memory op; skip it
break;
}
// check if the offset is not too high for implicit exception
{
intptr_t offset = 0;
const TypePtr *adr_type = NULL; // Do not need this return value here
const Node* base = mach->get_base_and_disp(offset, adr_type);
if (base == NULL || base == (Node*)-1) {
// cannot reason about it; is probably not implicit null exception
} else {
const TypePtr* tptr = base->bottom_type()->is_ptr();
// Give up if offset is not a compile-time constant
if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
continue;
offset += tptr->_offset; // correct if base is offseted
if( MacroAssembler::needs_explicit_null_check(offset) )
continue; // Give up is reference is beyond 4K page size
}
}
// Check ctrl input to see if the null-check dominates the memory op
Block *cb = bbs[mach->_idx];
cb = cb->_idom; // Always hoist at least 1 block
if( !was_store ) { // Stores can be hoisted only one block
while( cb->_dom_depth > _dom_depth )
cb = cb->_idom; // Hoist loads as far as we want
}
if( cb != this ) continue;
// Found a memory user; see if it can be hoisted to check-block
uint vidx = 0; // Capture index of value into memop
uint j;
for( j = mach->req()-1; j > 0; j-- ) {
if( mach->in(j) == val ) vidx = j;
// Block of memory-op input
Block *inb = bbs[mach->in(j)->_idx];
Block *b = this; // Start from nul check
while( b != inb && b->_dom_depth > inb->_dom_depth )
b = b->_idom; // search upwards for input
// See if input dominates null check
if( b != inb )
break;
}
if( j > 0 )
continue;
Block *mb = bbs[mach->_idx];
// Hoisting stores requires more checks for the anti-dependence case.
// Give up hoisting if we have to move the store past any load.
if( was_store ) {
Block *b = mb; // Start searching here for a local load
// mach use (faulting) trying to hoist
// n might be blocker to hoisting
while( b != this ) {
uint k;
for( k = 1; k < b->_nodes.size(); k++ ) {
Node *n = b->_nodes[k];
if( n->check_for_anti_dependence() &&
n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
break; // Found anti-dependent load
}
if( k < b->_nodes.size() )
break; // Found anti-dependent load
// Make sure control does not do a merge (would have to check allpaths)
if( b->num_preds() != 2 ) break;
b = bbs[b->pred(1)->_idx]; // Move up to predecessor block
}
if( b != this ) continue;
}
// Make sure this memory op is not already being used for a NullCheck
MachNode *e = mb->end()->is_Mach();
if( e && e->is_MachNullCheck() && e->in(1) == mach )
continue; // Already being used as a NULL check
// Found a candidate! Pick one with least dom depth - the highest
// in the dom tree should be closest to the null check.
if( !best ||
bbs[mach->_idx]->_dom_depth < bbs[best->_idx]->_dom_depth ) {
best = mach;
bidx = vidx;
}
}
// No candidate!
if( !best ) return;
// ---- Found an implicit null check
extern int implicit_null_checks;
implicit_null_checks++;
// Hoist the memory candidate up to the end of the test block.
Block *old_block = bbs[best->_idx];
old_block->find_remove(best);
add_inst(best);
bbs.map(best->_idx,this);
// Move the control dependence
if (best->in(0) && best->in(0) == old_block->_nodes[0])
best->set_req(0, _nodes[0]);
// Check for flag-killing projections that also need to be hoisted
// Should be DU safe because no edge updates.
for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
Node* n = best->fast_out(j);
if( n->Opcode() == Op_MachProj ) {
bbs[n->_idx]->find_remove(n);
add_inst(n);
bbs.map(n->_idx,this);
}
}
// proj==Op_True --> ne test; proj==Op_False --> eq test.
// One of two graph shapes got matched:
// (IfTrue (If (Bool NE (CmpP ptr NULL))))
// (IfFalse (If (Bool EQ (CmpP ptr NULL))))
// NULL checks are always branch-if-eq. If we see a IfTrue projection
// then we are replacing a 'ne' test with a 'eq' NULL check test.
// We need to flip the projections to keep the same semantics.
if( proj->Opcode() == Op_IfTrue ) {
// Swap order of projections in basic block to swap branch targets
Node *tmp1 = _nodes[end_idx()+1];
Node *tmp2 = _nodes[end_idx()+2];
_nodes.map(end_idx()+1, tmp2);
_nodes.map(end_idx()+2, tmp1);
Node *tmp = new (1) Node(1);
tmp1->replace_by(tmp);
tmp2->replace_by(tmp1);
tmp->replace_by(tmp2);
}
// Remove the existing null check; use a new implicit null check instead.
// Since schedule-local needs precise def-use info, we need to correct
// it as well.
Node *old_tst = proj->in(0);
MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
_nodes.map(end_idx(),nul_chk);
bbs.map(nul_chk->_idx,this);
// Redirect users of old_test to nul_chk
for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
old_tst->last_out(i2)->set_req(0, nul_chk);
// Clean-up any dead code
for (uint i3 = 0; i3 < old_tst->req(); i3++)
old_tst->set_req(i3, NULL);
latency.at_put_grow(nul_chk->_idx, nul_chk->latency_from_uses(bbs, latency));
latency.at_put_grow(best ->_idx, best ->latency_from_uses(bbs, latency));
#ifndef PRODUCT
if (TraceOptoPipelining) {
tty->print("# implicit_null_check: latency %4d for ", latency.at_grow(best->_idx));
best->fast_dump();
tty->print("# implicit_null_check: latency %4d for ", latency.at_grow(nul_chk->_idx));
nul_chk->fast_dump();
}
#endif
}
//------------------------------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;
}
//------------------------------implicit_null_check----------------------------
// Detect implicit-null-check opportunities. Basically, find NULL checks
// with suitable memory ops nearby. Use the memory op to do the NULL check.
// I can generate a memory op if there is not one nearby.
// The proj is the control projection for the not-null case.
// The val is the pointer being checked for nullness or
// decodeHeapOop_not_null node if it did not fold into address.
void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
// Assume if null check need for 0 offset then always needed
// Intel solaris doesn't support any null checks yet and no
// mechanism exists (yet) to set the switches at an os_cpu level
if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
// Make sure the ptr-is-null path appears to be uncommon!
float f = end()->as_MachIf()->_prob;
if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
if( f > PROB_UNLIKELY_MAG(4) ) return;
uint bidx = 0; // Capture index of value into memop
bool was_store; // Memory op is a store op
// Get the successor block for if the test ptr is non-null
Block* not_null_block; // this one goes with the proj
Block* null_block;
if (_nodes[_nodes.size()-1] == proj) {
null_block = _succs[0];
not_null_block = _succs[1];
} else {
assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
not_null_block = _succs[0];
null_block = _succs[1];
}
while (null_block->is_Empty() == Block::empty_with_goto) {
null_block = null_block->_succs[0];
}
// Search the exception block for an uncommon trap.
// (See Parse::do_if and Parse::do_ifnull for the reason
// we need an uncommon trap. Briefly, we need a way to
// detect failure of this optimization, as in 6366351.)
{
bool found_trap = false;
for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
Node* nn = null_block->_nodes[i1];
if (nn->is_MachCall() &&
nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
if (trtype->isa_int() && trtype->is_int()->is_con()) {
jint tr_con = trtype->is_int()->get_con();
Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
assert((int)reason < (int)BitsPerInt, "recode bit map");
if (is_set_nth_bit(allowed_reasons, (int) reason)
&& action != Deoptimization::Action_none) {
// This uncommon trap is sure to recompile, eventually.
// When that happens, C->too_many_traps will prevent
// this transformation from happening again.
found_trap = true;
}
}
break;
}
}
if (!found_trap) {
// We did not find an uncommon trap.
return;
}
}
// Check for decodeHeapOop_not_null node which did not fold into address
bool is_decoden = ((intptr_t)val) & 1;
val = (Node*)(((intptr_t)val) & ~1);
assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
(val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
// Search the successor block for a load or store who's base value is also
// the tested value. There may be several.
Node_List *out = new Node_List(Thread::current()->resource_area());
MachNode *best = NULL; // Best found so far
for (DUIterator i = val->outs(); val->has_out(i); i++) {
Node *m = val->out(i);
if( !m->is_Mach() ) continue;
MachNode *mach = m->as_Mach();
was_store = false;
int iop = mach->ideal_Opcode();
switch( iop ) {
case Op_LoadB:
case Op_LoadUS:
case Op_LoadD:
case Op_LoadF:
case Op_LoadI:
case Op_LoadL:
case Op_LoadP:
case Op_LoadN:
case Op_LoadS:
case Op_LoadKlass:
case Op_LoadNKlass:
case Op_LoadRange:
case Op_LoadD_unaligned:
case Op_LoadL_unaligned:
assert(mach->in(2) == val, "should be address");
break;
case Op_StoreB:
case Op_StoreC:
case Op_StoreCM:
case Op_StoreD:
case Op_StoreF:
case Op_StoreI:
case Op_StoreL:
case Op_StoreP:
case Op_StoreN:
was_store = true; // Memory op is a store op
// Stores will have their address in slot 2 (memory in slot 1).
// If the value being nul-checked is in another slot, it means we
// are storing the checked value, which does NOT check the value!
if( mach->in(2) != val ) continue;
break; // Found a memory op?
case Op_StrComp:
case Op_StrEquals:
case Op_StrIndexOf:
case Op_AryEq:
// Not a legit memory op for implicit null check regardless of
// embedded loads
continue;
default: // Also check for embedded loads
if( !mach->needs_anti_dependence_check() )
continue; // Not an memory op; skip it
if( must_clone[iop] ) {
// Do not move nodes which produce flags because
// RA will try to clone it to place near branch and
// it will cause recompilation, see clone_node().
continue;
}
{
// Check that value is used in memory address in
// instructions with embedded load (CmpP val1,(val2+off)).
Node* base;
Node* index;
const MachOper* oper = mach->memory_inputs(base, index);
if (oper == NULL || oper == (MachOper*)-1) {
continue; // Not an memory op; skip it
}
if (val == base ||
val == index && val->bottom_type()->isa_narrowoop()) {
break; // Found it
} else {
continue; // Skip it
}
}
break;
}
// check if the offset is not too high for implicit exception
{
intptr_t offset = 0;
const TypePtr *adr_type = NULL; // Do not need this return value here
const Node* base = mach->get_base_and_disp(offset, adr_type);
if (base == NULL || base == NodeSentinel) {
// Narrow oop address doesn't have base, only index
if( val->bottom_type()->isa_narrowoop() &&
MacroAssembler::needs_explicit_null_check(offset) )
continue; // Give up if offset is beyond page size
// cannot reason about it; is probably not implicit null exception
} else {
const TypePtr* tptr;
if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
// 32-bits narrow oop can be the base of address expressions
tptr = base->bottom_type()->make_ptr();
} else {
// only regular oops are expected here
tptr = base->bottom_type()->is_ptr();
}
// Give up if offset is not a compile-time constant
if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
continue;
offset += tptr->_offset; // correct if base is offseted
if( MacroAssembler::needs_explicit_null_check(offset) )
continue; // Give up is reference is beyond 4K page size
}
}
// Check ctrl input to see if the null-check dominates the memory op
Block *cb = cfg->_bbs[mach->_idx];
cb = cb->_idom; // Always hoist at least 1 block
if( !was_store ) { // Stores can be hoisted only one block
while( cb->_dom_depth > (_dom_depth + 1))
cb = cb->_idom; // Hoist loads as far as we want
// The non-null-block should dominate the memory op, too. Live
// range spilling will insert a spill in the non-null-block if it is
// needs to spill the memory op for an implicit null check.
if (cb->_dom_depth == (_dom_depth + 1)) {
if (cb != not_null_block) continue;
cb = cb->_idom;
}
}
if( cb != this ) continue;
// Found a memory user; see if it can be hoisted to check-block
uint vidx = 0; // Capture index of value into memop
uint j;
for( j = mach->req()-1; j > 0; j-- ) {
if( mach->in(j) == val ) {
vidx = j;
// Ignore DecodeN val which could be hoisted to where needed.
if( is_decoden ) continue;
}
// Block of memory-op input
Block *inb = cfg->_bbs[mach->in(j)->_idx];
Block *b = this; // Start from nul check
while( b != inb && b->_dom_depth > inb->_dom_depth )
b = b->_idom; // search upwards for input
// See if input dominates null check
if( b != inb )
break;
}
if( j > 0 )
continue;
Block *mb = cfg->_bbs[mach->_idx];
// Hoisting stores requires more checks for the anti-dependence case.
// Give up hoisting if we have to move the store past any load.
if( was_store ) {
Block *b = mb; // Start searching here for a local load
// mach use (faulting) trying to hoist
// n might be blocker to hoisting
while( b != this ) {
uint k;
for( k = 1; k < b->_nodes.size(); k++ ) {
Node *n = b->_nodes[k];
if( n->needs_anti_dependence_check() &&
n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
break; // Found anti-dependent load
}
if( k < b->_nodes.size() )
break; // Found anti-dependent load
// Make sure control does not do a merge (would have to check allpaths)
if( b->num_preds() != 2 ) break;
b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
}
if( b != this ) continue;
}
// Make sure this memory op is not already being used for a NullCheck
Node *e = mb->end();
if( e->is_MachNullCheck() && e->in(1) == mach )
continue; // Already being used as a NULL check
// Found a candidate! Pick one with least dom depth - the highest
// in the dom tree should be closest to the null check.
if( !best ||
cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
best = mach;
bidx = vidx;
}
}
// No candidate!
if( !best ) return;
// ---- Found an implicit null check
extern int implicit_null_checks;
implicit_null_checks++;
if( is_decoden ) {
// Check if we need to hoist decodeHeapOop_not_null first.
Block *valb = cfg->_bbs[val->_idx];
if( this != valb && this->_dom_depth < valb->_dom_depth ) {
// Hoist it up to the end of the test block.
valb->find_remove(val);
this->add_inst(val);
cfg->_bbs.map(val->_idx,this);
// DecodeN on x86 may kill flags. Check for flag-killing projections
// that also need to be hoisted.
for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
Node* n = val->fast_out(j);
if( n->is_MachProj() ) {
cfg->_bbs[n->_idx]->find_remove(n);
this->add_inst(n);
cfg->_bbs.map(n->_idx,this);
}
}
}
}
// Hoist the memory candidate up to the end of the test block.
Block *old_block = cfg->_bbs[best->_idx];
old_block->find_remove(best);
add_inst(best);
cfg->_bbs.map(best->_idx,this);
// Move the control dependence
if (best->in(0) && best->in(0) == old_block->_nodes[0])
best->set_req(0, _nodes[0]);
// Check for flag-killing projections that also need to be hoisted
// Should be DU safe because no edge updates.
for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
Node* n = best->fast_out(j);
if( n->is_MachProj() ) {
cfg->_bbs[n->_idx]->find_remove(n);
add_inst(n);
cfg->_bbs.map(n->_idx,this);
}
}
Compile *C = cfg->C;
// proj==Op_True --> ne test; proj==Op_False --> eq test.
// One of two graph shapes got matched:
// (IfTrue (If (Bool NE (CmpP ptr NULL))))
// (IfFalse (If (Bool EQ (CmpP ptr NULL))))
// NULL checks are always branch-if-eq. If we see a IfTrue projection
// then we are replacing a 'ne' test with a 'eq' NULL check test.
// We need to flip the projections to keep the same semantics.
if( proj->Opcode() == Op_IfTrue ) {
// Swap order of projections in basic block to swap branch targets
Node *tmp1 = _nodes[end_idx()+1];
Node *tmp2 = _nodes[end_idx()+2];
_nodes.map(end_idx()+1, tmp2);
_nodes.map(end_idx()+2, tmp1);
Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
tmp1->replace_by(tmp);
tmp2->replace_by(tmp1);
tmp->replace_by(tmp2);
tmp->destruct();
}
// Remove the existing null check; use a new implicit null check instead.
// Since schedule-local needs precise def-use info, we need to correct
// it as well.
Node *old_tst = proj->in(0);
MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
_nodes.map(end_idx(),nul_chk);
cfg->_bbs.map(nul_chk->_idx,this);
// Redirect users of old_test to nul_chk
for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
old_tst->last_out(i2)->set_req(0, nul_chk);
// Clean-up any dead code
for (uint i3 = 0; i3 < old_tst->req(); i3++)
old_tst->set_req(i3, NULL);
cfg->latency_from_uses(nul_chk);
cfg->latency_from_uses(best);
}
//------------------------------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]->is_MachProj() ||
!_nodes[beg-1]->in(0)->is_MachCall() ) {
beg--;
assert(beg > 0,"Catch cleanup walking beyond block boundary");
}
// Range of inserted instructions is [beg, end)
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-- ) {
// It is safe here to clone a node with anti_dependence
// since clones dominate on each path.
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.
Unique_Node_List *out = new Unique_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 ) {
Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
use->set_req(k, fixup);
}
} else {
if (this == buse) {
catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
} else {
catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
}
}
} // 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];
uint new_cnt = end - beg;
// Remove any newly created, but dead, nodes.
for( uint j = new_cnt; j > 0; j-- ) {
Node *n = sb->_nodes[j];
if (n->outcnt() == 0 &&
(!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
n->disconnect_inputs(NULL);
sb->_nodes.remove(j);
new_cnt--;
}
}
// If any newly created nodes remain, move the CreateEx node to the top
if (new_cnt > 0) {
Node *cex = sb->_nodes[1+new_cnt];
if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
sb->_nodes.remove(1+new_cnt);
sb->_nodes.insert(1,cex);
}
}
}
}
