//------------------------------sched_call-------------------------------------
uint Block::sched_call( Matcher &m, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
RegMask regs;
// Schedule all the users of the call right now. All the users are
// projection Nodes, so they must be scheduled next to the call.
// Collect all the defined registers.
for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
Node* n = mcall->fast_out(i);
assert( n->Opcode()==Op_MachProj, "" );
--ready_cnt[n->_idx];
assert( !ready_cnt[n->_idx], "" );
// Schedule next to call
_nodes.map(node_cnt++, n);
// Collect defined registers
regs.OR(n->out_RegMask());
// Check for scheduling the next control-definer
if( n->bottom_type() == Type::CONTROL )
// Warm up next pile of heuristic bits
needed_for_next_call(n, next_call, bbs);
// Children of projections are now all ready
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* m = n->fast_out(j); // Get user
if( bbs[m->_idx] != this ) continue;
if( m->is_Phi() ) continue;
if( !--ready_cnt[m->_idx] )
worklist.push(m);
}
}
// Act as if the call defines the Frame Pointer.
// Certainly the FP is alive and well after the call.
regs.Insert(m.c_frame_pointer());
// Set all registers killed and not already defined by the call.
uint r_cnt = mcall->tf()->range()->cnt();
int op = mcall->ideal_Opcode();
MachProjNode *proj = new (1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
bbs.map(proj->_idx,this);
_nodes.insert(node_cnt++, proj);
for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
if( !regs.Member(r) ) { // Not already defined by the call
// Save-on-call register?
if( (m._register_save_policy[r] == 'C') ||
(m._register_save_policy[r] == 'A') ||
((m._register_save_policy[r] == 'E') &&
(op == Op_CallRuntime ||
op == Op_CallNative ||
op == Op_CallInterpreter ||
op == Op_CallLeaf)) ) {
proj->_rout.Insert(r);
}
}
}
return node_cnt;
}
//------------------------------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);
}
}
}
//------------------------------catch_cleanup_one_use--------------------------
static void catch_cleanup_one_use( Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int beg, int n_clone_idx, int use_idx ) {
if( !use_blk ) return; // Can happen if the use is a precedence edge
// Check out 'use'. If it is in this block, then it must be in
// the cloned area. Go to the clone(s) and set them to use the
// cloned version of 'n'.
if( use_blk == def_blk ) {
uint use_idx = def_blk->find_node(use);
uint offset_idx = use_idx - beg;
for( uint k = 0; k < def_blk->_num_succs; k++ ) {
// Get clone in each successor block
Block *sb = def_blk->_succs[k];
Node *clone = sb->_nodes[offset_idx+1];
assert( clone->Opcode() == use->Opcode(), "" );
// Make use-clone use the def-clone
for( uint l = 0; l < use->len(); l++ ) {
if( clone->in(l) == def ) {
if( l < use->req() ) {
clone->set_req(l,sb->_nodes[n_clone_idx]);
} else {
clone->rm_prec(l);
clone->add_prec(sb->_nodes[n_clone_idx]);
l--;
}
}
}
}
}
// Else the use is some block below the Catch. Find the path the value
// takes to reach the Catch and make the use occur only on this path.
else {
// Find which successor block dominates this use. The successor
// blocks must all be single-entry (from the Catch only; I will have
// split blocks to make this so), hence they all dominate.
while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
use_blk = use_blk->_idom;
// Find the successor
Node *fixup;
uint j;
for( j = 0; j < def_blk->_num_succs; j++ )
if( use_blk == def_blk->_succs[j] )
break;
if( j == def_blk->_num_succs ) {
// Block at same level in dom-tree is not a successor. It needs a
// PhiNode, the PhiNode uses from the def and IT's uses need fixup.
Node *phi = PhiNode::make(use_blk->head(), def);
use_blk->_nodes.insert( 1, phi );
bbs.map(phi->_idx,use_blk);
for(uint j3 = 1; j3 < use_blk->num_preds(); j3++ )
catch_cleanup_one_use( phi, bbs[use_blk->pred(j3)->_idx], def, def_blk, bbs, beg, n_clone_idx, j3 );
fixup = phi;
} else {
// Found the use just below the Catch. Make it use the clone.
fixup = def_blk->_succs[j]->_nodes[n_clone_idx];
}
if( use_idx >= 0 ) {
use->set_req(use_idx,fixup);
} else {
for( uint l = 0; l < use->len(); l++ ) {
if( use->in(l) == def ) {
if( l < use->req() ) {
use->set_req(l,fixup);
} else {
use->rm_prec(l);
use->add_prec(fixup);
l--;
}
}
}
}
}
}
//------------------------------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
}
//------------------------------sched_call-------------------------------------
uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
RegMask regs;
// Schedule all the users of the call right now. All the users are
// projection Nodes, so they must be scheduled next to the call.
// Collect all the defined registers.
for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
Node* n = mcall->fast_out(i);
assert( n->is_MachProj(), "" );
int n_cnt = ready_cnt.at(n->_idx)-1;
ready_cnt.at_put(n->_idx, n_cnt);
assert( n_cnt == 0, "" );
// Schedule next to call
_nodes.map(node_cnt++, n);
// Collect defined registers
regs.OR(n->out_RegMask());
// Check for scheduling the next control-definer
if( n->bottom_type() == Type::CONTROL )
// Warm up next pile of heuristic bits
needed_for_next_call(n, next_call, bbs);
// Children of projections are now all ready
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* m = n->fast_out(j); // Get user
if( bbs[m->_idx] != this ) continue;
if( m->is_Phi() ) 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);
}
}
// Act as if the call defines the Frame Pointer.
// Certainly the FP is alive and well after the call.
regs.Insert(matcher.c_frame_pointer());
// Set all registers killed and not already defined by the call.
uint r_cnt = mcall->tf()->range()->cnt();
int op = mcall->ideal_Opcode();
MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
bbs.map(proj->_idx,this);
_nodes.insert(node_cnt++, proj);
// Select the right register save policy.
const char * save_policy;
switch (op) {
case Op_CallRuntime:
case Op_CallLeaf:
case Op_CallLeafNoFP:
// Calling C code so use C calling convention
save_policy = matcher._c_reg_save_policy;
break;
case Op_CallStaticJava:
case Op_CallDynamicJava:
// Calling Java code so use Java calling convention
save_policy = matcher._register_save_policy;
break;
default:
ShouldNotReachHere();
}
// When using CallRuntime mark SOE registers as killed by the call
// so values that could show up in the RegisterMap aren't live in a
// callee saved register since the register wouldn't know where to
// find them. CallLeaf and CallLeafNoFP are ok because they can't
// have debug info on them. Strictly speaking this only needs to be
// done for oops since idealreg2debugmask takes care of debug info
// references but there no way to handle oops differently than other
// pointers as far as the kill mask goes.
bool exclude_soe = op == Op_CallRuntime;
// If the call is a MethodHandle invoke, we need to exclude the
// register which is used to save the SP value over MH invokes from
// the mask. Otherwise this register could be used for
// deoptimization information.
if (op == Op_CallStaticJava) {
MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
if (mcallstaticjava->_method_handle_invoke)
proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
}
add_call_kills(proj, regs, save_policy, exclude_soe);
return node_cnt;
}
//------------------------------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);
}
}
}
}