Google Git
Sign in
android / toolchain / jdk / jdk21 / HEAD / . / src / hotspot / share / opto / loopopts.cpp
blob: be9f5599d535d32804c7d6933052d9390f341c13 [file] [log] [blame]
/*
* Copyright (c) 1999, 2024, 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 "gc/shared/barrierSet.hpp"
#include "gc/shared/c2/barrierSetC2.hpp"
#include "memory/allocation.inline.hpp"
#include "memory/resourceArea.hpp"
#include "opto/addnode.hpp"
#include "opto/callnode.hpp"
#include "opto/castnode.hpp"
#include "opto/connode.hpp"
#include "opto/castnode.hpp"
#include "opto/divnode.hpp"
#include "opto/loopnode.hpp"
#include "opto/matcher.hpp"
#include "opto/mulnode.hpp"
#include "opto/movenode.hpp"
#include "opto/opaquenode.hpp"
#include "opto/rootnode.hpp"
#include "opto/subnode.hpp"
#include "opto/subtypenode.hpp"
#include "opto/vectornode.hpp"
#include "utilities/macros.hpp"
//=============================================================================
//------------------------------split_thru_phi---------------------------------
// Split Node 'n' through merge point if there is enough win.
Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
// ConvI2L may have type information on it which is unsafe to push up
// so disable this for now
return nullptr;
}
// Splitting range check CastIIs through a loop induction Phi can
// cause new Phis to be created that are left unrelated to the loop
// induction Phi and prevent optimizations (vectorization)
if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
n->in(1) == region->as_CountedLoop()->phi()) {
return nullptr;
}
if (cannot_split_division(n, region)) {
return nullptr;
}
int wins = 0;
assert(!n->is_CFG(), "");
assert(region->is_Region(), "");
const Type* type = n->bottom_type();
const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
Node* phi;
if (t_oop != nullptr && t_oop->is_known_instance_field()) {
int iid = t_oop->instance_id();
int index = C->get_alias_index(t_oop);
int offset = t_oop->offset();
phi = new PhiNode(region, type, nullptr, iid, index, offset);
} else {
phi = PhiNode::make_blank(region, n);
}
uint old_unique = C->unique();
for (uint i = 1; i < region->req(); i++) {
Node* x;
Node* the_clone = nullptr;
if (region->in(i) == C->top()) {
x = C->top(); // Dead path? Use a dead data op
} else {
x = n->clone(); // Else clone up the data op
the_clone = x; // Remember for possible deletion.
// Alter data node to use pre-phi inputs
if (n->in(0) == region)
x->set_req( 0, region->in(i) );
for (uint j = 1; j < n->req(); j++) {
Node* in = n->in(j);
if (in->is_Phi() && in->in(0) == region)
x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
}
}
// Check for a 'win' on some paths
const Type* t = x->Value(&_igvn);
bool singleton = t->singleton();
// A TOP singleton indicates that there are no possible values incoming
// along a particular edge. In most cases, this is OK, and the Phi will
// be eliminated later in an Ideal call. However, we can't allow this to
// happen if the singleton occurs on loop entry, as the elimination of
// the PhiNode may cause the resulting node to migrate back to a previous
// loop iteration.
if (singleton && t == Type::TOP) {
// Is_Loop() == false does not confirm the absence of a loop (e.g., an
// irreducible loop may not be indicated by an affirmative is_Loop());
// therefore, the only top we can split thru a phi is on a backedge of
// a loop.
singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
}
if (singleton) {
wins++;
x = ((PhaseGVN&)_igvn).makecon(t);
} else {
// We now call Identity to try to simplify the cloned node.
// Note that some Identity methods call phase->type(this).
// Make sure that the type array is big enough for
// our new node, even though we may throw the node away.
// (Note: This tweaking with igvn only works because x is a new node.)
_igvn.set_type(x, t);
// If x is a TypeNode, capture any more-precise type permanently into Node
// otherwise it will be not updated during igvn->transform since
// igvn->type(x) is set to x->Value() already.
x->raise_bottom_type(t);
Node* y = x->Identity(&_igvn);
if (y != x) {
wins++;
x = y;
} else {
y = _igvn.hash_find(x);
if (y) {
wins++;
x = y;
} else {
// Else x is a new node we are keeping
// We do not need register_new_node_with_optimizer
// because set_type has already been called.
_igvn._worklist.push(x);
}
}
}
if (x != the_clone && the_clone != nullptr)
_igvn.remove_dead_node(the_clone);
phi->set_req( i, x );
}
// Too few wins?
if (wins <= policy) {
_igvn.remove_dead_node(phi);
return nullptr;
}
// Record Phi
register_new_node( phi, region );
for (uint i2 = 1; i2 < phi->req(); i2++) {
Node *x = phi->in(i2);
// If we commoned up the cloned 'x' with another existing Node,
// the existing Node picks up a new use. We need to make the
// existing Node occur higher up so it dominates its uses.
Node *old_ctrl;
IdealLoopTree *old_loop;
if (x->is_Con()) {
// Constant's control is always root.
set_ctrl(x, C->root());
continue;
}
// The occasional new node
if (x->_idx >= old_unique) { // Found a new, unplaced node?
old_ctrl = nullptr;
old_loop = nullptr; // Not in any prior loop
} else {
old_ctrl = get_ctrl(x);
old_loop = get_loop(old_ctrl); // Get prior loop
}
// New late point must dominate new use
Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
if (new_ctrl == old_ctrl) // Nothing is changed
continue;
IdealLoopTree *new_loop = get_loop(new_ctrl);
// Don't move x into a loop if its uses are
// outside of loop. Otherwise x will be cloned
// for each use outside of this loop.
IdealLoopTree *use_loop = get_loop(region);
if (!new_loop->is_member(use_loop) &&
(old_loop == nullptr || !new_loop->is_member(old_loop))) {
// Take early control, later control will be recalculated
// during next iteration of loop optimizations.
new_ctrl = get_early_ctrl(x);
new_loop = get_loop(new_ctrl);
}
// Set new location
set_ctrl(x, new_ctrl);
// If changing loop bodies, see if we need to collect into new body
if (old_loop != new_loop) {
if (old_loop && !old_loop->_child)
old_loop->_body.yank(x);
if (!new_loop->_child)
new_loop->_body.push(x); // Collect body info
}
}
return phi;
}
// Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
// of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
// this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
// a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
// include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
// We also need to check other loop phis as they could have been created in the same split-if pass when applying
// PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
const Type* zero;
switch (n->Opcode()) {
case Op_DivI:
case Op_ModI:
zero = TypeInt::ZERO;
break;
case Op_DivL:
case Op_ModL:
zero = TypeLong::ZERO;
break;
default:
return false;
}
if (n->in(0) != nullptr) {
// Cannot split through phi if Div or Mod node has a control dependency to a zero check.
return true;
}
Node* divisor = n->in(2);
return is_divisor_counted_loop_phi(divisor, region) &&
loop_phi_backedge_type_contains_zero(divisor, zero);
}
bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) {
return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop;
}
bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
}
//------------------------------dominated_by------------------------------------
// Replace the dominated test with an obvious true or false. Place it on the
// IGVN worklist for later cleanup. Move control-dependent data Nodes on the
// live path up to the dominating control.
void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
// prevdom is the dominating projection of the dominating test.
assert(iff->Opcode() == Op_If ||
iff->Opcode() == Op_CountedLoopEnd ||
iff->Opcode() == Op_LongCountedLoopEnd ||
iff->Opcode() == Op_RangeCheck ||
iff->Opcode() == Op_ParsePredicate,
"Check this code when new subtype is added");
int pop = prevdom->Opcode();
assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
if (flip) {
if (pop == Op_IfTrue)
pop = Op_IfFalse;
else
pop = Op_IfTrue;
}
// 'con' is set to true or false to kill the dominated test.
Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
set_ctrl(con, C->root()); // Constant gets a new use
// Hack the dominated test
_igvn.replace_input_of(iff, 1, con);
// If I dont have a reachable TRUE and FALSE path following the IfNode then
// I can assume this path reaches an infinite loop. In this case it's not
// important to optimize the data Nodes - either the whole compilation will
// be tossed or this path (and all data Nodes) will go dead.
if (iff->outcnt() != 2) return;
// Make control-dependent data Nodes on the live path (path that will remain
// once the dominated IF is removed) become control-dependent on the
// dominating projection.
Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
// Loop predicates may have depending checks which should not
// be skipped. For example, range check predicate has two checks
// for lower and upper bounds.
if (dp == nullptr)
return;
ProjNode* dp_proj = dp->as_Proj();
ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
if (exclude_loop_predicate &&
(unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr ||
unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr ||
unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != nullptr)) {
// If this is a range check (IfNode::is_range_check), do not
// reorder because Compile::allow_range_check_smearing might have
// changed the check.
return; // Let IGVN transformation change control dependence.
}
IdealLoopTree* old_loop = get_loop(dp);
for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
Node* cd = dp->fast_out(i); // Control-dependent node
// Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
assert(cd->in(0) == dp, "");
_igvn.replace_input_of(cd, 0, prevdom);
set_early_ctrl(cd, false);
IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
if (old_loop != new_loop) {
if (!old_loop->_child) {
old_loop->_body.yank(cd);
}
if (!new_loop->_child) {
new_loop->_body.push(cd);
}
}
--i;
--imax;
}
}
}
//------------------------------has_local_phi_input----------------------------
// Return TRUE if 'n' has Phi inputs from its local block and no other
// block-local inputs (all non-local-phi inputs come from earlier blocks)
Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
Node *n_ctrl = get_ctrl(n);
// See if some inputs come from a Phi in this block, or from before
// this block.
uint i;
for( i = 1; i < n->req(); i++ ) {
Node *phi = n->in(i);
if( phi->is_Phi() && phi->in(0) == n_ctrl )
break;
}
if( i >= n->req() )
return nullptr; // No Phi inputs; nowhere to clone thru
// Check for inputs created between 'n' and the Phi input. These
// must split as well; they have already been given the chance
// (courtesy of a post-order visit) and since they did not we must
// recover the 'cost' of splitting them by being very profitable
// when splitting 'n'. Since this is unlikely we simply give up.
for( i = 1; i < n->req(); i++ ) {
Node *m = n->in(i);
if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
// We allow the special case of AddP's with no local inputs.
// This allows us to split-up address expressions.
if (m->is_AddP() &&
get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
// Move the AddP up to the dominating point. That's fine because control of m's inputs
// must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
Node* c = find_non_split_ctrl(idom(n_ctrl));
if (c->is_OuterStripMinedLoop()) {
c->as_Loop()->verify_strip_mined(1);
c = c->in(LoopNode::EntryControl);
}
set_ctrl_and_loop(m, c);
continue;
}
return nullptr;
}
assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
}
return n_ctrl;
}
// Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
assert(bt == T_INT || bt == T_LONG, "only for integers");
int n_op = n->Opcode();
if (n_op == Op_LShift(bt)) {
// Scale is loop invariant
Node* scale = n->in(2);
Node* scale_ctrl = get_ctrl(scale);
IdealLoopTree* scale_loop = get_loop(scale_ctrl);
if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
return nullptr;
}
const TypeInt* scale_t = scale->bottom_type()->isa_int();
if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
return nullptr; // Dont bother with byte/short masking
}
// Add must vary with loop (else shift would be loop-invariant)
Node* add = n->in(1);
Node* add_ctrl = get_ctrl(add);
IdealLoopTree* add_loop = get_loop(add_ctrl);
if (n_loop != add_loop) {
return nullptr; // happens w/ evil ZKM loops
}
// Convert I-V into I+ (0-V); same for V-I
if (add->Opcode() == Op_Sub(bt) &&
_igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
Node* zero = _igvn.integercon(0, bt);
set_ctrl(zero, C->root());
Node* neg = SubNode::make(zero, add->in(2), bt);
register_new_node(neg, get_ctrl(add->in(2)));
add = AddNode::make(add->in(1), neg, bt);
register_new_node(add, add_ctrl);
}
if (add->Opcode() != Op_Add(bt)) return nullptr;
assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
// See if one add input is loop invariant
Node* add_var = add->in(1);
Node* add_var_ctrl = get_ctrl(add_var);
IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
Node* add_invar = add->in(2);
Node* add_invar_ctrl = get_ctrl(add_invar);
IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
if (add_invar_loop == n_loop) {
// Swap to find the invariant part
add_invar = add_var;
add_invar_ctrl = add_var_ctrl;
add_invar_loop = add_var_loop;
add_var = add->in(2);
} else if (add_var_loop != n_loop) { // Else neither input is loop invariant
return nullptr;
}
if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
return nullptr; // No invariant part of the add?
}
// Yes! Reshape address expression!
Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
Node* inv_scale_ctrl =
dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
add_invar_ctrl : scale_ctrl;
register_new_node(inv_scale, inv_scale_ctrl);
Node* var_scale = LShiftNode::make(add_var, scale, bt);
register_new_node(var_scale, n_ctrl);
Node* var_add = AddNode::make(var_scale, inv_scale, bt);
register_new_node(var_add, n_ctrl);
_igvn.replace_node(n, var_add);
return var_add;
}
return nullptr;
}
//------------------------------remix_address_expressions----------------------
// Rework addressing expressions to get the most loop-invariant stuff
// moved out. We'd like to do all associative operators, but it's especially
// important (common) to do address expressions.
Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
if (!has_ctrl(n)) return nullptr;
Node* n_ctrl = get_ctrl(n);
IdealLoopTree* n_loop = get_loop(n_ctrl);
// See if 'n' mixes loop-varying and loop-invariant inputs and
// itself is loop-varying.
// Only interested in binary ops (and AddP)
if (n->req() < 3 || n->req() > 4) return nullptr;
Node* n1_ctrl = get_ctrl(n->in( 1));
Node* n2_ctrl = get_ctrl(n->in( 2));
Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
IdealLoopTree* n1_loop = get_loop(n1_ctrl);
IdealLoopTree* n2_loop = get_loop(n2_ctrl);
IdealLoopTree* n3_loop = get_loop(n3_ctrl);
// Does one of my inputs spin in a tighter loop than self?
if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
(n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
(n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
return nullptr; // Leave well enough alone
}
// Is at least one of my inputs loop-invariant?
if (n1_loop == n_loop &&
n2_loop == n_loop &&
n3_loop == n_loop) {
return nullptr; // No loop-invariant inputs
}
Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
if (res != nullptr) {
return res;
}
res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
if (res != nullptr) {
return res;
}
int n_op = n->Opcode();
// Replace (I+V) with (V+I)
if (n_op == Op_AddI ||
n_op == Op_AddL ||
n_op == Op_AddF ||
n_op == Op_AddD ||
n_op == Op_MulI ||
n_op == Op_MulL ||
n_op == Op_MulF ||
n_op == Op_MulD) {
if (n2_loop == n_loop) {
assert(n1_loop != n_loop, "");
n->swap_edges(1, 2);
}
}
// Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
// but not if I2 is a constant. Skip for irreducible loops.
if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
if (n2_loop == n_loop && n3_loop != n_loop) {
if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
IdealLoopTree* n22loop = get_loop(n22_ctrl);
IdealLoopTree* n23_loop = get_loop(n23_ctrl);
if (n22loop != n_loop && n22loop->is_member(n_loop) &&
n23_loop == n_loop) {
Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
// Stuff new AddP in the loop preheader
register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
register_new_node(add2, n_ctrl);
_igvn.replace_node(n, add2);
return add2;
}
}
}
// Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
if (n2_loop != n_loop && n3_loop == n_loop) {
if (n->in(3)->Opcode() == Op_AddX) {
Node* V = n->in(3)->in(1);
Node* I = n->in(3)->in(2);
if (is_member(n_loop,get_ctrl(V))) {
} else {
Node *tmp = V; V = I; I = tmp;
}
if (!is_member(n_loop,get_ctrl(I))) {
Node* add1 = new AddPNode(n->in(1), n->in(2), I);
// Stuff new AddP in the loop preheader
register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
Node* add2 = new AddPNode(n->in(1), add1, V);
register_new_node(add2, n_ctrl);
_igvn.replace_node(n, add2);
return add2;
}
}
}
}
return nullptr;
}
// Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
assert(n->Opcode() == Op_AddI, "sanity");
Node * nn = nullptr;
Node * in1 = n->in(1);
Node * in2 = n->in(2);
if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
IdealLoopTree* loop_n = get_loop(get_ctrl(n));
if (loop_n->is_counted() &&
loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
Matcher::match_rule_supported(Op_MulAddVS2VI) &&
Matcher::match_rule_supported(Op_MulAddS2I)) {
Node* mul_in1 = in1->in(1);
Node* mul_in2 = in1->in(2);
Node* mul_in3 = in2->in(1);
Node* mul_in4 = in2->in(2);
if (mul_in1->Opcode() == Op_LoadS &&
mul_in2->Opcode() == Op_LoadS &&
mul_in3->Opcode() == Op_LoadS &&
mul_in4->Opcode() == Op_LoadS) {
IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
// All nodes should be in the same counted loop.
if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
Node* adr1 = mul_in1->in(MemNode::Address);
Node* adr2 = mul_in2->in(MemNode::Address);
Node* adr3 = mul_in3->in(MemNode::Address);
Node* adr4 = mul_in4->in(MemNode::Address);
if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
(adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
register_new_node(nn, get_ctrl(n));
_igvn.replace_node(n, nn);
return nn;
} else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
(adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
register_new_node(nn, get_ctrl(n));
_igvn.replace_node(n, nn);
return nn;
}
}
}
}
}
}
return nn;
}
//------------------------------conditional_move-------------------------------
// Attempt to replace a Phi with a conditional move. We have some pretty
// strict profitability requirements. All Phis at the merge point must
// be converted, so we can remove the control flow. We need to limit the
// number of c-moves to a small handful. All code that was in the side-arms
// of the CFG diamond is now speculatively executed. This code has to be
// "cheap enough". We are pretty much limited to CFG diamonds that merge
// 1 or 2 items with a total of 1 or 2 ops executed speculatively.
Node *PhaseIdealLoop::conditional_move( Node *region ) {
assert(region->is_Region(), "sanity check");
if (region->req() != 3) return nullptr;
// Check for CFG diamond
Node *lp = region->in(1);
Node *rp = region->in(2);
if (!lp || !rp) return nullptr;
Node *lp_c = lp->in(0);
if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
IfNode *iff = lp_c->as_If();
// Check for ops pinned in an arm of the diamond.
// Can't remove the control flow in this case
if (lp->outcnt() > 1) return nullptr;
if (rp->outcnt() > 1) return nullptr;
IdealLoopTree* r_loop = get_loop(region);
assert(r_loop == get_loop(iff), "sanity");
// Always convert to CMOVE if all results are used only outside this loop.
bool used_inside_loop = (r_loop == _ltree_root);
// Check profitability
int cost = 0;
int phis = 0;
for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
Node *out = region->fast_out(i);
if (!out->is_Phi()) continue; // Ignore other control edges, etc
phis++;
PhiNode* phi = out->as_Phi();
BasicType bt = phi->type()->basic_type();
switch (bt) {
case T_DOUBLE:
case T_FLOAT:
if (C->use_cmove()) {
continue; //TODO: maybe we want to add some cost
}
cost += Matcher::float_cmove_cost(); // Could be very expensive
break;
case T_LONG: {
cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
}
case T_INT: // These all CMOV fine
case T_ADDRESS: { // (RawPtr)
cost++;
break;
}
case T_NARROWOOP: // Fall through
case T_OBJECT: { // Base oops are OK, but not derived oops
const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
// Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
// CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
// CMOVE'ing a derived pointer requires we also CMOVE the base. If we
// have a Phi for the base here that we convert to a CMOVE all is well
// and good. But if the base is dead, we'll not make a CMOVE. Later
// the allocator will have to produce a base by creating a CMOVE of the
// relevant bases. This puts the allocator in the business of
// manufacturing expensive instructions, generally a bad plan.
// Just Say No to Conditionally-Moved Derived Pointers.
if (tp && tp->offset() != 0)
return nullptr;
cost++;
break;
}
default:
return nullptr; // In particular, can't do memory or I/O
}
// Add in cost any speculative ops
for (uint j = 1; j < region->req(); j++) {
Node *proj = region->in(j);
Node *inp = phi->in(j);
if (get_ctrl(inp) == proj) { // Found local op
cost++;
// Check for a chain of dependent ops; these will all become
// speculative in a CMOV.
for (uint k = 1; k < inp->req(); k++)
if (get_ctrl(inp->in(k)) == proj)
cost += ConditionalMoveLimit; // Too much speculative goo
}
}
// See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
// This will likely Split-If, a higher-payoff operation.
for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
Node* use = phi->fast_out(k);
if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
cost += ConditionalMoveLimit;
// Is there a use inside the loop?
// Note: check only basic types since CMoveP is pinned.
if (!used_inside_loop && is_java_primitive(bt)) {
IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
if (r_loop == u_loop || r_loop->is_member(u_loop)) {
used_inside_loop = true;
}
}
}
}//for
Node* bol = iff->in(1);
if (bol->Opcode() == Op_Opaque4) {
return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
}
assert(bol->Opcode() == Op_Bool, "Unexpected node");
int cmp_op = bol->in(1)->Opcode();
if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
return nullptr;
}
// It is expensive to generate flags from a float compare.
// Avoid duplicated float compare.
if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
float infrequent_prob = PROB_UNLIKELY_MAG(3);
// Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
if (used_inside_loop) {
if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
// BlockLayoutByFrequency optimization moves infrequent branch
// from hot path. No point in CMOV'ing in such case (110 is used
// instead of 100 to take into account not exactness of float value).
if (BlockLayoutByFrequency) {
infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
}
}
// Check for highly predictable branch. No point in CMOV'ing if
// we are going to predict accurately all the time.
if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
//keep going
} else if (iff->_prob < infrequent_prob ||
iff->_prob > (1.0f - infrequent_prob))
return nullptr;
// --------------
// Now replace all Phis with CMOV's
Node *cmov_ctrl = iff->in(0);
uint flip = (lp->Opcode() == Op_IfTrue);
Node_List wq;
while (1) {
PhiNode* phi = nullptr;
for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
Node *out = region->fast_out(i);
if (out->is_Phi()) {
phi = out->as_Phi();
break;
}
}
if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
break;
}
if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
// Move speculative ops
wq.push(phi);
while (wq.size() > 0) {
Node *n = wq.pop();
for (uint j = 1; j < n->req(); j++) {
Node* m = n->in(j);
if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
#ifndef PRODUCT
if (PrintOpto && VerifyLoopOptimizations) {
tty->print(" speculate: ");
m->dump();
}
#endif
set_ctrl(m, cmov_ctrl);
wq.push(m);
}
}
}
Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
register_new_node( cmov, cmov_ctrl );
_igvn.replace_node( phi, cmov );
#ifndef PRODUCT
if (TraceLoopOpts) {
tty->print("CMOV ");
r_loop->dump_head();
if (Verbose) {
bol->in(1)->dump(1);
cmov->dump(1);
}
}
DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
#endif
}
// The useless CFG diamond will fold up later; see the optimization in
// RegionNode::Ideal.
_igvn._worklist.push(region);
return iff->in(1);
}
static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
Node* u = m->fast_out(i);
if (u->is_CFG()) {
if (u->is_NeverBranch()) {
u = u->as_NeverBranch()->proj_out(0);
enqueue_cfg_uses(u, wq);
} else {
wq.push(u);
}
}
}
}
// Try moving a store out of a loop, right before the loop
Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
// Store has to be first in the loop body
IdealLoopTree *n_loop = get_loop(n_ctrl);
if (n->is_Store() && n_loop != _ltree_root &&
n_loop->is_loop() && n_loop->_head->is_Loop() &&
n->in(0) != nullptr) {
Node* address = n->in(MemNode::Address);
Node* value = n->in(MemNode::ValueIn);
Node* mem = n->in(MemNode::Memory);
IdealLoopTree* address_loop = get_loop(get_ctrl(address));
IdealLoopTree* value_loop = get_loop(get_ctrl(value));
// - address and value must be loop invariant
// - memory must be a memory Phi for the loop
// - Store must be the only store on this memory slice in the
// loop: if there's another store following this one then value
// written at iteration i by the second store could be overwritten
// at iteration i+n by the first store: it's not safe to move the
// first store out of the loop
// - nothing must observe the memory Phi: it guarantees no read
// before the store, we are also guaranteed the store post
// dominates the loop head (ignoring a possible early
// exit). Otherwise there would be extra Phi involved between the
// loop's Phi and the store.
// - there must be no early exit from the loop before the Store
// (such an exit most of the time would be an extra use of the
// memory Phi but sometimes is a bottom memory Phi that takes the
// store as input).
if (!n_loop->is_member(address_loop) &&
!n_loop->is_member(value_loop) &&
mem->is_Phi() && mem->in(0) == n_loop->_head &&
mem->outcnt() == 1 &&
mem->in(LoopNode::LoopBackControl) == n) {
assert(n_loop->_tail != nullptr, "need a tail");
assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
// Verify that there's no early exit of the loop before the store.
bool ctrl_ok = false;
{
// Follow control from loop head until n, we exit the loop or
// we reach the tail
ResourceMark rm;
Unique_Node_List wq;
wq.push(n_loop->_head);
for (uint next = 0; next < wq.size(); ++next) {
Node *m = wq.at(next);
if (m == n->in(0)) {
ctrl_ok = true;
continue;
}
assert(!has_ctrl(m), "should be CFG");
if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
ctrl_ok = false;
break;
}
enqueue_cfg_uses(m, wq);
if (wq.size() > 10) {
ctrl_ok = false;
break;
}
}
}
if (ctrl_ok) {
// move the Store
_igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
_igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
_igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
// Disconnect the phi now. An empty phi can confuse other
// optimizations in this pass of loop opts.
_igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
n_loop->_body.yank(mem);
set_ctrl_and_loop(n, n->in(0));
return n;
}
}
}
return nullptr;
}
// Try moving a store out of a loop, right after the loop
void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
if (n->is_Store() && n->in(0) != nullptr) {
Node *n_ctrl = get_ctrl(n);
IdealLoopTree *n_loop = get_loop(n_ctrl);
// Store must be in a loop
if (n_loop != _ltree_root && !n_loop->_irreducible) {
Node* address = n->in(MemNode::Address);
Node* value = n->in(MemNode::ValueIn);
IdealLoopTree* address_loop = get_loop(get_ctrl(address));
// address must be loop invariant
if (!n_loop->is_member(address_loop)) {
// Store must be last on this memory slice in the loop and
// nothing in the loop must observe it
Node* phi = nullptr;
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* u = n->fast_out(i);
if (has_ctrl(u)) { // control use?
IdealLoopTree *u_loop = get_loop(get_ctrl(u));
if (!n_loop->is_member(u_loop)) {
continue;
}
if (u->is_Phi() && u->in(0) == n_loop->_head) {
assert(_igvn.type(u) == Type::MEMORY, "bad phi");
// multiple phis on the same slice are possible
if (phi != nullptr) {
return;
}
phi = u;
continue;
}
}
return;
}
if (phi != nullptr) {
// Nothing in the loop before the store (next iteration)
// must observe the stored value
bool mem_ok = true;
{
ResourceMark rm;
Unique_Node_List wq;
wq.push(phi);
for (uint next = 0; next < wq.size() && mem_ok; ++next) {
Node *m = wq.at(next);
for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
Node* u = m->fast_out(i);
if (u->is_Store() || u->is_Phi()) {
if (u != n) {
wq.push(u);
mem_ok = (wq.size() <= 10);
}
} else {
mem_ok = false;
break;
}
}
}
}
if (mem_ok) {
// Move the store out of the loop if the LCA of all
// users (except for the phi) is outside the loop.
Node* hook = new Node(1);
hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
_igvn.rehash_node_delayed(phi);
int count = phi->replace_edge(n, hook, &_igvn);
assert(count > 0, "inconsistent phi");
// Compute latest point this store can go
Node* lca = get_late_ctrl(n, get_ctrl(n));
if (lca->is_OuterStripMinedLoop()) {
lca = lca->in(LoopNode::EntryControl);
}
if (n_loop->is_member(get_loop(lca))) {
// LCA is in the loop - bail out
_igvn.replace_node(hook, n);
return;
}
#ifdef ASSERT
if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
n_loop->_head->as_Loop()->verify_strip_mined(1);
Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
IdealLoopTree* outer_loop = get_loop(outer);
assert(n_loop->_parent == outer_loop, "broken loop tree");
assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
}
#endif
lca = place_outside_loop(lca, n_loop);
assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop");
// Move store out of the loop
_igvn.replace_node(hook, n->in(MemNode::Memory));
_igvn.replace_input_of(n, 0, lca);
set_ctrl_and_loop(n, lca);
// Disconnect the phi now. An empty phi can confuse other
// optimizations in this pass of loop opts..
if (phi->in(LoopNode::LoopBackControl) == phi) {
_igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
n_loop->_body.yank(phi);
}
}
}
}
}
}
}
//------------------------------split_if_with_blocks_pre-----------------------
// Do the real work in a non-recursive function. Data nodes want to be
// cloned in the pre-order so they can feed each other nicely.
Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
// Cloning these guys is unlikely to win
int n_op = n->Opcode();
if (n_op == Op_MergeMem) {
return n;
}
if (n->is_Proj()) {
return n;
}
// Do not clone-up CmpFXXX variations, as these are always
// followed by a CmpI
if (n->is_Cmp()) {
return n;
}
// Attempt to use a conditional move instead of a phi/branch
if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
Node *cmov = conditional_move( n );
if (cmov) {
return cmov;
}
}
if (n->is_CFG() || n->is_LoadStore()) {
return n;
}
if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
if (!C->major_progress()) { // If chance of no more loop opts...
_igvn._worklist.push(n); // maybe we'll remove them
}
return n;
}
if (n->is_Con()) {
return n; // No cloning for Con nodes
}
Node *n_ctrl = get_ctrl(n);
if (!n_ctrl) {
return n; // Dead node
}
Node* res = try_move_store_before_loop(n, n_ctrl);
if (res != nullptr) {
return n;
}
// Attempt to remix address expressions for loop invariants
Node *m = remix_address_expressions( n );
if( m ) return m;
if (n_op == Op_AddI) {
Node *nn = convert_add_to_muladd( n );
if ( nn ) return nn;
}
if (n->is_ConstraintCast()) {
Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
// ConstraintCastNode::dominating_cast() uses node control input to determine domination.
// Node control inputs don't necessarily agree with loop control info (due to
// transformations happened in between), thus additional dominance check is needed
// to keep loop info valid.
if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
_igvn.replace_node(n, dom_cast);
return dom_cast;
}
}
// Determine if the Node has inputs from some local Phi.
// Returns the block to clone thru.
Node *n_blk = has_local_phi_input( n );
if( !n_blk ) return n;
// Do not clone the trip counter through on a CountedLoop
// (messes up the canonical shape).
if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
(n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
return n;
}
// Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
return n;
}
// Check for having no control input; not pinned. Allow
// dominating control.
if (n->in(0)) {
Node *dom = idom(n_blk);
if (dom_lca(n->in(0), dom) != n->in(0)) {
return n;
}
}
// Policy: when is it profitable. You must get more wins than
// policy before it is considered profitable. Policy is usually 0,
// so 1 win is considered profitable. Big merges will require big
// cloning, so get a larger policy.
int policy = n_blk->req() >> 2;
// If the loop is a candidate for range check elimination,
// delay splitting through it's phi until a later loop optimization
if (n_blk->is_BaseCountedLoop()) {
IdealLoopTree *lp = get_loop(n_blk);
if (lp && lp->_rce_candidate) {
return n;
}
}
if (must_throttle_split_if()) return n;
// Split 'n' through the merge point if it is profitable
Node *phi = split_thru_phi( n, n_blk, policy );
if (!phi) return n;
// Found a Phi to split thru!
// Replace 'n' with the new phi
_igvn.replace_node( n, phi );
// Moved a load around the loop, 'en-registering' something.
if (n_blk->is_Loop() && n->is_Load() &&
!phi->in(LoopNode::LoopBackControl)->is_Load())
C->set_major_progress();
return phi;
}
static bool merge_point_too_heavy(Compile* C, Node* region) {
// Bail out if the region and its phis have too many users.
int weight = 0;
for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
weight += region->fast_out(i)->outcnt();
}
int nodes_left = C->max_node_limit() - C->live_nodes();
if (weight * 8 > nodes_left) {
if (PrintOpto) {
tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
}
return true;
} else {
return false;
}
}
static bool merge_point_safe(Node* region) {
// 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
// having a PhiNode input. This sidesteps the dangerous case where the split
// ConvI2LNode may become TOP if the input Value() does not
// overlap the ConvI2L range, leaving a node which may not dominate its
// uses.
// A better fix for this problem can be found in the BugTraq entry, but
// expediency for Mantis demands this hack.
#ifdef _LP64
for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
Node* n = region->fast_out(i);
if (n->is_Phi()) {
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* m = n->fast_out(j);
if (m->Opcode() == Op_ConvI2L)
return false;
if (m->is_CastII()) {
return false;
}
}
}
}
#endif
return true;
}
//------------------------------place_outside_loop---------------------------------
// Place some computation outside of this loop on the path to the use passed as argument
Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
Node* head = loop->_head;
assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
loop = loop->_parent;
assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
}
// Pick control right outside the loop
for (;;) {
Node* dom = idom(useblock);
if (loop->is_member(get_loop(dom)) ||
// NeverBranch nodes are not assigned to the loop when constructed
(dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) {
break;
}
useblock = dom;
}
assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
return useblock;
}
bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
return false;
}
if (!n->in(0)->is_Region()) {
return false;
}
Node* region = n->in(0);
Node* dom = idom(region);
if (!dom->is_If() || dom->in(1) != n->in(1)) {
return false;
}
IfNode* dom_if = dom->as_If();
Node* proj_true = dom_if->proj_out(1);
Node* proj_false = dom_if->proj_out(0);
for (uint i = 1; i < region->req(); i++) {
if (is_dominator(proj_true, region->in(i))) {
continue;
}
if (is_dominator(proj_false, region->in(i))) {
continue;
}
return false;
}
return true;
}
bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
if (must_throttle_split_if()) {
return false;
}
// Do not do 'split-if' if irreducible loops are present.
if (_has_irreducible_loops) {
return false;
}
if (merge_point_too_heavy(C, n_ctrl)) {
return false;
}
// Do not do 'split-if' if some paths are dead. First do dead code
// elimination and then see if its still profitable.
for (uint i = 1; i < n_ctrl->req(); i++) {
if (n_ctrl->in(i) == C->top()) {
return false;
}
}
// If trying to do a 'Split-If' at the loop head, it is only
// profitable if the cmp folds up on BOTH paths. Otherwise we
// risk peeling a loop forever.
// CNC - Disabled for now. Requires careful handling of loop
// body selection for the cloned code. Also, make sure we check
// for any input path not being in the same loop as n_ctrl. For
// irreducible loops we cannot check for 'n_ctrl->is_Loop()'
// because the alternative loop entry points won't be converted
// into LoopNodes.
IdealLoopTree *n_loop = get_loop(n_ctrl);
for (uint j = 1; j < n_ctrl->req(); j++) {
if (get_loop(n_ctrl->in(j)) != n_loop) {
return false;
}
}
// Check for safety of the merge point.
if (!merge_point_safe(n_ctrl)) {
return false;
}
return true;
}
// Detect if the node is the inner strip-mined loop
// Return: null if it's not the case, or the exit of outer strip-mined loop
static Node* is_inner_of_stripmined_loop(const Node* out) {
Node* out_le = nullptr;
if (out->is_CountedLoopEnd()) {
const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
if (loop != nullptr && loop->is_strip_mined()) {
out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
}
}
return out_le;
}
//------------------------------split_if_with_blocks_post----------------------
// Do the real work in a non-recursive function. CFG hackery wants to be
// in the post-order, so it can dirty the I-DOM info and not use the dirtied
// info.
void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
// Cloning Cmp through Phi's involves the split-if transform.
// FastLock is not used by an If
if (n->is_Cmp() && !n->is_FastLock()) {
Node *n_ctrl = get_ctrl(n);
// Determine if the Node has inputs from some local Phi.
// Returns the block to clone thru.
Node *n_blk = has_local_phi_input(n);
if (n_blk != n_ctrl) {
return;
}
if (!can_split_if(n_ctrl)) {
return;
}
if (n->outcnt() != 1) {
return; // Multiple bool's from 1 compare?
}
Node *bol = n->unique_out();
assert(bol->is_Bool(), "expect a bool here");
if (bol->outcnt() != 1) {
return;// Multiple branches from 1 compare?
}
Node *iff = bol->unique_out();
// Check some safety conditions
if (iff->is_If()) { // Classic split-if?
if (iff->in(0) != n_ctrl) {
return; // Compare must be in same blk as if
}
} else if (iff->is_CMove()) { // Trying to split-up a CMOVE
// Can't split CMove with different control.
if (get_ctrl(iff) != n_ctrl) {
return;
}
if (get_ctrl(iff->in(2)) == n_ctrl ||
get_ctrl(iff->in(3)) == n_ctrl) {
return; // Inputs not yet split-up
}
if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
return; // Loop-invar test gates loop-varying CMOVE
}
} else {
return; // some other kind of node, such as an Allocate
}
// When is split-if profitable? Every 'win' on means some control flow
// goes dead, so it's almost always a win.
int policy = 0;
// Split compare 'n' through the merge point if it is profitable
Node *phi = split_thru_phi( n, n_ctrl, policy);
if (!phi) {
return;
}
// Found a Phi to split thru!
// Replace 'n' with the new phi
_igvn.replace_node(n, phi);
// Now split the bool up thru the phi
Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
guarantee(bolphi != nullptr, "null boolean phi node");
_igvn.replace_node(bol, bolphi);
assert(iff->in(1) == bolphi, "");
if (bolphi->Value(&_igvn)->singleton()) {
return;
}
// Conditional-move? Must split up now
if (!iff->is_If()) {
Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
_igvn.replace_node(iff, cmovphi);
return;
}
// Now split the IF
do_split_if(iff);
return;
}
// Two identical ifs back to back can be merged
if (try_merge_identical_ifs(n)) {
return;
}
// Check for an IF ready to split; one that has its
// condition codes input coming from a Phi at the block start.
int n_op = n->Opcode();
// Check for an IF being dominated by another IF same test
if (n_op == Op_If ||
n_op == Op_RangeCheck) {
Node *bol = n->in(1);
uint max = bol->outcnt();
// Check for same test used more than once?
if (max > 1 && bol->is_Bool()) {
// Search up IDOMs to see if this IF is dominated.
Node *cutoff = get_ctrl(bol);
// Now search up IDOMs till cutoff, looking for a dominating test
Node *prevdom = n;
Node *dom = idom(prevdom);
while (dom != cutoff) {
if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
safe_for_if_replacement(dom)) {
// It's invalid to move control dependent data nodes in the inner
// strip-mined loop, because:
// 1) break validation of LoopNode::verify_strip_mined()
// 2) move code with side-effect in strip-mined loop
// Move to the exit of outer strip-mined loop in that case.
Node* out_le = is_inner_of_stripmined_loop(dom);
if (out_le != nullptr) {
prevdom = out_le;
}
// Replace the dominated test with an obvious true or false.
// Place it on the IGVN worklist for later cleanup.
C->set_major_progress();
dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
return;
}
prevdom = dom;
dom = idom(prevdom);
}
}
}
try_sink_out_of_loop(n);
try_move_store_after_loop(n);
}
// Transform:
//
// if (some_condition) {
// // body 1
// } else {
// // body 2
// }
// if (some_condition) {
// // body 3
// } else {
// // body 4
// }
//
// into:
//
//
// if (some_condition) {
// // body 1
// // body 3
// } else {
// // body 2
// // body 4
// }
bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
Node *n_ctrl = n->in(0);
IfNode* dom_if = idom(n_ctrl)->as_If();
ProjNode* dom_proj_true = dom_if->proj_out(1);
ProjNode* dom_proj_false = dom_if->proj_out(0);
// Now split the IF
RegionNode* new_false_region;
RegionNode* new_true_region;
do_split_if(n, &new_false_region, &new_true_region);
assert(new_false_region->req() == new_true_region->req(), "");
#ifdef ASSERT
for (uint i = 1; i < new_false_region->req(); ++i) {
assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
assert(i == new_false_region->req() - 1 || new_false_region->in(i)->in(0)->in(1) == new_false_region->in(i + 1)->in(0)->in(1), "unexpected shape following split if");
}
#endif
assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
// We now have:
// if (some_condition) {
// // body 1
// if (some_condition) {
// body3: // new_true_region
// // body3
// } else {
// goto body4;
// }
// } else {
// // body 2
// if (some_condition) {
// goto body3;
// } else {
// body4: // new_false_region
// // body4;
// }
// }
//
// clone pinned nodes thru the resulting regions
push_pinned_nodes_thru_region(dom_if, new_true_region);
push_pinned_nodes_thru_region(dom_if, new_false_region);
// Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
// on a projection of n to have the dom_if as a control dependency. We don't want the CastPP to end up with an
// unrelated control dependency.
for (uint i = 1; i < new_false_region->req(); i++) {
if (is_dominator(dom_proj_true, new_false_region->in(i))) {
dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
} else {
assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
}
}
return true;
}
return false;
}
void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
for (DUIterator i = region->outs(); region->has_out(i); i++) {
Node* u = region->out(i);
if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
continue;
}
assert(u->in(0) == region, "not a control dependent node?");
uint j = 1;
for (; j < u->req(); ++j) {
Node* in = u->in(j);
if (!is_dominator(ctrl_or_self(in), dom_if)) {
break;
}
}
if (j == u->req()) {
Node *phi = PhiNode::make_blank(region, u);
for (uint k = 1; k < region->req(); ++k) {
Node* clone = u->clone();
clone->set_req(0, region->in(k));
register_new_node(clone, region->in(k));
phi->init_req(k, clone);
}
register_new_node(phi, region);
_igvn.replace_node(u, phi);
--i;
}
}
}
bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
if (!dom->is_CountedLoopEnd()) {
return true;
}
CountedLoopEndNode* le = dom->as_CountedLoopEnd();
CountedLoopNode* cl = le->loopnode();
if (cl == nullptr) {
return true;
}
if (!cl->is_main_loop()) {
return true;
}
if (cl->is_canonical_loop_entry() == nullptr) {
return true;
}
// Further unrolling is possible so loop exit condition might change
return false;
}
// See if a shared loop-varying computation has no loop-varying uses.
// Happens if something is only used for JVM state in uncommon trap exits,
// like various versions of induction variable+offset. Clone the
// computation per usage to allow it to sink out of the loop.
void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
if (has_ctrl(n) &&
!n->is_Phi() &&
!n->is_Bool() &&
!n->is_Proj() &&
!n->is_MergeMem() &&
!n->is_CMove() &&
n->Opcode() != Op_Opaque4 &&
!n->is_Type()) {
Node *n_ctrl = get_ctrl(n);
IdealLoopTree *n_loop = get_loop(n_ctrl);
if (n->in(0) != nullptr) {
IdealLoopTree* loop_ctrl = get_loop(n->in(0));
if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
// n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
// for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
// Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
_igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
}
}
if (n_loop != _ltree_root && n->outcnt() > 1) {
// Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
// previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
Node* outer_loop_clone = nullptr;
for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
Node* u = n->last_out(j); // Clone private computation per use
_igvn.rehash_node_delayed(u);
Node* x = n->clone(); // Clone computation
Node* x_ctrl = nullptr;
if (u->is_Phi()) {
// Replace all uses of normal nodes. Replace Phi uses
// individually, so the separate Nodes can sink down
// different paths.
uint k = 1;
while (u->in(k) != n) k++;
u->set_req(k, x);
// x goes next to Phi input path
x_ctrl = u->in(0)->in(k);
// Find control for 'x' next to use but not inside inner loops.
x_ctrl = place_outside_loop(x_ctrl, n_loop);
--j;
} else { // Normal use
if (has_ctrl(u)) {
x_ctrl = get_ctrl(u);
} else {
x_ctrl = u->in(0);
}
// Find control for 'x' next to use but not inside inner loops.
x_ctrl = place_outside_loop(x_ctrl, n_loop);
// Replace all uses
if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
// If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
// anymore now that we're going to pin n as well
_igvn.replace_node(u, x);
--j;
} else {
int nb = u->replace_edge(n, x, &_igvn);
j -= nb;
}
}
if (n->is_Load()) {
// For loads, add a control edge to a CFG node outside of the loop
// to force them to not combine and return back inside the loop
// during GVN optimization (4641526).
assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
IdealLoopTree* x_loop = get_loop(x_ctrl);
Node* x_head = x_loop->_head;
if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
// Do not add duplicate LoadNodes to the outer strip mined loop
if (outer_loop_clone != nullptr) {
_igvn.replace_node(x, outer_loop_clone);
continue;
}
outer_loop_clone = x;
}
x->set_req(0, x_ctrl);
} else if (n->in(0) != nullptr){
x->set_req(0, x_ctrl);
}
assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
register_new_node(x, x_ctrl);
// Chain of AddP nodes: (AddP base (AddP base (AddP base )))
// All AddP nodes must keep the same base after sinking so:
// 1- We don't add a CastPP here until the last one of the chain is sunk: if part of the chain is not sunk,
// their bases remain the same.
// (see 2- below)
assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
!x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
!(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
assert(!x->is_Load(), "load should be pinned");
// Use a cast node to pin clone out of loop
Node* cast = nullptr;
for (uint k = 0; k < x->req(); k++) {
Node* in = x->in(k);
if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
const Type* in_t = _igvn.type(in);
cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
ConstraintCastNode::UnconditionalDependency, nullptr);
}
if (cast != nullptr) {
Node* prev = _igvn.hash_find_insert(cast);
if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
cast->destruct(&_igvn);
cast = prev;
} else {
register_new_node(cast, x_ctrl);
}
x->replace_edge(in, cast);
// Chain of AddP nodes:
// 2- A CastPP of the base is only added now that all AddP nodes are sunk
if (x->is_AddP() && k == AddPNode::Base) {
update_addp_chain_base(x, n->in(AddPNode::Base), cast);
}
break;
}
}
assert(cast != nullptr, "must have added a cast to pin the node");
}
}
_igvn.remove_dead_node(n);
}
_dom_lca_tags_round = 0;
}
}
}
void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
ResourceMark rm;
Node_List wq;
wq.push(x);
while (wq.size() != 0) {
Node* n = wq.pop();
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* u = n->fast_out(i);
if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
_igvn.replace_input_of(u, AddPNode::Base, new_base);
wq.push(u);
}
}
}
}
// Compute the early control of a node by following its inputs until we reach
// nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
Node* early_ctrl = nullptr;
ResourceMark rm;
Unique_Node_List wq;
wq.push(n);
for (uint i = 0; i < wq.size(); i++) {
Node* m = wq.at(i);
Node* c = nullptr;
if (m->is_CFG()) {
c = m;
} else if (m->pinned()) {
c = m->in(0);
} else {
for (uint j = 0; j < m->req(); j++) {
Node* in = m->in(j);
if (in != nullptr) {
wq.push(in);
}
}
}
if (c != nullptr) {
assert(is_dominator(c, n_ctrl), "control input must dominate current control");
if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
early_ctrl = c;
}
}
}
assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
return early_ctrl;
}
bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* u = n->fast_out(i);
if (u->is_Opaque1()) {
return false; // Found loop limit, bugfix for 4677003
}
// We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
// get_late_ctrl_with_anti_dep() use their own tag
_dom_lca_tags_round++;
assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
if (u->is_Phi()) {
for (uint j = 1; j < u->req(); ++j) {
if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
return false;
}
}
} else {
Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
return false;
}
}
}
return true;
}
bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
if (n->is_Load()) {
ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
}
IdealLoopTree *u_loop = get_loop(ctrl);
if (u_loop == n_loop) {
return false; // Found loop-varying use
}
if (n_loop->is_member(u_loop)) {
return false; // Found use in inner loop
}
// Sinking a node from a pre loop to its main loop pins the node between the pre and main loops. If that node is input
// to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
// test of the pre loop above the point in the graph where it's pinned.
if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop() &&
u_loop->_head->is_CountedLoop() && u_loop->_head->as_CountedLoop()->is_main_loop() &&
n_loop->_next == get_loop(u_loop->_head->as_CountedLoop()->skip_strip_mined())) {
return false;
}
return true;
}
//------------------------------split_if_with_blocks---------------------------
// Check for aggressive application of 'split-if' optimization,
// using basic block level info.
void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
Node* root = C->root();
visited.set(root->_idx); // first, mark root as visited
// Do pre-visit work for root
Node* n = split_if_with_blocks_pre(root);
uint cnt = n->outcnt();
uint i = 0;
while (true) {
// Visit all children
if (i < cnt) {
Node* use = n->raw_out(i);
++i;
if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
// Now do pre-visit work for this use
use = split_if_with_blocks_pre(use);
nstack.push(n, i); // Save parent and next use's index.
n = use; // Process all children of current use.
cnt = use->outcnt();
i = 0;
}
}
else {
// All of n's children have been processed, complete post-processing.
if (cnt != 0 && !n->is_Con()) {
assert(has_node(n), "no dead nodes");
split_if_with_blocks_post(n);
}
if (must_throttle_split_if()) {
nstack.clear();
}
if (nstack.is_empty()) {
// Finished all nodes on stack.
break;
}
// Get saved parent node and next use's index. Visit the rest of uses.
n = nstack.node();
cnt = n->outcnt();
i = nstack.index();
nstack.pop();
}
}
}
//=============================================================================
//
// C L O N E A L O O P B O D Y
//
//------------------------------clone_iff--------------------------------------
// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
// "Nearly" because all Nodes have been cloned from the original in the loop,
// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
// through the Phi recursively, and return a Bool.
Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
// Convert this Phi into a Phi merging Bools
uint i;
for (i = 1; i < phi->req(); i++) {
Node *b = phi->in(i);
if (b->is_Phi()) {
_igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
} else {
assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
}
}
Node* n = phi->in(1);
Node* sample_opaque = nullptr;
Node *sample_bool = nullptr;
if (n->Opcode() == Op_Opaque4) {
sample_opaque = n;
sample_bool = n->in(1);
assert(sample_bool->is_Bool(), "wrong type");
} else {
sample_bool = n;
}
Node *sample_cmp = sample_bool->in(1);
// Make Phis to merge the Cmp's inputs.
PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
for (i = 1; i < phi->req(); i++) {
Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
phi1->set_req(i, n1);
phi2->set_req(i, n2);
phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
}
// See if these Phis have been made before.
// Register with optimizer
Node *hit1 = _igvn.hash_find_insert(phi1);
if (hit1) { // Hit, toss just made Phi
_igvn.remove_dead_node(phi1); // Remove new phi
assert(hit1->is_Phi(), "" );
phi1 = (PhiNode*)hit1; // Use existing phi
} else { // Miss
_igvn.register_new_node_with_optimizer(phi1);
}
Node *hit2 = _igvn.hash_find_insert(phi2);
if (hit2) { // Hit, toss just made Phi
_igvn.remove_dead_node(phi2); // Remove new phi
assert(hit2->is_Phi(), "" );
phi2 = (PhiNode*)hit2; // Use existing phi
} else { // Miss
_igvn.register_new_node_with_optimizer(phi2);
}
// Register Phis with loop/block info
set_ctrl(phi1, phi->in(0));
set_ctrl(phi2, phi->in(0));
// Make a new Cmp
Node *cmp = sample_cmp->clone();
cmp->set_req(1, phi1);
cmp->set_req(2, phi2);
_igvn.register_new_node_with_optimizer(cmp);
set_ctrl(cmp, phi->in(0));
// Make a new Bool
Node *b = sample_bool->clone();
b->set_req(1,cmp);
_igvn.register_new_node_with_optimizer(b);
set_ctrl(b, phi->in(0));
if (sample_opaque != nullptr) {
Node* opaque = sample_opaque->clone();
opaque->set_req(1, b);
_igvn.register_new_node_with_optimizer(opaque);
set_ctrl(opaque, phi->in(0));
return opaque;
}
assert(b->is_Bool(), "");
return b;
}
//------------------------------clone_bool-------------------------------------
// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
// "Nearly" because all Nodes have been cloned from the original in the loop,
// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
// through the Phi recursively, and return a Bool.
CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
uint i;
// Convert this Phi into a Phi merging Bools
for( i = 1; i < phi->req(); i++ ) {
Node *b = phi->in(i);
if( b->is_Phi() ) {
_igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
} else {
assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
}
}
Node *sample_cmp = phi->in(1);
// Make Phis to merge the Cmp's inputs.
PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
for( uint j = 1; j < phi->req(); j++ ) {
Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
Node *n1, *n2;
if( cmp_top->is_Cmp() ) {
n1 = cmp_top->in(1);
n2 = cmp_top->in(2);
} else {
n1 = n2 = cmp_top;
}
phi1->set_req( j, n1 );
phi2->set_req( j, n2 );
phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
}
// See if these Phis have been made before.
// Register with optimizer
Node *hit1 = _igvn.hash_find_insert(phi1);
if( hit1 ) { // Hit, toss just made Phi
_igvn.remove_dead_node(phi1); // Remove new phi
assert( hit1->is_Phi(), "" );
phi1 = (PhiNode*)hit1; // Use existing phi
} else { // Miss
_igvn.register_new_node_with_optimizer(phi1);
}
Node *hit2 = _igvn.hash_find_insert(phi2);
if( hit2 ) { // Hit, toss just made Phi
_igvn.remove_dead_node(phi2); // Remove new phi
assert( hit2->is_Phi(), "" );
phi2 = (PhiNode*)hit2; // Use existing phi
} else { // Miss
_igvn.register_new_node_with_optimizer(phi2);
}
// Register Phis with loop/block info
set_ctrl(phi1, phi->in(0));
set_ctrl(phi2, phi->in(0));
// Make a new Cmp
Node *cmp = sample_cmp->clone();
cmp->set_req( 1, phi1 );
cmp->set_req( 2, phi2 );
_igvn.register_new_node_with_optimizer(cmp);
set_ctrl(cmp, phi->in(0));
assert( cmp->is_Cmp(), "" );
return (CmpNode*)cmp;
}
void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
IdealLoopTree* loop, IdealLoopTree* outer_loop,
Node_List*& split_if_set, Node_List*& split_bool_set,
Node_List*& split_cex_set, Node_List& worklist,
uint new_counter, CloneLoopMode mode) {
Node* nnn = old_new[old->_idx];
// Copy uses to a worklist, so I can munge the def-use info
// with impunity.
for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
worklist.push(old->fast_out(j));
while( worklist.size() ) {
Node *use = worklist.pop();
if (!has_node(use)) continue; // Ignore dead nodes
if (use->in(0) == C->top()) continue;
IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
// Check for data-use outside of loop - at least one of OLD or USE
// must not be a CFG node.
#ifdef ASSERT
if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
assert(mode != IgnoreStripMined, "incorrect cloning mode");
assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
}
#endif
if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
// If the Data use is an IF, that means we have an IF outside of the
// loop that is switching on a condition that is set inside of the
// loop. Happens if people set a loop-exit flag; then test the flag
// in the loop to break the loop, then test is again outside of the
// loop to determine which way the loop exited.
// Loop predicate If node connects to Bool node through Opaque1 node.
//
// If the use is an AllocateArray through its ValidLengthTest input,
// make sure the Bool/Cmp input is cloned down to avoid a Phi between
// the AllocateArray node and its ValidLengthTest input that could cause
// split if to break.
if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4 ||
(use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
// Since this code is highly unlikely, we lazily build the worklist
// of such Nodes to go split.
if (!split_if_set) {
split_if_set = new Node_List();
}
split_if_set->push(use);
}
if (use->is_Bool()) {
if (!split_bool_set) {
split_bool_set = new Node_List();
}
split_bool_set->push(use);
}
if (use->Opcode() == Op_CreateEx) {
if (!split_cex_set) {
split_cex_set = new Node_List();
}
split_cex_set->push(use);
}
// Get "block" use is in
uint idx = 0;
while( use->in(idx) != old ) idx++;
Node *prev = use->is_CFG() ? use : get_ctrl(use);
assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
? prev->in(2)
: idom(prev);
if( use->is_Phi() ) // Phi use is in prior block
cfg = prev->in(idx); // NOT in block of Phi itself
if (cfg->is_top()) { // Use is dead?
_igvn.replace_input_of(use, idx, C->top());
continue;
}
// If use is referenced through control edge... (idx == 0)
if (mode == IgnoreStripMined && idx == 0) {
LoopNode *head = loop->_head->as_Loop();
if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
// That node is outside the inner loop, leave it outside the
// outer loop as well to not confuse verification code.
assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
_igvn.replace_input_of(use, 0, head->outer_loop_exit());
continue;
}
}
while(!outer_loop->is_member(get_loop(cfg))) {
prev = cfg;
cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
}
// If the use occurs after merging several exits from the loop, then
// old value must have dominated all those exits. Since the same old
// value was used on all those exits we did not need a Phi at this
// merge point. NOW we do need a Phi here. Each loop exit value
// is now merged with the peeled body exit; each exit gets its own
// private Phi and those Phis need to be merged here.
Node *phi;
if( prev->is_Region() ) {
if( idx == 0 ) { // Updating control edge?
phi = prev; // Just use existing control
} else { // Else need a new Phi
phi = PhiNode::make( prev, old );
// Now recursively fix up the new uses of old!
for( uint i = 1; i < prev->req(); i++ ) {
worklist.push(phi); // Onto worklist once for each 'old' input
}
}
} else {
// Get new RegionNode merging old and new loop exits
prev = old_new[prev->_idx];
assert( prev, "just made this in step 7" );
if( idx == 0) { // Updating control edge?
phi = prev; // Just use existing control
} else { // Else need a new Phi
// Make a new Phi merging data values properly
phi = PhiNode::make( prev, old );
phi->set_req( 1, nnn );
}
}
// If inserting a new Phi, check for prior hits
if( idx != 0 ) {
Node *hit = _igvn.hash_find_insert(phi);
if( hit == nullptr ) {
_igvn.register_new_node_with_optimizer(phi); // Register new phi
} else { // or
// Remove the new phi from the graph and use the hit
_igvn.remove_dead_node(phi);
phi = hit; // Use existing phi
}
set_ctrl(phi, prev);
}
// Make 'use' use the Phi instead of the old loop body exit value
assert(use->in(idx) == old, "old is still input of use");
// We notify all uses of old, including use, and the indirect uses,
// that may now be optimized because we have replaced old with phi.
_igvn.add_users_to_worklist(old);
_igvn.replace_input_of(use, idx, phi);
if( use->_idx >= new_counter ) { // If updating new phis
// Not needed for correctness, but prevents a weak assert
// in AddPNode from tripping (when we end up with different
// base & derived Phis that will become the same after
// IGVN does CSE).
Node *hit = _igvn.hash_find_insert(use);
if( hit ) // Go ahead and re-hash for hits.
_igvn.replace_node( use, hit );
}
}
}
}
static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
bool check_old_new) {
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* u = n->fast_out(j);
assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
Node* c = phase->get_ctrl(u);
IdealLoopTree* u_loop = phase->get_loop(c);
assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
if (!loop->is_member(u_loop)) {
if (outer_loop->is_member(u_loop)) {
wq.push(u);
} else {
// nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
// the outer loop too
Node* u_c = u->in(0);
if (u_c != nullptr) {
IdealLoopTree* u_c_loop = phase->get_loop(u_c);
if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
wq.push(u);
}
}
}
}
}
}
}
void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
IdealLoopTree* outer_loop, int dd, Node_List &old_new,
Node_List& extra_data_nodes) {
if (head->is_strip_mined() && mode != IgnoreStripMined) {
CountedLoopNode* cl = head->as_CountedLoop();
Node* l = cl->outer_loop();
Node* tail = cl->outer_loop_tail();
IfNode* le = cl->outer_loop_end();
Node* sfpt = cl->outer_safepoint();
CountedLoopEndNode* cle = cl->loopexit();
CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
Node* cle_out = cle->proj_out(false);
Node* new_sfpt = nullptr;
Node* new_cle_out = cle_out->clone();
old_new.map(cle_out->_idx, new_cle_out);
if (mode == CloneIncludesStripMined) {
// clone outer loop body
Node* new_l = l->clone();
Node* new_tail = tail->clone();
IfNode* new_le = le->clone()->as_If();
new_sfpt = sfpt->clone();
set_loop(new_l, outer_loop->_parent);
set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
set_loop(new_cle_out, outer_loop->_parent);
set_idom(new_cle_out, new_cle, dd);
set_loop(new_sfpt, outer_loop->_parent);
set_idom(new_sfpt, new_cle_out, dd);
set_loop(new_le, outer_loop->_parent);
set_idom(new_le, new_sfpt, dd);
set_loop(new_tail, outer_loop->_parent);
set_idom(new_tail, new_le, dd);
set_idom(new_cl, new_l, dd);
old_new.map(l->_idx, new_l);
old_new.map(tail->_idx, new_tail);
old_new.map(le->_idx, new_le);
old_new.map(sfpt->_idx, new_sfpt);
new_l->set_req(LoopNode::LoopBackControl, new_tail);
new_l->set_req(0, new_l);
new_tail->set_req(0, new_le);
new_le->set_req(0, new_sfpt);
new_sfpt->set_req(0, new_cle_out);
new_cle_out->set_req(0, new_cle);
new_cl->set_req(LoopNode::EntryControl, new_l);
_igvn.register_new_node_with_optimizer(new_l);
_igvn.register_new_node_with_optimizer(new_tail);
_igvn.register_new_node_with_optimizer(new_le);
} else {
Node *newhead = old_new[loop->_head->_idx];
newhead->as_Loop()->clear_strip_mined();
_igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
}
// Look at data node that were assigned a control in the outer
// loop: they are kept in the outer loop by the safepoint so start
// from the safepoint node's inputs.
IdealLoopTree* outer_loop = get_loop(l);
Node_Stack stack(2);
stack.push(sfpt, 1);
uint new_counter = C->unique();
while (stack.size() > 0) {
Node* n = stack.node();
uint i = stack.index();
while (i < n->req() &&
(n->in(i) == nullptr ||
!has_ctrl(n->in(i)) ||
get_loop(get_ctrl(n->in(i))) != outer_loop ||
(old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
i++;
}
if (i < n->req()) {
stack.set_index(i+1);
stack.push(n->in(i), 0);
} else {
assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
Node* m = n == sfpt ? new_sfpt : n->clone();
if (m != nullptr) {
for (uint i = 0; i < n->req(); i++) {
if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
m->set_req(i, old_new[m->in(i)->_idx]);
}
}
} else {
assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
}
if (n != sfpt) {
extra_data_nodes.push(n);
_igvn.register_new_node_with_optimizer(m);
assert(get_ctrl(n) == cle_out, "what other control?");
set_ctrl(m, new_cle_out);
old_new.map(n->_idx, m);
}
stack.pop();
}
}
if (mode == CloneIncludesStripMined) {
_igvn.register_new_node_with_optimizer(new_sfpt);
_igvn.register_new_node_with_optimizer(new_cle_out);
}
// Some other transformation may have pessimistically assigned some
// data nodes to the outer loop. Set their control so they are out
// of the outer loop.
ResourceMark rm;
Unique_Node_List wq;
for (uint i = 0; i < extra_data_nodes.size(); i++) {
Node* old = extra_data_nodes.at(i);
collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
}
for (uint i = 0; i < loop->_body.size(); i++) {
Node* old = loop->_body.at(i);
collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
}
Node* inner_out = sfpt->in(0);
if (inner_out->outcnt() > 1) {
collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
}
Node* new_ctrl = cl->outer_loop_exit();
assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
for (uint i = 0; i < wq.size(); i++) {
Node* n = wq.at(i);
set_ctrl(n, new_ctrl);
if (n->in(0) != nullptr) {
_igvn.replace_input_of(n, 0, new_ctrl);
}
collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
}
} else {
Node *newhead = old_new[loop->_head->_idx];
set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
}
}
//------------------------------clone_loop-------------------------------------
//
// C L O N E A L O O P B O D Y
//
// This is the basic building block of the loop optimizations. It clones an
// entire loop body. It makes an old_new loop body mapping; with this mapping
// you can find the new-loop equivalent to an old-loop node. All new-loop
// nodes are exactly equal to their old-loop counterparts, all edges are the
// same. All exits from the old-loop now have a RegionNode that merges the
// equivalent new-loop path. This is true even for the normal "loop-exit"
// condition. All uses of loop-invariant old-loop values now come from (one
// or more) Phis that merge their new-loop equivalents.
//
// This operation leaves the graph in an illegal state: there are two valid
// control edges coming from the loop pre-header to both loop bodies. I'll
// definitely have to hack the graph after running this transform.
//
// From this building block I will further edit edges to perform loop peeling
// or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
//
// Parameter side_by_size_idom:
// When side_by_size_idom is null, the dominator tree is constructed for
// the clone loop to dominate the original. Used in construction of
// pre-main-post loop sequence.
// When nonnull, the clone and original are side-by-side, both are
// dominated by the side_by_side_idom node. Used in construction of
// unswitched loops.
void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
CloneLoopMode mode, Node* side_by_side_idom) {
LoopNode* head = loop->_head->as_Loop();
head->verify_strip_mined(1);
if (C->do_vector_loop() && PrintOpto) {
const char* mname = C->method()->name()->as_quoted_ascii();
if (mname != nullptr) {
tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
}
}
CloneMap& cm = C->clone_map();
if (C->do_vector_loop()) {
cm.set_clone_idx(cm.max_gen()+1);
#ifndef PRODUCT
if (PrintOpto) {
tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
loop->dump_head();
}
#endif
}
// Step 1: Clone the loop body. Make the old->new mapping.
clone_loop_body(loop->_body, old_new, &cm);
IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
// Step 2: Fix the edges in the new body. If the old input is outside the
// loop use it. If the old input is INside the loop, use the corresponding
// new node instead.
fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
Node_List extra_data_nodes; // data nodes in the outer strip mined loop
clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
// Step 3: Now fix control uses. Loop varying control uses have already
// been fixed up (as part of all input edges in Step 2). Loop invariant
// control uses must be either an IfFalse or an IfTrue. Make a merge
// point to merge the old and new IfFalse/IfTrue nodes; make the use
// refer to this.
Node_List worklist;
uint new_counter = C->unique();
fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
// Step 4: If loop-invariant use is not control, it must be dominated by a
// loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
// there if needed. Make a Phi there merging old and new used values.
Node_List *split_if_set = nullptr;
Node_List *split_bool_set = nullptr;
Node_List *split_cex_set = nullptr;
fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
for (uint i = 0; i < extra_data_nodes.size(); i++) {
Node* old = extra_data_nodes.at(i);
clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
split_bool_set, split_cex_set, worklist, new_counter,
mode);
}
// Check for IFs that need splitting/cloning. Happens if an IF outside of
// the loop uses a condition set in the loop. The original IF probably
// takes control from one or more OLD Regions (which in turn get from NEW
// Regions). In any case, there will be a set of Phis for each merge point
// from the IF up to where the original BOOL def exists the loop.
finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
}
void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
if (split_if_set) {
while (split_if_set->size()) {
Node *iff = split_if_set->pop();
uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
if (iff->in(input)->is_Phi()) {
Node *b = clone_iff(iff->in(input)->as_Phi());
_igvn.replace_input_of(iff, input, b);
}
}
}
if (split_bool_set) {
while (split_bool_set->size()) {
Node *b = split_bool_set->pop();
Node *phi = b->in(1);
assert(phi->is_Phi(), "");
CmpNode *cmp = clone_bool((PhiNode*) phi);
_igvn.replace_input_of(b, 1, cmp);
}
}
if (split_cex_set) {
while (split_cex_set->size()) {
Node *b = split_cex_set->pop();
assert(b->in(0)->is_Region(), "");
assert(b->in(1)->is_Phi(), "");
assert(b->in(0)->in(0) == b->in(1)->in(0), "");
split_up(b, b->in(0), nullptr);
}
}
}
void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
Node_List*& split_bool_set, Node_List*& split_cex_set) {
for(uint i = 0; i < body.size(); i++ ) {
Node* old = body.at(i);
clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
split_bool_set, split_cex_set, worklist, new_counter,
mode);
}
}
void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
LoopNode* head = loop->_head->as_Loop();
for(uint i = 0; i < body.size(); i++ ) {
Node* old = body.at(i);
if( !old->is_CFG() ) continue;
// Copy uses to a worklist, so I can munge the def-use info
// with impunity.
for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
worklist.push(old->fast_out(j));
}
while (worklist.size()) { // Visit all uses
Node *use = worklist.pop();
if (!has_node(use)) continue; // Ignore dead nodes
IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
if (!loop->is_member(use_loop) && use->is_CFG()) {
// Both OLD and USE are CFG nodes here.
assert(use->is_Proj(), "" );
Node* nnn = old_new[old->_idx];
Node* newuse = nullptr;
if (head->is_strip_mined() && mode != IgnoreStripMined) {
CountedLoopNode* cl = head->as_CountedLoop();
CountedLoopEndNode* cle = cl->loopexit();
Node* cle_out = cle->proj_out_or_null(false);
if (use == cle_out) {
IfNode* le = cl->outer_loop_end();
use = le->proj_out(false);
use_loop = get_loop(use);
if (mode == CloneIncludesStripMined) {
nnn = old_new[le->_idx];
} else {
newuse = old_new[cle_out->_idx];
}
}
}
if (newuse == nullptr) {
newuse = use->clone();
}
// Clone the loop exit control projection
if (C->do_vector_loop() && cm != nullptr) {
cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
}
newuse->set_req(0,nnn);
_igvn.register_new_node_with_optimizer(newuse);
set_loop(newuse, use_loop);
set_idom(newuse, nnn, dom_depth(nnn) + 1 );
// We need a Region to merge the exit from the peeled body and the
// exit from the old loop body.
RegionNode *r = new RegionNode(3);
uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
// The original user of 'use' uses 'r' instead.
for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
Node* useuse = use->last_out(l);
_igvn.rehash_node_delayed(useuse);
uint uses_found = 0;
if (useuse->in(0) == use) {
useuse->set_req(0, r);
uses_found++;
if (useuse->is_CFG()) {
// This is not a dom_depth > dd_r because when new
// control flow is constructed by a loop opt, a node and
// its dominator can end up at the same dom_depth
assert(dom_depth(useuse) >= dd_r, "");
set_idom(useuse, r, dom_depth(useuse));
}
}
for (uint k = 1; k < useuse->req(); k++) {
if( useuse->in(k) == use ) {
useuse->set_req(k, r);
uses_found++;
if (useuse->is_Loop() && k == LoopNode::EntryControl) {
// This is not a dom_depth > dd_r because when new
// control flow is constructed by a loop opt, a node
// and its dominator can end up at the same dom_depth
assert(dom_depth(useuse) >= dd_r , "");
set_idom(useuse, r, dom_depth(useuse));
}
}
}
l -= uses_found; // we deleted 1 or more copies of this edge
}
assert(use->is_Proj(), "loop exit should be projection");
// lazy_replace() below moves all nodes that are:
// - control dependent on the loop exit or
// - have control set to the loop exit
// below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
// possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
// becomes dead and is "replaced" by the region.
Node* use_clone = use->clone();
register_control(use_clone, use_loop, idom(use), dom_depth(use));
// Now finish up 'r'
r->set_req(1, newuse);
r->set_req(2, use_clone);
_igvn.register_new_node_with_optimizer(r);
set_loop(r, use_loop);
set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
lazy_replace(use, r);
// Map the (cloned) old use to the new merge point
old_new.map(use_clone->_idx, r);
} // End of if a loop-exit test
}
}
}
void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
IdealLoopTree* parent, bool partial) {
for(uint i = 0; i < body.size(); i++ ) {
Node *old = body.at(i);
Node *nnn = old_new[old->_idx];
// Fix CFG/Loop controlling the new node
if (has_ctrl(old)) {
set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
} else {
set_loop(nnn, parent);
if (old->outcnt() > 0) {
Node* dom = idom(old);
if (old_new[dom->_idx] != nullptr) {
dom = old_new[dom->_idx];
set_idom(nnn, dom, dd );
}
}
}
// Correct edges to the new node
for (uint j = 0; j < nnn->req(); j++) {
Node *n = nnn->in(j);
if (n != nullptr) {
IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
if (loop->is_member(old_in_loop)) {
if (old_new[n->_idx] != nullptr) {
nnn->set_req(j, old_new[n->_idx]);
} else {
assert(!body.contains(n), "");
assert(partial, "node not cloned");
}
}
}
}
_igvn.hash_find_insert(nnn);
}
}
void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
for (uint i = 0; i < body.size(); i++) {
Node* old = body.at(i);
Node* nnn = old->clone();
old_new.map(old->_idx, nnn);
if (C->do_vector_loop() && cm != nullptr) {
cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
}
_igvn.register_new_node_with_optimizer(nnn);
}
}
//---------------------- stride_of_possible_iv -------------------------------------
// Looks for an iff/bool/comp with one operand of the compare
// being a cycle involving an add and a phi,
// with an optional truncation (left-shift followed by a right-shift)
// of the add. Returns zero if not an iv.
int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
Node* trunc1 = nullptr;
Node* trunc2 = nullptr;
const TypeInteger* ttype = nullptr;
if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
return 0;
}
BoolNode* bl = iff->in(1)->as_Bool();
Node* cmp = bl->in(1);
if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
return 0;
}
// Must have an invariant operand
if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
return 0;
}
Node* add2 = nullptr;
Node* cmp1 = cmp->in(1);
if (cmp1->is_Phi()) {
// (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
Node* phi = cmp1;
for (uint i = 1; i < phi->req(); i++) {
Node* in = phi->in(i);
Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
&trunc1, &trunc2, &ttype, T_INT);
if (add && add->in(1) == phi) {
add2 = add->in(2);
break;
}
}
} else {
// (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
Node* addtrunc = cmp1;
Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
&trunc1, &trunc2, &ttype, T_INT);
if (add && add->in(1)->is_Phi()) {
Node* phi = add->in(1);
for (uint i = 1; i < phi->req(); i++) {
if (phi->in(i) == addtrunc) {
add2 = add->in(2);
break;
}
}
}
}
if (add2 != nullptr) {
const TypeInt* add2t = _igvn.type(add2)->is_int();
if (add2t->is_con()) {
return add2t->get_con();
}
}
return 0;
}
//---------------------- stay_in_loop -------------------------------------
// Return the (unique) control output node that's in the loop (if it exists.)
Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
Node* unique = nullptr;
if (!n) return nullptr;
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* use = n->fast_out(i);
if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
if (unique != nullptr) {
return nullptr;
}
unique = use;
}
}
return unique;
}
//------------------------------ register_node -------------------------------------
// Utility to register node "n" with PhaseIdealLoop
void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
_igvn.register_new_node_with_optimizer(n);
loop->_body.push(n);
if (n->is_CFG()) {
set_loop(n, loop);
set_idom(n, pred, ddepth);
} else {
set_ctrl(n, pred);
}
}
//------------------------------ proj_clone -------------------------------------
// Utility to create an if-projection
ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
ProjNode* c = p->clone()->as_Proj();
c->set_req(0, iff);
return c;
}
//------------------------------ short_circuit_if -------------------------------------
// Force the iff control output to be the live_proj
Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
guarantee(live_proj != nullptr, "null projection");
int proj_con = live_proj->_con;
assert(proj_con == 0 || proj_con == 1, "false or true projection");
Node *con = _igvn.intcon(proj_con);
set_ctrl(con, C->root());
if (iff) {
iff->set_req(1, con);
}
return con;
}
//------------------------------ insert_if_before_proj -------------------------------------
// Insert a new if before an if projection (* - new node)
//
// before
// if(test)
// / \
// v v
// other-proj proj (arg)
//
// after
// if(test)
// / \
// / v
// | * proj-clone
// v |
// other-proj v
// * new_if(relop(cmp[IU](left,right)))
// / \
// v v
// * new-proj proj
// (returned)
//
ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
IfNode* iff = proj->in(0)->as_If();
IdealLoopTree *loop = get_loop(proj);
ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
uint ddepth = dom_depth(proj);
_igvn.rehash_node_delayed(iff);
_igvn.rehash_node_delayed(proj);
proj->set_req(0, nullptr); // temporary disconnect
ProjNode* proj2 = proj_clone(proj, iff);
register_node(proj2, loop, iff, ddepth);
Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
register_node(cmp, loop, proj2, ddepth);
BoolNode* bol = new BoolNode(cmp, relop);
register_node(bol, loop, proj2, ddepth);
int opcode = iff->Opcode();
assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
register_node(new_if, loop, proj2, ddepth);
proj->set_req(0, new_if); // reattach
set_idom(proj, new_if, ddepth);
ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
guarantee(new_exit != nullptr, "null exit node");
register_node(new_exit, get_loop(other_proj), new_if, ddepth);
return new_exit;
}
//------------------------------ insert_region_before_proj -------------------------------------
// Insert a region before an if projection (* - new node)
//
// before
// if(test)
// / |
// v |
// proj v
// other-proj
//
// after
// if(test)
// / |
// v |
// * proj-clone v
// | other-proj
// v
// * new-region
// |
// v
// * dum_if
// / \
// v \
// * dum-proj v
// proj
//
RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
IfNode* iff = proj->in(0)->as_If();
IdealLoopTree *loop = get_loop(proj);
ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
uint ddepth = dom_depth(proj);
_igvn.rehash_node_delayed(iff);
_igvn.rehash_node_delayed(proj);
proj->set_req(0, nullptr); // temporary disconnect
ProjNode* proj2 = proj_clone(proj, iff);
register_node(proj2, loop, iff, ddepth);
RegionNode* reg = new RegionNode(2);
reg->set_req(1, proj2);
register_node(reg, loop, iff, ddepth);
IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
register_node(dum_if, loop, reg, ddepth);
proj->set_req(0, dum_if); // reattach
set_idom(proj, dum_if, ddepth);
ProjNode* dum_proj = proj_clone(other_proj, dum_if);
register_node(dum_proj, loop, dum_if, ddepth);
return reg;
}
//------------------------------ insert_cmpi_loop_exit -------------------------------------
// Clone a signed compare loop exit from an unsigned compare and
// insert it before the unsigned cmp on the stay-in-loop path.
// All new nodes inserted in the dominator tree between the original
// if and it's projections. The original if test is replaced with
// a constant to force the stay-in-loop path.
//
// This is done to make sure that the original if and it's projections
// still dominate the same set of control nodes, that the ctrl() relation
// from data nodes to them is preserved, and that their loop nesting is
// preserved.
//
// before
// if(i <u limit) unsigned compare loop exit
// / |
// v v
// exit-proj stay-in-loop-proj
//
// after
// if(stay-in-loop-const) original if
// / |
// / v
// / if(i < limit) new signed test
// / / |
// / / v
// / / if(i <u limit) new cloned unsigned test
// / / / |
// v v v |
// region |
// | |
// dum-if |
// / | |
// ether | |
// v v
// exit-proj stay-in-loop-proj
//
IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
const bool Signed = true;
const bool Unsigned = false;
BoolNode* bol = if_cmpu->in(1)->as_Bool();
if (bol->_test._test != BoolTest::lt) return nullptr;
CmpNode* cmpu = bol->in(1)->as_Cmp();
if (cmpu->Opcode() != Op_CmpU) return nullptr;
int stride = stride_of_possible_iv(if_cmpu);
if (stride == 0) return nullptr;
Node* lp_proj = stay_in_loop(if_cmpu, loop);
guarantee(lp_proj != nullptr, "null loop node");
ProjNode* lp_continue = lp_proj->as_Proj();
ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
if (!lp_exit->is_IfFalse()) {
// The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
// We therefore can't add a single exit condition.
return nullptr;
}
// The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
// Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
Node* limit = nullptr;
if (stride > 0) {
limit = cmpu->in(2);
} else {
limit = _igvn.makecon(TypeInt::ZERO);
set_ctrl(limit, C->root());
}
// Create a new region on the exit path
RegionNode* reg = insert_region_before_proj(lp_exit);
guarantee(reg != nullptr, "null region node");
// Clone the if-cmpu-true-false using a signed compare
BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
reg->add_req(cmpi_exit);
// Clone the if-cmpu-true-false
BoolTest::mask rel_u = bol->_test._test;
ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
reg->add_req(cmpu_exit);
// Force original if to stay in loop.
short_circuit_if(if_cmpu, lp_continue);
return cmpi_exit->in(0)->as_If();
}
//------------------------------ remove_cmpi_loop_exit -------------------------------------
// Remove a previously inserted signed compare loop exit.
void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
Node* lp_proj = stay_in_loop(if_cmp, loop);
assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
stay_in_loop(lp_proj, loop)->is_If() &&
stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
set_ctrl(con, C->root());
if_cmp->set_req(1, con);
}
//------------------------------ scheduled_nodelist -------------------------------------
// Create a post order schedule of nodes that are in the
// "member" set. The list is returned in "sched".
// The first node in "sched" is the loop head, followed by
// nodes which have no inputs in the "member" set, and then
// followed by the nodes that have an immediate input dependence
// on a node in "sched".
void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
assert(member.test(loop->_head->_idx), "loop head must be in member set");
VectorSet visited;
Node_Stack nstack(loop->_body.size());
Node* n = loop->_head; // top of stack is cached in "n"
uint idx = 0;
visited.set(n->_idx);
// Initially push all with no inputs from within member set
for(uint i = 0; i < loop->_body.size(); i++ ) {
Node *elt = loop->_body.at(i);
if (member.test(elt->_idx)) {
bool found = false;
for (uint j = 0; j < elt->req(); j++) {
Node* def = elt->in(j);
if (def && member.test(def->_idx) && def != elt) {
found = true;
break;
}
}
if (!found && elt != loop->_head) {
nstack.push(n, idx);
n = elt;
assert(!visited.test(n->_idx), "not seen yet");
visited.set(n->_idx);
}
}
}
// traverse out's that are in the member set
while (true) {
if (idx < n->outcnt()) {
Node* use = n->raw_out(idx);
idx++;
if (!visited.test_set(use->_idx)) {
if (member.test(use->_idx)) {
nstack.push(n, idx);
n = use;
idx = 0;
}
}
} else {
// All outputs processed
sched.push(n);
if (nstack.is_empty()) break;
n = nstack.node();
idx = nstack.index();
nstack.pop();
}
}
}
//------------------------------ has_use_in_set -------------------------------------
// Has a use in the vector set
bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* use = n->fast_out(j);
if (vset.test(use->_idx)) {
return true;
}
}
return false;
}
//------------------------------ has_use_internal_to_set -------------------------------------
// Has use internal to the vector set (ie. not in a phi at the loop head)
bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
Node* head = loop->_head;
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* use = n->fast_out(j);
if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
return true;
}
}
return false;
}
//------------------------------ clone_for_use_outside_loop -------------------------------------
// clone "n" for uses that are outside of loop
int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
int cloned = 0;
assert(worklist.size() == 0, "should be empty");
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* use = n->fast_out(j);
if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
worklist.push(use);
}
}
if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
"Too many clones required in clone_for_use_outside_loop in partial peeling")) {
return -1;
}
while( worklist.size() ) {
Node *use = worklist.pop();
if (!has_node(use) || use->in(0) == C->top()) continue;
uint j;
for (j = 0; j < use->req(); j++) {
if (use->in(j) == n) break;
}
assert(j < use->req(), "must be there");
// clone "n" and insert it between the inputs of "n" and the use outside the loop
Node* n_clone = n->clone();
_igvn.replace_input_of(use, j, n_clone);
cloned++;
Node* use_c;
if (!use->is_Phi()) {
use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
} else {
// Use in a phi is considered a use in the associated predecessor block
use_c = use->in(0)->in(j);
}
set_ctrl(n_clone, use_c);
assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
get_loop(use_c)->_body.push(n_clone);
_igvn.register_new_node_with_optimizer(n_clone);
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
}
#endif
}
return cloned;
}
//------------------------------ clone_for_special_use_inside_loop -------------------------------------
// clone "n" for special uses that are in the not_peeled region.
// If these def-uses occur in separate blocks, the code generator
// marks the method as not compilable. For example, if a "BoolNode"
// is in a different basic block than the "IfNode" that uses it, then
// the compilation is aborted in the code generator.
void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
if (n->is_Phi() || n->is_Load()) {
return;
}
assert(worklist.size() == 0, "should be empty");
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* use = n->fast_out(j);
if ( not_peel.test(use->_idx) &&
(use->is_If() || use->is_CMove() || use->is_Bool()) &&
use->in(1) == n) {
worklist.push(use);
}
}
if (worklist.size() > 0) {
// clone "n" and insert it between inputs of "n" and the use
Node* n_clone = n->clone();
loop->_body.push(n_clone);
_igvn.register_new_node_with_optimizer(n_clone);
set_ctrl(n_clone, get_ctrl(n));
sink_list.push(n_clone);
not_peel.set(n_clone->_idx);
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
}
#endif
while( worklist.size() ) {
Node *use = worklist.pop();
_igvn.rehash_node_delayed(use);
for (uint j = 1; j < use->req(); j++) {
if (use->in(j) == n) {
use->set_req(j, n_clone);
}
}
}
}
}
//------------------------------ insert_phi_for_loop -------------------------------------
// Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
Node *phi = PhiNode::make(lp, back_edge_val);
phi->set_req(LoopNode::EntryControl, lp_entry_val);
// Use existing phi if it already exists
Node *hit = _igvn.hash_find_insert(phi);
if( hit == nullptr ) {
_igvn.register_new_node_with_optimizer(phi);
set_ctrl(phi, lp);
} else {
// Remove the new phi from the graph and use the hit
_igvn.remove_dead_node(phi);
phi = hit;
}
_igvn.replace_input_of(use, idx, phi);
}
#ifdef ASSERT
//------------------------------ is_valid_loop_partition -------------------------------------
// Validate the loop partition sets: peel and not_peel
bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
VectorSet& not_peel ) {
uint i;
// Check that peel_list entries are in the peel set
for (i = 0; i < peel_list.size(); i++) {
if (!peel.test(peel_list.at(i)->_idx)) {
return false;
}
}
// Check at loop members are in one of peel set or not_peel set
for (i = 0; i < loop->_body.size(); i++ ) {
Node *def = loop->_body.at(i);
uint di = def->_idx;
// Check that peel set elements are in peel_list
if (peel.test(di)) {
if (not_peel.test(di)) {
return false;
}
// Must be in peel_list also
bool found = false;
for (uint j = 0; j < peel_list.size(); j++) {
if (peel_list.at(j)->_idx == di) {
found = true;
break;
}
}
if (!found) {
return false;
}
} else if (not_peel.test(di)) {
if (peel.test(di)) {
return false;
}
} else {
return false;
}
}
return true;
}
//------------------------------ is_valid_clone_loop_exit_use -------------------------------------
// Ensure a use outside of loop is of the right form
bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
return (use->is_Phi() &&
use_c->is_Region() && use_c->req() == 3 &&
(use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
}
//------------------------------ is_valid_clone_loop_form -------------------------------------
// Ensure that all uses outside of loop are of the right form
bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
uint orig_exit_idx, uint clone_exit_idx) {
uint len = peel_list.size();
for (uint i = 0; i < len; i++) {
Node *def = peel_list.at(i);
for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
Node *use = def->fast_out(j);
Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
if (!loop->is_member(get_loop(use_c))) {
// use is not in the loop, check for correct structure
if (use->in(0) == def) {
// Okay
} else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
return false;
}
}
}
}
return true;
}
#endif
//------------------------------ partial_peel -------------------------------------
// Partially peel (aka loop rotation) the top portion of a loop (called
// the peel section below) by cloning it and placing one copy just before
// the new loop head and the other copy at the bottom of the new loop.
//
// before after where it came from
//
// stmt1 stmt1
// loop: stmt2 clone
// stmt2 if condA goto exitA clone
// if condA goto exitA new_loop: new
// stmt3 stmt3 clone
// if !condB goto loop if condB goto exitB clone
// exitB: stmt2 orig
// stmt4 if !condA goto new_loop orig
// exitA: goto exitA
// exitB:
// stmt4
// exitA:
//
// Step 1: find the cut point: an exit test on probable
// induction variable.
// Step 2: schedule (with cloning) operations in the peel
// section that can be executed after the cut into
// the section that is not peeled. This may need
// to clone operations into exit blocks. For
// instance, a reference to A[i] in the not-peel
// section and a reference to B[i] in an exit block
// may cause a left-shift of i by 2 to be placed
// in the peel block. This step will clone the left
// shift into the exit block and sink the left shift
// from the peel to the not-peel section.
// Step 3: clone the loop, retarget the control, and insert
// phis for values that are live across the new loop
// head. This is very dependent on the graph structure
// from clone_loop. It creates region nodes for
// exit control and associated phi nodes for values
// flow out of the loop through that exit. The region
// node is dominated by the clone's control projection.
// So the clone's peel section is placed before the
// new loop head, and the clone's not-peel section is
// forms the top part of the new loop. The original
// peel section forms the tail of the new loop.
// Step 4: update the dominator tree and recompute the
// dominator depth.
//
// orig
//
// stmt1
// |
// v
// predicates
// |
// v
// loop<----+
// | |
// stmt2 |
// | |
// v |
// ifA |
// / | |
// v v |
// false true ^ <-- last_peel
// / | |
// / ===|==cut |
// / stmt3 | <-- first_not_peel
// / | |
// | v |
// v ifB |
// exitA: / \ |
// / \ |
// v v |
// false true |
// / \ |
// / ----+
// |
// v
// exitB:
// stmt4
//
//
// after clone loop
//
// stmt1
// |
// v
// predicates
// / \
// clone / \ orig
// / \
// / \
// v v
// +---->loop loop<----+
// | | | |
// | stmt2 stmt2 |
// | | | |
// | v v |
// | ifA ifA |
// | | \ / | |
// | v v v v |
// ^ true false false true ^ <-- last_peel
// | | ^ \ / | |
// | cut==|== \ \ / ===|==cut |
// | stmt3 \ \ / stmt3 | <-- first_not_peel
// | | dom | | | |
// | v \ 1v v2 v |
// | ifB regionA ifB |
// | / \ | / \ |
// | / \ v / \ |
// | v v exitA: v v |
// | true false false true |
// | / ^ \ / \ |
// +---- \ \ / ----+
// dom \ /
// \ 1v v2
// regionB
// |
// v
// exitB:
// stmt4
//
//
// after partial peel
//
// stmt1
// |
// v
// predicates
// /
// clone / orig
// / TOP
// / \
// v v
// TOP->loop loop----+
// | | |
// stmt2 stmt2 |
// | | |
// v v |
// ifA ifA |
// | \ / | |
// v v v v |
// true false false true | <-- last_peel
// | ^ \ / +------|---+
// +->newloop \ \ / === ==cut | |
// | stmt3 \ \ / TOP | |
// | | dom | | stmt3 | | <-- first_not_peel
// | v \ 1v v2 v | |
// | ifB regionA ifB ^ v
// | / \ | / \ | |
// | / \ v / \ | |
// | v v exitA: v v | |
// | true false false true | |
// | / ^ \ / \ | |
// | | \ \ / v | |
// | | dom \ / TOP | |
// | | \ 1v v2 | |
// ^ v regionB | |
// | | | | |
// | | v ^ v
// | | exitB: | |
// | | stmt4 | |
// | +------------>-----------------+ |
// | |
// +-----------------<---------------------+
//
//
// final graph
//
// stmt1
// |
// v
// predicates
// |
// v
// stmt2 clone
// |
// v
// ........> ifA clone
// : / |
// dom / |
// : v v
// : false true
// : | |
// : | v
// : | newloop<-----+
// : | | |
// : | stmt3 clone |
// : | | |
// : | v |
// : | ifB |
// : | / \ |
// : | v v |
// : | false true |
// : | | | |
// : | v stmt2 |
// : | exitB: | |
// : | stmt4 v |
// : | ifA orig |
// : | / \ |
// : | / \ |
// : | v v |
// : | false true |
// : | / \ |
// : v v -----+
// RegionA
// |
// v
// exitA
//
bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
if (!loop->_head->is_Loop()) {
return false;
}
LoopNode *head = loop->_head->as_Loop();
if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
return false;
}
// Check for complex exit control
for (uint ii = 0; ii < loop->_body.size(); ii++) {
Node *n = loop->_body.at(ii);
int opc = n->Opcode();
if (n->is_Call() ||
opc == Op_Catch ||
opc == Op_CatchProj ||
opc == Op_Jump ||
opc == Op_JumpProj) {
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
}
#endif
return false;
}
}
int dd = dom_depth(head);
// Step 1: find cut point
// Walk up dominators to loop head looking for first loop exit
// which is executed on every path thru loop.
IfNode *peel_if = nullptr;
IfNode *peel_if_cmpu = nullptr;
Node *iff = loop->tail();
while (iff != head) {
if (iff->is_If()) {
Node *ctrl = get_ctrl(iff->in(1));
if (ctrl->is_top()) return false; // Dead test on live IF.
// If loop-varying exit-test, check for induction variable
if (loop->is_member(get_loop(ctrl)) &&
loop->is_loop_exit(iff) &&
is_possible_iv_test(iff)) {
Node* cmp = iff->in(1)->in(1);
if (cmp->Opcode() == Op_CmpI) {
peel_if = iff->as_If();
} else {
assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
peel_if_cmpu = iff->as_If();
}
}
}
iff = idom(iff);
}
// Prefer signed compare over unsigned compare.
IfNode* new_peel_if = nullptr;
if (peel_if == nullptr) {
if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
return false; // No peel point found
}
new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
if (new_peel_if == nullptr) {
return false; // No peel point found
}
peel_if = new_peel_if;
}
Node* last_peel = stay_in_loop(peel_if, loop);
Node* first_not_peeled = stay_in_loop(last_peel, loop);
if (first_not_peeled == nullptr || first_not_peeled == head) {
return false;
}
#ifndef PRODUCT
if (TraceLoopOpts) {
tty->print("PartialPeel ");
loop->dump_head();
}
if (TracePartialPeeling) {
tty->print_cr("before partial peel one iteration");
Node_List wl;
Node* t = head->in(2);
while (true) {
wl.push(t);
if (t == head) break;
t = idom(t);
}
while (wl.size() > 0) {
Node* tt = wl.pop();
tt->dump();
if (tt == last_peel) tty->print_cr("-- cut --");
}
}
#endif
VectorSet peel;
VectorSet not_peel;
Node_List peel_list;
Node_List worklist;
Node_List sink_list;
uint estimate = loop->est_loop_clone_sz(1);
if (exceeding_node_budget(estimate)) {
return false;
}
// Set of cfg nodes to peel are those that are executable from
// the head through last_peel.
assert(worklist.size() == 0, "should be empty");
worklist.push(head);
peel.set(head->_idx);
while (worklist.size() > 0) {
Node *n = worklist.pop();
if (n != last_peel) {
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
Node* use = n->fast_out(j);
if (use->is_CFG() &&
loop->is_member(get_loop(use)) &&
!peel.test_set(use->_idx)) {
worklist.push(use);
}
}
}
}
// Set of non-cfg nodes to peel are those that are control
// dependent on the cfg nodes.
for (uint i = 0; i < loop->_body.size(); i++) {
Node *n = loop->_body.at(i);
Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
if (peel.test(n_c->_idx)) {
peel.set(n->_idx);
} else {
not_peel.set(n->_idx);
}
}
// Step 2: move operations from the peeled section down into the
// not-peeled section
// Get a post order schedule of nodes in the peel region
// Result in right-most operand.
scheduled_nodelist(loop, peel, peel_list);
assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
// For future check for too many new phis
uint old_phi_cnt = 0;
for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
Node* use = head->fast_out(j);
if (use->is_Phi()) old_phi_cnt++;
}
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("\npeeled list");
}
#endif
// Evacuate nodes in peel region into the not_peeled region if possible
bool too_many_clones = false;
uint new_phi_cnt = 0;
uint cloned_for_outside_use = 0;
for (uint i = 0; i < peel_list.size();) {
Node* n = peel_list.at(i);
#ifndef PRODUCT
if (TracePartialPeeling) n->dump();
#endif
bool incr = true;
if (!n->is_CFG()) {
if (has_use_in_set(n, not_peel)) {
// If not used internal to the peeled region,
// move "n" from peeled to not_peeled region.
if (!has_use_internal_to_set(n, peel, loop)) {
// if not pinned and not a load (which maybe anti-dependent on a store)
// and not a CMove (Matcher expects only bool->cmove).
if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
int new_clones = clone_for_use_outside_loop(loop, n, worklist);
if (new_clones == -1) {
too_many_clones = true;
break;
}
cloned_for_outside_use += new_clones;
sink_list.push(n);
peel.remove(n->_idx);
not_peel.set(n->_idx);
peel_list.remove(i);
incr = false;
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("sink to not_peeled region: %d newbb: %d",
n->_idx, get_ctrl(n)->_idx);
}
#endif
}
} else {
// Otherwise check for special def-use cases that span
// the peel/not_peel boundary such as bool->if
clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
new_phi_cnt++;
}
}
}
if (incr) i++;
}
estimate += cloned_for_outside_use + new_phi_cnt;
bool exceed_node_budget = !may_require_nodes(estimate);
bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
#ifndef PRODUCT
if (TracePartialPeeling && exceed_phi_limit) {
tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
}
#endif
if (new_peel_if != nullptr) {
remove_cmpi_loop_exit(new_peel_if, loop);
}
// Inhibit more partial peeling on this loop
assert(!head->is_partial_peel_loop(), "not partial peeled");
head->mark_partial_peel_failed();
if (cloned_for_outside_use > 0) {
// Terminate this round of loop opts because
// the graph outside this loop was changed.
C->set_major_progress();
return true;
}
return false;
}
// Step 3: clone loop, retarget control, and insert new phis
// Create new loop head for new phis and to hang
// the nodes being moved (sinked) from the peel region.
LoopNode* new_head = new LoopNode(last_peel, last_peel);
new_head->set_unswitch_count(head->unswitch_count()); // Preserve
_igvn.register_new_node_with_optimizer(new_head);
assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
_igvn.replace_input_of(first_not_peeled, 0, new_head);
set_loop(new_head, loop);
loop->_body.push(new_head);
not_peel.set(new_head->_idx);
set_idom(new_head, last_peel, dom_depth(first_not_peeled));
set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
while (sink_list.size() > 0) {
Node* n = sink_list.pop();
set_ctrl(n, new_head);
}
assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
clone_loop(loop, old_new, dd, IgnoreStripMined);
const uint clone_exit_idx = 1;
const uint orig_exit_idx = 2;
assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
Node* head_clone = old_new[head->_idx];
LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
Node* orig_tail_clone = head_clone->in(2);
// Add phi if "def" node is in peel set and "use" is not
for (uint i = 0; i < peel_list.size(); i++) {
Node *def = peel_list.at(i);
if (!def->is_CFG()) {
for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
Node *use = def->fast_out(j);
if (has_node(use) && use->in(0) != C->top() &&
(!peel.test(use->_idx) ||
(use->is_Phi() && use->in(0) == head)) ) {
worklist.push(use);
}
}
while( worklist.size() ) {
Node *use = worklist.pop();
for (uint j = 1; j < use->req(); j++) {
Node* n = use->in(j);
if (n == def) {
// "def" is in peel set, "use" is not in peel set
// or "use" is in the entry boundary (a phi) of the peel set
Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
if ( loop->is_member(get_loop( use_c )) ) {
// use is in loop
if (old_new[use->_idx] != nullptr) { // null for dead code
Node* use_clone = old_new[use->_idx];
_igvn.replace_input_of(use, j, C->top());
insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
}
} else {
assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
// use is not in the loop, check if the live range includes the cut
Node* lp_if = use_c->in(orig_exit_idx)->in(0);
if (not_peel.test(lp_if->_idx)) {
assert(j == orig_exit_idx, "use from original loop");
insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
}
}
}
}
}
}
}
// Step 3b: retarget control
// Redirect control to the new loop head if a cloned node in
// the not_peeled region has control that points into the peeled region.
// This necessary because the cloned peeled region will be outside
// the loop.
// from to
// cloned-peeled <---+
// new_head_clone: | <--+
// cloned-not_peeled in(0) in(0)
// orig-peeled
for (uint i = 0; i < loop->_body.size(); i++) {
Node *n = loop->_body.at(i);
if (!n->is_CFG() && n->in(0) != nullptr &&
not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
Node* n_clone = old_new[n->_idx];
_igvn.replace_input_of(n_clone, 0, new_head_clone);
}
}
// Backedge of the surviving new_head (the clone) is original last_peel
_igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
// Cut first node in original not_peel set
_igvn.rehash_node_delayed(new_head); // Multiple edge updates:
new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
// Copy head_clone back-branch info to original head
// and remove original head's loop entry and
// clone head's back-branch
_igvn.rehash_node_delayed(head); // Multiple edge updates
head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
head->set_req(LoopNode::LoopBackControl, C->top());
_igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
// Similarly modify the phis
for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
Node* use = head->fast_out(k);
if (use->is_Phi() && use->outcnt() > 0) {
Node* use_clone = old_new[use->_idx];
_igvn.rehash_node_delayed(use); // Multiple edge updates
use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
use->set_req(LoopNode::LoopBackControl, C->top());
_igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
}
}
// Step 4: update dominator tree and dominator depth
set_idom(head, orig_tail_clone, dd);
recompute_dom_depth();
// Inhibit more partial peeling on this loop
new_head_clone->set_partial_peel_loop();
C->set_major_progress();
loop->record_for_igvn();
#ifndef PRODUCT
if (TracePartialPeeling) {
tty->print_cr("\nafter partial peel one iteration");
Node_List wl;
Node* t = last_peel;
while (true) {
wl.push(t);
if (t == head_clone) break;
t = idom(t);
}
while (wl.size() > 0) {
Node* tt = wl.pop();
if (tt == head) tty->print_cr("orig head");
else if (tt == new_head_clone) tty->print_cr("new head");
else if (tt == head_clone) tty->print_cr("clone head");
tt->dump();
}
}
#endif
return true;
}
// Transform:
//
// loop<-----------------+
// | |
// stmt1 stmt2 .. stmtn |
// | | | |
// \ | / |
// v v v |
// region |
// | |
// shared_stmt |
// | |
// v |
// if |
// / \ |
// | -----------+
// v
//
// into:
//
// loop<-------------------+
// | |
// v |
// +->loop |
// | | |
// | stmt1 stmt2 .. stmtn |
// | | | | |
// | | \ / |
// | | v v |
// | | region1 |
// | | | |
// | shared_stmt shared_stmt |
// | | | |
// | v v |
// | if if |
// | /\ / \ |
// +-- | | -------+
// \ /
// v v
// region2
//
// (region2 is shown to merge mirrored projections of the loop exit
// ifs to make the diagram clearer but they really merge the same
// projection)
//
// Conditions for this transformation to trigger:
// - the path through stmt1 is frequent enough
// - the inner loop will be turned into a counted loop after transformation
bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
if (!DuplicateBackedge) {
return false;
}
assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
if (!loop->_head->is_Loop()) {
return false;
}
uint estimate = loop->est_loop_clone_sz(1);
if (exceeding_node_budget(estimate)) {
return false;
}
LoopNode *head = loop->_head->as_Loop();
Node* region = nullptr;
IfNode* exit_test = nullptr;
uint inner;
float f;
if (StressDuplicateBackedge) {
if (head->is_strip_mined()) {
return false;
}
Node* c = head->in(LoopNode::LoopBackControl);
while (c != head) {
if (c->is_Region()) {
region = c;
}
c = idom(c);
}
if (region == nullptr) {
return false;
}
inner = 1;
} else {
// Is the shape of the loop that of a counted loop...
Node* back_control = loop_exit_control(head, loop);
if (back_control == nullptr) {
return false;
}
BoolTest::mask bt = BoolTest::illegal;
float cl_prob = 0;
Node* incr = nullptr;
Node* limit = nullptr;
Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
return false;
}
// With an extra phi for the candidate iv?
// Or the region node is the loop head
if (!incr->is_Phi() || incr->in(0) == head) {
return false;
}
PathFrequency pf(head, this);
region = incr->in(0);
// Go over all paths for the extra phi's region and see if that
// path is frequent enough and would match the expected iv shape
// if the extra phi is removed
inner = 0;
for (uint i = 1; i < incr->req(); ++i) {
Node* in = incr->in(i);
Node* trunc1 = nullptr;
Node* trunc2 = nullptr;
const TypeInteger* iv_trunc_t = nullptr;
Node* orig_in = in;
if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
continue;
}
assert(in->Opcode() == Op_AddI, "wrong increment code");
Node* xphi = nullptr;
Node* stride = loop_iv_stride(in, loop, xphi);
if (stride == nullptr) {
continue;
}
PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
if (phi == nullptr ||
(trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
(trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
return false;
}
f = pf.to(region->in(i));
if (f > 0.5) {
inner = i;
break;
}
}
if (inner == 0) {
return false;
}
exit_test = back_control->in(0)->as_If();
}
if (idom(region)->is_Catch()) {
return false;
}
// Collect all control nodes that need to be cloned (shared_stmt in the diagram)
Unique_Node_List wq;
wq.push(head->in(LoopNode::LoopBackControl));
for (uint i = 0; i < wq.size(); i++) {
Node* c = wq.at(i);
assert(get_loop(c) == loop, "not in the right loop?");
if (c->is_Region()) {
if (c != region) {
for (uint j = 1; j < c->req(); ++j) {
wq.push(c->in(j));
}
}
} else {
wq.push(c->in(0));
}
assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
}
Node* region_dom = idom(region);
// Can't do the transformation if this would cause a membar pair to
// be split
for (uint i = 0; i < wq.size(); i++) {
Node* c = wq.at(i);
if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
if (!wq.member(c->as_MemBar()->leading_membar())) {
return false;
}
}
}
// Collect data nodes that need to be clones as well
int dd = dom_depth(head);
for (uint i = 0; i < loop->_body.size(); ++i) {
Node* n = loop->_body.at(i);
if (has_ctrl(n)) {
Node* c = get_ctrl(n);
if (wq.member(c)) {
wq.push(n);
}
} else {
set_idom(n, idom(n), dd);
}
}
// clone shared_stmt
clone_loop_body(wq, old_new, nullptr);
Node* region_clone = old_new[region->_idx];
region_clone->set_req(inner, C->top());
set_idom(region, region->in(inner), dd);
// Prepare the outer loop
Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
_igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
set_idom(head, outer_head, dd);
fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
// Make one of the shared_stmt copies only reachable from stmt1, the
// other only from stmt2..stmtn.
Node* dom = nullptr;
for (uint i = 1; i < region->req(); ++i) {
if (i != inner) {
_igvn.replace_input_of(region, i, C->top());
}
Node* in = region_clone->in(i);
if (in->is_top()) {
continue;
}
if (dom == nullptr) {
dom = in;
} else {
dom = dom_lca(dom, in);
}
}
set_idom(region_clone, dom, dd);
// Set up the outer loop
for (uint i = 0; i < head->outcnt(); i++) {
Node* u = head->raw_out(i);
if (u->is_Phi()) {
Node* outer_phi = u->clone();
outer_phi->set_req(0, outer_head);
Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
if (backedge == nullptr) {
backedge = u->in(LoopNode::LoopBackControl);
}
outer_phi->set_req(LoopNode::LoopBackControl, backedge);
register_new_node(outer_phi, outer_head);
_igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
}
}
// create control and data nodes for out of loop uses (including region2)
Node_List worklist;
uint new_counter = C->unique();
fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
Node_List *split_if_set = nullptr;
Node_List *split_bool_set = nullptr;
Node_List *split_cex_set = nullptr;
fix_data_uses(wq, loop, ControlAroundStripMined, head->is_strip_mined() ? loop->_parent : loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
if (exit_test != nullptr) {
float cnt = exit_test->_fcnt;
if (cnt != COUNT_UNKNOWN) {
exit_test->_fcnt = cnt * f;
old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
}
}
C->set_major_progress();
return true;
}
// Having ReductionNodes in the loop is expensive. They need to recursively
// fold together the vector values, for every vectorized loop iteration. If
// we encounter the following pattern, we can vector accumulate the values
// inside the loop, and only have a single UnorderedReduction after the loop.
//
// CountedLoop init
// | |
// +------+ | +-----------------------+
// | | | |
// PhiNode (s) |
// | |
// | Vector |
// | | |
// UnorderedReduction (first_ur) |
// | |
// ... Vector |
// | | |
// UnorderedReduction (last_ur) |
// | |
// +---------------------+
//
// We patch the graph to look like this:
//
// CountedLoop identity_vector
// | |
// +-------+ | +---------------+
// | | | |
// PhiNode (v) |
// | |
// | Vector |
// | | |
// VectorAccumulator |
// | |
// ... Vector |
// | | |
// init VectorAccumulator |
// | | | |
// UnorderedReduction +-----------+
//
// We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
// use vector_accumulators, which do the same reductions, but only element
// wise. This is a single operation per vector_accumulator, rather than many
// for a UnorderedReduction. We can then reduce the last vector_accumulator
// after the loop, and also reduce the init value into it.
// We can not do this with all reductions. Some reductions do not allow the
// reordering of operations (for example float addition).
void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
// Find all Phi nodes with UnorderedReduction on backedge.
CountedLoopNode* cl = loop->_head->as_CountedLoop();
for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
Node* phi = cl->fast_out(j);
// We have a phi with a single use, and a UnorderedReduction on the backedge.
if (!phi->is_Phi() || phi->outcnt() != 1 || !phi->in(2)->is_UnorderedReduction()) {
continue;
}
UnorderedReductionNode* last_ur = phi->in(2)->as_UnorderedReduction();
// Determine types
const TypeVect* vec_t = last_ur->vect_type();
uint vector_length = vec_t->length();
BasicType bt = vec_t->element_basic_type();
const Type* bt_t = Type::get_const_basic_type(bt);
// Convert opcode from vector-reduction -> scalar -> normal-vector-op
const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
const int vopc = VectorNode::opcode(sopc, bt);
if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
DEBUG_ONLY( last_ur->dump(); )
assert(false, "do not have normal vector op for this reduction");
continue; // not implemented -> fails
}
// Traverse up the chain of UnorderedReductions, checking that it loops back to
// the phi. Check that all UnorderedReductions only have a single use, except for
// the last (last_ur), which only has phi as a use in the loop, and all other uses
// are outside the loop.
UnorderedReductionNode* current = last_ur;
UnorderedReductionNode* first_ur = nullptr;
while (true) {
assert(current->is_UnorderedReduction(), "sanity");
// Expect no ctrl and a vector_input from within the loop.
Node* ctrl = current->in(0);
Node* vector_input = current->in(2);
if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
DEBUG_ONLY( current->dump(1); )
assert(false, "reduction has ctrl or bad vector_input");
break; // Chain traversal fails.
}
assert(current->vect_type() != nullptr, "must have vector type");
if (current->vect_type() != last_ur->vect_type()) {
// Reductions do not have the same vector type (length and element type).
break; // Chain traversal fails.
}
// Expect single use of UnorderedReduction, except for last_ur.
if (current == last_ur) {
// Expect all uses to be outside the loop, except phi.
for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
Node* use = current->fast_out(k);
if (use != phi && ctrl_or_self(use) == cl) {
DEBUG_ONLY( current->dump(-1); )
assert(false, "reduction has use inside loop");
// Should not be allowed by SuperWord::mark_reductions
return; // bail out of optimization
}
}
} else {
if (current->outcnt() != 1) {
break; // Chain traversal fails.
}
}
// Expect another UnorderedReduction or phi as the scalar input.
Node* scalar_input = current->in(1);
if (scalar_input->is_UnorderedReduction() &&
scalar_input->Opcode() == current->Opcode()) {
// Move up the UnorderedReduction chain.
current = scalar_input->as_UnorderedReduction();
} else if (scalar_input == phi) {
// Chain terminates at phi.
first_ur = current;
current = nullptr;
break; // Success.
} else {
// scalar_input is neither phi nor a matching reduction
// Can for example be scalar reduction when we have
// partial vectorization.
break; // Chain traversal fails.
}
}
if (current != nullptr) {
// Chain traversal was not successful.
continue;
}
assert(first_ur != nullptr, "must have successfully terminated chain traversal");
Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
set_ctrl(identity_scalar, C->root());
VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t);
register_new_node(identity_vector, C->root());
assert(vec_t == identity_vector->vect_type(), "matching vector type");
VectorNode::trace_new_vector(identity_vector, "UnorderedReduction");
// Turn the scalar phi into a vector phi.
_igvn.rehash_node_delayed(phi);
Node* init = phi->in(1); // Remember init before replacing it.
phi->set_req_X(1, identity_vector, &_igvn);
phi->as_Type()->set_type(vec_t);
_igvn.set_type(phi, vec_t);
// Traverse down the chain of UnorderedReductions, and replace them with vector_accumulators.
current = first_ur;
while (true) {
// Create vector_accumulator to replace current.
Node* last_vector_accumulator = current->in(1);
Node* vector_input = current->in(2);
VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
register_new_node(vector_accumulator, cl);
_igvn.replace_node(current, vector_accumulator);
VectorNode::trace_new_vector(vector_accumulator, "UnorderedReduction");
if (current == last_ur) {
break;
}
current = vector_accumulator->unique_out()->as_UnorderedReduction();
}
// Create post-loop reduction.
Node* last_accumulator = phi->in(2);
Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
// Take over uses of last_accumulator that are not in the loop.
for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
Node* use = last_accumulator->out(i);
if (use != phi && use != post_loop_reduction) {
assert(ctrl_or_self(use) != cl, "use must be outside loop");
use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
--i;
}
}
register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
VectorNode::trace_new_vector(post_loop_reduction, "UnorderedReduction");
assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
assert(phi->outcnt() == 1, "accumulator is the only use of phi");
}
}