6912521: System.arraycopy works slower than the simple loop for little lengths
Summary: convert small array copies to series of loads and stores
Reviewed-by: kvn, vlivanov
/*
* Copyright (c) 2015, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "opto/arraycopynode.hpp"
#include "opto/graphKit.hpp"
ArrayCopyNode::ArrayCopyNode(Compile* C, bool alloc_tightly_coupled)
: CallNode(arraycopy_type(), NULL, TypeRawPtr::BOTTOM),
_alloc_tightly_coupled(alloc_tightly_coupled),
_kind(None),
_arguments_validated(false) {
init_class_id(Class_ArrayCopy);
init_flags(Flag_is_macro);
C->add_macro_node(this);
}
uint ArrayCopyNode::size_of() const { return sizeof(*this); }
ArrayCopyNode* ArrayCopyNode::make(GraphKit* kit, bool may_throw,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* length,
bool alloc_tightly_coupled,
Node* src_klass, Node* dest_klass,
Node* src_length, Node* dest_length) {
ArrayCopyNode* ac = new ArrayCopyNode(kit->C, alloc_tightly_coupled);
Node* prev_mem = kit->set_predefined_input_for_runtime_call(ac);
ac->init_req(ArrayCopyNode::Src, src);
ac->init_req(ArrayCopyNode::SrcPos, src_offset);
ac->init_req(ArrayCopyNode::Dest, dest);
ac->init_req(ArrayCopyNode::DestPos, dest_offset);
ac->init_req(ArrayCopyNode::Length, length);
ac->init_req(ArrayCopyNode::SrcLen, src_length);
ac->init_req(ArrayCopyNode::DestLen, dest_length);
ac->init_req(ArrayCopyNode::SrcKlass, src_klass);
ac->init_req(ArrayCopyNode::DestKlass, dest_klass);
if (may_throw) {
ac->set_req(TypeFunc::I_O , kit->i_o());
kit->add_safepoint_edges(ac, false);
}
return ac;
}
void ArrayCopyNode::connect_outputs(GraphKit* kit) {
kit->set_all_memory_call(this, true);
kit->set_control(kit->gvn().transform(new ProjNode(this,TypeFunc::Control)));
kit->set_i_o(kit->gvn().transform(new ProjNode(this, TypeFunc::I_O)));
kit->make_slow_call_ex(this, kit->env()->Throwable_klass(), true);
kit->set_all_memory_call(this);
}
#ifndef PRODUCT
const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
void ArrayCopyNode::dump_spec(outputStream *st) const {
CallNode::dump_spec(st);
st->print(" (%s%s)", _kind_names[_kind], _alloc_tightly_coupled ? ", tightly coupled allocation" : "");
}
#endif
intptr_t ArrayCopyNode::get_length_if_constant(PhaseGVN *phase) const {
// check that length is constant
Node* length = in(ArrayCopyNode::Length);
const Type* length_type = phase->type(length);
if (length_type == Type::TOP) {
return -1;
}
assert(is_clonebasic() || is_arraycopy() || is_copyof() || is_copyofrange(), "unexpected array copy type");
return is_clonebasic() ? length->find_intptr_t_con(-1) : length->find_int_con(-1);
}
int ArrayCopyNode::get_count(PhaseGVN *phase) const {
Node* src = in(ArrayCopyNode::Src);
const Type* src_type = phase->type(src);
if (is_clonebasic()) {
if (src_type->isa_instptr()) {
const TypeInstPtr* inst_src = src_type->is_instptr();
ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
// ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
// fields into account. They are rare anyway so easier to simply
// skip instances with injected fields.
if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
return -1;
}
int nb_fields = ik->nof_nonstatic_fields();
return nb_fields;
} else {
const TypeAryPtr* ary_src = src_type->isa_aryptr();
assert (ary_src != NULL, "not an array or instance?");
// clone passes a length as a rounded number of longs. If we're
// cloning an array we'll do it element by element. If the
// length input to ArrayCopyNode is constant, length of input
// array must be too.
assert((get_length_if_constant(phase) == -1) == !ary_src->size()->is_con(), "inconsistent");
if (ary_src->size()->is_con()) {
return ary_src->size()->get_con();
}
return -1;
}
}
return get_length_if_constant(phase);
}
Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
if (!is_clonebasic()) {
return NULL;
}
Node* src = in(ArrayCopyNode::Src);
Node* dest = in(ArrayCopyNode::Dest);
Node* ctl = in(TypeFunc::Control);
Node* in_mem = in(TypeFunc::Memory);
const Type* src_type = phase->type(src);
const Type* dest_type = phase->type(dest);
assert(src->is_AddP(), "should be base + off");
assert(dest->is_AddP(), "should be base + off");
Node* base_src = src->in(AddPNode::Base);
Node* base_dest = dest->in(AddPNode::Base);
MergeMemNode* mem = MergeMemNode::make(in_mem);
const TypeInstPtr* inst_src = src_type->isa_instptr();
if (inst_src == NULL) {
return NULL;
}
if (!inst_src->klass_is_exact()) {
ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
assert(!ik->is_interface() && !ik->has_subklass(), "inconsistent klass hierarchy");
phase->C->dependencies()->assert_leaf_type(ik);
}
ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
assert(ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem, "too many fields");
for (int i = 0; i < count; i++) {
ciField* field = ik->nonstatic_field_at(i);
int fieldidx = phase->C->alias_type(field)->index();
const TypePtr* adr_type = phase->C->alias_type(field)->adr_type();
Node* off = phase->MakeConX(field->offset());
Node* next_src = phase->transform(new AddPNode(base_src,base_src,off));
Node* next_dest = phase->transform(new AddPNode(base_dest,base_dest,off));
BasicType bt = field->layout_type();
const Type *type;
if (bt == T_OBJECT) {
if (!field->type()->is_loaded()) {
type = TypeInstPtr::BOTTOM;
} else {
ciType* field_klass = field->type();
type = TypeOopPtr::make_from_klass(field_klass->as_klass());
}
} else {
type = Type::get_const_basic_type(bt);
}
Node* v = LoadNode::make(*phase, ctl, mem->memory_at(fieldidx), next_src, adr_type, type, bt, MemNode::unordered);
v = phase->transform(v);
Node* s = StoreNode::make(*phase, ctl, mem->memory_at(fieldidx), next_dest, adr_type, v, bt, MemNode::unordered);
s = phase->transform(s);
mem->set_memory_at(fieldidx, s);
}
if (!finish_transform(phase, can_reshape, ctl, mem)) {
return NULL;
}
return mem;
}
bool ArrayCopyNode::prepare_array_copy(PhaseGVN *phase, bool can_reshape,
Node*& adr_src,
Node*& base_src,
Node*& adr_dest,
Node*& base_dest,
BasicType& copy_type,
const Type*& value_type,
bool& disjoint_bases) {
Node* src = in(ArrayCopyNode::Src);
Node* dest = in(ArrayCopyNode::Dest);
const Type* src_type = phase->type(src);
const TypeAryPtr* ary_src = src_type->isa_aryptr();
if (is_arraycopy() || is_copyofrange() || is_copyof()) {
const Type* dest_type = phase->type(dest);
const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
Node* src_offset = in(ArrayCopyNode::SrcPos);
Node* dest_offset = in(ArrayCopyNode::DestPos);
// newly allocated object is guaranteed to not overlap with source object
disjoint_bases = is_alloc_tightly_coupled();
if (ary_src == NULL || ary_src->klass() == NULL ||
ary_dest == NULL || ary_dest->klass() == NULL) {
// We don't know if arguments are arrays
return false;
}
BasicType src_elem = ary_src->klass()->as_array_klass()->element_type()->basic_type();
BasicType dest_elem = ary_dest->klass()->as_array_klass()->element_type()->basic_type();
if (src_elem == T_ARRAY) src_elem = T_OBJECT;
if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
if (src_elem != dest_elem || dest_elem == T_VOID) {
// We don't know if arguments are arrays of the same type
return false;
}
if (dest_elem == T_OBJECT && (!is_alloc_tightly_coupled() || !GraphKit::use_ReduceInitialCardMarks())) {
// It's an object array copy but we can't emit the card marking
// that is needed
return false;
}
value_type = ary_src->elem();
base_src = src;
base_dest = dest;
uint shift = exact_log2(type2aelembytes(dest_elem));
uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
adr_src = src;
adr_dest = dest;
src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
adr_src = phase->transform(new AddPNode(base_src, adr_src, src_scale));
adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, dest_scale));
adr_src = new AddPNode(base_src, adr_src, phase->MakeConX(header));
adr_dest = new AddPNode(base_dest, adr_dest, phase->MakeConX(header));
adr_src = phase->transform(adr_src);
adr_dest = phase->transform(adr_dest);
copy_type = dest_elem;
} else {
assert (is_clonebasic(), "should be");
disjoint_bases = true;
assert(src->is_AddP(), "should be base + off");
assert(dest->is_AddP(), "should be base + off");
adr_src = src;
base_src = src->in(AddPNode::Base);
adr_dest = dest;
base_dest = dest->in(AddPNode::Base);
assert(phase->type(src->in(AddPNode::Offset))->is_intptr_t()->get_con() == phase->type(dest->in(AddPNode::Offset))->is_intptr_t()->get_con(), "same start offset?");
BasicType elem = ary_src->klass()->as_array_klass()->element_type()->basic_type();
if (elem == T_ARRAY) elem = T_OBJECT;
int diff = arrayOopDesc::base_offset_in_bytes(elem) - phase->type(src->in(AddPNode::Offset))->is_intptr_t()->get_con();
assert(diff >= 0, "clone should not start after 1st array element");
if (diff > 0) {
adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(diff)));
adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(diff)));
}
copy_type = elem;
value_type = ary_src->elem();
}
return true;
}
const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN *phase, Node* n) {
const Type* at = phase->type(n);
assert(at != Type::TOP, "unexpected type");
const TypePtr* atp = at->isa_ptr();
// adjust atp to be the correct array element address type
atp = atp->add_offset(Type::OffsetBot);
return atp;
}
void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
Node* ctl = in(TypeFunc::Control);
if (!disjoint_bases && count > 1) {
Node* src_offset = in(ArrayCopyNode::SrcPos);
Node* dest_offset = in(ArrayCopyNode::DestPos);
assert(src_offset != NULL && dest_offset != NULL, "should be");
Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
phase->transform(iff);
forward_ctl = phase->transform(new IfFalseNode(iff));
backward_ctl = phase->transform(new IfTrueNode(iff));
} else {
forward_ctl = ctl;
}
}
Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
bool can_reshape,
Node* forward_ctl,
Node* start_mem_src,
Node* start_mem_dest,
const TypePtr* atp_src,
const TypePtr* atp_dest,
Node* adr_src,
Node* base_src,
Node* adr_dest,
Node* base_dest,
BasicType copy_type,
const Type* value_type,
int count) {
Node* mem = phase->C->top();
if (!forward_ctl->is_top()) {
// copy forward
mem = start_mem_dest;
if (count > 0) {
Node* v = LoadNode::make(*phase, forward_ctl, start_mem_src, adr_src, atp_src, value_type, copy_type, MemNode::unordered);
v = phase->transform(v);
mem = StoreNode::make(*phase, forward_ctl, mem, adr_dest, atp_dest, v, copy_type, MemNode::unordered);
mem = phase->transform(mem);
for (int i = 1; i < count; i++) {
Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
v = LoadNode::make(*phase, forward_ctl, mem, next_src, atp_src, value_type, copy_type, MemNode::unordered);
v = phase->transform(v);
mem = StoreNode::make(*phase, forward_ctl,mem,next_dest,atp_dest,v, copy_type, MemNode::unordered);
mem = phase->transform(mem);
}
} else if(can_reshape) {
PhaseIterGVN* igvn = phase->is_IterGVN();
igvn->_worklist.push(adr_src);
igvn->_worklist.push(adr_dest);
}
}
return mem;
}
Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
bool can_reshape,
Node* backward_ctl,
Node* start_mem_src,
Node* start_mem_dest,
const TypePtr* atp_src,
const TypePtr* atp_dest,
Node* adr_src,
Node* base_src,
Node* adr_dest,
Node* base_dest,
BasicType copy_type,
const Type* value_type,
int count) {
Node* mem = phase->C->top();
if (!backward_ctl->is_top()) {
// copy backward
mem = start_mem_dest;
if (count > 0) {
for (int i = count-1; i >= 1; i--) {
Node* off = phase->MakeConX(type2aelembytes(copy_type) * i);
Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
Node* v = LoadNode::make(*phase, backward_ctl, mem, next_src, atp_src, value_type, copy_type, MemNode::unordered);
v = phase->transform(v);
mem = StoreNode::make(*phase, backward_ctl,mem,next_dest,atp_dest,v, copy_type, MemNode::unordered);
mem = phase->transform(mem);
}
Node* v = LoadNode::make(*phase, backward_ctl, mem, adr_src, atp_src, value_type, copy_type, MemNode::unordered);
v = phase->transform(v);
mem = StoreNode::make(*phase, backward_ctl, mem, adr_dest, atp_dest, v, copy_type, MemNode::unordered);
mem = phase->transform(mem);
} else if(can_reshape) {
PhaseIterGVN* igvn = phase->is_IterGVN();
igvn->_worklist.push(adr_src);
igvn->_worklist.push(adr_dest);
}
}
return mem;
}
bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
Node* ctl, Node *mem) {
if (can_reshape) {
PhaseIterGVN* igvn = phase->is_IterGVN();
igvn->set_delay_transform(false);
if (is_clonebasic()) {
Node* out_mem = proj_out(TypeFunc::Memory);
if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
return false;
}
igvn->replace_node(out_mem->raw_out(0), mem);
Node* out_ctl = proj_out(TypeFunc::Control);
igvn->replace_node(out_ctl, ctl);
} else {
// replace fallthrough projections of the ArrayCopyNode by the
// new memory, control and the input IO.
CallProjections callprojs;
extract_projections(&callprojs, true);
igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
igvn->replace_node(callprojs.fallthrough_memproj, mem);
igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
// The ArrayCopyNode is not disconnected. It still has the
// projections for the exception case. Replace current
// ArrayCopyNode with a dummy new one with a top() control so
// that this part of the graph stays consistent but is
// eventually removed.
set_req(0, phase->C->top());
remove_dead_region(phase, can_reshape);
}
} else {
if (in(TypeFunc::Control) != ctl) {
// we can't return new memory and control from Ideal at parse time
assert(!is_clonebasic(), "added control for clone?");
return NULL;
}
}
return true;
}
Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (remove_dead_region(phase, can_reshape)) return this;
if (StressArrayCopyMacroNode && !can_reshape) {
phase->record_for_igvn(this);
return NULL;
}
// See if it's a small array copy and we can inline it as
// loads/stores
// Here we can only do:
// - arraycopy if all arguments were validated before and we don't
// need card marking
// - clone for which we don't need to do card marking
if (!is_clonebasic() && !is_arraycopy_validated() &&
!is_copyofrange_validated() && !is_copyof_validated()) {
return NULL;
}
if (in(TypeFunc::Control)->is_top() || in(TypeFunc::Memory)->is_top()) {
return NULL;
}
int count = get_count(phase);
if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
return NULL;
}
Node* mem = try_clone_instance(phase, can_reshape, count);
if (mem != NULL) {
return mem;
}
Node* adr_src = NULL;
Node* base_src = NULL;
Node* adr_dest = NULL;
Node* base_dest = NULL;
BasicType copy_type = T_ILLEGAL;
const Type* value_type = NULL;
bool disjoint_bases = false;
if (!prepare_array_copy(phase, can_reshape,
adr_src, base_src, adr_dest, base_dest,
copy_type, value_type, disjoint_bases)) {
return NULL;
}
Node* src = in(ArrayCopyNode::Src);
Node* dest = in(ArrayCopyNode::Dest);
const TypePtr* atp_src = get_address_type(phase, src);
const TypePtr* atp_dest = get_address_type(phase, dest);
uint alias_idx_src = phase->C->get_alias_index(atp_src);
uint alias_idx_dest = phase->C->get_alias_index(atp_dest);
Node *in_mem = in(TypeFunc::Memory);
Node *start_mem_src = in_mem;
Node *start_mem_dest = in_mem;
if (in_mem->is_MergeMem()) {
start_mem_src = in_mem->as_MergeMem()->memory_at(alias_idx_src);
start_mem_dest = in_mem->as_MergeMem()->memory_at(alias_idx_dest);
}
if (can_reshape) {
assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
phase->is_IterGVN()->set_delay_transform(true);
}
Node* backward_ctl = phase->C->top();
Node* forward_ctl = phase->C->top();
array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
start_mem_src, start_mem_dest,
atp_src, atp_dest,
adr_src, base_src, adr_dest, base_dest,
copy_type, value_type, count);
Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
start_mem_src, start_mem_dest,
atp_src, atp_dest,
adr_src, base_src, adr_dest, base_dest,
copy_type, value_type, count);
Node* ctl = NULL;
if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
ctl = new RegionNode(3);
mem = new PhiNode(ctl, Type::MEMORY, atp_dest);
ctl->init_req(1, forward_ctl);
mem->init_req(1, forward_mem);
ctl->init_req(2, backward_ctl);
mem->init_req(2, backward_mem);
ctl = phase->transform(ctl);
mem = phase->transform(mem);
} else if (!forward_ctl->is_top()) {
ctl = forward_ctl;
mem = forward_mem;
} else {
assert(!backward_ctl->is_top(), "no copy?");
ctl = backward_ctl;
mem = backward_mem;
}
if (can_reshape) {
assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
phase->is_IterGVN()->set_delay_transform(false);
}
MergeMemNode* out_mem = MergeMemNode::make(in_mem);
out_mem->set_memory_at(alias_idx_dest, mem);
mem = out_mem;
if (!finish_transform(phase, can_reshape, ctl, mem)) {
return NULL;
}
return mem;
}