/* for purely static "profiles" for selecting static superinstructions */
Inst *vm_disassemble_inst(Inst *ip, Inst vm_prim[])
{
#define return do {fputs(" \n 0\t",vm_out); goto _endif2_; } while (0)
#include <java-profile.i>
#undef return
/* else */
{
add_inst(b,"unknown");
ip++;
}
_endif_:
fputc(' ',vm_out);
_endif2_:
return ip;
}
//------------------------------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
}
//------------------------------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);
}
