// These two phis are pre-filled with copies of of the fast IO and Memory
PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
result_rgn->init_req(slow_result_path, control());
result_io ->init_req(slow_result_path, i_o());
result_mem->init_req(slow_result_path, reset_memory());
result_val->init_req(slow_result_path, slow_val);
set_all_memory(_gvn.transform(result_mem));
set_i_o( _gvn.transform(result_io));
}
C->set_has_split_ifs(true); // Has chance for split-if optimization
set_result(result_rgn, result_val);
return true;
}
//---------------------------load_mirror_from_klass----------------------------
// Given a klass oop, load its java mirror (a java.lang.Class oop).
Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
}
//-----------------------load_klass_from_mirror_common-------------------------
// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
// Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
// and branch to the given path on the region.
// If never_see_null, take an uncommon trap on null, so we can optimistically
// compile for the non-null case.
// If the region is NULL, force never_see_null = true.
Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
bool never_see_null,
RegionNode* region,
int null_path,
int offset) {
if (region == NULL) never_see_null = true;
Node* p = basic_plus_adr(mirror, offset);
const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
Node* null_ctl = top();
kls = null_check_oop(kls, &null_ctl, never_see_null);
if (region != NULL) {
// Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
region->init_req(null_path, null_ctl);
} else {
assert(null_ctl == top(), "no loose ends");
}
return kls;
}
//--------------------(inline_native_Class_query helpers)---------------------
// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
// Fall through if (mods & mask) == bits, take the guard otherwise.
Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
// Branch around if the given klass has the given modifier bit set.
// Like generate_guard, adds a new path onto the region.
Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
Node* mask = intcon(modifier_mask);
Node* bits = intcon(modifier_bits);
Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
return generate_fair_guard(bol, region);
}
Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
}
//-------------------------inline_native_Class_query-------------------
bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
const Type* return_type = TypeInt::BOOL;
Node* prim_return_value = top(); // what happens if it's a primitive class?
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
bool expect_prim = false; // most of these guys expect to work on refs
enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
Node* mirror = argument(0);
Node* obj = top();
switch (id) {
case vmIntrinsics::_isInstance:
// nothing is an instance of a primitive type
prim_return_value = intcon(0);
obj = argument(1);
break;
case vmIntrinsics::_getModifiers:
prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
break;
case vmIntrinsics::_isInterface:
prim_return_value = intcon(0);
break;
case vmIntrinsics::_isArray:
prim_return_value = intcon(0);
expect_prim = true; // cf. ObjectStreamClass.getClassSignature
break;
case vmIntrinsics::_isPrimitive:
prim_return_value = intcon(1);
expect_prim = true; // obviously
break;
case vmIntrinsics::_getSuperclass:
prim_return_value = null();
return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
break;
case vmIntrinsics::_getComponentType:
prim_return_value = null();
return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
break;
case vmIntrinsics::_getClassAccessFlags:
prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
return_type = TypeInt::INT; // not bool! 6297094
break;
default:
fatal_unexpected_iid(id);
break;
}
const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
if (mirror_con == NULL) return false; // cannot happen?
#ifndef PRODUCT
if (C->print_intrinsics() || C->print_inlining()) {
ciType* k = mirror_con->java_mirror_type();
if (k) {
tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
k->print_name();
tty->cr();
}
}
#endif
// Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
RegionNode* region = new (C) RegionNode(PATH_LIMIT);
record_for_igvn(region);
PhiNode* phi = new (C) PhiNode(region, return_type);
// The mirror will never be null of Reflection.getClassAccessFlags, however
// it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
// if it is. See bug 4774291.
// For Reflection.getClassAccessFlags(), the null check occurs in
// the wrong place; see inline_unsafe_access(), above, for a similar
// situation.
mirror = null_check(mirror);
// If mirror or obj is dead, only null-path is taken.
if (stopped()) return true;
if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
// Now load the mirror's klass metaobject, and null-check it.
// Side-effects region with the control path if the klass is null.
Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
// If kls is null, we have a primitive mirror.
phi->init_req(_prim_path, prim_return_value);
if (stopped()) { set_result(region, phi); return true; }
bool safe_for_replace = (region->in(_prim_path) == top());
Node* p; // handy temp
Node* null_ctl;
// Now that we have the non-null klass, we can perform the real query.
// For constant classes, the query will constant-fold in LoadNode::Value.
Node* query_value = top();
switch (id) {
case vmIntrinsics::_isInstance:
// nothing is an instance of a primitive type
query_value = gen_instanceof(obj, kls, safe_for_replace);
break;
case vmIntrinsics::_getModifiers:
p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
break;
case vmIntrinsics::_isInterface:
// (To verify this code sequence, check the asserts in JVM_IsInterface.)
if (generate_interface_guard(kls, region) != NULL)
// A guard was added. If the guard is taken, it was an interface.
phi->add_req(intcon(1));
// If we fall through, it's a plain class.
query_value = intcon(0);
break;
case vmIntrinsics::_isArray:
// (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
if (generate_array_guard(kls, region) != NULL)
// A guard was added. If the guard is taken, it was an array.
phi->add_req(intcon(1));
// If we fall through, it's a plain class.
query_value = intcon(0);
break;
case vmIntrinsics::_isPrimitive:
query_value = intcon(0); // "normal" path produces false
break;
case vmIntrinsics::_getSuperclass:
// The rules here are somewhat unfortunate, but we can still do better
// with random logic than with a JNI call.
// Interfaces store null or Object as _super, but must report null.
// Arrays store an intermediate super as _super, but must report Object.
// Other types can report the actual _super.
// (To verify this code sequence, check the asserts in JVM_IsInterface.)
if (generate_interface_guard(kls, region) != NULL)
// A guard was added. If the guard is taken, it was an interface.
phi->add_req(null());
if (generate_array_guard(kls, region) != NULL)
// A guard was added. If the guard is taken, it was an array.
phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
// If we fall through, it's a plain class. Get its _super.
p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
null_ctl = top();
kls = null_check_oop(kls, &null_ctl);
if (null_ctl != top()) {
// If the guard is taken, Object.superClass is null (both klass and mirror).
region->add_req(null_ctl);
phi ->add_req(null());
}
if (!stopped()) {
query_value = load_mirror_from_klass(kls);
}
break;
case vmIntrinsics::_getComponentType:
if (generate_array_guard(kls, region) != NULL) {
// Be sure to pin the oop load to the guard edge just created:
Node* is_array_ctrl = region->in(region->req()-1);
Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
phi->add_req(cmo);
}
query_value = null(); // non-array case is null
break;
case vmIntrinsics::_getClassAccessFlags:
p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
break;
default:
fatal_unexpected_iid(id);
break;
}
// Fall-through is the normal case of a query to a real class.
phi->init_req(1, query_value);
region->init_req(1, control());
C->set_has_split_ifs(true); // Has chance for split-if optimization
set_result(region, phi);
return true;
}
//--------------------------inline_native_subtype_check------------------------
// This intrinsic takes the JNI calls out of the heart of
// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
bool LibraryCallKit::inline_native_subtype_check() {
// Pull both arguments off the stack.
Node* args[2]; // two java.lang.Class mirrors: superc, subc
args[0] = argument(0);
args[1] = argument(1);
Node* klasses[2]; // corresponding Klasses: superk, subk
klasses[0] = klasses[1] = top();
enum {
// A full decision tree on {superc is prim, subc is prim}:
_prim_0_path = 1, // {P,N} => false
// {P,P} & superc!=subc => false
_prim_same_path, // {P,P} & superc==subc => true
_prim_1_path, // {N,P} => false
_ref_subtype_path, // {N,N} & subtype check wins => true
_both_ref_path, // {N,N} & subtype check loses => false
PATH_LIMIT
};
RegionNode* region = new (C) RegionNode(PATH_LIMIT);
Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
record_for_igvn(region);
const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
// First null-check both mirrors and load each mirror's klass metaobject.
int which_arg;
for (which_arg = 0; which_arg <= 1; which_arg++) {
Node* arg = args[which_arg];
arg = null_check(arg);
if (stopped()) break;
args[which_arg] = arg;
Node* p = basic_plus_adr(arg, class_klass_offset);
Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
klasses[which_arg] = _gvn.transform(kls);
}
// Having loaded both klasses, test each for null.
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
for (which_arg = 0; which_arg <= 1; which_arg++) {
Node* kls = klasses[which_arg];
Node* null_ctl = top();
kls = null_check_oop(kls, &null_ctl, never_see_null);
int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
region->init_req(prim_path, null_ctl);
if (stopped()) break;
klasses[which_arg] = kls;
}
if (!stopped()) {
// now we have two reference types, in klasses[0..1]
Node* subk = klasses[1]; // the argument to isAssignableFrom
Node* superk = klasses[0]; // the receiver
region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
// now we have a successful reference subtype check
region->set_req(_ref_subtype_path, control());
}
// If both operands are primitive (both klasses null), then
// we must return true when they are identical primitives.
// It is convenient to test this after the first null klass check.
set_control(region->in(_prim_0_path)); // go back to first null check
if (!stopped()) {
// Since superc is primitive, make a guard for the superc==subc case.
Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
generate_guard(bol_eq, region, PROB_FAIR);
if (region->req() == PATH_LIMIT+1) {
// A guard was added. If the added guard is taken, superc==subc.
region->swap_edges(PATH_LIMIT, _prim_same_path);
region->del_req(PATH_LIMIT);
}
region->set_req(_prim_0_path, control()); // Not equal after all.
}
// these are the only paths that produce 'true':
phi->set_req(_prim_same_path, intcon(1));
phi->set_req(_ref_subtype_path, intcon(1));
// pull together the cases:
assert(region->req() == PATH_LIMIT, "sane region");
for (uint i = 1; i < region->req(); i++) {
Node* ctl = region->in(i);
if (ctl == NULL || ctl == top()) {
region->set_req(i, top());
phi ->set_req(i, top());
} else if (phi->in(i) == NULL) {
phi->set_req(i, intcon(0)); // all other paths produce 'false'
}
}
set_control(_gvn.transform(region));
set_result(_gvn.transform(phi));
return true;
}
//---------------------generate_array_guard_common------------------------
Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
bool obj_array, bool not_array) {
// If obj_array/non_array==false/false:
// Branch around if the given klass is in fact an array (either obj or prim).
// If obj_array/non_array==false/true:
// Branch around if the given klass is not an array klass of any kind.
// If obj_array/non_array==true/true:
// Branch around if the kls is not an oop array (kls is int[], String, etc.)
// If obj_array/non_array==true/false:
// Branch around if the kls is an oop array (Object[] or subtype)
//
// Like generate_guard, adds a new path onto the region.
jint layout_con = 0;
Node* layout_val = get_layout_helper(kls, layout_con);
if (layout_val == NULL) {
bool query = (obj_array
? Klass::layout_helper_is_objArray(layout_con)
: Klass::layout_helper_is_array(layout_con));
if (query == not_array) {
return NULL; // never a branch
} else { // always a branch
Node* always_branch = control();
if (region != NULL)
region->add_req(always_branch);
set_control(top());
return always_branch;
}
}
// Now test the correct condition.
jint nval = (obj_array
? (jint)(Klass::_lh_array_tag_type_value
<< Klass::_lh_array_tag_shift)
: Klass::_lh_neutral_value);
Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
// invert the test if we are looking for a non-array
if (not_array) btest = BoolTest(btest).negate();
Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
return generate_fair_guard(bol, region);
}
//-----------------------inline_native_newArray--------------------------
// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
bool LibraryCallKit::inline_native_newArray() {
Node* mirror = argument(0);
Node* count_val = argument(1);
mirror = null_check(mirror);
// If mirror or obj is dead, only null-path is taken.
if (stopped()) return true;
enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
PhiNode* result_val = new(C) PhiNode(result_reg,
TypeInstPtr::NOTNULL);
PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
TypePtr::BOTTOM);
bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
result_reg, _slow_path);
Node* normal_ctl = control();
Node* no_array_ctl = result_reg->in(_slow_path);
// Generate code for the slow case. We make a call to newArray().
set_control(no_array_ctl);
if (!stopped()) {
// Either the input type is void.class, or else the
// array klass has not yet been cached. Either the
// ensuing call will throw an exception, or else it
// will cache the array klass for next time.
PreserveJVMState pjvms(this);
CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
Node* slow_result = set_results_for_java_call(slow_call);
// this->control() comes from set_results_for_java_call
result_reg->set_req(_slow_path, control());
result_val->set_req(_slow_path, slow_result);
result_io ->set_req(_slow_path, i_o());
result_mem->set_req(_slow_path, reset_memory());
}
set_control(normal_ctl);
if (!stopped()) {
// Normal case: The array type has been cached in the java.lang.Class.
// The following call works fine even if the array type is polymorphic.
// It could be a dynamic mix of int[], boolean[], Object[], etc.
Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
result_reg->init_req(_normal_path, control());
result_val->init_req(_normal_path, obj);
result_io ->init_req(_normal_path, i_o());
result_mem->init_req(_normal_path, reset_memory());
}
// Return the combined state.
set_i_o( _gvn.transform(result_io) );
set_all_memory( _gvn.transform(result_mem));
C->set_has_split_ifs(true); // Has chance for split-if optimization
set_result(result_reg, result_val);
return true;
}
//----------------------inline_native_getLength--------------------------
// public static native int java.lang.reflect.Array.getLength(Object array);
bool LibraryCallKit::inline_native_getLength() {
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
Node* array = null_check(argument(0));
// If array is dead, only null-path is taken.
if (stopped()) return true;
// Deoptimize if it is a non-array.
Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
if (non_array != NULL) {
PreserveJVMState pjvms(this);
set_control(non_array);
uncommon_trap(Deoptimization::Reason_intrinsic,
Deoptimization::Action_maybe_recompile);
}
// If control is dead, only non-array-path is taken.
if (stopped()) return true;
// The works fine even if the array type is polymorphic.
// It could be a dynamic mix of int[], boolean[], Object[], etc.
Node* result = load_array_length(array);
C->set_has_split_ifs(true); // Has chance for split-if optimization
set_result(result);
return true;
}
//------------------------inline_array_copyOf----------------------------
// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
// Get the arguments.
Node* original = argument(0);
Node* start = is_copyOfRange? argument(1): intcon(0);
Node* end = is_copyOfRange? argument(2): argument(1);
Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
Node* newcopy = NULL;
// Set the original stack and the reexecute bit for the interpreter to reexecute
// the bytecode that invokes Arrays.copyOf if deoptimization happens.
{ PreserveReexecuteState preexecs(this);
jvms()->set_should_reexecute(true);
array_type_mirror = null_check(array_type_mirror);
original = null_check(original);
// Check if a null path was taken unconditionally.
if (stopped()) return true;
Node* orig_length = load_array_length(original);
Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
klass_node = null_check(klass_node);
RegionNode* bailout = new (C) RegionNode(1);
record_for_igvn(bailout);
// Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
// Bail out if that is so.
Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
if (not_objArray != NULL) {
// Improve the klass node's type from the new optimistic assumption:
ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
Node* cast = new (C) CastPPNode(klass_node, akls);
cast->init_req(0, control());
klass_node = _gvn.transform(cast);
}
// Bail out if either start or end is negative.
generate_negative_guard(start, bailout, &start);
generate_negative_guard(end, bailout, &end);
Node* length = end;
if (_gvn.type(start) != TypeInt::ZERO) {
length = _gvn.transform(new (C) SubINode(end, start));
}
// Bail out if length is negative.
// Without this the new_array would throw
// NegativeArraySizeException but IllegalArgumentException is what
// should be thrown
generate_negative_guard(length, bailout, &length);
if (bailout->req() > 1) {
PreserveJVMState pjvms(this);
set_control(_gvn.transform(bailout));
uncommon_trap(Deoptimization::Reason_intrinsic,
Deoptimization::Action_maybe_recompile);
}
if (!stopped()) {
// How many elements will we copy from the original?
// The answer is MinI(orig_length - start, length).
Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
newcopy = new_array(klass_node, length, 0); // no argments to push
// Generate a direct call to the right arraycopy function(s).
// We know the copy is disjoint but we might not know if the
// oop stores need checking.
// Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
// This will fail a store-check if x contains any non-nulls.
bool disjoint_bases = true;
// if start > orig_length then the length of the copy may be
// negative.
bool length_never_negative = !is_copyOfRange;
generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
original, start, newcopy, intcon(0), moved,
disjoint_bases, length_never_negative);
}
} // original reexecute is set back here
C->set_has_split_ifs(true); // Has chance for split-if optimization
if (!stopped()) {
set_result(newcopy);
}
return true;
}
//----------------------generate_virtual_guard---------------------------
// Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
RegionNode* slow_region) {
ciMethod* method = callee();
int vtable_index = method->vtable_index();
assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
err_msg_res("bad index %d", vtable_index));
// Get the Method* out of the appropriate vtable entry.
int entry_offset = (InstanceKlass::vtable_start_offset() +
vtable_index*vtableEntry::size()) * wordSize +
vtableEntry::method_offset_in_bytes();
Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
// Compare the target method with the expected method (e.g., Object.hashCode).
const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
Node* native_call = makecon(native_call_addr);
Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
return generate_slow_guard(test_native, slow_region);
}
//-----------------------generate_method_call----------------------------
// Use generate_method_call to make a slow-call to the real
// method if the fast path fails. An alternative would be to
// use a stub like OptoRuntime::slow_arraycopy_Java.
// This only works for expanding the current library call,
// not another intrinsic. (E.g., don't use this for making an
// arraycopy call inside of the copyOf intrinsic.)
CallJavaNode*
LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
// When compiling the intrinsic method itself, do not use this technique.
guarantee(callee() != C->method(), "cannot make slow-call to self");
ciMethod* method = callee();
// ensure the JVMS we have will be correct for this call
guarantee(method_id == method->intrinsic_id(), "must match");
const TypeFunc* tf = TypeFunc::make(method);
CallJavaNode* slow_call;
if (is_static) {
assert(!is_virtual, "");
slow_call = new(C) CallStaticJavaNode(C, tf,
SharedRuntime::get_resolve_static_call_stub(),
method, bci());
} else if (is_virtual) {
null_check_receiver();
int vtable_index = Method::invalid_vtable_index;
if (UseInlineCaches) {
// Suppress the vtable call
} else {
// hashCode and clone are not a miranda methods,
// so the vtable index is fixed.
// No need to use the linkResolver to get it.
vtable_index = method->vtable_index();
assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
err_msg_res("bad index %d", vtable_index));
}
slow_call = new(C) CallDynamicJavaNode(tf,
SharedRuntime::get_resolve_virtual_call_stub(),
method, vtable_index, bci());
} else { // neither virtual nor static: opt_virtual
null_check_receiver();
slow_call = new(C) CallStaticJavaNode(C, tf,
SharedRuntime::get_resolve_opt_virtual_call_stub(),
method, bci());
slow_call->set_optimized_virtual(true);
}
set_arguments_for_java_call(slow_call);
set_edges_for_java_call(slow_call);
return slow_call;
}
/**
* Build special case code for calls to hashCode on an object. This call may
* be virtual (invokevirtual) or bound (invokespecial). For each case we generate
* slightly different code.
*/
bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
assert(is_static == callee()->is_static(), "correct intrinsic selection");
assert(!(is_virtual && is_static), "either virtual, special, or static");
enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
Node* obj = NULL;
if (!is_static) {
// Check for hashing null object
obj = null_check_receiver();
if (stopped()) return true; // unconditionally null
result_reg->init_req(_null_path, top());
result_val->init_req(_null_path, top());
} else {
// Do a null check, and return zero if null.
// System.identityHashCode(null) == 0
obj = argument(0);
Node* null_ctl = top();
obj = null_check_oop(obj, &null_ctl);
result_reg->init_req(_null_path, null_ctl);
result_val->init_req(_null_path, _gvn.intcon(0));
}
// Unconditionally null? Then return right away.
if (stopped()) {
set_control( result_reg->in(_null_path));
if (!stopped())
set_result(result_val->in(_null_path));
return true;
}
// We only go to the fast case code if we pass a number of guards. The
// paths which do not pass are accumulated in the slow_region.
RegionNode* slow_region = new (C) RegionNode(1);
record_for_igvn(slow_region);
// If this is a virtual call, we generate a funny guard. We pull out
// the vtable entry corresponding to hashCode() from the target object.
// If the target method which we are calling happens to be the native
// Object hashCode() method, we pass the guard. We do not need this
// guard for non-virtual calls -- the caller is known to be the native
// Object hashCode().
if (is_virtual) {
// After null check, get the object's klass.
Node* obj_klass = load_object_klass(obj);
generate_virtual_guard(obj_klass, slow_region);
}
// Get the header out of the object, use LoadMarkNode when available
Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
// The control of the load must be NULL. Otherwise, the load can move before
// the null check after castPP removal.
Node* no_ctrl = NULL;
Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
// Test the header to see if it is unlocked.
Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
generate_slow_guard(test_unlocked, slow_region);
// Get the hash value and check to see that it has been properly assigned.
// We depend on hash_mask being at most 32 bits and avoid the use of
// hash_mask_in_place because it could be larger than 32 bits in a 64-bit
// vm: see markOop.hpp.
Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
// This hack lets the hash bits live anywhere in the mark object now, as long
// as the shift drops the relevant bits into the low 32 bits. Note that
// Java spec says that HashCode is an int so there's no point in capturing
// an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
hshifted_header = ConvX2I(hshifted_header);
Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
generate_slow_guard(test_assigned, slow_region);
Node* init_mem = reset_memory();
// fill in the rest of the null path:
result_io ->init_req(_null_path, i_o());
result_mem->init_req(_null_path, init_mem);
result_val->init_req(_fast_path, hash_val);
result_reg->init_req(_fast_path, control());
result_io ->init_req(_fast_path, i_o());
result_mem->init_req(_fast_path, init_mem);
// Generate code for the slow case. We make a call to hashCode().
set_control(_gvn.transform(slow_region));
if (!stopped()) {
// No need for PreserveJVMState, because we're using up the present state.
set_all_memory(init_mem);
vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
Node* slow_result = set_results_for_java_call(slow_call);
// this->control() comes from set_results_for_java_call
result_reg->init_req(_slow_path, control());
result_val->init_req(_slow_path, slow_result);
result_io ->set_req(_slow_path, i_o());
result_mem ->set_req(_slow_path, reset_memory());
}
// Return the combined state.
set_i_o( _gvn.transform(result_io) );
set_all_memory( _gvn.transform(result_mem));
set_result(result_reg, result_val);
return true;
}
//---------------------------inline_native_getClass----------------------------
// public final native Class<?> java.lang.Object.getClass();
//
// Build special case code for calls to getClass on an object.
bool LibraryCallKit::inline_native_getClass() {
Node* obj = null_check_receiver();
if (stopped()) return true;
set_result(load_mirror_from_klass(load_object_klass(obj)));
return true;
}
//-----------------inline_native_Reflection_getCallerClass---------------------
// public static native Class<?> sun.reflect.Reflection.getCallerClass();
//
// In the presence of deep enough inlining, getCallerClass() becomes a no-op.
//
// NOTE: This code must perform the same logic as JVM_GetCallerClass
// in that it must skip particular security frames and checks for
// caller sensitive methods.
bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
#ifndef PRODUCT
if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
}
#endif
if (!jvms()->has_method()) {
#ifndef PRODUCT
if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
tty->print_cr(" Bailing out because intrinsic was inlined at top level");
}
#endif
return false;
}
// Walk back up the JVM state to find the caller at the required
// depth.
JVMState* caller_jvms = jvms();
// Cf. JVM_GetCallerClass
// NOTE: Start the loop at depth 1 because the current JVM state does
// not include the Reflection.getCallerClass() frame.
for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
ciMethod* m = caller_jvms->method();
switch (n) {
case 0:
fatal("current JVM state does not include the Reflection.getCallerClass frame");
break;
case 1:
// Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
if (!m->caller_sensitive()) {
#ifndef PRODUCT
if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
}
#endif
return false; // bail-out; let JVM_GetCallerClass do the work
}
break;
default:
if (!m->is_ignored_by_security_stack_walk()) {
// We have reached the desired frame; return the holder class.
// Acquire method holder as java.lang.Class and push as constant.
ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
ciInstance* caller_mirror = caller_klass->java_mirror();
set_result(makecon(TypeInstPtr::make(caller_mirror)));
#ifndef PRODUCT
if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
tty->print_cr(" JVM state at this point:");
for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
ciMethod* m = jvms()->of_depth(i)->method();
tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
}
}
#endif
return true;
}
break;
}
}
#ifndef PRODUCT
if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
tty->print_cr(" JVM state at this point:");
for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
ciMethod* m = jvms()->of_depth(i)->method();
tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
}
}
#endif
return false; // bail-out; let JVM_GetCallerClass do the work
}
bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
Node* arg = argument(0);
Node* result = NULL;
switch (id) {
case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
case vmIntrinsics::_doubleToLongBits: {
// two paths (plus control) merge in a wood
RegionNode *r = new (C) RegionNode(3);
Node *phi = new (C) PhiNode(r, TypeLong::LONG);
Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
// Build the boolean node
Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
// Branch either way.
// NaN case is less traveled, which makes all the difference.
IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
Node *opt_isnan = _gvn.transform(ifisnan);
assert( opt_isnan->is_If(), "Expect an IfNode");
IfNode *opt_ifisnan = (IfNode*)opt_isnan;
Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
set_control(iftrue);
static const jlong nan_bits = CONST64(0x7ff8000000000000);
Node *slow_result = longcon(nan_bits); // return NaN
phi->init_req(1, _gvn.transform( slow_result ));
r->init_req(1, iftrue);
// Else fall through
Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
set_control(iffalse);
phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
r->init_req(2, iffalse);
// Post merge
set_control(_gvn.transform(r));
record_for_igvn(r);
C->set_has_split_ifs(true); // Has chance for split-if optimization
result = phi;
assert(result->bottom_type()->isa_long(), "must be");
break;
}
case vmIntrinsics::_floatToIntBits: {
// two paths (plus control) merge in a wood
RegionNode *r = new (C) RegionNode(3);
Node *phi = new (C) PhiNode(r, TypeInt::INT);
Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
// Build the boolean node
Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
// Branch either way.
// NaN case is less traveled, which makes all the difference.
IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
Node *opt_isnan = _gvn.transform(ifisnan);
assert( opt_isnan->is_If(), "Expect an IfNode");
IfNode *opt_ifisnan = (IfNode*)opt_isnan;
Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
set_control(iftrue);
static const jint nan_bits = 0x7fc00000;
Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
phi->init_req(1, _gvn.transform( slow_result ));
r->init_req(1, iftrue);
// Else fall through
Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
set_control(iffalse);
phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
r->init_req(2, iffalse);
// Post merge
set_control(_gvn.transform(r));
record_for_igvn(r);
C->set_has_split_ifs(true); // Has chance for split-if optimization
result = phi;
assert(result->bottom_type()->isa_int(), "must be");
break;
}
default:
fatal_unexpected_iid(id);
break;
}
set_result(_gvn.transform(result));
return true;
}
#ifdef _LP64
#define XTOP ,top() /*additional argument*/
#else //_LP64
#define XTOP /*no additional argument*/
#endif //_LP64
//----------------------inline_unsafe_copyMemory-------------------------
// public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
bool LibraryCallKit::inline_unsafe_copyMemory() {
if (callee()->is_static()) return false; // caller must have the capability!
null_check_receiver(); // null-check receiver
if (stopped()) return true;
C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
Node* src_ptr = argument(1); // type: oop
Node* src_off = ConvL2X(argument(2)); // type: long
Node* dst_ptr = argument(4); // type: oop
Node* dst_off = ConvL2X(argument(5)); // type: long
Node* size = ConvL2X(argument(7)); // type: long
assert(Unsafe_field_offset_to_byte_offset(11) == 11,
"fieldOffset must be byte-scaled");
Node* src = make_unsafe_address(src_ptr, src_off);
Node* dst = make_unsafe_address(dst_ptr, dst_off);
// Conservatively insert a memory barrier on all memory slices.
// Do not let writes of the copy source or destination float below the copy.
insert_mem_bar(Op_MemBarCPUOrder);
// Call it. Note that the length argument is not scaled.
make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::fast_arraycopy_Type(),
StubRoutines::unsafe_arraycopy(),
"unsafe_arraycopy",
TypeRawPtr::BOTTOM,
src, dst, size XTOP);
// Do not let reads of the copy destination float above the copy.
insert_mem_bar(Op_MemBarCPUOrder);
return true;
}
//------------------------clone_coping-----------------------------------
// Helper function for inline_native_clone.
void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
assert(obj_size != NULL, "");
Node* raw_obj = alloc_obj->in(1);
assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
AllocateNode* alloc = NULL;
if (ReduceBulkZeroing) {
// We will be completely responsible for initializing this object -
// mark Initialize node as complete.
alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
// The object was just allocated - there should be no any stores!
guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
// Mark as complete_with_arraycopy so that on AllocateNode
// expansion, we know this AllocateNode is initialized by an array
// copy and a StoreStore barrier exists after the array copy.
alloc->initialization()->set_complete_with_arraycopy();
}
// Copy the fastest available way.
// TODO: generate fields copies for small objects instead.
Node* src = obj;
Node* dest = alloc_obj;
Node* size = _gvn.transform(obj_size);
// Exclude the header but include array length to copy by 8 bytes words.
// Can't use base_offset_in_bytes(bt) since basic type is unknown.
int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
instanceOopDesc::base_offset_in_bytes();
// base_off:
// 8 - 32-bit VM
// 12 - 64-bit VM, compressed klass
// 16 - 64-bit VM, normal klass
if (base_off % BytesPerLong != 0) {
assert(UseCompressedClassPointers, "");
if (is_array) {
// Exclude length to copy by 8 bytes words.
base_off += sizeof(int);
} else {
// Include klass to copy by 8 bytes words.
base_off = instanceOopDesc::klass_offset_in_bytes();
}
assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
}
src = basic_plus_adr(src, base_off);
dest = basic_plus_adr(dest, base_off);
// Compute the length also, if needed:
Node* countx = size;
countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
bool disjoint_bases = true;
generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
src, NULL, dest, NULL, countx,
/*dest_uninitialized*/true);
// If necessary, emit some card marks afterwards. (Non-arrays only.)
if (card_mark) {
assert(!is_array, "");
// Put in store barrier for any and all oops we are sticking
// into this object. (We could avoid this if we could prove
// that the object type contains no oop fields at all.)
Node* no_particular_value = NULL;
Node* no_particular_field = NULL;
int raw_adr_idx = Compile::AliasIdxRaw;
post_barrier(control(),
memory(raw_adr_type),
alloc_obj,
no_particular_field,
raw_adr_idx,
no_particular_value,
T_OBJECT,
false);
}
// Do not let reads from the cloned object float above the arraycopy.
if (alloc != NULL) {
// Do not let stores that initialize this object be reordered with
// a subsequent store that would make this object accessible by
// other threads.
// Record what AllocateNode this StoreStore protects so that
// escape analysis can go from the MemBarStoreStoreNode to the
// AllocateNode and eliminate the MemBarStoreStoreNode if possible
// based on the escape status of the AllocateNode.
insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
} else {
insert_mem_bar(Op_MemBarCPUOrder);
}
}
//------------------------inline_native_clone----------------------------
// protected native Object java.lang.Object.clone();
//
// Here are the simple edge cases:
// null receiver => normal trap
// virtual and clone was overridden => slow path to out-of-line clone
// not cloneable or finalizer => slow path to out-of-line Object.clone
//
// The general case has two steps, allocation and copying.
// Allocation has two cases, and uses GraphKit::new_instance or new_array.
//
// Copying also has two cases, oop arrays and everything else.
// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
// Everything else uses the tight inline loop supplied by CopyArrayNode.
//
// These steps fold up nicely if and when the cloned object's klass
// can be sharply typed as an object array, a type array, or an instance.
//
bool LibraryCallKit::inline_native_clone(bool is_virtual) {
PhiNode* result_val;
// Set the reexecute bit for the interpreter to reexecute
// the bytecode that invokes Object.clone if deoptimization happens.
{ PreserveReexecuteState preexecs(this);
jvms()->set_should_reexecute(true);
Node* obj = null_check_receiver();
if (stopped()) return true;
Node* obj_klass = load_object_klass(obj);
const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
const TypeOopPtr* toop = ((tklass != NULL)
? tklass->as_instance_type()
: TypeInstPtr::NOTNULL);
// Conservatively insert a memory barrier on all memory slices.
// Do not let writes into the original float below the clone.
insert_mem_bar(Op_MemBarCPUOrder);
// paths into result_reg:
enum {
_slow_path = 1, // out-of-line call to clone method (virtual or not)
_objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
_array_path, // plain array allocation, plus arrayof_long_arraycopy
_instance_path, // plain instance allocation, plus arrayof_long_arraycopy
PATH_LIMIT
};
RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
result_val = new(C) PhiNode(result_reg,
TypeInstPtr::NOTNULL);
PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
TypePtr::BOTTOM);
record_for_igvn(result_reg);
const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
int raw_adr_idx = Compile::AliasIdxRaw;
Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
if (array_ctl != NULL) {
// It's an array.
PreserveJVMState pjvms(this);
set_control(array_ctl);
Node* obj_length = load_array_length(obj);
Node* obj_size = NULL;
Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
if (!use_ReduceInitialCardMarks()) {
// If it is an oop array, it requires very special treatment,
// because card marking is required on each card of the array.
Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
if (is_obja != NULL) {
PreserveJVMState pjvms2(this);
set_control(is_obja);
// Generate a direct call to the right arraycopy function(s).
bool disjoint_bases = true;
bool length_never_negative = true;
generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
obj, intcon(0), alloc_obj, intcon(0),
obj_length,
disjoint_bases, length_never_negative);
result_reg->init_req(_objArray_path, control());
result_val->init_req(_objArray_path, alloc_obj);
result_i_o ->set_req(_objArray_path, i_o());
result_mem ->set_req(_objArray_path, reset_memory());
}
}
// Otherwise, there are no card marks to worry about.
// (We can dispense with card marks if we know the allocation
// comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
// causes the non-eden paths to take compensating steps to
// simulate a fresh allocation, so that no further
// card marks are required in compiled code to initialize
// the object.)
if (!stopped()) {
copy_to_clone(obj, alloc_obj, obj_size, true, false);
// Present the results of the copy.
result_reg->init_req(_array_path, control());
result_val->init_req(_array_path, alloc_obj);
result_i_o ->set_req(_array_path, i_o());
result_mem ->set_req(_array_path, reset_memory());
}
}
// We only go to the instance fast case code if we pass a number of guards.
// The paths which do not pass are accumulated in the slow_region.
RegionNode* slow_region = new (C) RegionNode(1);
record_for_igvn(slow_region);
if (!stopped()) {
// It's an instance (we did array above). Make the slow-path tests.
// If this is a virtual call, we generate a funny guard. We grab
// the vtable entry corresponding to clone() from the target object.
// If the target method which we are calling happens to be the
// Object clone() method, we pass the guard. We do not need this
// guard for non-virtual calls; the caller is known to be the native
// Object clone().
if (is_virtual) {
generate_virtual_guard(obj_klass, slow_region);
}
// The object must be cloneable and must not have a finalizer.
// Both of these conditions may be checked in a single test.
// We could optimize the cloneable test further, but we don't care.
generate_access_flags_guard(obj_klass,
// Test both conditions:
JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
// Must be cloneable but not finalizer:
JVM_ACC_IS_CLONEABLE,
slow_region);
}
if (!stopped()) {
// It's an instance, and it passed the slow-path tests.
PreserveJVMState pjvms(this);
Node* obj_size = NULL;
// Need to deoptimize on exception from allocation since Object.clone intrinsic
// is reexecuted if deoptimization occurs and there could be problems when merging
// exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
// Present the results of the slow call.
result_reg->init_req(_instance_path, control());
result_val->init_req(_instance_path, alloc_obj);
result_i_o ->set_req(_instance_path, i_o());
result_mem ->set_req(_instance_path, reset_memory());
}
// Generate code for the slow case. We make a call to clone().
set_control(_gvn.transform(slow_region));
if (!stopped()) {
PreserveJVMState pjvms(this);
CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
Node* slow_result = set_results_for_java_call(slow_call);
// this->control() comes from set_results_for_java_call
result_reg->init_req(_slow_path, control());
result_val->init_req(_slow_path, slow_result);
result_i_o ->set_req(_slow_path, i_o());
result_mem ->set_req(_slow_path, reset_memory());
}
// Return the combined state.
set_control( _gvn.transform(result_reg));
set_i_o( _gvn.transform(result_i_o));
set_all_memory( _gvn.transform(result_mem));
} // original reexecute is set back here
set_result(_gvn.transform(result_val));
return true;
}
//------------------------------basictype2arraycopy----------------------------
address LibraryCallKit::basictype2arraycopy(BasicType t,
Node* src_offset,
Node* dest_offset,
bool disjoint_bases,
const char* &name,
bool dest_uninitialized) {
const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
bool aligned = false;
bool disjoint = disjoint_bases;
// if the offsets are the same, we can treat the memory regions as
// disjoint, because either the memory regions are in different arrays,
// or they are identical (which we can treat as disjoint.) We can also
// treat a copy with a destination index less that the source index
// as disjoint since a low->high copy will work correctly in this case.
if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
// both indices are constants
int s_offs = src_offset_inttype->get_con();
int d_offs = dest_offset_inttype->get_con();
int element_size = type2aelembytes(t);
aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
if (s_offs >= d_offs) disjoint = true;
} else if (src_offset == dest_offset && src_offset != NULL) {
// This can occur if the offsets are identical non-constants.
disjoint = true;
}
return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
}
//------------------------------inline_arraycopy-----------------------
// public static native void java.lang.System.arraycopy(Object src, int srcPos,
// Object dest, int destPos,
// int length);
bool LibraryCallKit::inline_arraycopy() {
// Get the arguments.
Node* src = argument(0); // type: oop
Node* src_offset = argument(1); // type: int
Node* dest = argument(2); // type: oop
Node* dest_offset = argument(3); // type: int
Node* length = argument(4); // type: int
// Compile time checks. If any of these checks cannot be verified at compile time,
// we do not make a fast path for this call. Instead, we let the call remain as it
// is. The checks we choose to mandate at compile time are:
//
// (1) src and dest are arrays.
const Type* src_type = src->Value(&_gvn);
const Type* dest_type = dest->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
const TypeAryPtr* top_dest = dest_type->isa_aryptr();
// Do we have the type of src?
bool has_src = (top_src != NULL && top_src->klass() != NULL);
// Do we have the type of dest?
bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
// Is the type for src from speculation?
bool src_spec = false;
// Is the type for dest from speculation?
bool dest_spec = false;
if (!has_src || !has_dest) {
// We don't have sufficient type information, let's see if
// speculative types can help. We need to have types for both src
// and dest so that it pays off.
// Do we already have or could we have type information for src
bool could_have_src = has_src;
// Do we already have or could we have type information for dest
bool could_have_dest = has_dest;
ciKlass* src_k = NULL;
if (!has_src) {
src_k = src_type->speculative_type();
if (src_k != NULL && src_k->is_array_klass()) {
could_have_src = true;
}
}
ciKlass* dest_k = NULL;
if (!has_dest) {
dest_k = dest_type->speculative_type();
if (dest_k != NULL && dest_k->is_array_klass()) {
could_have_dest = true;
}
}
if (could_have_src && could_have_dest) {
// This is going to pay off so emit the required guards
if (!has_src) {
src = maybe_cast_profiled_obj(src, src_k);
src_type = _gvn.type(src);
top_src = src_type->isa_aryptr();
has_src = (top_src != NULL && top_src->klass() != NULL);
src_spec = true;
}
if (!has_dest) {
dest = maybe_cast_profiled_obj(dest, dest_k);
dest_type = _gvn.type(dest);
top_dest = dest_type->isa_aryptr();
has_dest = (top_dest != NULL && top_dest->klass() != NULL);
dest_spec = true;
}
}
}
if (!has_src || !has_dest) {
// Conservatively insert a memory barrier on all memory slices.
// Do not let writes into the source float below the arraycopy.
insert_mem_bar(Op_MemBarCPUOrder);
// Call StubRoutines::generic_arraycopy stub.
generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
src, src_offset, dest, dest_offset, length);
// Do not let reads from the destination float above the arraycopy.
// Since we cannot type the arrays, we don't know which slices
// might be affected. We could restrict this barrier only to those
// memory slices which pertain to array elements--but don't bother.
if (!InsertMemBarAfterArraycopy)
// (If InsertMemBarAfterArraycopy, there is already one in place.)
insert_mem_bar(Op_MemBarCPUOrder);
return true;
}
// (2) src and dest arrays must have elements of the same BasicType
// Figure out the size and type of the elements we will be copying.
BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
BasicType dest_elem = top_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) {
// The component types are not the same or are not recognized. Punt.
// (But, avoid the native method wrapper to JVM_ArrayCopy.)
generate_slow_arraycopy(TypePtr::BOTTOM,
src, src_offset, dest, dest_offset, length,
/*dest_uninitialized*/false);
return true;
}
if (src_elem == T_OBJECT) {
// If both arrays are object arrays then having the exact types
// for both will remove the need for a subtype check at runtime
// before the call and may make it possible to pick a faster copy
// routine (without a subtype check on every element)
// Do we have the exact type of src?
bool could_have_src = src_spec;
// Do we have the exact type of dest?
bool could_have_dest = dest_spec;
ciKlass* src_k = top_src->klass();
ciKlass* dest_k = top_dest->klass();
if (!src_spec) {
src_k = src_type->speculative_type();
if (src_k != NULL && src_k->is_array_klass()) {
could_have_src = true;
}
}
if (!dest_spec) {
dest_k = dest_type->speculative_type();
if (dest_k != NULL && dest_k->is_array_klass()) {
could_have_dest = true;
}
}
if (could_have_src && could_have_dest) {
// If we can have both exact types, emit the missing guards
if (could_have_src && !src_spec) {
src = maybe_cast_profiled_obj(src, src_k);
}
if (could_have_dest && !dest_spec) {
dest = maybe_cast_profiled_obj(dest, dest_k);
}
}
}
//---------------------------------------------------------------------------
// We will make a fast path for this call to arraycopy.
// We have the following tests left to perform:
//
// (3) src and dest must not be null.
// (4) src_offset must not be negative.
// (5) dest_offset must not be negative.
// (6) length must not be negative.
// (7) src_offset + length must not exceed length of src.
// (8) dest_offset + length must not exceed length of dest.
// (9) each element of an oop array must be assignable
RegionNode* slow_region = new (C) RegionNode(1);
record_for_igvn(slow_region);
// (3) operands must not be null
// We currently perform our null checks with the null_check routine.
// This means that the null exceptions will be reported in the caller
// rather than (correctly) reported inside of the native arraycopy call.
// This should be corrected, given time. We do our null check with the
// stack pointer restored.
src = null_check(src, T_ARRAY);
dest = null_check(dest, T_ARRAY);
// (4) src_offset must not be negative.
generate_negative_guard(src_offset, slow_region);
// (5) dest_offset must not be negative.
generate_negative_guard(dest_offset, slow_region);
// (6) length must not be negative (moved to generate_arraycopy()).
// generate_negative_guard(length, slow_region);
// (7) src_offset + length must not exceed length of src.
generate_limit_guard(src_offset, length,
load_array_length(src),
slow_region);
// (8) dest_offset + length must not exceed length of dest.
generate_limit_guard(dest_offset, length,
load_array_length(dest),
slow_region);
// (9) each element of an oop array must be assignable
// The generate_arraycopy subroutine checks this.
// This is where the memory effects are placed:
const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
generate_arraycopy(adr_type, dest_elem,
src, src_offset, dest, dest_offset, length,
false, false, slow_region);
return true;
}
//-----------------------------generate_arraycopy----------------------
// Generate an optimized call to arraycopy.
// Caller must guard against non-arrays.
// Caller must determine a common array basic-type for both arrays.
// Caller must validate offsets against array bounds.
// The slow_region has already collected guard failure paths
// (such as out of bounds length or non-conformable array types).
// The generated code has this shape, in general:
//
// if (length == 0) return // via zero_path
// slowval = -1
// if (types unknown) {
// slowval = call generic copy loop
// if (slowval == 0) return // via checked_path
// } else if (indexes in bounds) {
// if ((is object array) && !(array type check)) {
// slowval = call checked copy loop
// if (slowval == 0) return // via checked_path
// } else {
// call bulk copy loop
// return // via fast_path
// }
// }
// // adjust params for remaining work:
// if (slowval != -1) {
// n = -1^slowval; src_offset += n; dest_offset += n; length -= n
// }
// slow_region:
// call slow arraycopy(src, src_offset, dest, dest_offset, length)
// return // via slow_call_path
//
// This routine is used from several intrinsics: System.arraycopy,
// Object.clone (the array subcase), and Arrays.copyOf[Range].
//
void
LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
BasicType basic_elem_type,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* copy_length,
bool disjoint_bases,
bool length_never_negative,
RegionNode* slow_region) {
if (slow_region == NULL) {
slow_region = new(C) RegionNode(1);
record_for_igvn(slow_region);
}
Node* original_dest = dest;
AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
bool dest_uninitialized = false;
// See if this is the initialization of a newly-allocated array.
// If so, we will take responsibility here for initializing it to zero.
// (Note: Because tightly_coupled_allocation performs checks on the
// out-edges of the dest, we need to avoid making derived pointers
// from it until we have checked its uses.)
if (ReduceBulkZeroing
&& !ZeroTLAB // pointless if already zeroed
&& basic_elem_type != T_CONFLICT // avoid corner case
&& !src->eqv_uncast(dest)
&& ((alloc = tightly_coupled_allocation(dest, slow_region))
!= NULL)
&& _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
&& alloc->maybe_set_complete(&_gvn)) {
// "You break it, you buy it."
InitializeNode* init = alloc->initialization();
assert(init->is_complete(), "we just did this");
init->set_complete_with_arraycopy();
assert(dest->is_CheckCastPP(), "sanity");
assert(dest->in(0)->in(0) == init, "dest pinned");
adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
// From this point on, every exit path is responsible for
// initializing any non-copied parts of the object to zero.
// Also, if this flag is set we make sure that arraycopy interacts properly
// with G1, eliding pre-barriers. See CR 6627983.
dest_uninitialized = true;
} else {
// No zeroing elimination here.
alloc = NULL;
//original_dest = dest;
//dest_uninitialized = false;
}
// Results are placed here:
enum { fast_path = 1, // normal void-returning assembly stub
checked_path = 2, // special assembly stub with cleanup
slow_call_path = 3, // something went wrong; call the VM
zero_path = 4, // bypass when length of copy is zero
bcopy_path = 5, // copy primitive array by 64-bit blocks
PATH_LIMIT = 6
};
RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
record_for_igvn(result_region);
_gvn.set_type_bottom(result_i_o);
_gvn.set_type_bottom(result_memory);
assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
// The slow_control path:
Node* slow_control;
Node* slow_i_o = i_o();
Node* slow_mem = memory(adr_type);
debug_only(slow_control = (Node*) badAddress);
// Checked control path:
Node* checked_control = top();
Node* checked_mem = NULL;
Node* checked_i_o = NULL;
Node* checked_value = NULL;
if (basic_elem_type == T_CONFLICT) {
assert(!dest_uninitialized, "");
Node* cv = generate_generic_arraycopy(adr_type,
src, src_offset, dest, dest_offset,
copy_length, dest_uninitialized);
if (cv == NULL) cv = intcon(-1); // failure (no stub available)
checked_control = control();
checked_i_o = i_o();
checked_mem = memory(adr_type);
checked_value = cv;
set_control(top()); // no fast path
}
Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
if (not_pos != NULL) {
PreserveJVMState pjvms(this);
set_control(not_pos);
// (6) length must not be negative.
if (!length_never_negative) {
generate_negative_guard(copy_length, slow_region);
}
// copy_length is 0.
if (!stopped() && dest_uninitialized) {
Node* dest_length = alloc->in(AllocateNode::ALength);
if (copy_length->eqv_uncast(dest_length)
|| _gvn.find_int_con(dest_length, 1) <= 0) {
// There is no zeroing to do. No need for a secondary raw memory barrier.
} else {
// Clear the whole thing since there are no source elements to copy.
generate_clear_array(adr_type, dest, basic_elem_type,
intcon(0), NULL,
alloc->in(AllocateNode::AllocSize));
// Use a secondary InitializeNode as raw memory barrier.
// Currently it is needed only on this path since other
// paths have stub or runtime calls as raw memory barriers.
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
Compile::AliasIdxRaw,
top())->as_Initialize();
init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
}
}
// Present the results of the fast call.
result_region->init_req(zero_path, control());
result_i_o ->init_req(zero_path, i_o());
result_memory->init_req(zero_path, memory(adr_type));
}
if (!stopped() && dest_uninitialized) {
// We have to initialize the *uncopied* part of the array to zero.
// The copy destination is the slice dest[off..off+len]. The other slices
// are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
Node* dest_size = alloc->in(AllocateNode::AllocSize);
Node* dest_length = alloc->in(AllocateNode::ALength);
Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
copy_length));
// If there is a head section that needs zeroing, do it now.
if (find_int_con(dest_offset, -1) != 0) {
generate_clear_array(adr_type, dest, basic_elem_type,
intcon(0), dest_offset,
NULL);
}
// Next, perform a dynamic check on the tail length.
// It is often zero, and we can win big if we prove this.
// There are two wins: Avoid generating the ClearArray
// with its attendant messy index arithmetic, and upgrade
// the copy to a more hardware-friendly word size of 64 bits.
Node* tail_ctl = NULL;
if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
tail_ctl = generate_slow_guard(bol_lt, NULL);
assert(tail_ctl != NULL || !stopped(), "must be an outcome");
}
// At this point, let's assume there is no tail.
if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
// There is no tail. Try an upgrade to a 64-bit copy.
bool didit = false;
{ PreserveJVMState pjvms(this);
didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
src, src_offset, dest, dest_offset,
dest_size, dest_uninitialized);
if (didit) {
// Present the results of the block-copying fast call.
result_region->init_req(bcopy_path, control());
result_i_o ->init_req(bcopy_path, i_o());
result_memory->init_req(bcopy_path, memory(adr_type));
}
}
if (didit)
set_control(top()); // no regular fast path
}
// Clear the tail, if any.
if (tail_ctl != NULL) {
Node* notail_ctl = stopped() ? NULL : control();
set_control(tail_ctl);
if (notail_ctl == NULL) {
generate_clear_array(adr_type, dest, basic_elem_type,
dest_tail, NULL,
dest_size);
} else {
// Make a local merge.
Node* done_ctl = new(C) RegionNode(3);
Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
done_ctl->init_req(1, notail_ctl);
done_mem->init_req(1, memory(adr_type));
generate_clear_array(adr_type, dest, basic_elem_type,
dest_tail, NULL,
dest_size);
done_ctl->init_req(2, control());
done_mem->init_req(2, memory(adr_type));
set_control( _gvn.transform(done_ctl));
set_memory( _gvn.transform(done_mem), adr_type );
}
}
}
BasicType copy_type = basic_elem_type;
assert(basic_elem_type != T_ARRAY, "caller must fix this");
if (!stopped() && copy_type == T_OBJECT) {
// If src and dest have compatible element types, we can copy bits.
// Types S[] and D[] are compatible if D is a supertype of S.
//
// If they are not, we will use checked_oop_disjoint_arraycopy,
// which performs a fast optimistic per-oop check, and backs off
// further to JVM_ArrayCopy on the first per-oop check that fails.
// (Actually, we don't move raw bits only; the GC requires card marks.)
// Get the Klass* for both src and dest
Node* src_klass = load_object_klass(src);
Node* dest_klass = load_object_klass(dest);
// Generate the subtype check.
// This might fold up statically, or then again it might not.
//
// Non-static example: Copying List<String>.elements to a new String[].
// The backing store for a List<String> is always an Object[],
// but its elements are always type String, if the generic types
// are correct at the source level.
//
// Test S[] against D[], not S against D, because (probably)
// the secondary supertype cache is less busy for S[] than S.
// This usually only matters when D is an interface.
Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
// Plug failing path into checked_oop_disjoint_arraycopy
if (not_subtype_ctrl != top()) {
PreserveJVMState pjvms(this);
set_control(not_subtype_ctrl);
// (At this point we can assume disjoint_bases, since types differ.)
int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
Node* p1 = basic_plus_adr(dest_klass, ek_offset);
Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM);
Node* dest_elem_klass = _gvn.transform(n1);
Node* cv = generate_checkcast_arraycopy(adr_type,
dest_elem_klass,
src, src_offset, dest, dest_offset,
ConvI2X(copy_length), dest_uninitialized);
if (cv == NULL) cv = intcon(-1); // failure (no stub available)
checked_control = control();
checked_i_o = i_o();
checked_mem = memory(adr_type);
checked_value = cv;
}
// At this point we know we do not need type checks on oop stores.
// Let's see if we need card marks:
if (alloc != NULL && use_ReduceInitialCardMarks()) {
// If we do not need card marks, copy using the jint or jlong stub.
copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
"sizes agree");
}
}
if (!stopped()) {
// Generate the fast path, if possible.
PreserveJVMState pjvms(this);
generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
src, src_offset, dest, dest_offset,
ConvI2X(copy_length), dest_uninitialized);
// Present the results of the fast call.
result_region->init_req(fast_path, control());
result_i_o ->init_req(fast_path, i_o());
result_memory->init_req(fast_path, memory(adr_type));
}
// Here are all the slow paths up to this point, in one bundle:
slow_control = top();
if (slow_region != NULL)
slow_control = _gvn.transform(slow_region);
DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
set_control(checked_control);
if (!stopped()) {
// Clean up after the checked call.
// The returned value is either 0 or -1^K,
// where K = number of partially transferred array elements.
Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
// If it is 0, we are done, so transfer to the end.
Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
result_region->init_req(checked_path, checks_done);
result_i_o ->init_req(checked_path, checked_i_o);
result_memory->init_req(checked_path, checked_mem);
// If it is not zero, merge into the slow call.
set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
RegionNode* slow_reg2 = new(C) RegionNode(3);
PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
record_for_igvn(slow_reg2);
slow_reg2 ->init_req(1, slow_control);
slow_i_o2 ->init_req(1, slow_i_o);
slow_mem2 ->init_req(1, slow_mem);
slow_reg2 ->init_req(2, control());
slow_i_o2 ->init_req(2, checked_i_o);
slow_mem2 ->init_req(2, checked_mem);
slow_control = _gvn.transform(slow_reg2);
slow_i_o = _gvn.transform(slow_i_o2);
slow_mem = _gvn.transform(slow_mem2);
if (alloc != NULL) {
// We'll restart from the very beginning, after zeroing the whole thing.
// This can cause double writes, but that's OK since dest is brand new.
// So we ignore the low 31 bits of the value returned from the stub.
} else {
// We must continue the copy exactly where it failed, or else
// another thread might see the wrong number of writes to dest.
Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
slow_offset->init_req(1, intcon(0));
slow_offset->init_req(2, checked_offset);
slow_offset = _gvn.transform(slow_offset);
// Adjust the arguments by the conditionally incoming offset.
Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
// Tweak the node variables to adjust the code produced below:
src_offset = src_off_plus;
dest_offset = dest_off_plus;
copy_length = length_minus;
}
}
set_control(slow_control);
if (!stopped()) {
// Generate the slow path, if needed.
PreserveJVMState pjvms(this); // replace_in_map may trash the map
set_memory(slow_mem, adr_type);
set_i_o(slow_i_o);
if (dest_uninitialized) {
generate_clear_array(adr_type, dest, basic_elem_type,
intcon(0), NULL,
alloc->in(AllocateNode::AllocSize));
}
generate_slow_arraycopy(adr_type,
src, src_offset, dest, dest_offset,
copy_length, /*dest_uninitialized*/false);
result_region->init_req(slow_call_path, control());
result_i_o ->init_req(slow_call_path, i_o());
result_memory->init_req(slow_call_path, memory(adr_type));
}
// Remove unused edges.
for (uint i = 1; i < result_region->req(); i++) {
if (result_region->in(i) == NULL)
result_region->init_req(i, top());
}
// Finished; return the combined state.
set_control( _gvn.transform(result_region));
set_i_o( _gvn.transform(result_i_o) );
set_memory( _gvn.transform(result_memory), adr_type );
// The memory edges above are precise in order to model effects around
// array copies accurately to allow value numbering of field loads around
// arraycopy. Such field loads, both before and after, are common in Java
// collections and similar classes involving header/array data structures.
//
// But with low number of register or when some registers are used or killed
// by arraycopy calls it causes registers spilling on stack. See 6544710.
// The next memory barrier is added to avoid it. If the arraycopy can be
// optimized away (which it can, sometimes) then we can manually remove
// the membar also.
//
// Do not let reads from the cloned object float above the arraycopy.
if (alloc != NULL) {
// Do not let stores that initialize this object be reordered with
// a subsequent store that would make this object accessible by
// other threads.
// Record what AllocateNode this StoreStore protects so that
// escape analysis can go from the MemBarStoreStoreNode to the
// AllocateNode and eliminate the MemBarStoreStoreNode if possible
// based on the escape status of the AllocateNode.
insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
} else if (InsertMemBarAfterArraycopy)
insert_mem_bar(Op_MemBarCPUOrder);
}
// Helper function which determines if an arraycopy immediately follows
// an allocation, with no intervening tests or other escapes for the object.
AllocateArrayNode*
LibraryCallKit::tightly_coupled_allocation(Node* ptr,
RegionNode* slow_region) {
if (stopped()) return NULL; // no fast path
if (C->AliasLevel() == 0) return NULL; // no MergeMems around
AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
if (alloc == NULL) return NULL;
Node* rawmem = memory(Compile::AliasIdxRaw);
// Is the allocation's memory state untouched?
if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
// Bail out if there have been raw-memory effects since the allocation.
// (Example: There might have been a call or safepoint.)
return NULL;
}
rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
return NULL;
}
// There must be no unexpected observers of this allocation.
for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
Node* obs = ptr->fast_out(i);
if (obs != this->map()) {
return NULL;
}
}
// This arraycopy must unconditionally follow the allocation of the ptr.
Node* alloc_ctl = ptr->in(0);
assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
Node* ctl = control();
while (ctl != alloc_ctl) {
// There may be guards which feed into the slow_region.
// Any other control flow means that we might not get a chance
// to finish initializing the allocated object.
if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
IfNode* iff = ctl->in(0)->as_If();
Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
ctl = iff->in(0); // This test feeds the known slow_region.
continue;
}
// One more try: Various low-level checks bottom out in
// uncommon traps. If the debug-info of the trap omits
// any reference to the allocation, as we've already
// observed, then there can be no objection to the trap.
bool found_trap = false;
for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
Node* obs = not_ctl->fast_out(j);
if (obs->in(0) == not_ctl && obs->is_Call() &&
(obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
found_trap = true; break;
}
}
if (found_trap) {
ctl = iff->in(0); // This test feeds a harmless uncommon trap.
continue;
}
}
return NULL;
}
// If we get this far, we have an allocation which immediately
// precedes the arraycopy, and we can take over zeroing the new object.
// The arraycopy will finish the initialization, and provide
// a new control state to which we will anchor the destination pointer.
return alloc;
}
// Helper for initialization of arrays, creating a ClearArray.
// It writes zero bits in [start..end), within the body of an array object.
// The memory effects are all chained onto the 'adr_type' alias category.
//
// Since the object is otherwise uninitialized, we are free
// to put a little "slop" around the edges of the cleared area,
// as long as it does not go back into the array's header,
// or beyond the array end within the heap.
//
// The lower edge can be rounded down to the nearest jint and the
// upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
//
// Arguments:
// adr_type memory slice where writes are generated
// dest oop of the destination array
// basic_elem_type element type of the destination
// slice_idx array index of first element to store
// slice_len number of elements to store (or NULL)
// dest_size total size in bytes of the array object
//
// Exactly one of slice_len or dest_size must be non-NULL.
// If dest_size is non-NULL, zeroing extends to the end of the object.
// If slice_len is non-NULL, the slice_idx value must be a constant.
void
LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
Node* dest,
BasicType basic_elem_type,
Node* slice_idx,
Node* slice_len,
Node* dest_size) {
// one or the other but not both of slice_len and dest_size:
assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
if (slice_len == NULL) slice_len = top();
if (dest_size == NULL) dest_size = top();
// operate on this memory slice:
Node* mem = memory(adr_type); // memory slice to operate on
// scaling and rounding of indexes:
int scale = exact_log2(type2aelembytes(basic_elem_type));
int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
int clear_low = (-1 << scale) & (BytesPerInt - 1);
int bump_bit = (-1 << scale) & BytesPerInt;
// determine constant starts and ends
const intptr_t BIG_NEG = -128;
assert(BIG_NEG + 2*abase < 0, "neg enough");
intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
if (slice_len_con == 0) {
return; // nothing to do here
}
intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
intptr_t end_con = find_intptr_t_con(dest_size, -1);
if (slice_idx_con >= 0 && slice_len_con >= 0) {
assert(end_con < 0, "not two cons");
end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
BytesPerLong);
}
if (start_con >= 0 && end_con >= 0) {
// Constant start and end. Simple.
mem = ClearArrayNode::clear_memory(control(), mem, dest,
start_con, end_con, &_gvn);
} else if (start_con >= 0 && dest_size != top()) {
// Constant start, pre-rounded end after the tail of the array.
Node* end = dest_size;
mem = ClearArrayNode::clear_memory(control(), mem, dest,
start_con, end, &_gvn);
} else if (start_con >= 0 && slice_len != top()) {
// Constant start, non-constant end. End needs rounding up.
// End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
intptr_t end_base = abase + (slice_idx_con << scale);
int end_round = (-1 << scale) & (BytesPerLong - 1);
Node* end = ConvI2X(slice_len);
if (scale != 0)
end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
end_base += end_round;
end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
mem = ClearArrayNode::clear_memory(control(), mem, dest,
start_con, end, &_gvn);
} else if (start_con < 0 && dest_size != top()) {
// Non-constant start, pre-rounded end after the tail of the array.
// This is almost certainly a "round-to-end" operation.
Node* start = slice_idx;
start = ConvI2X(start);
if (scale != 0)
start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
if ((bump_bit | clear_low) != 0) {
int to_clear = (bump_bit | clear_low);
// Align up mod 8, then store a jint zero unconditionally
// just before the mod-8 boundary.
if (((abase + bump_bit) & ~to_clear) - bump_bit
< arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
bump_bit = 0;
assert((abase & to_clear) == 0, "array base must be long-aligned");
} else {
// Bump 'start' up to (or past) the next jint boundary:
start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
assert((abase & clear_low) == 0, "array base must be int-aligned");
}
// Round bumped 'start' down to jlong boundary in body of array.
start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
if (bump_bit != 0) {
// Store a zero to the immediately preceding jint:
Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
Node* p1 = basic_plus_adr(dest, x1);
mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
mem = _gvn.transform(mem);
}
}
Node* end = dest_size; // pre-rounded
mem = ClearArrayNode::clear_memory(control(), mem, dest,
start, end, &_gvn);
} else {
// Non-constant start, unrounded non-constant end.
// (Nobody zeroes a random midsection of an array using this routine.)
ShouldNotReachHere(); // fix caller
}
// Done.
set_memory(mem, adr_type);
}
bool
LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
BasicType basic_elem_type,
AllocateNode* alloc,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* dest_size, bool dest_uninitialized) {
// See if there is an advantage from block transfer.
int scale = exact_log2(type2aelembytes(basic_elem_type));
if (scale >= LogBytesPerLong)
return false; // it is already a block transfer
// Look at the alignment of the starting offsets.
int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
if (src_off_con < 0 || dest_off_con < 0)
// At present, we can only understand constants.
return false;
intptr_t src_off = abase + (src_off_con << scale);
intptr_t dest_off = abase + (dest_off_con << scale);
if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
// Non-aligned; too bad.
// One more chance: Pick off an initial 32-bit word.
// This is a common case, since abase can be odd mod 8.
if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
Node* sptr = basic_plus_adr(src, src_off);
Node* dptr = basic_plus_adr(dest, dest_off);
const TypePtr* s_adr_type = _gvn.type(sptr)->is_ptr();
assert(s_adr_type->isa_aryptr(), "impossible slice");
Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, s_adr_type, MemNode::unordered);
store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
src_off += BytesPerInt;
dest_off += BytesPerInt;
} else {
return false;
}
}
assert(src_off % BytesPerLong == 0, "");
assert(dest_off % BytesPerLong == 0, "");
// Do this copy by giant steps.
Node* sptr = basic_plus_adr(src, src_off);
Node* dptr = basic_plus_adr(dest, dest_off);
Node* countx = dest_size;
countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
bool disjoint_bases = true; // since alloc != NULL
generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
sptr, NULL, dptr, NULL, countx, dest_uninitialized);
return true;
}
// Helper function; generates code for the slow case.
// We make a call to a runtime method which emulates the native method,
// but without the native wrapper overhead.
void
LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* copy_length, bool dest_uninitialized) {
assert(!dest_uninitialized, "Invariant");
Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
OptoRuntime::slow_arraycopy_Type(),
OptoRuntime::slow_arraycopy_Java(),
"slow_arraycopy", adr_type,
src, src_offset, dest, dest_offset,
copy_length);
// Handle exceptions thrown by this fellow:
make_slow_call_ex(call, env()->Throwable_klass(), false);
}
// Helper function; generates code for cases requiring runtime checks.
Node*
LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
Node* dest_elem_klass,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* copy_length, bool dest_uninitialized) {
if (stopped()) return NULL;
address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
return NULL;
}
// Pick out the parameters required to perform a store-check
// for the target array. This is an optimistic check. It will
// look in each non-null element's class, at the desired klass's
// super_check_offset, for the desired klass.
int sco_offset = in_bytes(Klass::super_check_offset_offset());
Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
Node* check_offset = ConvI2X(_gvn.transform(n3));
Node* check_value = dest_elem_klass;
Node* src_start = array_element_address(src, src_offset, T_OBJECT);
Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
// (We know the arrays are never conjoint, because their types differ.)
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::checkcast_arraycopy_Type(),
copyfunc_addr, "checkcast_arraycopy", adr_type,
// five arguments, of which two are
// intptr_t (jlong in LP64)
src_start, dest_start,
copy_length XTOP,
check_offset XTOP,
check_value);
return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
}
// Helper function; generates code for cases requiring runtime checks.
Node*
LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* copy_length, bool dest_uninitialized) {
assert(!dest_uninitialized, "Invariant");
if (stopped()) return NULL;
address copyfunc_addr = StubRoutines::generic_arraycopy();
if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
return NULL;
}
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::generic_arraycopy_Type(),
copyfunc_addr, "generic_arraycopy", adr_type,
src, src_offset, dest, dest_offset, copy_length);
return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
}
// Helper function; generates the fast out-of-line call to an arraycopy stub.
void
LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
BasicType basic_elem_type,
bool disjoint_bases,
Node* src, Node* src_offset,
Node* dest, Node* dest_offset,
Node* copy_length, bool dest_uninitialized) {
if (stopped()) return; // nothing to do
Node* src_start = src;
Node* dest_start = dest;
if (src_offset != NULL || dest_offset != NULL) {
assert(src_offset != NULL && dest_offset != NULL, "");
src_start = array_element_address(src, src_offset, basic_elem_type);
dest_start = array_element_address(dest, dest_offset, basic_elem_type);
}
// Figure out which arraycopy runtime method to call.
const char* copyfunc_name = "arraycopy";
address copyfunc_addr =
basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
disjoint_bases, copyfunc_name, dest_uninitialized);
// Call it. Note that the count_ix value is not scaled to a byte-size.
make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::fast_arraycopy_Type(),
copyfunc_addr, copyfunc_name, adr_type,
src_start, dest_start, copy_length XTOP);
}
//-------------inline_encodeISOArray-----------------------------------
// encode char[] to byte[] in ISO_8859_1
bool LibraryCallKit::inline_encodeISOArray() {
assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
// no receiver since it is static method
Node *src = argument(0);
Node *src_offset = argument(1);
Node *dst = argument(2);
Node *dst_offset = argument(3);
Node *length = argument(4);
const Type* src_type = src->Value(&_gvn);
const Type* dst_type = dst->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
const TypeAryPtr* top_dest = dst_type->isa_aryptr();
if (top_src == NULL || top_src->klass() == NULL ||
top_dest == NULL || top_dest->klass() == NULL) {
// failed array check
return false;
}
// Figure out the size and type of the elements we will be copying.
BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (src_elem != T_CHAR || dst_elem != T_BYTE) {
return false;
}
Node* src_start = array_element_address(src, src_offset, src_elem);
Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
// 'src_start' points to src array + scaled offset
// 'dst_start' points to dst array + scaled offset
const TypeAryPtr* mtype = TypeAryPtr::BYTES;
Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
enc = _gvn.transform(enc);
Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
set_memory(res_mem, mtype);
set_result(enc);
return true;
}
//-------------inline_multiplyToLen-----------------------------------
bool LibraryCallKit::inline_multiplyToLen() {
assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
address stubAddr = StubRoutines::multiplyToLen();
if (stubAddr == NULL) {
return false; // Intrinsic's stub is not implemented on this platform
}
const char* stubName = "multiplyToLen";
assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
// no receiver because it is a static method
Node* x = argument(0);
Node* xlen = argument(1);
Node* y = argument(2);
Node* ylen = argument(3);
Node* z = argument(4);
const Type* x_type = x->Value(&_gvn);
const Type* y_type = y->Value(&_gvn);
const TypeAryPtr* top_x = x_type->isa_aryptr();
const TypeAryPtr* top_y = y_type->isa_aryptr();
if (top_x == NULL || top_x->klass() == NULL ||
top_y == NULL || top_y->klass() == NULL) {
// failed array check
return false;
}
BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (x_elem != T_INT || y_elem != T_INT) {
return false;
}
// Set the original stack and the reexecute bit for the interpreter to reexecute
// the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
// on the return from z array allocation in runtime.
{ PreserveReexecuteState preexecs(this);
jvms()->set_should_reexecute(true);
Node* x_start = array_element_address(x, intcon(0), x_elem);
Node* y_start = array_element_address(y, intcon(0), y_elem);
// 'x_start' points to x array + scaled xlen
// 'y_start' points to y array + scaled ylen
// Allocate the result array
Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen));
ciKlass* klass = ciTypeArrayKlass::make(T_INT);
Node* klass_node = makecon(TypeKlassPtr::make(klass));
IdealKit ideal(this);
#define __ ideal.
Node* one = __ ConI(1);
Node* zero = __ ConI(0);
IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
__ set(need_alloc, zero);
__ set(z_alloc, z);
__ if_then(z, BoolTest::eq, null()); {
__ increment (need_alloc, one);
} __ else_(); {
// Update graphKit memory and control from IdealKit.
sync_kit(ideal);
Node* zlen_arg = load_array_length(z);
// Update IdealKit memory and control from graphKit.
__ sync_kit(this);
__ if_then(zlen_arg, BoolTest::lt, zlen); {
__ increment (need_alloc, one);
} __ end_if();
} __ end_if();
__ if_then(__ value(need_alloc), BoolTest::ne, zero); {
// Update graphKit memory and control from IdealKit.
sync_kit(ideal);
Node * narr = new_array(klass_node, zlen, 1);
// Update IdealKit memory and control from graphKit.
__ sync_kit(this);
__ set(z_alloc, narr);
} __ end_if();
sync_kit(ideal);
z = __ value(z_alloc);
// Can't use TypeAryPtr::INTS which uses Bottom offset.
_gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
// Final sync IdealKit and GraphKit.
final_sync(ideal);
#undef __
Node* z_start = array_element_address(z, intcon(0), T_INT);
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::multiplyToLen_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
x_start, xlen, y_start, ylen, z_start, zlen);
} // original reexecute is set back here
C->set_has_split_ifs(true); // Has chance for split-if optimization
set_result(z);
return true;
}
//-------------inline_squareToLen------------------------------------
bool LibraryCallKit::inline_squareToLen() {
assert(UseSquareToLenIntrinsic, "not implementated on this platform");
address stubAddr = StubRoutines::squareToLen();
if (stubAddr == NULL) {
return false; // Intrinsic's stub is not implemented on this platform
}
const char* stubName = "squareToLen";
assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
Node* x = argument(0);
Node* len = argument(1);
Node* z = argument(2);
Node* zlen = argument(3);
const Type* x_type = x->Value(&_gvn);
const Type* z_type = z->Value(&_gvn);
const TypeAryPtr* top_x = x_type->isa_aryptr();
const TypeAryPtr* top_z = z_type->isa_aryptr();
if (top_x == NULL || top_x->klass() == NULL ||
top_z == NULL || top_z->klass() == NULL) {
// failed array check
return false;
}
BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (x_elem != T_INT || z_elem != T_INT) {
return false;
}
Node* x_start = array_element_address(x, intcon(0), x_elem);
Node* z_start = array_element_address(z, intcon(0), z_elem);
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::squareToLen_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
x_start, len, z_start, zlen);
set_result(z);
return true;
}
//-------------inline_mulAdd------------------------------------------
bool LibraryCallKit::inline_mulAdd() {
assert(UseMulAddIntrinsic, "not implementated on this platform");
address stubAddr = StubRoutines::mulAdd();
if (stubAddr == NULL) {
return false; // Intrinsic's stub is not implemented on this platform
}
const char* stubName = "mulAdd";
assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
Node* out = argument(0);
Node* in = argument(1);
Node* offset = argument(2);
Node* len = argument(3);
Node* k = argument(4);
const Type* out_type = out->Value(&_gvn);
const Type* in_type = in->Value(&_gvn);
const TypeAryPtr* top_out = out_type->isa_aryptr();
const TypeAryPtr* top_in = in_type->isa_aryptr();
if (top_out == NULL || top_out->klass() == NULL ||
top_in == NULL || top_in->klass() == NULL) {
// failed array check
return false;
}
BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (out_elem != T_INT || in_elem != T_INT) {
return false;
}
Node* outlen = load_array_length(out);
Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset));
Node* out_start = array_element_address(out, intcon(0), out_elem);
Node* in_start = array_element_address(in, intcon(0), in_elem);
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::mulAdd_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
out_start,in_start, new_offset, len, k);
Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
set_result(result);
return true;
}
//-------------inline_montgomeryMultiply-----------------------------------
bool LibraryCallKit::inline_montgomeryMultiply() {
address stubAddr = StubRoutines::montgomeryMultiply();
if (stubAddr == NULL) {
return false; // Intrinsic's stub is not implemented on this platform
}
assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
const char* stubName = "montgomery_multiply";
assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
Node* a = argument(0);
Node* b = argument(1);
Node* n = argument(2);
Node* len = argument(3);
Node* inv = argument(4);
Node* m = argument(6);
const Type* a_type = a->Value(&_gvn);
const TypeAryPtr* top_a = a_type->isa_aryptr();
const Type* b_type = b->Value(&_gvn);
const TypeAryPtr* top_b = b_type->isa_aryptr();
const Type* n_type = a->Value(&_gvn);
const TypeAryPtr* top_n = n_type->isa_aryptr();
const Type* m_type = a->Value(&_gvn);
const TypeAryPtr* top_m = m_type->isa_aryptr();
if (top_a == NULL || top_a->klass() == NULL ||
top_b == NULL || top_b->klass() == NULL ||
top_n == NULL || top_n->klass() == NULL ||
top_m == NULL || top_m->klass() == NULL) {
// failed array check
return false;
}
BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
return false;
}
// Make the call
{
Node* a_start = array_element_address(a, intcon(0), a_elem);
Node* b_start = array_element_address(b, intcon(0), b_elem);
Node* n_start = array_element_address(n, intcon(0), n_elem);
Node* m_start = array_element_address(m, intcon(0), m_elem);
Node* call = NULL;
if (CCallingConventionRequiresIntsAsLongs) {
Node* len_I2L = ConvI2L(len);
call = make_runtime_call(RC_LEAF,
OptoRuntime::montgomeryMultiply_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
a_start, b_start, n_start, len_I2L XTOP, inv,
top(), m_start);
} else {
call = make_runtime_call(RC_LEAF,
OptoRuntime::montgomeryMultiply_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
a_start, b_start, n_start, len, inv, top(),
m_start);
}
set_result(m);
}
return true;
}
bool LibraryCallKit::inline_montgomerySquare() {
address stubAddr = StubRoutines::montgomerySquare();
if (stubAddr == NULL) {
return false; // Intrinsic's stub is not implemented on this platform
}
assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
const char* stubName = "montgomery_square";
assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
Node* a = argument(0);
Node* n = argument(1);
Node* len = argument(2);
Node* inv = argument(3);
Node* m = argument(5);
const Type* a_type = a->Value(&_gvn);
const TypeAryPtr* top_a = a_type->isa_aryptr();
const Type* n_type = a->Value(&_gvn);
const TypeAryPtr* top_n = n_type->isa_aryptr();
const Type* m_type = a->Value(&_gvn);
const TypeAryPtr* top_m = m_type->isa_aryptr();
if (top_a == NULL || top_a->klass() == NULL ||
top_n == NULL || top_n->klass() == NULL ||
top_m == NULL || top_m->klass() == NULL) {
// failed array check
return false;
}
BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
return false;
}
// Make the call
{
Node* a_start = array_element_address(a, intcon(0), a_elem);
Node* n_start = array_element_address(n, intcon(0), n_elem);
Node* m_start = array_element_address(m, intcon(0), m_elem);
Node* call = NULL;
if (CCallingConventionRequiresIntsAsLongs) {
Node* len_I2L = ConvI2L(len);
call = make_runtime_call(RC_LEAF,
OptoRuntime::montgomerySquare_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
a_start, n_start, len_I2L XTOP, inv, top(),
m_start);
} else {
call = make_runtime_call(RC_LEAF,
OptoRuntime::montgomerySquare_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
a_start, n_start, len, inv, top(),
m_start);
}
set_result(m);
}
return true;
}
/**
* Calculate CRC32 for byte.
* int java.util.zip.CRC32.update(int crc, int b)
*/
bool LibraryCallKit::inline_updateCRC32() {
assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
assert(callee()->signature()->size() == 2, "update has 2 parameters");
// no receiver since it is static method
Node* crc = argument(0); // type: int
Node* b = argument(1); // type: int
/*
* int c = ~ crc;
* b = timesXtoThe32[(b ^ c) & 0xFF];
* b = b ^ (c >>> 8);
* crc = ~b;
*/
Node* M1 = intcon(-1);
crc = _gvn.transform(new (C) XorINode(crc, M1));
Node* result = _gvn.transform(new (C) XorINode(crc, b));
result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
result = _gvn.transform(new (C) XorINode(crc, result));
result = _gvn.transform(new (C) XorINode(result, M1));
set_result(result);
return true;
}
/**
* Calculate CRC32 for byte[] array.
* int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
*/
bool LibraryCallKit::inline_updateBytesCRC32() {
assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
// no receiver since it is static method
Node* crc = argument(0); // type: int
Node* src = argument(1); // type: oop
Node* offset = argument(2); // type: int
Node* length = argument(3); // type: int
const Type* src_type = src->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
if (top_src == NULL || top_src->klass() == NULL) {
// failed array check
return false;
}
// Figure out the size and type of the elements we will be copying.
BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (src_elem != T_BYTE) {
return false;
}
// 'src_start' points to src array + scaled offset
Node* src_start = array_element_address(src, offset, src_elem);
// We assume that range check is done by caller.
// TODO: generate range check (offset+length < src.length) in debug VM.
// Call the stub.
address stubAddr = StubRoutines::updateBytesCRC32();
const char *stubName = "updateBytesCRC32";
Node* call;
if (CCallingConventionRequiresIntsAsLongs) {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
crc XTOP, src_start, length XTOP);
} else {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
crc, src_start, length);
}
Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
set_result(result);
return true;
}
/**
* Calculate CRC32 for ByteBuffer.
* int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
*/
bool LibraryCallKit::inline_updateByteBufferCRC32() {
assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
// no receiver since it is static method
Node* crc = argument(0); // type: int
Node* src = argument(1); // type: long
Node* offset = argument(3); // type: int
Node* length = argument(4); // type: int
src = ConvL2X(src); // adjust Java long to machine word
Node* base = _gvn.transform(new (C) CastX2PNode(src));
offset = ConvI2X(offset);
// 'src_start' points to src array + scaled offset
Node* src_start = basic_plus_adr(top(), base, offset);
// Call the stub.
address stubAddr = StubRoutines::updateBytesCRC32();
const char *stubName = "updateBytesCRC32";
Node* call;
if (CCallingConventionRequiresIntsAsLongs) {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
crc XTOP, src_start, length XTOP);
} else {
call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
crc, src_start, length);
}
Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
set_result(result);
return true;
}
//----------------------------inline_reference_get----------------------------
// public T java.lang.ref.Reference.get();
bool LibraryCallKit::inline_reference_get() {
const int referent_offset = java_lang_ref_Reference::referent_offset;
guarantee(referent_offset > 0, "should have already been set");
// Get the argument:
Node* reference_obj = null_check_receiver();
if (stopped()) return true;
Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
ciInstanceKlass* klass = env()->Object_klass();
const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
Node* no_ctrl = NULL;
Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
// Use the pre-barrier to record the value in the referent field
pre_barrier(false /* do_load */,
control(),
NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
result /* pre_val */,
T_OBJECT);
// Add memory barrier to prevent commoning reads from this field
// across safepoint since GC can change its value.
insert_mem_bar(Op_MemBarCPUOrder);
set_result(result);
return true;
}
Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
bool is_exact=true, bool is_static=false) {
const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
assert(tinst != NULL, "obj is null");
assert(tinst->klass()->is_loaded(), "obj is not loaded");
assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
ciSymbol::make(fieldTypeString),
is_static);
if (field == NULL) return (Node *) NULL;
assert (field != NULL, "undefined field");
// Next code copied from Parse::do_get_xxx():
// Compute address and memory type.
int offset = field->offset_in_bytes();
bool is_vol = field->is_volatile();
ciType* field_klass = field->type();
assert(field_klass->is_loaded(), "should be loaded");
const TypePtr* adr_type = C->alias_type(field)->adr_type();
Node *adr = basic_plus_adr(fromObj, fromObj, offset);
BasicType bt = field->layout_type();
// Build the resultant type of the load
const Type *type;
if (bt == T_OBJECT) {
type = TypeOopPtr::make_from_klass(field_klass->as_klass());
} else {
type = Type::get_const_basic_type(bt);
}
Node* leading_membar = NULL;
if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
leading_membar = insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
}
// Build the load.
MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
// If reference is volatile, prevent following memory ops from
// floating up past the volatile read. Also prevents commoning
// another volatile read.
if (is_vol) {
// Memory barrier includes bogus read of value to force load BEFORE membar
Node* mb = insert_mem_bar(Op_MemBarAcquire, loadedField);
mb->as_MemBar()->set_trailing_load();
}
return loadedField;
}
//------------------------------inline_aescrypt_Block-----------------------
bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
address stubAddr = NULL;
const char *stubName;
assert(UseAES, "need AES instruction support");
switch(id) {
case vmIntrinsics::_aescrypt_encryptBlock:
stubAddr = StubRoutines::aescrypt_encryptBlock();
stubName = "aescrypt_encryptBlock";
break;
case vmIntrinsics::_aescrypt_decryptBlock:
stubAddr = StubRoutines::aescrypt_decryptBlock();
stubName = "aescrypt_decryptBlock";
break;
}
if (stubAddr == NULL) return false;
Node* aescrypt_object = argument(0);
Node* src = argument(1);
Node* src_offset = argument(2);
Node* dest = argument(3);
Node* dest_offset = argument(4);
// (1) src and dest are arrays.
const Type* src_type = src->Value(&_gvn);
const Type* dest_type = dest->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
const TypeAryPtr* top_dest = dest_type->isa_aryptr();
assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
// for the quick and dirty code we will skip all the checks.
// we are just trying to get the call to be generated.
Node* src_start = src;
Node* dest_start = dest;
if (src_offset != NULL || dest_offset != NULL) {
assert(src_offset != NULL && dest_offset != NULL, "");
src_start = array_element_address(src, src_offset, T_BYTE);
dest_start = array_element_address(dest, dest_offset, T_BYTE);
}
// now need to get the start of its expanded key array
// this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
if (k_start == NULL) return false;
if (Matcher::pass_original_key_for_aes()) {
// on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
// compatibility issues between Java key expansion and SPARC crypto instructions
Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
if (original_k_start == NULL) return false;
// Call the stub.
make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, dest_start, k_start, original_k_start);
} else {
// Call the stub.
make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, dest_start, k_start);
}
return true;
}
//------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
address stubAddr = NULL;
const char *stubName = NULL;
assert(UseAES, "need AES instruction support");
switch(id) {
case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
stubName = "cipherBlockChaining_encryptAESCrypt";
break;
case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
stubName = "cipherBlockChaining_decryptAESCrypt";
break;
}
if (stubAddr == NULL) return false;
Node* cipherBlockChaining_object = argument(0);
Node* src = argument(1);
Node* src_offset = argument(2);
Node* len = argument(3);
Node* dest = argument(4);
Node* dest_offset = argument(5);
// (1) src and dest are arrays.
const Type* src_type = src->Value(&_gvn);
const Type* dest_type = dest->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
const TypeAryPtr* top_dest = dest_type->isa_aryptr();
assert (top_src != NULL && top_src->klass() != NULL
&& top_dest != NULL && top_dest->klass() != NULL, "args are strange");
// checks are the responsibility of the caller
Node* src_start = src;
Node* dest_start = dest;
if (src_offset != NULL || dest_offset != NULL) {
assert(src_offset != NULL && dest_offset != NULL, "");
src_start = array_element_address(src, src_offset, T_BYTE);
dest_start = array_element_address(dest, dest_offset, T_BYTE);
}
// if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
// (because of the predicated logic executed earlier).
// so we cast it here safely.
// this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
if (embeddedCipherObj == NULL) return false;
// cast it to what we know it will be at runtime
const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
assert(tinst != NULL, "CBC obj is null");
assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
const TypeOopPtr* xtype = aklass->as_instance_type();
Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
aescrypt_object = _gvn.transform(aescrypt_object);
// we need to get the start of the aescrypt_object's expanded key array
Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
if (k_start == NULL) return false;
// similarly, get the start address of the r vector
Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
if (objRvec == NULL) return false;
Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
Node* cbcCrypt;
if (Matcher::pass_original_key_for_aes()) {
// on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
// compatibility issues between Java key expansion and SPARC crypto instructions
Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
if (original_k_start == NULL) return false;
// Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::cipherBlockChaining_aescrypt_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, dest_start, k_start, r_start, len, original_k_start);
} else {
// Call the stub, passing src_start, dest_start, k_start, r_start and src_len
cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::cipherBlockChaining_aescrypt_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, dest_start, k_start, r_start, len);
}
// return cipher length (int)
Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
set_result(retvalue);
return true;
}
//------------------------------get_key_start_from_aescrypt_object-----------------------
Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
#ifdef PPC64
// MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
// Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
// However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
// The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
if (objSessionK == NULL) {
return (Node *) NULL;
}
Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
#else
Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
#endif // PPC64
assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
if (objAESCryptKey == NULL) return (Node *) NULL;
// now have the array, need to get the start address of the K array
Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
return k_start;
}
//------------------------------get_original_key_start_from_aescrypt_object-----------------------
Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
if (objAESCryptKey == NULL) return (Node *) NULL;
// now have the array, need to get the start address of the lastKey array
Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
return original_k_start;
}
//----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
// Return node representing slow path of predicate check.
// the pseudo code we want to emulate with this predicate is:
// for encryption:
// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
// for decryption:
// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
// note cipher==plain is more conservative than the original java code but that's OK
//
Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
// The receiver was checked for NULL already.
Node* objCBC = argument(0);
// Load embeddedCipher field of CipherBlockChaining object.
Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
// get AESCrypt klass for instanceOf check
// AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
// will have same classloader as CipherBlockChaining object
const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
assert(tinst != NULL, "CBCobj is null");
assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
// we want to do an instanceof comparison against the AESCrypt class
ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
if (!klass_AESCrypt->is_loaded()) {
// if AESCrypt is not even loaded, we never take the intrinsic fast path
Node* ctrl = control();
set_control(top()); // no regular fast path
return ctrl;
}
ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
// for encryption, we are done
if (!decrypting)
return instof_false; // even if it is NULL
// for decryption, we need to add a further check to avoid
// taking the intrinsic path when cipher and plain are the same
// see the original java code for why.
RegionNode* region = new(C) RegionNode(3);
region->init_req(1, instof_false);
Node* src = argument(1);
Node* dest = argument(4);
Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
region->init_req(2, src_dest_conjoint);
record_for_igvn(region);
return _gvn.transform(region);
}
//------------------------------inline_ghash_processBlocks
bool LibraryCallKit::inline_ghash_processBlocks() {
address stubAddr;
const char *stubName;
assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
stubAddr = StubRoutines::ghash_processBlocks();
stubName = "ghash_processBlocks";
Node* data = argument(0);
Node* offset = argument(1);
Node* len = argument(2);
Node* state = argument(3);
Node* subkeyH = argument(4);
Node* state_start = array_element_address(state, intcon(0), T_LONG);
assert(state_start, "state is NULL");
Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
assert(subkeyH_start, "subkeyH is NULL");
Node* data_start = array_element_address(data, offset, T_BYTE);
assert(data_start, "data is NULL");
Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::ghash_processBlocks_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
state_start, subkeyH_start, data_start, len);
return true;
}
//------------------------------inline_sha_implCompress-----------------------
//
// Calculate SHA (i.e., SHA-1) for single-block byte[] array.
// void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
//
// Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
// void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
//
// Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
// void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
//
bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
Node* sha_obj = argument(0);
Node* src = argument(1); // type oop
Node* ofs = argument(2); // type int
const Type* src_type = src->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
if (top_src == NULL || top_src->klass() == NULL) {
// failed array check
return false;
}
// Figure out the size and type of the elements we will be copying.
BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (src_elem != T_BYTE) {
return false;
}
// 'src_start' points to src array + offset
Node* src_start = array_element_address(src, ofs, src_elem);
Node* state = NULL;
address stubAddr;
const char *stubName;
switch(id) {
case vmIntrinsics::_sha_implCompress:
assert(UseSHA1Intrinsics, "need SHA1 instruction support");
state = get_state_from_sha_object(sha_obj);
stubAddr = StubRoutines::sha1_implCompress();
stubName = "sha1_implCompress";
break;
case vmIntrinsics::_sha2_implCompress:
assert(UseSHA256Intrinsics, "need SHA256 instruction support");
state = get_state_from_sha_object(sha_obj);
stubAddr = StubRoutines::sha256_implCompress();
stubName = "sha256_implCompress";
break;
case vmIntrinsics::_sha5_implCompress:
assert(UseSHA512Intrinsics, "need SHA512 instruction support");
state = get_state_from_sha5_object(sha_obj);
stubAddr = StubRoutines::sha512_implCompress();
stubName = "sha512_implCompress";
break;
default:
fatal_unexpected_iid(id);
return false;
}
if (state == NULL) return false;
// Call the stub.
Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state);
return true;
}
//------------------------------inline_digestBase_implCompressMB-----------------------
//
// Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
// int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
//
bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
"need SHA1/SHA256/SHA512 instruction support");
assert((uint)predicate < 3, "sanity");
assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
Node* src = argument(1); // byte[] array
Node* ofs = argument(2); // type int
Node* limit = argument(3); // type int
const Type* src_type = src->Value(&_gvn);
const TypeAryPtr* top_src = src_type->isa_aryptr();
if (top_src == NULL || top_src->klass() == NULL) {
// failed array check
return false;
}
// Figure out the size and type of the elements we will be copying.
BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
if (src_elem != T_BYTE) {
return false;
}
// 'src_start' points to src array + offset
Node* src_start = array_element_address(src, ofs, src_elem);
const char* klass_SHA_name = NULL;
const char* stub_name = NULL;
address stub_addr = NULL;
bool long_state = false;
switch (predicate) {
case 0:
if (UseSHA1Intrinsics) {
klass_SHA_name = "sun/security/provider/SHA";
stub_name = "sha1_implCompressMB";
stub_addr = StubRoutines::sha1_implCompressMB();
}
break;
case 1:
if (UseSHA256Intrinsics) {
klass_SHA_name = "sun/security/provider/SHA2";
stub_name = "sha256_implCompressMB";
stub_addr = StubRoutines::sha256_implCompressMB();
}
break;
case 2:
if (UseSHA512Intrinsics) {
klass_SHA_name = "sun/security/provider/SHA5";
stub_name = "sha512_implCompressMB";
stub_addr = StubRoutines::sha512_implCompressMB();
long_state = true;
}
break;
default:
fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
}
if (klass_SHA_name != NULL) {
// get DigestBase klass to lookup for SHA klass
const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
assert(tinst != NULL, "digestBase_obj is not instance???");
assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
}
return false;
}
//------------------------------inline_sha_implCompressMB-----------------------
bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
bool long_state, address stubAddr, const char *stubName,
Node* src_start, Node* ofs, Node* limit) {
const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
const TypeOopPtr* xtype = aklass->as_instance_type();
Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype);
sha_obj = _gvn.transform(sha_obj);
Node* state;
if (long_state) {
state = get_state_from_sha5_object(sha_obj);
} else {
state = get_state_from_sha_object(sha_obj);
}
if (state == NULL) return false;
// Call the stub.
Node *call;
if (CCallingConventionRequiresIntsAsLongs) {
call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::digestBase_implCompressMB_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state, ofs XTOP, limit XTOP);
} else {
call = make_runtime_call(RC_LEAF|RC_NO_FP,
OptoRuntime::digestBase_implCompressMB_Type(),
stubAddr, stubName, TypePtr::BOTTOM,
src_start, state, ofs, limit);
}
// return ofs (int)
Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
set_result(result);
return true;
}
//------------------------------get_state_from_sha_object-----------------------
Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
if (sha_state == NULL) return (Node *) NULL;
// now have the array, need to get the start address of the state array
Node* state = array_element_address(sha_state, intcon(0), T_INT);
return state;
}
//------------------------------get_state_from_sha5_object-----------------------
Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
if (sha_state == NULL) return (Node *) NULL;
// now have the array, need to get the start address of the state array
Node* state = array_element_address(sha_state, intcon(0), T_LONG);
return state;
}
//----------------------------inline_digestBase_implCompressMB_predicate----------------------------
// Return node representing slow path of predicate check.
// the pseudo code we want to emulate with this predicate is:
// if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
//
Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
"need SHA1/SHA256/SHA512 instruction support");
assert((uint)predicate < 3, "sanity");
// The receiver was checked for NULL already.
Node* digestBaseObj = argument(0);
// get DigestBase klass for instanceOf check
const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
assert(tinst != NULL, "digestBaseObj is null");
assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
const char* klass_SHA_name = NULL;
switch (predicate) {
case 0:
if (UseSHA1Intrinsics) {
// we want to do an instanceof comparison against the SHA class
klass_SHA_name = "sun/security/provider/SHA";
}
break;
case 1:
if (UseSHA256Intrinsics) {
// we want to do an instanceof comparison against the SHA2 class
klass_SHA_name = "sun/security/provider/SHA2";
}
break;
case 2:
if (UseSHA512Intrinsics) {
// we want to do an instanceof comparison against the SHA5 class
klass_SHA_name = "sun/security/provider/SHA5";
}
break;
default:
fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
}
ciKlass* klass_SHA = NULL;
if (klass_SHA_name != NULL) {
klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
}
if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
// if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
Node* ctrl = control();
set_control(top()); // no intrinsic path
return ctrl;
}
ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1)));
Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
return instof_false; // even if it is NULL
}
bool LibraryCallKit::inline_profileBoolean() {
Node* counts = argument(1);
const TypeAryPtr* ary = NULL;
ciArray* aobj = NULL;
if (counts->is_Con()
&& (ary = counts->bottom_type()->isa_aryptr()) != NULL
&& (aobj = ary->const_oop()->as_array()) != NULL
&& (aobj->length() == 2)) {
// Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
jint false_cnt = aobj->element_value(0).as_int();
jint true_cnt = aobj->element_value(1).as_int();
if (C->log() != NULL) {
C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
false_cnt, true_cnt);
}
if (false_cnt + true_cnt == 0) {
// According to profile, never executed.
uncommon_trap_exact(Deoptimization::Reason_intrinsic,
Deoptimization::Action_reinterpret);
return true;
}
// result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
// is a number of each value occurrences.
Node* result = argument(0);
if (false_cnt == 0 || true_cnt == 0) {
// According to profile, one value has been never seen.
int expected_val = (false_cnt == 0) ? 1 : 0;
Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val)));
Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
Node* fast_path = _gvn.transform(new (C) IfTrueNode(check));
Node* slow_path = _gvn.transform(new (C) IfFalseNode(check));
{ // Slow path: uncommon trap for never seen value and then reexecute
// MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
// the value has been seen at least once.
PreserveJVMState pjvms(this);
PreserveReexecuteState preexecs(this);
jvms()->set_should_reexecute(true);
set_control(slow_path);
set_i_o(i_o());
uncommon_trap_exact(Deoptimization::Reason_intrinsic,
Deoptimization::Action_reinterpret);
}
// The guard for never seen value enables sharpening of the result and
// returning a constant. It allows to eliminate branches on the same value
// later on.
set_control(fast_path);
result = intcon(expected_val);
}
// Stop profiling.
// MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
// By replacing method body with profile data (represented as ProfileBooleanNode
// on IR level) we effectively disable profiling.
// It enables full speed execution once optimized code is generated.
Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt));
C->record_for_igvn(profile);
set_result(profile);
return true;
} else {
// Continue profiling.
// Profile data isn't available at the moment. So, execute method's bytecode version.
// Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
// is compiled and counters aren't available since corresponding MethodHandle
// isn't a compile-time constant.
return false;
}
}
C:\hotspot-69087d08d473\src\share\vm/opto/live.cpp
/*
* Copyright (c) 1997, 2013, 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 "memory/allocation.inline.hpp"
#include "opto/callnode.hpp"
#include "opto/chaitin.hpp"
#include "opto/live.hpp"
#include "opto/machnode.hpp"
// Compute live-in/live-out. We use a totally incremental algorithm. The LIVE
// problem is monotonic. The steady-state solution looks like this: pull a
// block from the worklist. It has a set of delta's - values which are newly
// live-in from the block. Push these to the live-out sets of all predecessor
// blocks. At each predecessor, the new live-out values are ANDed with what is
// already live-out (extra stuff is added to the live-out sets). Then the
// remaining new live-out values are ANDed with what is locally defined.
// Leftover bits become the new live-in for the predecessor block, and the pred
// block is put on the worklist.
// The locally live-in stuff is computed once and added to predecessor
// live-out sets. This separate compilation is done in the outer loop below.
PhaseLive::PhaseLive( const PhaseCFG &cfg, const LRG_List &names, Arena *arena ) : Phase(LIVE), _cfg(cfg), _names(names), _arena(arena), _live(0) {
}
void PhaseLive::compute(uint maxlrg) {
_maxlrg = maxlrg;
_worklist = new (_arena) Block_List();
// Init the sparse live arrays. This data is live on exit from here!
// The _live info is the live-out info.
_live = (IndexSet*)_arena->Amalloc(sizeof(IndexSet) * _cfg.number_of_blocks());
uint i;
for (i = 0; i < _cfg.number_of_blocks(); i++) {
_live[i].initialize(_maxlrg);
}
// Init the sparse arrays for delta-sets.
ResourceMark rm; // Nuke temp storage on exit
// Does the memory used by _defs and _deltas get reclaimed? Does it matter? TT
// Array of values defined locally in blocks
_defs = NEW_RESOURCE_ARRAY(IndexSet,_cfg.number_of_blocks());
for (i = 0; i < _cfg.number_of_blocks(); i++) {
_defs[i].initialize(_maxlrg);
}
// Array of delta-set pointers, indexed by block pre_order-1.
_deltas = NEW_RESOURCE_ARRAY(IndexSet*,_cfg.number_of_blocks());
memset( _deltas, 0, sizeof(IndexSet*)* _cfg.number_of_blocks());
_free_IndexSet = NULL;
// Blocks having done pass-1
VectorSet first_pass(Thread::current()->resource_area());
// Outer loop: must compute local live-in sets and push into predecessors.
for (uint j = _cfg.number_of_blocks(); j > 0; j--) {
Block* block = _cfg.get_block(j - 1);
// Compute the local live-in set. Start with any new live-out bits.
IndexSet* use = getset(block);
IndexSet* def = &_defs[block->_pre_order-1];
DEBUG_ONLY(IndexSet *def_outside = getfreeset();)
uint i;
for (i = block->number_of_nodes(); i > 1; i--) {
Node* n = block->get_node(i-1);
if (n->is_Phi()) {
break;
}
uint r = _names.at(n->_idx);
assert(!def_outside->member(r), "Use of external LRG overlaps the same LRG defined in this block");
def->insert( r );
use->remove( r );
uint cnt = n->req();
for (uint k = 1; k < cnt; k++) {
Node *nk = n->in(k);
uint nkidx = nk->_idx;
if (_cfg.get_block_for_node(nk) != block) {
uint u = _names.at(nkidx);
use->insert(u);
DEBUG_ONLY(def_outside->insert(u);)
}
}
}
#ifdef ASSERT
def_outside->set_next(_free_IndexSet);
_free_IndexSet = def_outside; // Drop onto free list
#endif
// Remove anything defined by Phis and the block start instruction
for (uint k = i; k > 0; k--) {
uint r = _names.at(block->get_node(k - 1)->_idx);
def->insert(r);
use->remove(r);
}
// Push these live-in things to predecessors
for (uint l = 1; l < block->num_preds(); l++) {
Block* p = _cfg.get_block_for_node(block->pred(l));
add_liveout(p, use, first_pass);
// PhiNode uses go in the live-out set of prior blocks.
for (uint k = i; k > 0; k--) {
add_liveout(p, _names.at(block->get_node(k-1)->in(l)->_idx), first_pass);
}
}
freeset(block);
first_pass.set(block->_pre_order);
// Inner loop: blocks that picked up new live-out values to be propagated
while (_worklist->size()) {
Block* block = _worklist->pop();
IndexSet *delta = getset(block);
assert( delta->count(), "missing delta set" );
// Add new-live-in to predecessors live-out sets
for (uint l = 1; l < block->num_preds(); l++) {
Block* predecessor = _cfg.get_block_for_node(block->pred(l));
add_liveout(predecessor, delta, first_pass);
}
freeset(block);
} // End of while-worklist-not-empty
} // End of for-all-blocks-outer-loop
// We explicitly clear all of the IndexSets which we are about to release.
// This allows us to recycle their internal memory into IndexSet's free list.
for (i = 0; i < _cfg.number_of_blocks(); i++) {
_defs[i].clear();
if (_deltas[i]) {
// Is this always true?
_deltas[i]->clear();
}
}
IndexSet *free = _free_IndexSet;
while (free != NULL) {
IndexSet *temp = free;
free = free->next();
temp->clear();
}
}
#ifndef PRODUCT
void PhaseLive::stats(uint iters) const {
}
#endif
// Get an IndexSet for a block. Return existing one, if any. Make a new
// empty one if a prior one does not exist.
IndexSet *PhaseLive::getset( Block *p ) {
IndexSet *delta = _deltas[p->_pre_order-1];
if( !delta ) // Not on worklist?
// Get a free set; flag as being on worklist
delta = _deltas[p->_pre_order-1] = getfreeset();
return delta; // Return set of new live-out items
}
// Pull from free list, or allocate. Internal allocation on the returned set
// is always from thread local storage.
IndexSet *PhaseLive::getfreeset( ) {
IndexSet *f = _free_IndexSet;
if( !f ) {
f = new IndexSet;
// f->set_arena(Thread::current()->resource_area());
f->initialize(_maxlrg, Thread::current()->resource_area());
} else {
// Pull from free list
_free_IndexSet = f->next();
//f->_cnt = 0; // Reset to empty
// f->set_arena(Thread::current()->resource_area());
f->initialize(_maxlrg, Thread::current()->resource_area());
}
return f;
}
// Free an IndexSet from a block.
void PhaseLive::freeset( const Block *p ) {
IndexSet *f = _deltas[p->_pre_order-1];
f->set_next(_free_IndexSet);
_free_IndexSet = f; // Drop onto free list
_deltas[p->_pre_order-1] = NULL;
}
// Add a live-out value to a given blocks live-out set. If it is new, then
// also add it to the delta set and stick the block on the worklist.
void PhaseLive::add_liveout( Block *p, uint r, VectorSet &first_pass ) {
IndexSet *live = &_live[p->_pre_order-1];
if( live->insert(r) ) { // If actually inserted...
// We extended the live-out set. See if the value is generated locally.
// If it is not, then we must extend the live-in set.
if( !_defs[p->_pre_order-1].member( r ) ) {
if( !_deltas[p->_pre_order-1] && // Not on worklist?
first_pass.test(p->_pre_order) )
_worklist->push(p); // Actually go on worklist if already 1st pass
getset(p)->insert(r);
}
}
}
// Add a vector of live-out values to a given blocks live-out set.
void PhaseLive::add_liveout( Block *p, IndexSet *lo, VectorSet &first_pass ) {
IndexSet *live = &_live[p->_pre_order-1];
IndexSet *defs = &_defs[p->_pre_order-1];
IndexSet *on_worklist = _deltas[p->_pre_order-1];
IndexSet *delta = on_worklist ? on_worklist : getfreeset();
IndexSetIterator elements(lo);
uint r;
while ((r = elements.next()) != 0) {
if( live->insert(r) && // If actually inserted...
!defs->member( r ) ) // and not defined locally
delta->insert(r); // Then add to live-in set
}
if( delta->count() ) { // If actually added things
_deltas[p->_pre_order-1] = delta; // Flag as on worklist now
if( !on_worklist && // Not on worklist?
first_pass.test(p->_pre_order) )
_worklist->push(p); // Actually go on worklist if already 1st pass
} else { // Nothing there; just free it
delta->set_next(_free_IndexSet);
_free_IndexSet = delta; // Drop onto free list
}
}
#ifndef PRODUCT
// Dump the live-out set for a block
void PhaseLive::dump( const Block *b ) const {
tty->print("Block %d: ",b->_pre_order);
tty->print("LiveOut: "); _live[b->_pre_order-1].dump();
uint cnt = b->number_of_nodes();
for( uint i=0; i<cnt; i++ ) {
tty->print("L%d/", _names.at(b->get_node(i)->_idx));
b->get_node(i)->dump();
}
tty->print("\n");
}
// Verify that base pointers and derived pointers are still sane.
void PhaseChaitin::verify_base_ptrs( ResourceArea *a ) const {
#ifdef ASSERT
Unique_Node_List worklist(a);
for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
Block* block = _cfg.get_block(i);
for (uint j = block->end_idx() + 1; j > 1; j--) {
Node* n = block->get_node(j-1);
if (n->is_Phi()) {
break;
}
// Found a safepoint?
if (n->is_MachSafePoint()) {
MachSafePointNode *sfpt = n->as_MachSafePoint();
JVMState* jvms = sfpt->jvms();
if (jvms != NULL) {
// Now scan for a live derived pointer
if (jvms->oopoff() < sfpt->req()) {
// Check each derived/base pair
for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) {
Node *check = sfpt->in(idx);
bool is_derived = ((idx - jvms->oopoff()) & 1) == 0;
// search upwards through spills and spill phis for AddP
worklist.clear();
worklist.push(check);
uint k = 0;
while( k < worklist.size() ) {
check = worklist.at(k);
assert(check,"Bad base or derived pointer");
// See PhaseChaitin::find_base_for_derived() for all cases.
int isc = check->is_Copy();
if( isc ) {
worklist.push(check->in(isc));
} else if( check->is_Phi() ) {
for (uint m = 1; m < check->req(); m++)
worklist.push(check->in(m));
} else if( check->is_Con() ) {
if (is_derived) {
// Derived is NULL+offset
assert(!is_derived || check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad derived pointer");
} else {
assert(check->bottom_type()->is_ptr()->_offset == 0,"Bad base pointer");
// Base either ConP(NULL) or loadConP
if (check->is_Mach()) {
assert(check->as_Mach()->ideal_Opcode() == Op_ConP,"Bad base pointer");
} else {
assert(check->Opcode() == Op_ConP &&
check->bottom_type()->is_ptr()->ptr() == TypePtr::Null,"Bad base pointer");
}
}
} else if( check->bottom_type()->is_ptr()->_offset == 0 ) {
if(check->is_Proj() || check->is_Mach() &&
(check->as_Mach()->ideal_Opcode() == Op_CreateEx ||
check->as_Mach()->ideal_Opcode() == Op_ThreadLocal ||
check->as_Mach()->ideal_Opcode() == Op_CMoveP ||
check->as_Mach()->ideal_Opcode() == Op_CheckCastPP ||
#ifdef _LP64
UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP ||
UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN ||
UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass ||
#endif
check->as_Mach()->ideal_Opcode() == Op_LoadP ||
check->as_Mach()->ideal_Opcode() == Op_LoadKlass)) {
// Valid nodes
} else {
check->dump();
assert(false,"Bad base or derived pointer");
}
} else {
assert(is_derived,"Bad base pointer");
assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP,"Bad derived pointer");
}
k++;
assert(k < 100000,"Derived pointer checking in infinite loop");
} // End while
}
} // End of check for derived pointers
} // End of Kcheck for debug info
} // End of if found a safepoint
} // End of forall instructions in block
} // End of forall blocks
#endif
}
// Verify that graphs and base pointers are still sane.
void PhaseChaitin::verify( ResourceArea *a, bool verify_ifg ) const {
#ifdef ASSERT
if( VerifyOpto || VerifyRegisterAllocator ) {
_cfg.verify();
verify_base_ptrs(a);
if(verify_ifg)
_ifg->verify(this);
}
#endif
}
#endif
C:\hotspot-69087d08d473\src\share\vm/opto/live.hpp
/*
* Copyright (c) 1997, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_LIVE_HPP
#define SHARE_VM_OPTO_LIVE_HPP
#include "libadt/port.hpp"
#include "libadt/vectset.hpp"
#include "opto/block.hpp"
#include "opto/indexSet.hpp"
#include "opto/phase.hpp"
#include "opto/regmask.hpp"
class Block;
class PhaseCFG;
class VectorSet;
class IndexSet;
//------------------------------LRG_List---------------------------------------
// Map Node indices to Live RanGe indices.
// Array lookup in the optimized case.
typedef GrowableArray<uint> LRG_List;
//------------------------------PhaseLive--------------------------------------
// Compute live-in/live-out
class PhaseLive : public Phase {
// Array of Sets of values live at the start of a block.
// Indexed by block pre-order number.
IndexSet *_live;
// Array of Sets of values defined locally in the block
// Indexed by block pre-order number.
IndexSet *_defs;
// Array of delta-set pointers, indexed by block pre-order number
IndexSet **_deltas;
IndexSet *_free_IndexSet; // Free list of same
Block_List *_worklist; // Worklist for iterative solution
const PhaseCFG &_cfg; // Basic blocks
const LRG_List &_names; // Mapping from Nodes to live ranges
uint _maxlrg; // Largest live-range number
Arena *_arena;
IndexSet *getset( Block *p );
IndexSet *getfreeset( );
void freeset( const Block *p );
void add_liveout( Block *p, uint r, VectorSet &first_pass );
void add_liveout( Block *p, IndexSet *lo, VectorSet &first_pass );
public:
PhaseLive(const PhaseCFG &cfg, const LRG_List &names, Arena *arena);
~PhaseLive() {}
// Compute liveness info
void compute(uint maxlrg);
// Reset arena storage
void reset() { _live = NULL; }
// Return the live-out set for this block
IndexSet *live( const Block * b ) { return &_live[b->_pre_order-1]; }
#ifndef PRODUCT
void dump( const Block *b ) const;
void stats(uint iters) const;
#endif
};
#endif // SHARE_VM_OPTO_LIVE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/locknode.cpp
/*
* Copyright (c) 1999, 2012, 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/locknode.hpp"
#include "opto/parse.hpp"
#include "opto/rootnode.hpp"
#include "opto/runtime.hpp"
//=============================================================================
const RegMask &BoxLockNode::in_RegMask(uint i) const {
return _inmask;
}
const RegMask &BoxLockNode::out_RegMask() const {
return *Matcher::idealreg2regmask[Op_RegP];
}
uint BoxLockNode::size_of() const { return sizeof(*this); }
BoxLockNode::BoxLockNode( int slot ) : Node( Compile::current()->root() ),
_slot(slot), _is_eliminated(false) {
init_class_id(Class_BoxLock);
init_flags(Flag_rematerialize);
OptoReg::Name reg = OptoReg::stack2reg(_slot);
_inmask.Insert(reg);
}
//-----------------------------hash--------------------------------------------
uint BoxLockNode::hash() const {
if (EliminateNestedLocks)
return NO_HASH; // Each locked region has own BoxLock node
return Node::hash() + _slot + (_is_eliminated ? Compile::current()->fixed_slots() : 0);
}
//------------------------------cmp--------------------------------------------
uint BoxLockNode::cmp( const Node &n ) const {
if (EliminateNestedLocks)
return (&n == this); // Always fail except on self
const BoxLockNode &bn = (const BoxLockNode &)n;
return bn._slot == _slot && bn._is_eliminated == _is_eliminated;
}
BoxLockNode* BoxLockNode::box_node(Node* box) {
// Chase down the BoxNode after RA which may spill box nodes.
while (!box->is_BoxLock()) {
// if (box_node->is_SpillCopy()) {
// Node *m = box_node->in(1);
// if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_StoreP) {
// box_node = m->in(m->as_Mach()->operand_index(2));
// continue;
// }
// }
assert(box->is_SpillCopy() || box->is_Phi(), "Bad spill of Lock.");
// Only BoxLock nodes with the same stack slot are merged.
// So it is enough to trace one path to find the slot value.
box = box->in(1);
}
return box->as_BoxLock();
}
OptoReg::Name BoxLockNode::reg(Node* box) {
return box_node(box)->in_RegMask(0).find_first_elem();
}
// Is BoxLock node used for one simple lock region (same box and obj)?
bool BoxLockNode::is_simple_lock_region(LockNode** unique_lock, Node* obj) {
LockNode* lock = NULL;
bool has_one_lock = false;
for (uint i = 0; i < this->outcnt(); i++) {
Node* n = this->raw_out(i);
assert(!n->is_Phi(), "should not merge BoxLock nodes");
if (n->is_AbstractLock()) {
AbstractLockNode* alock = n->as_AbstractLock();
// Check lock's box since box could be referenced by Lock's debug info.
if (alock->box_node() == this) {
if (alock->obj_node()->eqv_uncast(obj)) {
if ((unique_lock != NULL) && alock->is_Lock()) {
if (lock == NULL) {
lock = alock->as_Lock();
has_one_lock = true;
} else if (lock != alock->as_Lock()) {
has_one_lock = false;
}
}
} else {
return false; // Different objects
}
}
}
}
#ifdef ASSERT
// Verify that FastLock and Safepoint reference only this lock region.
for (uint i = 0; i < this->outcnt(); i++) {
Node* n = this->raw_out(i);
if (n->is_FastLock()) {
FastLockNode* flock = n->as_FastLock();
assert((flock->box_node() == this) && flock->obj_node()->eqv_uncast(obj),"");
}
// Don't check monitor info in safepoints since the referenced object could
// be different from the locked object. It could be Phi node of different
// cast nodes which point to this locked object.
// We assume that no other objects could be referenced in monitor info
// associated with this BoxLock node because all associated locks and
// unlocks are reference only this one object.
}
#endif
if (unique_lock != NULL && has_one_lock) {
*unique_lock = lock;
}
return true;
}
//=============================================================================
//-----------------------------hash--------------------------------------------
uint FastLockNode::hash() const { return NO_HASH; }
uint FastLockNode::size_of() const { return sizeof(*this); }
//------------------------------cmp--------------------------------------------
uint FastLockNode::cmp( const Node &n ) const {
return (&n == this); // Always fail except on self
}
//=============================================================================
//-----------------------------hash--------------------------------------------
uint FastUnlockNode::hash() const { return NO_HASH; }
//------------------------------cmp--------------------------------------------
uint FastUnlockNode::cmp( const Node &n ) const {
return (&n == this); // Always fail except on self
}
//
// Create a counter which counts the number of times this lock is acquired
//
void FastLockNode::create_lock_counter(JVMState* state) {
BiasedLockingNamedCounter* blnc = (BiasedLockingNamedCounter*)
OptoRuntime::new_named_counter(state, NamedCounter::BiasedLockingCounter);
_counters = blnc->counters();
}
void FastLockNode::create_rtm_lock_counter(JVMState* state) {
#if INCLUDE_RTM_OPT
Compile* C = Compile::current();
if (C->profile_rtm() || (PrintPreciseRTMLockingStatistics && C->use_rtm())) {
RTMLockingNamedCounter* rlnc = (RTMLockingNamedCounter*)
OptoRuntime::new_named_counter(state, NamedCounter::RTMLockingCounter);
_rtm_counters = rlnc->counters();
if (UseRTMForStackLocks) {
rlnc = (RTMLockingNamedCounter*)
OptoRuntime::new_named_counter(state, NamedCounter::RTMLockingCounter);
_stack_rtm_counters = rlnc->counters();
}
}
#endif
}
//=============================================================================
//------------------------------do_monitor_enter-------------------------------
void Parse::do_monitor_enter() {
kill_dead_locals();
// Null check; get casted pointer.
Node* obj = null_check(peek());
// Check for locking null object
if (stopped()) return;
// the monitor object is not part of debug info expression stack
pop();
// Insert a FastLockNode which takes as arguments the current thread pointer,
// the obj pointer & the address of the stack slot pair used for the lock.
shared_lock(obj);
}
//------------------------------do_monitor_exit--------------------------------
void Parse::do_monitor_exit() {
kill_dead_locals();
pop(); // Pop oop to unlock
// Because monitors are guaranteed paired (else we bail out), we know
// the matching Lock for this Unlock. Hence we know there is no need
// for a null check on Unlock.
shared_unlock(map()->peek_monitor_box(), map()->peek_monitor_obj());
}
C:\hotspot-69087d08d473\src\share\vm/opto/locknode.hpp
/*
* Copyright (c) 1999, 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.
*
*/
#ifndef SHARE_VM_OPTO_LOCKNODE_HPP
#define SHARE_VM_OPTO_LOCKNODE_HPP
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/subnode.hpp"
#if defined AD_MD_HPP
# include AD_MD_HPP
#elif defined TARGET_ARCH_MODEL_x86_32
# include "adfiles/ad_x86_32.hpp"
#elif defined TARGET_ARCH_MODEL_x86_64
# include "adfiles/ad_x86_64.hpp"
#elif defined TARGET_ARCH_MODEL_aarch64
# include "adfiles/ad_aarch64.hpp"
#elif defined TARGET_ARCH_MODEL_sparc
# include "adfiles/ad_sparc.hpp"
#elif defined TARGET_ARCH_MODEL_zero
# include "adfiles/ad_zero.hpp"
#elif defined TARGET_ARCH_MODEL_ppc_64
# include "adfiles/ad_ppc_64.hpp"
#endif
//------------------------------BoxLockNode------------------------------------
class BoxLockNode : public Node {
const int _slot; // stack slot
RegMask _inmask; // OptoReg corresponding to stack slot
bool _is_eliminated; // Associated locks were safely eliminated
public:
BoxLockNode( int lock );
virtual int Opcode() const;
virtual void emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const;
virtual uint size(PhaseRegAlloc *ra_) const;
virtual const RegMask &in_RegMask(uint) const;
virtual const RegMask &out_RegMask() const;
virtual uint size_of() const;
virtual uint hash() const;
virtual uint cmp( const Node &n ) const;
virtual const class Type *bottom_type() const { return TypeRawPtr::BOTTOM; }
virtual uint ideal_reg() const { return Op_RegP; }
static OptoReg::Name reg(Node* box_node);
static BoxLockNode* box_node(Node* box_node);
static bool same_slot(Node* box1, Node* box2) {
return box1->as_BoxLock()->_slot == box2->as_BoxLock()->_slot;
}
int stack_slot() const { return _slot; }
bool is_eliminated() const { return _is_eliminated; }
// mark lock as eliminated.
void set_eliminated() { _is_eliminated = true; }
// Is BoxLock node used for one simple lock region?
bool is_simple_lock_region(LockNode** unique_lock, Node* obj);
#ifndef PRODUCT
virtual void format( PhaseRegAlloc *, outputStream *st ) const;
virtual void dump_spec(outputStream *st) const { st->print(" Lock %d",_slot); }
#endif
};
//------------------------------FastLockNode-----------------------------------
class FastLockNode: public CmpNode {
private:
BiasedLockingCounters* _counters;
RTMLockingCounters* _rtm_counters; // RTM lock counters for inflated locks
RTMLockingCounters* _stack_rtm_counters; // RTM lock counters for stack locks
public:
FastLockNode(Node *ctrl, Node *oop, Node *box) : CmpNode(oop,box) {
init_req(0,ctrl);
init_class_id(Class_FastLock);
_counters = NULL;
_rtm_counters = NULL;
_stack_rtm_counters = NULL;
}
Node* obj_node() const { return in(1); }
Node* box_node() const { return in(2); }
void set_box_node(Node* box) { set_req(2, box); }
// FastLock and FastUnlockNode do not hash, we need one for each correspoding
// LockNode/UnLockNode to avoid creating Phi's.
virtual uint hash() const ; // { return NO_HASH; }
virtual uint size_of() const;
virtual uint cmp( const Node &n ) const ; // Always fail, except on self
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const { return TypeInt::CC; }
const Type *sub(const Type *t1, const Type *t2) const { return TypeInt::CC;}
void create_lock_counter(JVMState* s);
void create_rtm_lock_counter(JVMState* state);
BiasedLockingCounters* counters() const { return _counters; }
RTMLockingCounters* rtm_counters() const { return _rtm_counters; }
RTMLockingCounters* stack_rtm_counters() const { return _stack_rtm_counters; }
};
//------------------------------FastUnlockNode---------------------------------
class FastUnlockNode: public CmpNode {
public:
FastUnlockNode(Node *ctrl, Node *oop, Node *box) : CmpNode(oop,box) {
init_req(0,ctrl);
init_class_id(Class_FastUnlock);
}
Node* obj_node() const { return in(1); }
Node* box_node() const { return in(2); }
// FastLock and FastUnlockNode do not hash, we need one for each correspoding
// LockNode/UnLockNode to avoid creating Phi's.
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const ; // Always fail, except on self
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const { return TypeInt::CC; }
const Type *sub(const Type *t1, const Type *t2) const { return TypeInt::CC;}
};
#endif // SHARE_VM_OPTO_LOCKNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/loopnode.cpp
/*
* Copyright (c) 1998, 2020, 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 "ci/ciMethodData.hpp"
#include "compiler/compileLog.hpp"
#include "libadt/vectset.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/callnode.hpp"
#include "opto/connode.hpp"
#include "opto/divnode.hpp"
#include "opto/idealGraphPrinter.hpp"
#include "opto/loopnode.hpp"
#include "opto/mulnode.hpp"
#include "opto/rootnode.hpp"
#include "opto/superword.hpp"
//=============================================================================
//------------------------------is_loop_iv-------------------------------------
// Determine if a node is Counted loop induction variable.
// The method is declared in node.hpp.
const Node* Node::is_loop_iv() const {
if (this->is_Phi() && !this->as_Phi()->is_copy() &&
this->as_Phi()->region()->is_CountedLoop() &&
this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
return this;
} else {
return NULL;
}
}
//=============================================================================
//------------------------------dump_spec--------------------------------------
// Dump special per-node info
#ifndef PRODUCT
void LoopNode::dump_spec(outputStream *st) const {
if (is_inner_loop()) st->print( "inner " );
if (is_partial_peel_loop()) st->print( "partial_peel " );
if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
}
#endif
//------------------------------is_valid_counted_loop-------------------------
bool LoopNode::is_valid_counted_loop() const {
if (is_CountedLoop()) {
CountedLoopNode* l = as_CountedLoop();
CountedLoopEndNode* le = l->loopexit();
if (le != NULL &&
le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
Node* phi = l->phi();
Node* exit = le->proj_out(0 /* false */);
if (exit != NULL && exit->Opcode() == Op_IfFalse &&
phi != NULL && phi->is_Phi() &&
phi->in(LoopNode::LoopBackControl) == l->incr() &&
le->loopnode() == l && le->stride_is_con()) {
return true;
}
}
}
return false;
}
//------------------------------get_early_ctrl---------------------------------
// Compute earliest legal control
Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
uint i;
Node *early;
if (n->in(0) && !n->is_expensive()) {
early = n->in(0);
if (!early->is_CFG()) // Might be a non-CFG multi-def
early = get_ctrl(early); // So treat input as a straight data input
i = 1;
} else {
early = get_ctrl(n->in(1));
i = 2;
}
uint e_d = dom_depth(early);
assert( early, "" );
for (; i < n->req(); i++) {
Node *cin = get_ctrl(n->in(i));
assert( cin, "" );
// Keep deepest dominator depth
uint c_d = dom_depth(cin);
if (c_d > e_d) { // Deeper guy?
early = cin; // Keep deepest found so far
e_d = c_d;
} else if (c_d == e_d && // Same depth?
early != cin) { // If not equal, must use slower algorithm
// If same depth but not equal, one _must_ dominate the other
// and we want the deeper (i.e., dominated) guy.
Node *n1 = early;
Node *n2 = cin;
while (1) {
n1 = idom(n1); // Walk up until break cycle
n2 = idom(n2);
if (n1 == cin || // Walked early up to cin
dom_depth(n2) < c_d)
break; // early is deeper; keep him
if (n2 == early || // Walked cin up to early
dom_depth(n1) < c_d) {
early = cin; // cin is deeper; keep him
break;
}
}
e_d = dom_depth(early); // Reset depth register cache
}
}
// Return earliest legal location
assert(early == find_non_split_ctrl(early), "unexpected early control");
if (n->is_expensive()) {
assert(n->in(0), "should have control input");
early = get_early_ctrl_for_expensive(n, early);
}
return early;
}
//------------------------------get_early_ctrl_for_expensive---------------------------------
// Move node up the dominator tree as high as legal while still beneficial
Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
assert(OptimizeExpensiveOps, "optimization off?");
Node* ctl = n->in(0);
assert(ctl->is_CFG(), "expensive input 0 must be cfg");
uint min_dom_depth = dom_depth(earliest);
#ifdef ASSERT
if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
}
#endif
if (dom_depth(ctl) < min_dom_depth) {
return earliest;
}
while (1) {
Node *next = ctl;
// Moving the node out of a loop on the projection of a If
// confuses loop predication. So once we hit a Loop in a If branch
// that doesn't branch to an UNC, we stop. The code that process
// expensive nodes will notice the loop and skip over it to try to
// move the node further up.
if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
break;
}
next = idom(ctl->in(1)->in(0));
} else if (ctl->is_Proj()) {
// We only move it up along a projection if the projection is
// the single control projection for its parent: same code path,
// if it's a If with UNC or fallthrough of a call.
Node* parent_ctl = ctl->in(0);
if (parent_ctl == NULL) {
break;
} else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
} else if (parent_ctl->is_If()) {
if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
break;
}
assert(idom(ctl) == parent_ctl, "strange");
next = idom(parent_ctl);
} else if (ctl->is_CatchProj()) {
if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
break;
}
assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
next = parent_ctl->in(0)->in(0)->in(0);
} else {
// Check if parent control has a single projection (this
// control is the only possible successor of the parent
// control). If so, we can try to move the node above the
// parent control.
int nb_ctl_proj = 0;
for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
Node *p = parent_ctl->fast_out(i);
if (p->is_Proj() && p->is_CFG()) {
nb_ctl_proj++;
if (nb_ctl_proj > 1) {
break;
}
}
}
if (nb_ctl_proj > 1) {
break;
}
assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
assert(idom(ctl) == parent_ctl, "strange");
next = idom(parent_ctl);
}
} else {
next = idom(ctl);
}
if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
break;
}
ctl = next;
}
if (ctl != n->in(0)) {
_igvn.hash_delete(n);
n->set_req(0, ctl);
_igvn.hash_insert(n);
}
return ctl;
}
//------------------------------set_early_ctrl---------------------------------
// Set earliest legal control
void PhaseIdealLoop::set_early_ctrl( Node *n ) {
Node *early = get_early_ctrl(n);
// Record earliest legal location
set_ctrl(n, early);
}
//------------------------------set_subtree_ctrl-------------------------------
// set missing _ctrl entries on new nodes
void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
// Already set? Get out.
if( _nodes[n->_idx] ) return;
// Recursively set _nodes array to indicate where the Node goes
uint i;
for( i = 0; i < n->req(); ++i ) {
Node *m = n->in(i);
if( m && m != C->root() )
set_subtree_ctrl( m );
}
// Fixup self
set_early_ctrl( n );
}
//------------------------------is_counted_loop--------------------------------
bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
PhaseGVN *gvn = &_igvn;
// Counted loop head must be a good RegionNode with only 3 not NULL
// control input edges: Self, Entry, LoopBack.
if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
return false;
}
Node *init_control = x->in(LoopNode::EntryControl);
Node *back_control = x->in(LoopNode::LoopBackControl);
if (init_control == NULL || back_control == NULL) // Partially dead
return false;
// Must also check for TOP when looking for a dead loop
if (init_control->is_top() || back_control->is_top())
return false;
// Allow funny placement of Safepoint
if (back_control->Opcode() == Op_SafePoint) {
if (UseCountedLoopSafepoints) {
// Leaving the safepoint on the backedge and creating a
// CountedLoop will confuse optimizations. We can't move the
// safepoint around because its jvm state wouldn't match a new
// location. Give up on that loop.
return false;
}
back_control = back_control->in(TypeFunc::Control);
}
// Controlling test for loop
Node *iftrue = back_control;
uint iftrue_op = iftrue->Opcode();
if (iftrue_op != Op_IfTrue &&
iftrue_op != Op_IfFalse)
// I have a weird back-control. Probably the loop-exit test is in
// the middle of the loop and I am looking at some trailing control-flow
// merge point. To fix this I would have to partially peel the loop.
return false; // Obscure back-control
// Get boolean guarding loop-back test
Node *iff = iftrue->in(0);
if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
return false;
BoolNode *test = iff->in(1)->as_Bool();
BoolTest::mask bt = test->_test._test;
float cl_prob = iff->as_If()->_prob;
if (iftrue_op == Op_IfFalse) {
bt = BoolTest(bt).negate();
cl_prob = 1.0 - cl_prob;
}
// Get backedge compare
Node *cmp = test->in(1);
int cmp_op = cmp->Opcode();
if (cmp_op != Op_CmpI)
return false; // Avoid pointer & float compares
// Find the trip-counter increment & limit. Limit must be loop invariant.
Node *incr = cmp->in(1);
Node *limit = cmp->in(2);
// ---------
// need 'loop()' test to tell if limit is loop invariant
// ---------
if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
Node *tmp = incr; // Then reverse order into the CmpI
incr = limit;
limit = tmp;
bt = BoolTest(bt).commute(); // And commute the exit test
}
if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
return false;
if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
return false;
Node* phi_incr = NULL;
// Trip-counter increment must be commutative & associative.
if (incr->is_Phi()) {
if (incr->as_Phi()->region() != x || incr->req() != 3)
return false; // Not simple trip counter expression
phi_incr = incr;
incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
return false;
}
Node* trunc1 = NULL;
Node* trunc2 = NULL;
const TypeInt* iv_trunc_t = NULL;
if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
return false; // Funny increment opcode
}
assert(incr->Opcode() == Op_AddI, "wrong increment code");
// Get merge point
Node *xphi = incr->in(1);
Node *stride = incr->in(2);
if (!stride->is_Con()) { // Oops, swap these
if (!xphi->is_Con()) // Is the other guy a constant?
return false; // Nope, unknown stride, bail out
Node *tmp = xphi; // 'incr' is commutative, so ok to swap
xphi = stride;
stride = tmp;
}
// Stride must be constant
int stride_con = stride->get_int();
if (stride_con == 0)
return false; // missed some peephole opt
if (!xphi->is_Phi())
return false; // Too much math on the trip counter
if (phi_incr != NULL && phi_incr != xphi)
return false;
PhiNode *phi = xphi->as_Phi();
// Phi must be of loop header; backedge must wrap to increment
if (phi->region() != x)
return false;
if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
return false;
}
Node *init_trip = phi->in(LoopNode::EntryControl);
// If iv trunc type is smaller than int, check for possible wrap.
if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
assert(trunc1 != NULL, "must have found some truncation");
// Get a better type for the phi (filtered thru if's)
const TypeInt* phi_ft = filtered_type(phi);
// Can iv take on a value that will wrap?
//
// Ensure iv's limit is not within "stride" of the wrap value.
//
// Example for "short" type
// Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
// If the stride is +10, then the last value of the induction
// variable before the increment (phi_ft->_hi) must be
// <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
// ensure no truncation occurs after the increment.
if (stride_con > 0) {
if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
iv_trunc_t->_lo > phi_ft->_lo) {
return false; // truncation may occur
}
} else if (stride_con < 0) {
if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
iv_trunc_t->_hi < phi_ft->_hi) {
return false; // truncation may occur
}
}
// No possibility of wrap so truncation can be discarded
// Promote iv type to Int
} else {
assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
}
// If the condition is inverted and we will be rolling
// through MININT to MAXINT, then bail out.
if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
// Odd stride
bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
// Count down loop rolls through MAXINT
(bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
// Count up loop rolls through MININT
(bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
return false; // Bail out
}
const TypeInt* init_t = gvn->type(init_trip)->is_int();
const TypeInt* limit_t = gvn->type(limit)->is_int();
if (stride_con > 0) {
jlong init_p = (jlong)init_t->_lo + stride_con;
if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
return false; // cyclic loop or this loop trips only once
} else {
jlong init_p = (jlong)init_t->_hi + stride_con;
if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
return false; // cyclic loop or this loop trips only once
}
if (phi_incr != NULL) {
// check if there is a possiblity of IV overflowing after the first increment
if (stride_con > 0) {
if (init_t->_hi > max_jint - stride_con) {
return false;
}
} else {
if (init_t->_lo < min_jint - stride_con) {
return false;
}
}
}
// =================================================
// ---- SUCCESS! Found A Trip-Counted Loop! -----
//
assert(x->Opcode() == Op_Loop, "regular loops only");
C->print_method(PHASE_BEFORE_CLOOPS, 3);
Node *hook = new (C) Node(6);
if (LoopLimitCheck) {
// ===================================================
// Generate loop limit check to avoid integer overflow
// in cases like next (cyclic loops):
//
// for (i=0; i <= max_jint; i++) {}
// for (i=0; i < max_jint; i+=2) {}
//
//
// Limit check predicate depends on the loop test:
//
// for(;i != limit; i++) --> limit <= (max_jint)
// for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
// for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
//
// Check if limit is excluded to do more precise int overflow check.
bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
// If compare points directly to the phi we need to adjust
// the compare so that it points to the incr. Limit have
// to be adjusted to keep trip count the same and the
// adjusted limit should be checked for int overflow.
if (phi_incr != NULL) {
stride_m += stride_con;
}
if (limit->is_Con()) {
int limit_con = limit->get_int();
if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
(stride_con < 0 && limit_con < (min_jint - stride_m))) {
// Bailout: it could be integer overflow.
return false;
}
} else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
(stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
// Limit's type may satisfy the condition, for example,
// when it is an array length.
} else {
// Generate loop's limit check.
// Loop limit check predicate should be near the loop.
ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
if (!limit_check_proj) {
// The limit check predicate is not generated if this method trapped here before.
#ifdef ASSERT
if (TraceLoopLimitCheck) {
tty->print("missing loop limit check:");
loop->dump_head();
x->dump(1);
}
#endif
return false;
}
IfNode* check_iff = limit_check_proj->in(0)->as_If();
Node* cmp_limit;
Node* bol;
if (stride_con > 0) {
cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
bol = new (C) BoolNode(cmp_limit, BoolTest::le);
} else {
cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
}
cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
bol = _igvn.register_new_node_with_optimizer(bol);
set_subtree_ctrl(bol);
// Replace condition in original predicate but preserve Opaque node
// so that previous predicates could be found.
assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
Node* opq = check_iff->in(1)->in(1);
_igvn.hash_delete(opq);
opq->set_req(1, bol);
// Update ctrl.
set_ctrl(opq, check_iff->in(0));
set_ctrl(check_iff->in(1), check_iff->in(0));
#ifndef PRODUCT
// report that the loop predication has been actually performed
// for this loop
if (TraceLoopLimitCheck) {
tty->print_cr("Counted Loop Limit Check generated:");
debug_only( bol->dump(2); )
}
#endif
}
if (phi_incr != NULL) {
// If compare points directly to the phi we need to adjust
// the compare so that it points to the incr. Limit have
// to be adjusted to keep trip count the same and we
// should avoid int overflow.
//
// i = init; do {} while(i++ < limit);
// is converted to
// i = init; do {} while(++i < limit+1);
//
limit = gvn->transform(new (C) AddINode(limit, stride));
}
// Now we need to canonicalize loop condition.
if (bt == BoolTest::ne) {
assert(stride_con == 1 || stride_con == -1, "simple increment only");
// 'ne' can be replaced with 'lt' only when init < limit.
if (stride_con > 0 && init_t->_hi < limit_t->_lo)
bt = BoolTest::lt;
// 'ne' can be replaced with 'gt' only when init > limit.
if (stride_con < 0 && init_t->_lo > limit_t->_hi)
bt = BoolTest::gt;
}
if (incl_limit) {
// The limit check guaranties that 'limit <= (max_jint - stride)' so
// we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
//
Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
limit = gvn->transform(new (C) AddINode(limit, one));
if (bt == BoolTest::le)
bt = BoolTest::lt;
else if (bt == BoolTest::ge)
bt = BoolTest::gt;
else
ShouldNotReachHere();
}
set_subtree_ctrl( limit );
} else { // LoopLimitCheck
// If compare points to incr, we are ok. Otherwise the compare
// can directly point to the phi; in this case adjust the compare so that
// it points to the incr by adjusting the limit.
if (cmp->in(1) == phi || cmp->in(2) == phi)
limit = gvn->transform(new (C) AddINode(limit,stride));
// trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
// Final value for iterator should be: trip_count * stride + init_trip.
Node *one_p = gvn->intcon( 1);
Node *one_m = gvn->intcon(-1);
Node *trip_count = NULL;
switch( bt ) {
case BoolTest::eq:
ShouldNotReachHere();
case BoolTest::ne: // Ahh, the case we desire
if (stride_con == 1)
trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
else if (stride_con == -1)
trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
else
ShouldNotReachHere();
set_subtree_ctrl(trip_count);
//_loop.map(trip_count->_idx,loop(limit));
break;
case BoolTest::le: // Maybe convert to '<' case
limit = gvn->transform(new (C) AddINode(limit,one_p));
set_subtree_ctrl( limit );
hook->init_req(4, limit);
bt = BoolTest::lt;
// Make the new limit be in the same loop nest as the old limit
//_loop.map(limit->_idx,limit_loop);
// Fall into next case
case BoolTest::lt: { // Maybe convert to '!=' case
if (stride_con < 0) // Count down loop rolls through MAXINT
ShouldNotReachHere();
Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
set_subtree_ctrl( range );
hook->init_req(0, range);
Node *bias = gvn->transform(new (C) AddINode(range,stride));
set_subtree_ctrl( bias );
hook->init_req(1, bias);
Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
set_subtree_ctrl( bias1 );
hook->init_req(2, bias1);
trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
set_subtree_ctrl( trip_count );
hook->init_req(3, trip_count);
break;
}
case BoolTest::ge: // Maybe convert to '>' case
limit = gvn->transform(new (C) AddINode(limit,one_m));
set_subtree_ctrl( limit );
hook->init_req(4 ,limit);
bt = BoolTest::gt;
// Make the new limit be in the same loop nest as the old limit
//_loop.map(limit->_idx,limit_loop);
// Fall into next case
case BoolTest::gt: { // Maybe convert to '!=' case
if (stride_con > 0) // count up loop rolls through MININT
ShouldNotReachHere();
Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
set_subtree_ctrl( range );
hook->init_req(0, range);
Node *bias = gvn->transform(new (C) AddINode(range,stride));
set_subtree_ctrl( bias );
hook->init_req(1, bias);
Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
set_subtree_ctrl( bias1 );
hook->init_req(2, bias1);
trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
set_subtree_ctrl( trip_count );
hook->init_req(3, trip_count);
break;
}
} // switch( bt )
Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
set_subtree_ctrl( span );
hook->init_req(5, span);
limit = gvn->transform(new (C) AddINode(span,init_trip));
set_subtree_ctrl( limit );
} // LoopLimitCheck
if (!UseCountedLoopSafepoints) {
// Check for SafePoint on backedge and remove
Node *sfpt = x->in(LoopNode::LoopBackControl);
if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
lazy_replace( sfpt, iftrue );
if (loop->_safepts != NULL) {
loop->_safepts->yank(sfpt);
}
loop->_tail = iftrue;
}
}
// Build a canonical trip test.
// Clone code, as old values may be in use.
incr = incr->clone();
incr->set_req(1,phi);
incr->set_req(2,stride);
incr = _igvn.register_new_node_with_optimizer(incr);
set_early_ctrl( incr );
_igvn.hash_delete(phi);
phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
// If phi type is more restrictive than Int, raise to
// Int to prevent (almost) infinite recursion in igvn
// which can only handle integer types for constants or minint..maxint.
if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
nphi = _igvn.register_new_node_with_optimizer(nphi);
set_ctrl(nphi, get_ctrl(phi));
_igvn.replace_node(phi, nphi);
phi = nphi->as_Phi();
}
cmp = cmp->clone();
cmp->set_req(1,incr);
cmp->set_req(2,limit);
cmp = _igvn.register_new_node_with_optimizer(cmp);
set_ctrl(cmp, iff->in(0));
test = test->clone()->as_Bool();
(*(BoolTest*)&test->_test)._test = bt;
test->set_req(1,cmp);
_igvn.register_new_node_with_optimizer(test);
set_ctrl(test, iff->in(0));
// Replace the old IfNode with a new LoopEndNode
Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
IfNode *le = lex->as_If();
uint dd = dom_depth(iff);
set_idom(le, le->in(0), dd); // Update dominance for loop exit
set_loop(le, loop);
// Get the loop-exit control
Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
// Need to swap loop-exit and loop-back control?
if (iftrue_op == Op_IfFalse) {
Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
loop->_tail = back_control = ift2;
set_loop(ift2, loop);
set_loop(iff2, get_loop(iffalse));
// Lazy update of 'get_ctrl' mechanism.
lazy_replace(iffalse, iff2);
lazy_replace(iftrue, ift2);
// Swap names
iffalse = iff2;
iftrue = ift2;
} else {
_igvn.hash_delete(iffalse);
_igvn.hash_delete(iftrue);
iffalse->set_req_X( 0, le, &_igvn );
iftrue ->set_req_X( 0, le, &_igvn );
}
set_idom(iftrue, le, dd+1);
set_idom(iffalse, le, dd+1);
assert(iff->outcnt() == 0, "should be dead now");
lazy_replace( iff, le ); // fix 'get_ctrl'
// Now setup a new CountedLoopNode to replace the existing LoopNode
CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
// The following assert is approximately true, and defines the intention
// of can_be_counted_loop. It fails, however, because phase->type
// is not yet initialized for this loop and its parts.
//assert(l->can_be_counted_loop(this), "sanity");
_igvn.register_new_node_with_optimizer(l);
set_loop(l, loop);
loop->_head = l;
// Fix all data nodes placed at the old loop head.
// Uses the lazy-update mechanism of 'get_ctrl'.
lazy_replace( x, l );
set_idom(l, init_control, dom_depth(x));
if (!UseCountedLoopSafepoints) {
// Check for immediately preceding SafePoint and remove
Node *sfpt2 = le->in(0);
if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
if (loop->_safepts != NULL) {
loop->_safepts->yank(sfpt2);
}
}
}
// Free up intermediate goo
_igvn.remove_dead_node(hook);
#ifdef ASSERT
assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
#endif
#ifndef PRODUCT
if (TraceLoopOpts) {
tty->print("Counted ");
loop->dump_head();
}
#endif
C->print_method(PHASE_AFTER_CLOOPS, 3);
return true;
}
//----------------------exact_limit-------------------------------------------
Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
assert(loop->_head->is_CountedLoop(), "");
CountedLoopNode *cl = loop->_head->as_CountedLoop();
assert(cl->is_valid_counted_loop(), "");
if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
cl->limit()->Opcode() == Op_LoopLimit) {
// Old code has exact limit (it could be incorrect in case of int overflow).
// Loop limit is exact with stride == 1. And loop may already have exact limit.
return cl->limit();
}
Node *limit = NULL;
#ifdef ASSERT
BoolTest::mask bt = cl->loopexit()->test_trip();
assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
#endif
if (cl->has_exact_trip_count()) {
// Simple case: loop has constant boundaries.
// Use jlongs to avoid integer overflow.
int stride_con = cl->stride_con();
jlong init_con = cl->init_trip()->get_int();
jlong limit_con = cl->limit()->get_int();
julong trip_cnt = cl->trip_count();
jlong final_con = init_con + trip_cnt*stride_con;
int final_int = (int)final_con;
// The final value should be in integer range since the loop
// is counted and the limit was checked for overflow.
assert(final_con == (jlong)final_int, "final value should be integer");
limit = _igvn.intcon(final_int);
} else {
// Create new LoopLimit node to get exact limit (final iv value).
limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
register_new_node(limit, cl->in(LoopNode::EntryControl));
}
assert(limit != NULL, "sanity");
return limit;
}
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Attempt to convert into a counted-loop.
Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (!can_be_counted_loop(phase)) {
phase->C->set_major_progress();
}
return RegionNode::Ideal(phase, can_reshape);
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Attempt to convert into a counted-loop.
Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
return RegionNode::Ideal(phase, can_reshape);
}
//------------------------------dump_spec--------------------------------------
// Dump special per-node info
#ifndef PRODUCT
void CountedLoopNode::dump_spec(outputStream *st) const {
LoopNode::dump_spec(st);
if (stride_is_con()) {
st->print("stride: %d ",stride_con());
}
if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
if (is_main_loop()) st->print("main of N%d", _idx);
if (is_post_loop()) st->print("post of N%d", _main_idx);
}
#endif
//=============================================================================
int CountedLoopEndNode::stride_con() const {
return stride()->bottom_type()->is_int()->get_con();
}
//=============================================================================
//------------------------------Value-----------------------------------------
const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
const Type* init_t = phase->type(in(Init));
const Type* limit_t = phase->type(in(Limit));
const Type* stride_t = phase->type(in(Stride));
// Either input is TOP ==> the result is TOP
if (init_t == Type::TOP) return Type::TOP;
if (limit_t == Type::TOP) return Type::TOP;
if (stride_t == Type::TOP) return Type::TOP;
int stride_con = stride_t->is_int()->get_con();
if (stride_con == 1)
return NULL; // Identity
if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
// Use jlongs to avoid integer overflow.
jlong init_con = init_t->is_int()->get_con();
jlong limit_con = limit_t->is_int()->get_con();
int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
jlong final_con = init_con + stride_con*trip_count;
int final_int = (int)final_con;
// The final value should be in integer range since the loop
// is counted and the limit was checked for overflow.
assert(final_con == (jlong)final_int, "final value should be integer");
return TypeInt::make(final_int);
}
return bottom_type(); // TypeInt::INT
}
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (phase->type(in(Init)) == Type::TOP ||
phase->type(in(Limit)) == Type::TOP ||
phase->type(in(Stride)) == Type::TOP)
return NULL; // Dead
int stride_con = phase->type(in(Stride))->is_int()->get_con();
if (stride_con == 1)
return NULL; // Identity
if (in(Init)->is_Con() && in(Limit)->is_Con())
return NULL; // Value
// Delay following optimizations until all loop optimizations
// done to keep Ideal graph simple.
if (!can_reshape || phase->C->major_progress())
return NULL;
const TypeInt* init_t = phase->type(in(Init) )->is_int();
const TypeInt* limit_t = phase->type(in(Limit))->is_int();
int stride_p;
jlong lim, ini;
julong max;
if (stride_con > 0) {
stride_p = stride_con;
lim = limit_t->_hi;
ini = init_t->_lo;
max = (julong)max_jint;
} else {
stride_p = -stride_con;
lim = init_t->_hi;
ini = limit_t->_lo;
max = (julong)min_jint;
}
julong range = lim - ini + stride_p;
if (range <= max) {
// Convert to integer expression if it is not overflow.
Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
return new (phase->C) AddINode(span, in(Init)); // exact limit
}
if (is_power_of_2(stride_p) || // divisor is 2^n
!Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
// Convert to long expression to avoid integer overflow
// and let igvn optimizer convert this division.
//
Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
Node* stride = phase->longcon(stride_con);
Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
Node *span;
if (stride_con > 0 && is_power_of_2(stride_p)) {
// bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
// and avoid generating rounding for division. Zero trip guard should
// guarantee that init < limit but sometimes the guard is missing and
// we can get situation when init > limit. Note, for the empty loop
// optimization zero trip guard is generated explicitly which leaves
// only RCE predicate where exact limit is used and the predicate
// will simply fail forcing recompilation.
Node* neg_stride = phase->longcon(-stride_con);
span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
} else {
Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
span = phase->transform(new (phase->C) MulLNode(trip, stride));
}
// Convert back to int
Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
return new (phase->C) AddINode(span_int, in(Init)); // exact limit
}
return NULL; // No progress
}
//------------------------------Identity---------------------------------------
// If stride == 1 return limit node.
Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
int stride_con = phase->type(in(Stride))->is_int()->get_con();
if (stride_con == 1 || stride_con == -1)
return in(Limit);
return this;
}
//=============================================================================
//----------------------match_incr_with_optional_truncation--------------------
// Match increment with optional truncation:
// CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
// Return NULL for failure. Success returns the increment node.
Node* CountedLoopNode::match_incr_with_optional_truncation(
Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
// Quick cutouts:
if (expr == NULL || expr->req() != 3) return NULL;
Node *t1 = NULL;
Node *t2 = NULL;
const TypeInt* trunc_t = TypeInt::INT;
Node* n1 = expr;
int n1op = n1->Opcode();
// Try to strip (n1 & M) or (n1 << N >> N) from n1.
if (n1op == Op_AndI &&
n1->in(2)->is_Con() &&
n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
// %%% This check should match any mask of 2**K-1.
t1 = n1;
n1 = t1->in(1);
n1op = n1->Opcode();
trunc_t = TypeInt::CHAR;
} else if (n1op == Op_RShiftI &&
n1->in(1) != NULL &&
n1->in(1)->Opcode() == Op_LShiftI &&
n1->in(2) == n1->in(1)->in(2) &&
n1->in(2)->is_Con()) {
jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
// %%% This check should match any shift in [1..31].
if (shift == 16 || shift == 8) {
t1 = n1;
t2 = t1->in(1);
n1 = t2->in(1);
n1op = n1->Opcode();
if (shift == 16) {
trunc_t = TypeInt::SHORT;
} else if (shift == 8) {
trunc_t = TypeInt::BYTE;
}
}
}
// If (maybe after stripping) it is an AddI, we won:
if (n1op == Op_AddI) {
*trunc1 = t1;
*trunc2 = t2;
*trunc_type = trunc_t;
return n1;
}
// failed
return NULL;
}
//------------------------------filtered_type--------------------------------
// Return a type based on condition control flow
// A successful return will be a type that is restricted due
// to a series of dominating if-tests, such as:
// if (i < 10) {
// if (i > 0) {
// here: "i" type is [1..10)
// }
// }
// or a control flow merge
// if (i < 10) {
// do {
// phi( , ) -- at top of loop type is [min_int..10)
// i = ?
// } while ( i < 10)
//
const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
assert(n && n->bottom_type()->is_int(), "must be int");
const TypeInt* filtered_t = NULL;
if (!n->is_Phi()) {
assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
filtered_t = filtered_type_from_dominators(n, n_ctrl);
} else {
Node* phi = n->as_Phi();
Node* region = phi->in(0);
assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
if (region && region != C->top()) {
for (uint i = 1; i < phi->req(); i++) {
Node* val = phi->in(i);
Node* use_c = region->in(i);
const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
if (val_t != NULL) {
if (filtered_t == NULL) {
filtered_t = val_t;
} else {
filtered_t = filtered_t->meet(val_t)->is_int();
}
}
}
}
}
const TypeInt* n_t = _igvn.type(n)->is_int();
if (filtered_t != NULL) {
n_t = n_t->join(filtered_t)->is_int();
}
return n_t;
}
//------------------------------filtered_type_from_dominators--------------------------------
// Return a possibly more restrictive type for val based on condition control flow of dominators
const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
if (val->is_Con()) {
return val->bottom_type()->is_int();
}
uint if_limit = 10; // Max number of dominating if's visited
const TypeInt* rtn_t = NULL;
if (use_ctrl && use_ctrl != C->top()) {
Node* val_ctrl = get_ctrl(val);
uint val_dom_depth = dom_depth(val_ctrl);
Node* pred = use_ctrl;
uint if_cnt = 0;
while (if_cnt < if_limit) {
if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
if_cnt++;
const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
if (if_t != NULL) {
if (rtn_t == NULL) {
rtn_t = if_t;
} else {
rtn_t = rtn_t->join(if_t)->is_int();
}
}
}
pred = idom(pred);
if (pred == NULL || pred == C->top()) {
break;
}
// Stop if going beyond definition block of val
if (dom_depth(pred) < val_dom_depth) {
break;
}
}
}
return rtn_t;
}
//------------------------------dump_spec--------------------------------------
// Dump special per-node info
#ifndef PRODUCT
void CountedLoopEndNode::dump_spec(outputStream *st) const {
if( in(TestValue)->is_Bool() ) {
BoolTest bt( test_trip()); // Added this for g++.
st->print("[");
bt.dump_on(st);
st->print("]");
}
st->print(" ");
IfNode::dump_spec(st);
}
#endif
//=============================================================================
//------------------------------is_member--------------------------------------
// Is 'l' a member of 'this'?
int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
while( l->_nest > _nest ) l = l->_parent;
return l == this;
}
//------------------------------set_nest---------------------------------------
// Set loop tree nesting depth. Accumulate _has_call bits.
int IdealLoopTree::set_nest( uint depth ) {
_nest = depth;
int bits = _has_call;
if( _child ) bits |= _child->set_nest(depth+1);
if( bits ) _has_call = 1;
if( _next ) bits |= _next ->set_nest(depth );
return bits;
}
//------------------------------split_fall_in----------------------------------
// Split out multiple fall-in edges from the loop header. Move them to a
// private RegionNode before the loop. This becomes the loop landing pad.
void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
PhaseIterGVN &igvn = phase->_igvn;
uint i;
// Make a new RegionNode to be the landing pad.
Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
phase->set_loop(landing_pad,_parent);
// Gather all the fall-in control paths into the landing pad
uint icnt = fall_in_cnt;
uint oreq = _head->req();
for( i = oreq-1; i>0; i-- )
if( !phase->is_member( this, _head->in(i) ) )
landing_pad->set_req(icnt--,_head->in(i));
// Peel off PhiNode edges as well
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *oj = _head->fast_out(j);
if( oj->is_Phi() ) {
PhiNode* old_phi = oj->as_Phi();
assert( old_phi->region() == _head, "" );
igvn.hash_delete(old_phi); // Yank from hash before hacking edges
Node *p = PhiNode::make_blank(landing_pad, old_phi);
uint icnt = fall_in_cnt;
for( i = oreq-1; i>0; i-- ) {
if( !phase->is_member( this, _head->in(i) ) ) {
p->init_req(icnt--, old_phi->in(i));
// Go ahead and clean out old edges from old phi
old_phi->del_req(i);
}
}
// Search for CSE's here, because ZKM.jar does a lot of
// loop hackery and we need to be a little incremental
// with the CSE to avoid O(N^2) node blow-up.
Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
if( p2 ) { // Found CSE
p->destruct(); // Recover useless new node
p = p2; // Use old node
} else {
igvn.register_new_node_with_optimizer(p, old_phi);
}
// Make old Phi refer to new Phi.
old_phi->add_req(p);
// Check for the special case of making the old phi useless and
// disappear it. In JavaGrande I have a case where this useless
// Phi is the loop limit and prevents recognizing a CountedLoop
// which in turn prevents removing an empty loop.
Node *id_old_phi = old_phi->Identity( &igvn );
if( id_old_phi != old_phi ) { // Found a simple identity?
// Note that I cannot call 'replace_node' here, because
// that will yank the edge from old_phi to the Region and
// I'm mid-iteration over the Region's uses.
for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
Node* use = old_phi->last_out(i);
igvn.rehash_node_delayed(use);
uint uses_found = 0;
for (uint j = 0; j < use->len(); j++) {
if (use->in(j) == old_phi) {
if (j < use->req()) use->set_req (j, id_old_phi);
else use->set_prec(j, id_old_phi);
uses_found++;
}
}
i -= uses_found; // we deleted 1 or more copies of this edge
}
}
igvn._worklist.push(old_phi);
}
}
// Finally clean out the fall-in edges from the RegionNode
for( i = oreq-1; i>0; i-- ) {
if( !phase->is_member( this, _head->in(i) ) ) {
_head->del_req(i);
}
}
// Transform landing pad
igvn.register_new_node_with_optimizer(landing_pad, _head);
// Insert landing pad into the header
_head->add_req(landing_pad);
}
//------------------------------split_outer_loop-------------------------------
// Split out the outermost loop from this shared header.
void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
PhaseIterGVN &igvn = phase->_igvn;
// Find index of outermost loop; it should also be my tail.
uint outer_idx = 1;
while( _head->in(outer_idx) != _tail ) outer_idx++;
// Make a LoopNode for the outermost loop.
Node *ctl = _head->in(LoopNode::EntryControl);
Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
outer = igvn.register_new_node_with_optimizer(outer, _head);
phase->set_created_loop_node();
// Outermost loop falls into '_head' loop
_head->set_req(LoopNode::EntryControl, outer);
_head->del_req(outer_idx);
// Split all the Phis up between '_head' loop and 'outer' loop.
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *out = _head->fast_out(j);
if( out->is_Phi() ) {
PhiNode *old_phi = out->as_Phi();
assert( old_phi->region() == _head, "" );
Node *phi = PhiNode::make_blank(outer, old_phi);
phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
phi = igvn.register_new_node_with_optimizer(phi, old_phi);
// Make old Phi point to new Phi on the fall-in path
igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
old_phi->del_req(outer_idx);
}
}
// Use the new loop head instead of the old shared one
_head = outer;
phase->set_loop(_head, this);
}
//------------------------------fix_parent-------------------------------------
static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
loop->_parent = parent;
if( loop->_child ) fix_parent( loop->_child, loop );
if( loop->_next ) fix_parent( loop->_next , parent );
}
//------------------------------estimate_path_freq-----------------------------
static float estimate_path_freq( Node *n ) {
// Try to extract some path frequency info
IfNode *iff;
for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
uint nop = n->Opcode();
if( nop == Op_SafePoint ) { // Skip any safepoint
n = n->in(0);
continue;
}
if( nop == Op_CatchProj ) { // Get count from a prior call
// Assume call does not always throw exceptions: means the call-site
// count is also the frequency of the fall-through path.
assert( n->is_CatchProj(), "" );
if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
return 0.0f; // Assume call exception path is rare
Node *call = n->in(0)->in(0)->in(0);
assert( call->is_Call(), "expect a call here" );
const JVMState *jvms = ((CallNode*)call)->jvms();
ciMethodData* methodData = jvms->method()->method_data();
if (!methodData->is_mature()) return 0.0f; // No call-site data
ciProfileData* data = methodData->bci_to_data(jvms->bci());
if ((data == NULL) || !data->is_CounterData()) {
// no call profile available, try call's control input
n = n->in(0);
continue;
}
return data->as_CounterData()->count()/FreqCountInvocations;
}
// See if there's a gating IF test
Node *n_c = n->in(0);
if( !n_c->is_If() ) break; // No estimate available
iff = n_c->as_If();
if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
// Compute how much count comes on this path
return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
// Have no count info. Skip dull uncommon-trap like branches.
if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
(nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
break;
// Skip through never-taken branch; look for a real loop exit.
n = iff->in(0);
}
return 0.0f; // No estimate available
}
//------------------------------merge_many_backedges---------------------------
// Merge all the backedges from the shared header into a private Region.
// Feed that region as the one backedge to this loop.
void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
uint i;
// Scan for the top 2 hottest backedges
float hotcnt = 0.0f;
float warmcnt = 0.0f;
uint hot_idx = 0;
// Loop starts at 2 because slot 1 is the fall-in path
for( i = 2; i < _head->req(); i++ ) {
float cnt = estimate_path_freq(_head->in(i));
if( cnt > hotcnt ) { // Grab hottest path
warmcnt = hotcnt;
hotcnt = cnt;
hot_idx = i;
} else if( cnt > warmcnt ) { // And 2nd hottest path
warmcnt = cnt;
}
}
// See if the hottest backedge is worthy of being an inner loop
// by being much hotter than the next hottest backedge.
if( hotcnt <= 0.0001 ||
hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
// Peel out the backedges into a private merge point; peel
// them all except optionally hot_idx.
PhaseIterGVN &igvn = phase->_igvn;
Node *hot_tail = NULL;
// Make a Region for the merge point
Node *r = new (phase->C) RegionNode(1);
for( i = 2; i < _head->req(); i++ ) {
if( i != hot_idx )
r->add_req( _head->in(i) );
else hot_tail = _head->in(i);
}
igvn.register_new_node_with_optimizer(r, _head);
// Plug region into end of loop _head, followed by hot_tail
while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
_head->set_req(2, r);
if( hot_idx ) _head->add_req(hot_tail);
// Split all the Phis up between '_head' loop and the Region 'r'
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *out = _head->fast_out(j);
if( out->is_Phi() ) {
PhiNode* n = out->as_Phi();
igvn.hash_delete(n); // Delete from hash before hacking edges
Node *hot_phi = NULL;
Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
// Check all inputs for the ones to peel out
uint j = 1;
for( uint i = 2; i < n->req(); i++ ) {
if( i != hot_idx )
phi->set_req( j++, n->in(i) );
else hot_phi = n->in(i);
}
// Register the phi but do not transform until whole place transforms
igvn.register_new_node_with_optimizer(phi, n);
// Add the merge phi to the old Phi
while( n->req() > 3 ) n->del_req( n->req()-1 );
n->set_req(2, phi);
if( hot_idx ) n->add_req(hot_phi);
}
}
// Insert a new IdealLoopTree inserted below me. Turn it into a clone
// of self loop tree. Turn self into a loop headed by _head and with
// tail being the new merge point.
IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
phase->set_loop(_tail,ilt); // Adjust tail
_tail = r; // Self's tail is new merge point
phase->set_loop(r,this);
ilt->_child = _child; // New guy has my children
_child = ilt; // Self has new guy as only child
ilt->_parent = this; // new guy has self for parent
ilt->_nest = _nest; // Same nesting depth (for now)
// Starting with 'ilt', look for child loop trees using the same shared
// header. Flatten these out; they will no longer be loops in the end.
IdealLoopTree **pilt = &_child;
while( ilt ) {
if( ilt->_head == _head ) {
uint i;
for( i = 2; i < _head->req(); i++ )
if( _head->in(i) == ilt->_tail )
break; // Still a loop
if( i == _head->req() ) { // No longer a loop
// Flatten ilt. Hang ilt's "_next" list from the end of
// ilt's '_child' list. Move the ilt's _child up to replace ilt.
IdealLoopTree **cp = &ilt->_child;
while( *cp ) cp = &(*cp)->_next; // Find end of child list
*cp = ilt->_next; // Hang next list at end of child list
*pilt = ilt->_child; // Move child up to replace ilt
ilt->_head = NULL; // Flag as a loop UNIONED into parent
ilt = ilt->_child; // Repeat using new ilt
continue; // do not advance over ilt->_child
}
assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
phase->set_loop(_head,ilt);
}
pilt = &ilt->_child; // Advance to next
ilt = *pilt;
}
if( _child ) fix_parent( _child, this );
}
//------------------------------beautify_loops---------------------------------
// Split shared headers and insert loop landing pads.
// Insert a LoopNode to replace the RegionNode.
// Return TRUE if loop tree is structurally changed.
bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
bool result = false;
// Cache parts in locals for easy
PhaseIterGVN &igvn = phase->_igvn;
igvn.hash_delete(_head); // Yank from hash before hacking edges
// Check for multiple fall-in paths. Peel off a landing pad if need be.
int fall_in_cnt = 0;
for( uint i = 1; i < _head->req(); i++ )
if( !phase->is_member( this, _head->in(i) ) )
fall_in_cnt++;
assert( fall_in_cnt, "at least 1 fall-in path" );
if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
split_fall_in( phase, fall_in_cnt );
// Swap inputs to the _head and all Phis to move the fall-in edge to
// the left.
fall_in_cnt = 1;
while( phase->is_member( this, _head->in(fall_in_cnt) ) )
fall_in_cnt++;
if( fall_in_cnt > 1 ) {
// Since I am just swapping inputs I do not need to update def-use info
Node *tmp = _head->in(1);
_head->set_req( 1, _head->in(fall_in_cnt) );
_head->set_req( fall_in_cnt, tmp );
// Swap also all Phis
for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
Node* phi = _head->fast_out(i);
if( phi->is_Phi() ) {
igvn.hash_delete(phi); // Yank from hash before hacking edges
tmp = phi->in(1);
phi->set_req( 1, phi->in(fall_in_cnt) );
phi->set_req( fall_in_cnt, tmp );
}
}
}
assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
// If I am a shared header (multiple backedges), peel off the many
// backedges into a private merge point and use the merge point as
// the one true backedge.
if (_head->req() > 3) {
// Merge the many backedges into a single backedge but leave
// the hottest backedge as separate edge for the following peel.
if (!_irreducible) {
merge_many_backedges( phase );
}
// When recursively beautify my children, split_fall_in can change
// loop tree structure when I am an irreducible loop. Then the head
// of my children has a req() not bigger than 3. Here we need to set
// result to true to catch that case in order to tell the caller to
// rebuild loop tree. See issue JDK-8244407 for details.
result = true;
}
// If I have one hot backedge, peel off myself loop.
// I better be the outermost loop.
if (_head->req() > 3 && !_irreducible) {
split_outer_loop( phase );
result = true;
} else if (!_head->is_Loop() && !_irreducible) {
// Make a new LoopNode to replace the old loop head
Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
l = igvn.register_new_node_with_optimizer(l, _head);
phase->set_created_loop_node();
// Go ahead and replace _head
phase->_igvn.replace_node( _head, l );
_head = l;
phase->set_loop(_head, this);
}
// Now recursively beautify nested loops
if( _child ) result |= _child->beautify_loops( phase );
if( _next ) result |= _next ->beautify_loops( phase );
return result;
}
//------------------------------allpaths_check_safepts----------------------------
// Allpaths backwards scan from loop tail, terminating each path at first safepoint
// encountered. Helper for check_safepts.
void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
assert(stack.size() == 0, "empty stack");
stack.push(_tail);
visited.Clear();
visited.set(_tail->_idx);
while (stack.size() > 0) {
Node* n = stack.pop();
if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
// Terminate this path
} else if (n->Opcode() == Op_SafePoint) {
if (_phase->get_loop(n) != this) {
if (_required_safept == NULL) _required_safept = new Node_List();
_required_safept->push(n); // save the one closest to the tail
}
// Terminate this path
} else {
uint start = n->is_Region() ? 1 : 0;
uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
for (uint i = start; i < end; i++) {
Node* in = n->in(i);
assert(in->is_CFG(), "must be");
if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
stack.push(in);
}
}
}
}
}
//------------------------------check_safepts----------------------------
// Given dominators, try to find loops with calls that must always be
// executed (call dominates loop tail). These loops do not need non-call
// safepoints (ncsfpt).
//
// A complication is that a safepoint in a inner loop may be needed
// by an outer loop. In the following, the inner loop sees it has a
// call (block 3) on every path from the head (block 2) to the
// backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
// in block 2, _but_ this leaves the outer loop without a safepoint.
//
// entry 0
// |
// v
// outer 1,2 +->1
// | |
// | v
// | 2<---+ ncsfpt in 2
// |_/|\ |
// | v |
// inner 2,3 / 3 | call in 3
// / | |
// v +--+
// exit 4
//
//
// This method creates a list (_required_safept) of ncsfpt nodes that must
// be protected is created for each loop. When a ncsfpt maybe deleted, it
// is first looked for in the lists for the outer loops of the current loop.
//
// The insights into the problem:
// A) counted loops are okay
// B) innermost loops are okay (only an inner loop can delete
// a ncsfpt needed by an outer loop)
// C) a loop is immune from an inner loop deleting a safepoint
// if the loop has a call on the idom-path
// D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
// idom-path that is not in a nested loop
// E) otherwise, an ncsfpt on the idom-path that is nested in an inner
// loop needs to be prevented from deletion by an inner loop
//
// There are two analyses:
// 1) The first, and cheaper one, scans the loop body from
// tail to head following the idom (immediate dominator)
// chain, looking for the cases (C,D,E) above.
// Since inner loops are scanned before outer loops, there is summary
// information about inner loops. Inner loops can be skipped over
// when the tail of an inner loop is encountered.
//
// 2) The second, invoked if the first fails to find a call or ncsfpt on
// the idom path (which is rare), scans all predecessor control paths
// from the tail to the head, terminating a path when a call or sfpt
// is encountered, to find the ncsfpt's that are closest to the tail.
//
void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
// Bottom up traversal
IdealLoopTree* ch = _child;
if (_child) _child->check_safepts(visited, stack);
if (_next) _next ->check_safepts(visited, stack);
if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
bool has_call = false; // call on dom-path
bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
// Scan the dom-path nodes from tail to head
for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
has_call = true;
_has_sfpt = 1; // Then no need for a safept!
break;
} else if (n->Opcode() == Op_SafePoint) {
if (_phase->get_loop(n) == this) {
has_local_ncsfpt = true;
break;
}
if (nonlocal_ncsfpt == NULL) {
nonlocal_ncsfpt = n; // save the one closest to the tail
}
} else {
IdealLoopTree* nlpt = _phase->get_loop(n);
if (this != nlpt) {
// If at an inner loop tail, see if the inner loop has already
// recorded seeing a call on the dom-path (and stop.) If not,
// jump to the head of the inner loop.
assert(is_member(nlpt), "nested loop");
Node* tail = nlpt->_tail;
if (tail->in(0)->is_If()) tail = tail->in(0);
if (n == tail) {
// If inner loop has call on dom-path, so does outer loop
if (nlpt->_has_sfpt) {
has_call = true;
_has_sfpt = 1;
break;
}
// Skip to head of inner loop
assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
n = nlpt->_head;
}
}
}
}
// Record safept's that this loop needs preserved when an
// inner loop attempts to delete it's safepoints.
if (_child != NULL && !has_call && !has_local_ncsfpt) {
if (nonlocal_ncsfpt != NULL) {
if (_required_safept == NULL) _required_safept = new Node_List();
_required_safept->push(nonlocal_ncsfpt);
} else {
// Failed to find a suitable safept on the dom-path. Now use
// an all paths walk from tail to head, looking for safepoints to preserve.
allpaths_check_safepts(visited, stack);
}
}
}
}
//---------------------------is_deleteable_safept----------------------------
// Is safept not required by an outer loop?
bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
assert(sfpt->Opcode() == Op_SafePoint, "");
IdealLoopTree* lp = get_loop(sfpt)->_parent;
while (lp != NULL) {
Node_List* sfpts = lp->_required_safept;
if (sfpts != NULL) {
for (uint i = 0; i < sfpts->size(); i++) {
if (sfpt == sfpts->at(i))
return false;
}
}
lp = lp->_parent;
}
return true;
}
//---------------------------replace_parallel_iv-------------------------------
// Replace parallel induction variable (parallel to trip counter)
void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
assert(loop->_head->is_CountedLoop(), "");
CountedLoopNode *cl = loop->_head->as_CountedLoop();
if (!cl->is_valid_counted_loop())
return; // skip malformed counted loop
Node *incr = cl->incr();
if (incr == NULL)
return; // Dead loop?
Node *init = cl->init_trip();
Node *phi = cl->phi();
int stride_con = cl->stride_con();
// Visit all children, looking for Phis
for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
Node *out = cl->out(i);
// Look for other phis (secondary IVs). Skip dead ones
if (!out->is_Phi() || out == phi || !has_node(out))
continue;
PhiNode* phi2 = out->as_Phi();
Node *incr2 = phi2->in( LoopNode::LoopBackControl );
// Look for induction variables of the form: X += constant
if (phi2->region() != loop->_head ||
incr2->req() != 3 ||
incr2->in(1) != phi2 ||
incr2 == incr ||
incr2->Opcode() != Op_AddI ||
!incr2->in(2)->is_Con())
continue;
// Check for parallel induction variable (parallel to trip counter)
// via an affine function. In particular, count-down loops with
// count-up array indices are common. We only RCE references off
// the trip-counter, so we need to convert all these to trip-counter
// expressions.
Node *init2 = phi2->in( LoopNode::EntryControl );
int stride_con2 = incr2->in(2)->get_int();
// The ratio of the two strides cannot be represented as an int
// if stride_con2 is min_int and stride_con is -1.
if (stride_con2 == min_jint && stride_con == -1) {
continue;
}
// The general case here gets a little tricky. We want to find the
// GCD of all possible parallel IV's and make a new IV using this
// GCD for the loop. Then all possible IVs are simple multiples of
// the GCD. In practice, this will cover very few extra loops.
// Instead we require 'stride_con2' to be a multiple of 'stride_con',
// where +/-1 is the common case, but other integer multiples are
// also easy to handle.
int ratio_con = stride_con2/stride_con;
if ((ratio_con * stride_con) == stride_con2) { // Check for exact
#ifndef PRODUCT
if (TraceLoopOpts) {
tty->print("Parallel IV: %d ", phi2->_idx);
loop->dump_head();
}
#endif
// Convert to using the trip counter. The parallel induction
// variable differs from the trip counter by a loop-invariant
// amount, the difference between their respective initial values.
// It is scaled by the 'ratio_con'.
Node* ratio = _igvn.intcon(ratio_con);
set_ctrl(ratio, C->root());
Node* ratio_init = new (C) MulINode(init, ratio);
_igvn.register_new_node_with_optimizer(ratio_init, init);
set_early_ctrl(ratio_init);
Node* diff = new (C) SubINode(init2, ratio_init);
_igvn.register_new_node_with_optimizer(diff, init2);
set_early_ctrl(diff);
Node* ratio_idx = new (C) MulINode(phi, ratio);
_igvn.register_new_node_with_optimizer(ratio_idx, phi);
set_ctrl(ratio_idx, cl);
Node* add = new (C) AddINode(ratio_idx, diff);
_igvn.register_new_node_with_optimizer(add);
set_ctrl(add, cl);
_igvn.replace_node( phi2, add );
// Sometimes an induction variable is unused
if (add->outcnt() == 0) {
_igvn.remove_dead_node(add);
}
--i; // deleted this phi; rescan starting with next position
continue;
}
}
}
void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
Node* keep = NULL;
if (keep_one) {
// Look for a safepoint on the idom-path.
for (Node* i = tail(); i != _head; i = phase->idom(i)) {
if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
keep = i;
break; // Found one
}
}
}
// Don't remove any safepoints if it is requested to keep a single safepoint and
// no safepoint was found on idom-path. It is not safe to remove any safepoint
// in this case since there's no safepoint dominating all paths in the loop body.
bool prune = !keep_one || keep != NULL;
// Delete other safepoints in this loop.
Node_List* sfpts = _safepts;
if (prune && sfpts != NULL) {
assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
for (uint i = 0; i < sfpts->size(); i++) {
Node* n = sfpts->at(i);
assert(phase->get_loop(n) == this, "");
if (n != keep && phase->is_deleteable_safept(n)) {
phase->lazy_replace(n, n->in(TypeFunc::Control));
}
}
}
}
//------------------------------counted_loop-----------------------------------
// Convert to counted loops where possible
void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
// For grins, set the inner-loop flag here
if (!_child) {
if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
}
if (_head->is_CountedLoop() ||
phase->is_counted_loop(_head, this)) {
if (!UseCountedLoopSafepoints) {
// Indicate we do not need a safepoint here
_has_sfpt = 1;
}
// Remove safepoints
bool keep_one_sfpt = !(_has_call || _has_sfpt);
remove_safepoints(phase, keep_one_sfpt);
// Look for induction variables
phase->replace_parallel_iv(this);
} else if (_parent != NULL && !_irreducible) {
// Not a counted loop. Keep one safepoint.
bool keep_one_sfpt = true;
remove_safepoints(phase, keep_one_sfpt);
}
// Recursively
if (_child) _child->counted_loop( phase );
if (_next) _next ->counted_loop( phase );
}
#ifndef PRODUCT
//------------------------------dump_head--------------------------------------
// Dump 1 liner for loop header info
void IdealLoopTree::dump_head( ) const {
for (uint i=0; i<_nest; i++)
tty->print(" ");
tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
if (_irreducible) tty->print(" IRREDUCIBLE");
Node* entry = _head->in(LoopNode::EntryControl);
if (LoopLimitCheck) {
Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
if (predicate != NULL ) {
tty->print(" limit_check");
entry = entry->in(0)->in(0);
}
}
if (UseLoopPredicate) {
entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
if (entry != NULL) {
tty->print(" predicated");
}
}
if (_head->is_CountedLoop()) {
CountedLoopNode *cl = _head->as_CountedLoop();
tty->print(" counted");
Node* init_n = cl->init_trip();
if (init_n != NULL && init_n->is_Con())
tty->print(" [%d,", cl->init_trip()->get_int());
else
tty->print(" [int,");
Node* limit_n = cl->limit();
if (limit_n != NULL && limit_n->is_Con())
tty->print("%d),", cl->limit()->get_int());
else
tty->print("int),");
int stride_con = cl->stride_con();
if (stride_con > 0) tty->print("+");
tty->print("%d", stride_con);
tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
if (cl->is_pre_loop ()) tty->print(" pre" );
if (cl->is_main_loop()) tty->print(" main");
if (cl->is_post_loop()) tty->print(" post");
}
if (_has_call) tty->print(" has_call");
if (_has_sfpt) tty->print(" has_sfpt");
if (_rce_candidate) tty->print(" rce");
if (_safepts != NULL && _safepts->size() > 0) {
tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
}
if (_required_safept != NULL && _required_safept->size() > 0) {
tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
}
tty->cr();
}
//------------------------------dump-------------------------------------------
// Dump loops by loop tree
void IdealLoopTree::dump( ) const {
dump_head();
if (_child) _child->dump();
if (_next) _next ->dump();
}
#endif
static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
if (loop == root) {
if (loop->_child != NULL) {
log->begin_head("loop_tree");
log->end_head();
if( loop->_child ) log_loop_tree(root, loop->_child, log);
log->tail("loop_tree");
assert(loop->_next == NULL, "what?");
}
} else {
Node* head = loop->_head;
log->begin_head("loop");
log->print(" idx='%d' ", head->_idx);
if (loop->_irreducible) log->print("irreducible='1' ");
if (head->is_Loop()) {
if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
}
if (head->is_CountedLoop()) {
CountedLoopNode* cl = head->as_CountedLoop();
if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
}
log->end_head();
if( loop->_child ) log_loop_tree(root, loop->_child, log);
log->tail("loop");
if( loop->_next ) log_loop_tree(root, loop->_next, log);
}
}
//---------------------collect_potentially_useful_predicates-----------------------
// Helper function to collect potentially useful predicates to prevent them from
// being eliminated by PhaseIdealLoop::eliminate_useless_predicates
void PhaseIdealLoop::collect_potentially_useful_predicates(
IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
if (loop->_child) { // child
collect_potentially_useful_predicates(loop->_child, useful_predicates);
}
// self (only loops that we can apply loop predication may use their predicates)
if (loop->_head->is_Loop() &&
!loop->_irreducible &&
!loop->tail()->is_top()) {
LoopNode* lpn = loop->_head->as_Loop();
Node* entry = lpn->in(LoopNode::EntryControl);
Node* predicate_proj = find_predicate(entry); // loop_limit_check first
if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
entry = entry->in(0)->in(0);
}
predicate_proj = find_predicate(entry); // Predicate
if (predicate_proj != NULL ) {
useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
}
}
if (loop->_next) { // sibling
collect_potentially_useful_predicates(loop->_next, useful_predicates);
}
}
//------------------------eliminate_useless_predicates-----------------------------
// Eliminate all inserted predicates if they could not be used by loop predication.
// Note: it will also eliminates loop limits check predicate since it also uses
// Opaque1 node (see Parse::add_predicate()).
void PhaseIdealLoop::eliminate_useless_predicates() {
if (C->predicate_count() == 0)
return; // no predicate left
Unique_Node_List useful_predicates; // to store useful predicates
if (C->has_loops()) {
collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
}
for (int i = C->predicate_count(); i > 0; i--) {
Node * n = C->predicate_opaque1_node(i-1);
assert(n->Opcode() == Op_Opaque1, "must be");
if (!useful_predicates.member(n)) { // not in the useful list
_igvn.replace_node(n, n->in(1));
}
}
}
//------------------------process_expensive_nodes-----------------------------
// Expensive nodes have their control input set to prevent the GVN
// from commoning them and as a result forcing the resulting node to
// be in a more frequent path. Use CFG information here, to change the
// control inputs so that some expensive nodes can be commoned while
// not executed more frequently.
bool PhaseIdealLoop::process_expensive_nodes() {
assert(OptimizeExpensiveOps, "optimization off?");
// Sort nodes to bring similar nodes together
C->sort_expensive_nodes();
bool progress = false;
for (int i = 0; i < C->expensive_count(); ) {
Node* n = C->expensive_node(i);
int start = i;
// Find nodes similar to n
i++;
for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
int end = i;
// And compare them two by two
for (int j = start; j < end; j++) {
Node* n1 = C->expensive_node(j);
if (is_node_unreachable(n1)) {
continue;
}
for (int k = j+1; k < end; k++) {
Node* n2 = C->expensive_node(k);
if (is_node_unreachable(n2)) {
continue;
}
assert(n1 != n2, "should be pair of nodes");
Node* c1 = n1->in(0);
Node* c2 = n2->in(0);
Node* parent_c1 = c1;
Node* parent_c2 = c2;
// The call to get_early_ctrl_for_expensive() moves the
// expensive nodes up but stops at loops that are in a if
// branch. See whether we can exit the loop and move above the
// If.
if (c1->is_Loop()) {
parent_c1 = c1->in(1);
}
if (c2->is_Loop()) {
parent_c2 = c2->in(1);
}
if (parent_c1 == parent_c2) {
_igvn._worklist.push(n1);
_igvn._worklist.push(n2);
continue;
}
// Look for identical expensive node up the dominator chain.
if (is_dominator(c1, c2)) {
c2 = c1;
} else if (is_dominator(c2, c1)) {
c1 = c2;
} else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
// Both branches have the same expensive node so move it up
// before the if.
c1 = c2 = idom(parent_c1->in(0));
}
// Do the actual moves
if (n1->in(0) != c1) {
_igvn.hash_delete(n1);
n1->set_req(0, c1);
_igvn.hash_insert(n1);
_igvn._worklist.push(n1);
progress = true;
}
if (n2->in(0) != c2) {
_igvn.hash_delete(n2);
n2->set_req(0, c2);
_igvn.hash_insert(n2);
_igvn._worklist.push(n2);
progress = true;
}
}
}
}
return progress;
}
//=============================================================================
//----------------------------build_and_optimize-------------------------------
// Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
// its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
ResourceMark rm;
int old_progress = C->major_progress();
uint orig_worklist_size = _igvn._worklist.size();
// Reset major-progress flag for the driver's heuristics
C->clear_major_progress();
#ifndef PRODUCT
// Capture for later assert
uint unique = C->unique();
_loop_invokes++;
_loop_work += unique;
#endif
// True if the method has at least 1 irreducible loop
_has_irreducible_loops = false;
_created_loop_node = false;
Arena *a = Thread::current()->resource_area();
VectorSet visited(a);
// Pre-grow the mapping from Nodes to IdealLoopTrees.
_nodes.map(C->unique(), NULL);
memset(_nodes.adr(), 0, wordSize * C->unique());
// Pre-build the top-level outermost loop tree entry
_ltree_root = new IdealLoopTree( this, C->root(), C->root() );
// Do not need a safepoint at the top level
_ltree_root->_has_sfpt = 1;
// Initialize Dominators.
// Checked in clone_loop_predicate() during beautify_loops().
_idom_size = 0;
_idom = NULL;
_dom_depth = NULL;
_dom_stk = NULL;
// Empty pre-order array
allocate_preorders();
// Build a loop tree on the fly. Build a mapping from CFG nodes to
// IdealLoopTree entries. Data nodes are NOT walked.
build_loop_tree();
// Check for bailout, and return
if (C->failing()) {
return;
}
// No loops after all
if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
// There should always be an outer loop containing the Root and Return nodes.
// If not, we have a degenerate empty program. Bail out in this case.
if (!has_node(C->root())) {
if (!_verify_only) {
C->clear_major_progress();
C->record_method_not_compilable("empty program detected during loop optimization");
}
return;
}
// Nothing to do, so get out
bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
if (stop_early && !do_expensive_nodes) {
_igvn.optimize(); // Cleanup NeverBranches
return;
}
// Set loop nesting depth
_ltree_root->set_nest( 0 );
// Split shared headers and insert loop landing pads.
// Do not bother doing this on the Root loop of course.
if( !_verify_me && !_verify_only && _ltree_root->_child ) {
C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
if( _ltree_root->_child->beautify_loops( this ) ) {
// Re-build loop tree!
_ltree_root->_child = NULL;
_nodes.clear();
reallocate_preorders();
build_loop_tree();
// Check for bailout, and return
if (C->failing()) {
return;
}
// Reset loop nesting depth
_ltree_root->set_nest( 0 );
C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
}
}
// Build Dominators for elision of NULL checks & loop finding.
// Since nodes do not have a slot for immediate dominator, make
// a persistent side array for that info indexed on node->_idx.
_idom_size = C->unique();
_idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
_dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
_dom_stk = NULL; // Allocated on demand in recompute_dom_depth
memset( _dom_depth, 0, _idom_size * sizeof(uint) );
Dominators();
if (!_verify_only) {
// As a side effect, Dominators removed any unreachable CFG paths
// into RegionNodes. It doesn't do this test against Root, so
// we do it here.
for( uint i = 1; i < C->root()->req(); i++ ) {
if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
_igvn.delete_input_of(C->root(), i);
i--; // Rerun same iteration on compressed edges
}
}
// Given dominators, try to find inner loops with calls that must
// always be executed (call dominates loop tail). These loops do
// not need a separate safepoint.
Node_List cisstack(a);
_ltree_root->check_safepts(visited, cisstack);
}
// Walk the DATA nodes and place into loops. Find earliest control
// node. For CFG nodes, the _nodes array starts out and remains
// holding the associated IdealLoopTree pointer. For DATA nodes, the
// _nodes array holds the earliest legal controlling CFG node.
// Allocate stack with enough space to avoid frequent realloc
int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
Node_Stack nstack( a, stack_size );
visited.Clear();
Node_List worklist(a);
// Don't need C->root() on worklist since
// it will be processed among C->top() inputs
worklist.push( C->top() );
visited.set( C->top()->_idx ); // Set C->top() as visited now
build_loop_early( visited, worklist, nstack );
// Given early legal placement, try finding counted loops. This placement
// is good enough to discover most loop invariants.
if( !_verify_me && !_verify_only )
_ltree_root->counted_loop( this );
// Find latest loop placement. Find ideal loop placement.
visited.Clear();
init_dom_lca_tags();
// Need C->root() on worklist when processing outs
worklist.push( C->root() );
NOT_PRODUCT( C->verify_graph_edges(); )
worklist.push( C->top() );
build_loop_late( visited, worklist, nstack );
if (_verify_only) {
// restore major progress flag
for (int i = 0; i < old_progress; i++)
C->set_major_progress();
assert(C->unique() == unique, "verification mode made Nodes? ? ?");
assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
return;
}
// clear out the dead code after build_loop_late
while (_deadlist.size()) {
_igvn.remove_globally_dead_node(_deadlist.pop());
}
if (stop_early) {
assert(do_expensive_nodes, "why are we here?");
if (process_expensive_nodes()) {
// If we made some progress when processing expensive nodes then
// the IGVN may modify the graph in a way that will allow us to
// make some more progress: we need to try processing expensive
// nodes again.
C->set_major_progress();
}
_igvn.optimize();
return;
}
// Some parser-inserted loop predicates could never be used by loop
// predication or they were moved away from loop during some optimizations.
// For example, peeling. Eliminate them before next loop optimizations.
if (UseLoopPredicate || LoopLimitCheck) {
eliminate_useless_predicates();
}
#ifndef PRODUCT
C->verify_graph_edges();
if (_verify_me) { // Nested verify pass?
// Check to see if the verify mode is broken
assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
return;
}
if(VerifyLoopOptimizations) verify();
if(TraceLoopOpts && C->has_loops()) {
_ltree_root->dump();
}
#endif
if (skip_loop_opts) {
// restore major progress flag
for (int i = 0; i < old_progress; i++) {
C->set_major_progress();
}
// Cleanup any modified bits
_igvn.optimize();
if (C->log() != NULL) {
log_loop_tree(_ltree_root, _ltree_root, C->log());
}
return;
}
if (ReassociateInvariants) {
// Reassociate invariants and prep for split_thru_phi
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current();
if (!lpt->is_counted() || !lpt->is_inner()) continue;
lpt->reassociate_invariants(this);
// Because RCE opportunities can be masked by split_thru_phi,
// look for RCE candidates and inhibit split_thru_phi
// on just their loop-phi's for this pass of loop opts
if (SplitIfBlocks && do_split_ifs) {
if (lpt->policy_range_check(this)) {
lpt->_rce_candidate = 1; // = true
}
}
}
}
// Check for aggressive application of split-if and other transforms
// that require basic-block info (like cloning through Phi's)
if( SplitIfBlocks && do_split_ifs ) {
visited.Clear();
split_if_with_blocks( visited, nstack );
NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
}
if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
C->set_major_progress();
}
// Perform loop predication before iteration splitting
if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
_ltree_root->_child->loop_predication(this);
}
if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
if (do_intrinsify_fill()) {
C->set_major_progress();
}
}
// Perform iteration-splitting on inner loops. Split iterations to avoid
// range checks or one-shot null checks.
// If split-if's didn't hack the graph too bad (no CFG changes)
// then do loop opts.
if (C->has_loops() && !C->major_progress()) {
memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
_ltree_root->_child->iteration_split( this, worklist );
// No verify after peeling! GCM has hoisted code out of the loop.
// After peeling, the hoisted code could sink inside the peeled area.
// The peeling code does not try to recompute the best location for
// all the code before the peeled area, so the verify pass will always
// complain about it.
}
// Do verify graph edges in any case
NOT_PRODUCT( C->verify_graph_edges(); );
if (!do_split_ifs) {
// We saw major progress in Split-If to get here. We forced a
// pass with unrolling and not split-if, however more split-if's
// might make progress. If the unrolling didn't make progress
// then the major-progress flag got cleared and we won't try
// another round of Split-If. In particular the ever-common
// instance-of/check-cast pattern requires at least 2 rounds of
// Split-If to clear out.
C->set_major_progress();
}
// Repeat loop optimizations if new loops were seen
if (created_loop_node()) {
C->set_major_progress();
}
// Keep loop predicates and perform optimizations with them
// until no more loop optimizations could be done.
// After that switch predicates off and do more loop optimizations.
if (!C->major_progress() && (C->predicate_count() > 0)) {
C->cleanup_loop_predicates(_igvn);
#ifndef PRODUCT
if (TraceLoopOpts) {
tty->print_cr("PredicatesOff");
}
#endif
C->set_major_progress();
}
// Convert scalar to superword operations at the end of all loop opts.
if (UseSuperWord && C->has_loops() && !C->major_progress()) {
// SuperWord transform
SuperWord sw(this);
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current();
if (lpt->is_counted()) {
sw.transform_loop(lpt);
}
}
}
// Cleanup any modified bits
_igvn.optimize();
// disable assert until issue with split_flow_path is resolved (6742111)
// assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
// "shouldn't introduce irreducible loops");
if (C->log() != NULL) {
log_loop_tree(_ltree_root, _ltree_root, C->log());
}
}
#ifndef PRODUCT
//------------------------------print_statistics-------------------------------
int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
void PhaseIdealLoop::print_statistics() {
tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
}
//------------------------------verify-----------------------------------------
// Build a verify-only PhaseIdealLoop, and see that it agrees with me.
static int fail; // debug only, so its multi-thread dont care
void PhaseIdealLoop::verify() const {
int old_progress = C->major_progress();
ResourceMark rm;
PhaseIdealLoop loop_verify( _igvn, this );
VectorSet visited(Thread::current()->resource_area());
fail = 0;
verify_compare( C->root(), &loop_verify, visited );
assert( fail == 0, "verify loops failed" );
// Verify loop structure is the same
_ltree_root->verify_tree(loop_verify._ltree_root, NULL);
// Reset major-progress. It was cleared by creating a verify version of
// PhaseIdealLoop.
for( int i=0; i<old_progress; i++ )
C->set_major_progress();
}
//------------------------------verify_compare---------------------------------
// Make sure me and the given PhaseIdealLoop agree on key data structures
void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
if( !n ) return;
if( visited.test_set( n->_idx ) ) return;
if( !_nodes[n->_idx] ) { // Unreachable
assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
return;
}
uint i;
for( i = 0; i < n->req(); i++ )
verify_compare( n->in(i), loop_verify, visited );
// Check the '_nodes' block/loop structure
i = n->_idx;
if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
if( _nodes[i] != loop_verify->_nodes[i] &&
get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
tty->print("Mismatched control setting for: ");
n->dump();
if( fail++ > 10 ) return;
Node *c = get_ctrl_no_update(n);
tty->print("We have it as: ");
if( c->in(0) ) c->dump();
else tty->print_cr("N%d",c->_idx);
tty->print("Verify thinks: ");
if( loop_verify->has_ctrl(n) )
loop_verify->get_ctrl_no_update(n)->dump();
else
loop_verify->get_loop_idx(n)->dump();
tty->cr();
}
} else { // We have a loop
IdealLoopTree *us = get_loop_idx(n);
if( loop_verify->has_ctrl(n) ) {
tty->print("Mismatched loop setting for: ");
n->dump();
if( fail++ > 10 ) return;
tty->print("We have it as: ");
us->dump();
tty->print("Verify thinks: ");
loop_verify->get_ctrl_no_update(n)->dump();
tty->cr();
} else if (!C->major_progress()) {
// Loop selection can be messed up if we did a major progress
// operation, like split-if. Do not verify in that case.
IdealLoopTree *them = loop_verify->get_loop_idx(n);
if( us->_head != them->_head || us->_tail != them->_tail ) {
tty->print("Unequals loops for: ");
n->dump();
if( fail++ > 10 ) return;
tty->print("We have it as: ");
us->dump();
tty->print("Verify thinks: ");
them->dump();
tty->cr();
}
}
}
// Check for immediate dominators being equal
if( i >= _idom_size ) {
if( !n->is_CFG() ) return;
tty->print("CFG Node with no idom: ");
n->dump();
return;
}
if( !n->is_CFG() ) return;
if( n == C->root() ) return; // No IDOM here
assert(n->_idx == i, "sanity");
Node *id = idom_no_update(n);
if( id != loop_verify->idom_no_update(n) ) {
tty->print("Unequals idoms for: ");
n->dump();
if( fail++ > 10 ) return;
tty->print("We have it as: ");
id->dump();
tty->print("Verify thinks: ");
loop_verify->idom_no_update(n)->dump();
tty->cr();
}
}
//------------------------------verify_tree------------------------------------
// Verify that tree structures match. Because the CFG can change, siblings
// within the loop tree can be reordered. We attempt to deal with that by
// reordering the verify's loop tree if possible.
void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
assert( _parent == parent, "Badly formed loop tree" );
// Siblings not in same order? Attempt to re-order.
if( _head != loop->_head ) {
// Find _next pointer to update
IdealLoopTree **pp = &loop->_parent->_child;
while( *pp != loop )
pp = &((*pp)->_next);
// Find proper sibling to be next
IdealLoopTree **nn = &loop->_next;
while( (*nn) && (*nn)->_head != _head )
nn = &((*nn)->_next);
// Check for no match.
if( !(*nn) ) {
// Annoyingly, irreducible loops can pick different headers
// after a major_progress operation, so the rest of the loop
// tree cannot be matched.
if (_irreducible && Compile::current()->major_progress()) return;
assert( 0, "failed to match loop tree" );
}
// Move (*nn) to (*pp)
IdealLoopTree *hit = *nn;
*nn = hit->_next;
hit->_next = loop;
*pp = loop;
loop = hit;
// Now try again to verify
}
assert( _head == loop->_head , "mismatched loop head" );
Node *tail = _tail; // Inline a non-updating version of
while( !tail->in(0) ) // the 'tail()' call.
tail = tail->in(1);
assert( tail == loop->_tail, "mismatched loop tail" );
ssssssss67
最新推荐文章于 2024-05-11 15:00:28 发布