1 /*
2 * Copyright (c) 1999, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/inlinetypenode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/matcher.hpp"
39 #include "opto/mulnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/subnode.hpp"
44 #include "opto/subtypenode.hpp"
45 #include "opto/superword.hpp"
46 #include "opto/vectornode.hpp"
47 #include "utilities/macros.hpp"
48
49 //=============================================================================
50 //------------------------------split_thru_phi---------------------------------
51 // Split Node 'n' through merge point if there is enough win.
52 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
53 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
54 // ConvI2L may have type information on it which is unsafe to push up
55 // so disable this for now
56 return nullptr;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
63 n->in(1) == region->as_CountedLoop()->phi()) {
64 return nullptr;
65 }
66
67 // Inline types should not be split through Phis because they cannot be merged
68 // through Phi nodes but each value input needs to be merged individually.
69 if (n->is_InlineType()) {
70 return nullptr;
71 }
72
73 if (cannot_split_division(n, region)) {
74 return nullptr;
75 }
76
77 int wins = 0;
78 assert(!n->is_CFG(), "");
79 assert(region->is_Region(), "");
80
81 const Type* type = n->bottom_type();
82 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
83 Node* phi;
84 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
85 int iid = t_oop->instance_id();
86 int index = C->get_alias_index(t_oop);
87 int offset = t_oop->offset();
88 phi = new PhiNode(region, type, nullptr, iid, index, offset);
89 } else {
90 phi = PhiNode::make_blank(region, n);
91 }
92 uint old_unique = C->unique();
93 for (uint i = 1; i < region->req(); i++) {
94 Node* x;
95 Node* the_clone = nullptr;
96 if (region->in(i) == C->top()) {
97 x = C->top(); // Dead path? Use a dead data op
98 } else {
99 x = n->clone(); // Else clone up the data op
100 the_clone = x; // Remember for possible deletion.
101 // Alter data node to use pre-phi inputs
102 if (n->in(0) == region)
103 x->set_req( 0, region->in(i) );
104 for (uint j = 1; j < n->req(); j++) {
105 Node* in = n->in(j);
106 if (in->is_Phi() && in->in(0) == region)
107 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
108 }
109 }
110 // Check for a 'win' on some paths
111 const Type* t = x->Value(&_igvn);
112
113 bool singleton = t->singleton();
114
115 // A TOP singleton indicates that there are no possible values incoming
116 // along a particular edge. In most cases, this is OK, and the Phi will
117 // be eliminated later in an Ideal call. However, we can't allow this to
118 // happen if the singleton occurs on loop entry, as the elimination of
119 // the PhiNode may cause the resulting node to migrate back to a previous
120 // loop iteration.
121 if (singleton && t == Type::TOP) {
122 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
123 // irreducible loop may not be indicated by an affirmative is_Loop());
124 // therefore, the only top we can split thru a phi is on a backedge of
125 // a loop.
126 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
127 }
128
129 if (singleton) {
130 wins++;
131 x = ((PhaseGVN&)_igvn).makecon(t);
132 } else {
133 // We now call Identity to try to simplify the cloned node.
134 // Note that some Identity methods call phase->type(this).
135 // Make sure that the type array is big enough for
136 // our new node, even though we may throw the node away.
137 // (Note: This tweaking with igvn only works because x is a new node.)
138 _igvn.set_type(x, t);
139 // If x is a TypeNode, capture any more-precise type permanently into Node
140 // otherwise it will be not updated during igvn->transform since
141 // igvn->type(x) is set to x->Value() already.
142 x->raise_bottom_type(t);
143 Node* y = x->Identity(&_igvn);
144 if (y != x) {
145 wins++;
146 x = y;
147 } else {
148 y = _igvn.hash_find(x);
149 if (y == nullptr) {
150 y = similar_subtype_check(x, region->in(i));
151 }
152 if (y) {
153 wins++;
154 x = y;
155 } else {
156 // Else x is a new node we are keeping
157 // We do not need register_new_node_with_optimizer
158 // because set_type has already been called.
159 _igvn._worklist.push(x);
160 }
161 }
162 }
163
164 phi->set_req( i, x );
165
166 if (the_clone == nullptr) {
167 continue;
168 }
169
170 if (the_clone != x) {
171 _igvn.remove_dead_node(the_clone);
172 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
173 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
174 // it is not a win if 'x' moved from an outer to an inner loop
175 // this edge case can only happen for Load nodes
176 wins = 0;
177 break;
178 }
179 }
180 // Too few wins?
181 if (wins <= policy) {
182 _igvn.remove_dead_node(phi);
183 return nullptr;
184 }
185
186 // Record Phi
187 register_new_node( phi, region );
188
189 for (uint i2 = 1; i2 < phi->req(); i2++) {
190 Node *x = phi->in(i2);
191 // If we commoned up the cloned 'x' with another existing Node,
192 // the existing Node picks up a new use. We need to make the
193 // existing Node occur higher up so it dominates its uses.
194 Node *old_ctrl;
195 IdealLoopTree *old_loop;
196
197 if (x->is_Con()) {
198 // Constant's control is always root.
199 set_ctrl(x, C->root());
200 continue;
201 }
202 // The occasional new node
203 if (x->_idx >= old_unique) { // Found a new, unplaced node?
204 old_ctrl = nullptr;
205 old_loop = nullptr; // Not in any prior loop
206 } else {
207 old_ctrl = get_ctrl(x);
208 old_loop = get_loop(old_ctrl); // Get prior loop
209 }
210 // New late point must dominate new use
211 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
212 if (new_ctrl == old_ctrl) // Nothing is changed
213 continue;
214
215 IdealLoopTree *new_loop = get_loop(new_ctrl);
216
217 // Don't move x into a loop if its uses are
218 // outside of loop. Otherwise x will be cloned
219 // for each use outside of this loop.
220 IdealLoopTree *use_loop = get_loop(region);
221 if (!new_loop->is_member(use_loop) &&
222 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
223 // Take early control, later control will be recalculated
224 // during next iteration of loop optimizations.
225 new_ctrl = get_early_ctrl(x);
226 new_loop = get_loop(new_ctrl);
227 }
228 // Set new location
229 set_ctrl(x, new_ctrl);
230 // If changing loop bodies, see if we need to collect into new body
231 if (old_loop != new_loop) {
232 if (old_loop && !old_loop->_child)
233 old_loop->_body.yank(x);
234 if (!new_loop->_child)
235 new_loop->_body.push(x); // Collect body info
236 }
237 }
238
239 return phi;
240 }
241
242 // Test whether node 'x' can move into an inner loop relative to node 'n'.
243 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
244 // BUT it can also return true and 'x' is in the outer loop
245 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
246 IdealLoopTree* n_loop_tree = get_loop(n_loop);
247 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
248 // x_loop_tree should be outer or same loop as n_loop_tree
249 return !x_loop_tree->is_member(n_loop_tree);
250 }
251
252 // Subtype checks that carry profile data don't common so look for a replacement by following edges
253 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
254 if (x->is_SubTypeCheck()) {
255 Node* in1 = x->in(1);
256 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
257 Node* u = in1->fast_out(i);
258 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
259 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
260 Node* bol = u->fast_out(j);
261 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
262 Node* iff = bol->fast_out(k);
263 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
264 // unrelated profile
265 if (iff->is_If() && is_dominator(iff, r_in)) {
266 return u;
267 }
268 }
269 }
270 }
271 }
272 }
273 return nullptr;
274 }
275
276 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
277 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
278 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
279 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
280 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
281 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
282 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
283 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
284 const Type* zero;
285 switch (n->Opcode()) {
286 case Op_DivI:
287 case Op_ModI:
288 zero = TypeInt::ZERO;
289 break;
290 case Op_DivL:
291 case Op_ModL:
292 zero = TypeLong::ZERO;
293 break;
294 default:
295 return false;
296 }
297
298 if (n->in(0) != nullptr) {
299 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
300 return true;
301 }
302
303 Node* divisor = n->in(2);
304 return is_divisor_counted_loop_phi(divisor, region) &&
305 loop_phi_backedge_type_contains_zero(divisor, zero);
306 }
307
308 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) {
309 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop;
310 }
311
312 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
313 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
314 }
315
316 //------------------------------dominated_by------------------------------------
317 // Replace the dominated test with an obvious true or false. Place it on the
318 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
319 // live path up to the dominating control.
320 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
321 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
322
323 // prevdom is the dominating projection of the dominating test.
324 assert(iff->Opcode() == Op_If ||
325 iff->Opcode() == Op_CountedLoopEnd ||
326 iff->Opcode() == Op_LongCountedLoopEnd ||
327 iff->Opcode() == Op_RangeCheck ||
328 iff->Opcode() == Op_ParsePredicate,
329 "Check this code when new subtype is added");
330
331 int pop = prevdom->Opcode();
332 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
333 if (flip) {
334 if (pop == Op_IfTrue)
335 pop = Op_IfFalse;
336 else
337 pop = Op_IfTrue;
338 }
339 // 'con' is set to true or false to kill the dominated test.
340 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
341 set_ctrl(con, C->root()); // Constant gets a new use
342 // Hack the dominated test
343 _igvn.replace_input_of(iff, 1, con);
344
345 // If I don't have a reachable TRUE and FALSE path following the IfNode then
346 // I can assume this path reaches an infinite loop. In this case it's not
347 // important to optimize the data Nodes - either the whole compilation will
348 // be tossed or this path (and all data Nodes) will go dead.
349 if (iff->outcnt() != 2) return;
350
351 // Make control-dependent data Nodes on the live path (path that will remain
352 // once the dominated IF is removed) become control-dependent on the
353 // dominating projection.
354 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
355
356 if (dp == nullptr)
357 return;
358
359 IdealLoopTree* old_loop = get_loop(dp);
360
361 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
362 Node* cd = dp->fast_out(i); // Control-dependent node
363 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
364 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
365 assert(cd->in(0) == dp, "");
366 _igvn.replace_input_of(cd, 0, prevdom);
367 if (pin_array_access_nodes) {
368 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
369 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
370 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
371 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
372 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
373 // dominating check.
374 Node* clone = cd->pin_array_access_node();
375 if (clone != nullptr) {
376 clone = _igvn.register_new_node_with_optimizer(clone, cd);
377 _igvn.replace_node(cd, clone);
378 cd = clone;
379 }
380 }
381 set_early_ctrl(cd, false);
382 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
383 if (old_loop != new_loop) {
384 if (!old_loop->_child) {
385 old_loop->_body.yank(cd);
386 }
387 if (!new_loop->_child) {
388 new_loop->_body.push(cd);
389 }
390 }
391 --i;
392 --imax;
393 }
394 }
395 }
396
397 //------------------------------has_local_phi_input----------------------------
398 // Return TRUE if 'n' has Phi inputs from its local block and no other
399 // block-local inputs (all non-local-phi inputs come from earlier blocks)
400 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
401 Node *n_ctrl = get_ctrl(n);
402 // See if some inputs come from a Phi in this block, or from before
403 // this block.
404 uint i;
405 for( i = 1; i < n->req(); i++ ) {
406 Node *phi = n->in(i);
407 if( phi->is_Phi() && phi->in(0) == n_ctrl )
408 break;
409 }
410 if( i >= n->req() )
411 return nullptr; // No Phi inputs; nowhere to clone thru
412
413 // Check for inputs created between 'n' and the Phi input. These
414 // must split as well; they have already been given the chance
415 // (courtesy of a post-order visit) and since they did not we must
416 // recover the 'cost' of splitting them by being very profitable
417 // when splitting 'n'. Since this is unlikely we simply give up.
418 for( i = 1; i < n->req(); i++ ) {
419 Node *m = n->in(i);
420 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
421 // We allow the special case of AddP's with no local inputs.
422 // This allows us to split-up address expressions.
423 if (m->is_AddP() &&
424 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
425 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
426 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
427 // Move the AddP up to the dominating point. That's fine because control of m's inputs
428 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
429 Node* c = find_non_split_ctrl(idom(n_ctrl));
430 if (c->is_OuterStripMinedLoop()) {
431 c->as_Loop()->verify_strip_mined(1);
432 c = c->in(LoopNode::EntryControl);
433 }
434 set_ctrl_and_loop(m, c);
435 continue;
436 }
437 return nullptr;
438 }
439 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
440 }
441
442 return n_ctrl;
443 }
444
445 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
446 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
447 assert(bt == T_INT || bt == T_LONG, "only for integers");
448 int n_op = n->Opcode();
449
450 if (n_op == Op_LShift(bt)) {
451 // Scale is loop invariant
452 Node* scale = n->in(2);
453 Node* scale_ctrl = get_ctrl(scale);
454 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
455 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
456 return nullptr;
457 }
458 const TypeInt* scale_t = scale->bottom_type()->isa_int();
459 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
460 return nullptr; // Dont bother with byte/short masking
461 }
462 // Add must vary with loop (else shift would be loop-invariant)
463 Node* add = n->in(1);
464 Node* add_ctrl = get_ctrl(add);
465 IdealLoopTree* add_loop = get_loop(add_ctrl);
466 if (n_loop != add_loop) {
467 return nullptr; // happens w/ evil ZKM loops
468 }
469
470 // Convert I-V into I+ (0-V); same for V-I
471 if (add->Opcode() == Op_Sub(bt) &&
472 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
473 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
474 Node* zero = _igvn.integercon(0, bt);
475 set_ctrl(zero, C->root());
476 Node* neg = SubNode::make(zero, add->in(2), bt);
477 register_new_node(neg, get_ctrl(add->in(2)));
478 add = AddNode::make(add->in(1), neg, bt);
479 register_new_node(add, add_ctrl);
480 }
481 if (add->Opcode() != Op_Add(bt)) return nullptr;
482 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
483 // See if one add input is loop invariant
484 Node* add_var = add->in(1);
485 Node* add_var_ctrl = get_ctrl(add_var);
486 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
487 Node* add_invar = add->in(2);
488 Node* add_invar_ctrl = get_ctrl(add_invar);
489 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
490 if (add_invar_loop == n_loop) {
491 // Swap to find the invariant part
492 add_invar = add_var;
493 add_invar_ctrl = add_var_ctrl;
494 add_invar_loop = add_var_loop;
495 add_var = add->in(2);
496 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
497 return nullptr;
498 }
499 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
500 return nullptr; // No invariant part of the add?
501 }
502
503 // Yes! Reshape address expression!
504 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
505 Node* inv_scale_ctrl =
506 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
507 add_invar_ctrl : scale_ctrl;
508 register_new_node(inv_scale, inv_scale_ctrl);
509 Node* var_scale = LShiftNode::make(add_var, scale, bt);
510 register_new_node(var_scale, n_ctrl);
511 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
512 register_new_node(var_add, n_ctrl);
513 _igvn.replace_node(n, var_add);
514 return var_add;
515 }
516 return nullptr;
517 }
518
519 //------------------------------remix_address_expressions----------------------
520 // Rework addressing expressions to get the most loop-invariant stuff
521 // moved out. We'd like to do all associative operators, but it's especially
522 // important (common) to do address expressions.
523 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
524 if (!has_ctrl(n)) return nullptr;
525 Node* n_ctrl = get_ctrl(n);
526 IdealLoopTree* n_loop = get_loop(n_ctrl);
527
528 // See if 'n' mixes loop-varying and loop-invariant inputs and
529 // itself is loop-varying.
530
531 // Only interested in binary ops (and AddP)
532 if (n->req() < 3 || n->req() > 4) return nullptr;
533
534 Node* n1_ctrl = get_ctrl(n->in( 1));
535 Node* n2_ctrl = get_ctrl(n->in( 2));
536 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
537 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
538 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
539 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
540
541 // Does one of my inputs spin in a tighter loop than self?
542 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
543 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
544 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
545 return nullptr; // Leave well enough alone
546 }
547
548 // Is at least one of my inputs loop-invariant?
549 if (n1_loop == n_loop &&
550 n2_loop == n_loop &&
551 n3_loop == n_loop) {
552 return nullptr; // No loop-invariant inputs
553 }
554
555 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
556 if (res != nullptr) {
557 return res;
558 }
559 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
560 if (res != nullptr) {
561 return res;
562 }
563
564 int n_op = n->Opcode();
565 // Replace (I+V) with (V+I)
566 if (n_op == Op_AddI ||
567 n_op == Op_AddL ||
568 n_op == Op_AddF ||
569 n_op == Op_AddD ||
570 n_op == Op_MulI ||
571 n_op == Op_MulL ||
572 n_op == Op_MulF ||
573 n_op == Op_MulD) {
574 if (n2_loop == n_loop) {
575 assert(n1_loop != n_loop, "");
576 n->swap_edges(1, 2);
577 }
578 }
579
580 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
581 // but not if I2 is a constant. Skip for irreducible loops.
582 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
583 if (n2_loop == n_loop && n3_loop != n_loop) {
584 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
585 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
586 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
587 IdealLoopTree* n22loop = get_loop(n22_ctrl);
588 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
589 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
590 n23_loop == n_loop) {
591 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
592 // Stuff new AddP in the loop preheader
593 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
594 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
595 register_new_node(add2, n_ctrl);
596 _igvn.replace_node(n, add2);
597 return add2;
598 }
599 }
600 }
601
602 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
603 if (n2_loop != n_loop && n3_loop == n_loop) {
604 if (n->in(3)->Opcode() == Op_AddX) {
605 Node* V = n->in(3)->in(1);
606 Node* I = n->in(3)->in(2);
607 if (is_member(n_loop,get_ctrl(V))) {
608 } else {
609 Node *tmp = V; V = I; I = tmp;
610 }
611 if (!is_member(n_loop,get_ctrl(I))) {
612 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
613 // Stuff new AddP in the loop preheader
614 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
615 Node* add2 = new AddPNode(n->in(1), add1, V);
616 register_new_node(add2, n_ctrl);
617 _igvn.replace_node(n, add2);
618 return add2;
619 }
620 }
621 }
622 }
623
624 return nullptr;
625 }
626
627 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
628 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
629 assert(n->Opcode() == Op_AddI, "sanity");
630 Node * nn = nullptr;
631 Node * in1 = n->in(1);
632 Node * in2 = n->in(2);
633 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
634 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
635 if (loop_n->is_counted() &&
636 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
637 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
638 Matcher::match_rule_supported(Op_MulAddS2I)) {
639 Node* mul_in1 = in1->in(1);
640 Node* mul_in2 = in1->in(2);
641 Node* mul_in3 = in2->in(1);
642 Node* mul_in4 = in2->in(2);
643 if (mul_in1->Opcode() == Op_LoadS &&
644 mul_in2->Opcode() == Op_LoadS &&
645 mul_in3->Opcode() == Op_LoadS &&
646 mul_in4->Opcode() == Op_LoadS) {
647 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
648 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
649 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
650 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
651 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
652 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
653 // All nodes should be in the same counted loop.
654 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
655 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
656 Node* adr1 = mul_in1->in(MemNode::Address);
657 Node* adr2 = mul_in2->in(MemNode::Address);
658 Node* adr3 = mul_in3->in(MemNode::Address);
659 Node* adr4 = mul_in4->in(MemNode::Address);
660 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
661 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
662 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
663 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
664 register_new_node(nn, get_ctrl(n));
665 _igvn.replace_node(n, nn);
666 return nn;
667 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
668 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
669 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
670 register_new_node(nn, get_ctrl(n));
671 _igvn.replace_node(n, nn);
672 return nn;
673 }
674 }
675 }
676 }
677 }
678 }
679 return nn;
680 }
681
682 //------------------------------conditional_move-------------------------------
683 // Attempt to replace a Phi with a conditional move. We have some pretty
684 // strict profitability requirements. All Phis at the merge point must
685 // be converted, so we can remove the control flow. We need to limit the
686 // number of c-moves to a small handful. All code that was in the side-arms
687 // of the CFG diamond is now speculatively executed. This code has to be
688 // "cheap enough". We are pretty much limited to CFG diamonds that merge
689 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
690 Node *PhaseIdealLoop::conditional_move( Node *region ) {
691
692 assert(region->is_Region(), "sanity check");
693 if (region->req() != 3) return nullptr;
694
695 // Check for CFG diamond
696 Node *lp = region->in(1);
697 Node *rp = region->in(2);
698 if (!lp || !rp) return nullptr;
699 Node *lp_c = lp->in(0);
700 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
701 IfNode *iff = lp_c->as_If();
702
703 // Check for ops pinned in an arm of the diamond.
704 // Can't remove the control flow in this case
705 if (lp->outcnt() > 1) return nullptr;
706 if (rp->outcnt() > 1) return nullptr;
707
708 IdealLoopTree* r_loop = get_loop(region);
709 assert(r_loop == get_loop(iff), "sanity");
710 // Always convert to CMOVE if all results are used only outside this loop.
711 bool used_inside_loop = (r_loop == _ltree_root);
712
713 // Check profitability
714 int cost = 0;
715 int phis = 0;
716 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
717 Node *out = region->fast_out(i);
718 if (!out->is_Phi()) continue; // Ignore other control edges, etc
719 phis++;
720 PhiNode* phi = out->as_Phi();
721 BasicType bt = phi->type()->basic_type();
722 switch (bt) {
723 case T_DOUBLE:
724 case T_FLOAT:
725 if (C->use_cmove()) {
726 continue; //TODO: maybe we want to add some cost
727 }
728 cost += Matcher::float_cmove_cost(); // Could be very expensive
729 break;
730 case T_LONG: {
731 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
732 }
733 case T_INT: // These all CMOV fine
734 case T_ADDRESS: { // (RawPtr)
735 cost++;
736 break;
737 }
738 case T_NARROWOOP: // Fall through
739 case T_OBJECT: { // Base oops are OK, but not derived oops
740 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
741 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
742 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
743 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
744 // have a Phi for the base here that we convert to a CMOVE all is well
745 // and good. But if the base is dead, we'll not make a CMOVE. Later
746 // the allocator will have to produce a base by creating a CMOVE of the
747 // relevant bases. This puts the allocator in the business of
748 // manufacturing expensive instructions, generally a bad plan.
749 // Just Say No to Conditionally-Moved Derived Pointers.
750 if (tp && tp->offset() != 0)
751 return nullptr;
752 cost++;
753 break;
754 }
755 default:
756 return nullptr; // In particular, can't do memory or I/O
757 }
758 // Add in cost any speculative ops
759 for (uint j = 1; j < region->req(); j++) {
760 Node *proj = region->in(j);
761 Node *inp = phi->in(j);
762 if (inp->isa_InlineType()) {
763 // TODO 8302217 This prevents PhiNode::push_inline_types_through
764 return nullptr;
765 }
766 if (get_ctrl(inp) == proj) { // Found local op
767 cost++;
768 // Check for a chain of dependent ops; these will all become
769 // speculative in a CMOV.
770 for (uint k = 1; k < inp->req(); k++)
771 if (get_ctrl(inp->in(k)) == proj)
772 cost += ConditionalMoveLimit; // Too much speculative goo
773 }
774 }
775 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
776 // This will likely Split-If, a higher-payoff operation.
777 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
778 Node* use = phi->fast_out(k);
779 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
780 cost += ConditionalMoveLimit;
781 // Is there a use inside the loop?
782 // Note: check only basic types since CMoveP is pinned.
783 if (!used_inside_loop && is_java_primitive(bt)) {
784 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
785 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
786 used_inside_loop = true;
787 }
788 }
789 }
790 }//for
791 Node* bol = iff->in(1);
792 if (bol->Opcode() == Op_Opaque4) {
793 return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
794 }
795 assert(bol->Opcode() == Op_Bool, "Unexpected node");
796 int cmp_op = bol->in(1)->Opcode();
797 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
798 return nullptr;
799 }
800 // It is expensive to generate flags from a float compare.
801 // Avoid duplicated float compare.
802 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
803
804 // Ignore cost if CMOVE can be moved outside the loop.
805 if (used_inside_loop && cost >= ConditionalMoveLimit) {
806 return nullptr;
807 }
808 // Check for highly predictable branch. No point in CMOV'ing if
809 // we are going to predict accurately all the time.
810 constexpr float infrequent_prob = PROB_UNLIKELY_MAG(2);
811 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
812 //keep going
813 } else if (iff->_prob < infrequent_prob || iff->_prob > (1.0f - infrequent_prob)) {
814 return nullptr;
815 }
816
817 // --------------
818 // Now replace all Phis with CMOV's
819 Node *cmov_ctrl = iff->in(0);
820 uint flip = (lp->Opcode() == Op_IfTrue);
821 Node_List wq;
822 while (1) {
823 PhiNode* phi = nullptr;
824 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
825 Node *out = region->fast_out(i);
826 if (out->is_Phi()) {
827 phi = out->as_Phi();
828 break;
829 }
830 }
831 if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
832 break;
833 }
834 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
835 // Move speculative ops
836 wq.push(phi);
837 while (wq.size() > 0) {
838 Node *n = wq.pop();
839 for (uint j = 1; j < n->req(); j++) {
840 Node* m = n->in(j);
841 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
842 #ifndef PRODUCT
843 if (PrintOpto && VerifyLoopOptimizations) {
844 tty->print(" speculate: ");
845 m->dump();
846 }
847 #endif
848 set_ctrl(m, cmov_ctrl);
849 wq.push(m);
850 }
851 }
852 }
853 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
854 register_new_node( cmov, cmov_ctrl );
855 _igvn.replace_node( phi, cmov );
856 #ifndef PRODUCT
857 if (TraceLoopOpts) {
858 tty->print("CMOV ");
859 r_loop->dump_head();
860 if (Verbose) {
861 bol->in(1)->dump(1);
862 cmov->dump(1);
863 }
864 }
865 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
866 #endif
867 }
868
869 // The useless CFG diamond will fold up later; see the optimization in
870 // RegionNode::Ideal.
871 _igvn._worklist.push(region);
872
873 return iff->in(1);
874 }
875
876 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
877 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
878 Node* u = m->fast_out(i);
879 if (u->is_CFG()) {
880 if (u->is_NeverBranch()) {
881 u = u->as_NeverBranch()->proj_out(0);
882 enqueue_cfg_uses(u, wq);
883 } else {
884 wq.push(u);
885 }
886 }
887 }
888 }
889
890 // Try moving a store out of a loop, right before the loop
891 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
892 // Store has to be first in the loop body
893 IdealLoopTree *n_loop = get_loop(n_ctrl);
894 if (n->is_Store() && n_loop != _ltree_root &&
895 n_loop->is_loop() && n_loop->_head->is_Loop() &&
896 n->in(0) != nullptr) {
897 Node* address = n->in(MemNode::Address);
898 Node* value = n->in(MemNode::ValueIn);
899 Node* mem = n->in(MemNode::Memory);
900 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
901 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
902
903 // - address and value must be loop invariant
904 // - memory must be a memory Phi for the loop
905 // - Store must be the only store on this memory slice in the
906 // loop: if there's another store following this one then value
907 // written at iteration i by the second store could be overwritten
908 // at iteration i+n by the first store: it's not safe to move the
909 // first store out of the loop
910 // - nothing must observe the memory Phi: it guarantees no read
911 // before the store, we are also guaranteed the store post
912 // dominates the loop head (ignoring a possible early
913 // exit). Otherwise there would be extra Phi involved between the
914 // loop's Phi and the store.
915 // - there must be no early exit from the loop before the Store
916 // (such an exit most of the time would be an extra use of the
917 // memory Phi but sometimes is a bottom memory Phi that takes the
918 // store as input).
919
920 if (!n_loop->is_member(address_loop) &&
921 !n_loop->is_member(value_loop) &&
922 mem->is_Phi() && mem->in(0) == n_loop->_head &&
923 mem->outcnt() == 1 &&
924 mem->in(LoopNode::LoopBackControl) == n) {
925
926 assert(n_loop->_tail != nullptr, "need a tail");
927 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
928
929 // Verify that there's no early exit of the loop before the store.
930 bool ctrl_ok = false;
931 {
932 // Follow control from loop head until n, we exit the loop or
933 // we reach the tail
934 ResourceMark rm;
935 Unique_Node_List wq;
936 wq.push(n_loop->_head);
937
938 for (uint next = 0; next < wq.size(); ++next) {
939 Node *m = wq.at(next);
940 if (m == n->in(0)) {
941 ctrl_ok = true;
942 continue;
943 }
944 assert(!has_ctrl(m), "should be CFG");
945 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
946 ctrl_ok = false;
947 break;
948 }
949 enqueue_cfg_uses(m, wq);
950 if (wq.size() > 10) {
951 ctrl_ok = false;
952 break;
953 }
954 }
955 }
956 if (ctrl_ok) {
957 // move the Store
958 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
959 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
960 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
961 // Disconnect the phi now. An empty phi can confuse other
962 // optimizations in this pass of loop opts.
963 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
964 n_loop->_body.yank(mem);
965
966 set_ctrl_and_loop(n, n->in(0));
967
968 return n;
969 }
970 }
971 }
972 return nullptr;
973 }
974
975 // Try moving a store out of a loop, right after the loop
976 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
977 if (n->is_Store() && n->in(0) != nullptr) {
978 Node *n_ctrl = get_ctrl(n);
979 IdealLoopTree *n_loop = get_loop(n_ctrl);
980 // Store must be in a loop
981 if (n_loop != _ltree_root && !n_loop->_irreducible) {
982 Node* address = n->in(MemNode::Address);
983 Node* value = n->in(MemNode::ValueIn);
984 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
985 // address must be loop invariant
986 if (!n_loop->is_member(address_loop)) {
987 // Store must be last on this memory slice in the loop and
988 // nothing in the loop must observe it
989 Node* phi = nullptr;
990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
991 Node* u = n->fast_out(i);
992 if (has_ctrl(u)) { // control use?
993 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
994 if (!n_loop->is_member(u_loop)) {
995 continue;
996 }
997 if (u->is_Phi() && u->in(0) == n_loop->_head) {
998 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
999 // multiple phis on the same slice are possible
1000 if (phi != nullptr) {
1001 return;
1002 }
1003 phi = u;
1004 continue;
1005 }
1006 }
1007 return;
1008 }
1009 if (phi != nullptr) {
1010 // Nothing in the loop before the store (next iteration)
1011 // must observe the stored value
1012 bool mem_ok = true;
1013 {
1014 ResourceMark rm;
1015 Unique_Node_List wq;
1016 wq.push(phi);
1017 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1018 Node *m = wq.at(next);
1019 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1020 Node* u = m->fast_out(i);
1021 if (u->is_Store() || u->is_Phi()) {
1022 if (u != n) {
1023 wq.push(u);
1024 mem_ok = (wq.size() <= 10);
1025 }
1026 } else {
1027 mem_ok = false;
1028 break;
1029 }
1030 }
1031 }
1032 }
1033 if (mem_ok) {
1034 // Move the store out of the loop if the LCA of all
1035 // users (except for the phi) is outside the loop.
1036 Node* hook = new Node(1);
1037 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1038 _igvn.rehash_node_delayed(phi);
1039 int count = phi->replace_edge(n, hook, &_igvn);
1040 assert(count > 0, "inconsistent phi");
1041
1042 // Compute latest point this store can go
1043 Node* lca = get_late_ctrl(n, get_ctrl(n));
1044 if (lca->is_OuterStripMinedLoop()) {
1045 lca = lca->in(LoopNode::EntryControl);
1046 }
1047 if (n_loop->is_member(get_loop(lca))) {
1048 // LCA is in the loop - bail out
1049 _igvn.replace_node(hook, n);
1050 return;
1051 }
1052 #ifdef ASSERT
1053 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1054 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1055 n_loop->_head->as_Loop()->verify_strip_mined(1);
1056 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1057 IdealLoopTree* outer_loop = get_loop(outer);
1058 assert(n_loop->_parent == outer_loop, "broken loop tree");
1059 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1060 }
1061 #endif
1062 lca = place_outside_loop(lca, n_loop);
1063 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1064 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop");
1065
1066 // Move store out of the loop
1067 _igvn.replace_node(hook, n->in(MemNode::Memory));
1068 _igvn.replace_input_of(n, 0, lca);
1069 set_ctrl_and_loop(n, lca);
1070
1071 // Disconnect the phi now. An empty phi can confuse other
1072 // optimizations in this pass of loop opts..
1073 if (phi->in(LoopNode::LoopBackControl) == phi) {
1074 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1075 n_loop->_body.yank(phi);
1076 }
1077 }
1078 }
1079 }
1080 }
1081 }
1082 }
1083
1084 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1085 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1086 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1087 // flat array checks out of loops, mainly to enable loop unswitching.
1088 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1089 // Skip checks for more than one array
1090 if (n->req() > 3) {
1091 return;
1092 }
1093 Node* mem = n->in(FlatArrayCheckNode::Memory);
1094 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1095 IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1096 IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1097
1098 // Check if array is loop invariant
1099 if (!check_loop->is_member(ary_loop)) {
1100 // Walk up memory graph from the check until we leave the loop
1101 VectorSet wq;
1102 wq.set(mem->_idx);
1103 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1104 if (mem->is_Phi()) {
1105 mem = mem->in(1);
1106 } else if (mem->is_MergeMem()) {
1107 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1108 } else if (mem->is_Proj()) {
1109 mem = mem->in(0);
1110 } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1111 mem = mem->in(TypeFunc::Memory);
1112 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1113 mem = mem->in(MemNode::Memory);
1114 } else {
1115 #ifdef ASSERT
1116 mem->dump();
1117 #endif
1118 ShouldNotReachHere();
1119 }
1120 if (wq.test_set(mem->_idx)) {
1121 return;
1122 }
1123 }
1124 // Replace memory input and re-compute ctrl to move the check out of the loop
1125 _igvn.replace_input_of(n, 1, mem);
1126 set_ctrl_and_loop(n, get_early_ctrl(n));
1127 Node* bol = n->unique_out();
1128 set_ctrl_and_loop(bol, get_early_ctrl(bol));
1129 }
1130 }
1131
1132 //------------------------------split_if_with_blocks_pre-----------------------
1133 // Do the real work in a non-recursive function. Data nodes want to be
1134 // cloned in the pre-order so they can feed each other nicely.
1135 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1136 // Cloning these guys is unlikely to win
1137 int n_op = n->Opcode();
1138 if (n_op == Op_MergeMem) {
1139 return n;
1140 }
1141 if (n->is_Proj()) {
1142 return n;
1143 }
1144
1145 if (n->isa_FlatArrayCheck()) {
1146 move_flat_array_check_out_of_loop(n);
1147 return n;
1148 }
1149
1150 // Do not clone-up CmpFXXX variations, as these are always
1151 // followed by a CmpI
1152 if (n->is_Cmp()) {
1153 return n;
1154 }
1155 // Attempt to use a conditional move instead of a phi/branch
1156 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1157 Node *cmov = conditional_move( n );
1158 if (cmov) {
1159 return cmov;
1160 }
1161 }
1162 if (n->is_CFG() || n->is_LoadStore()) {
1163 return n;
1164 }
1165 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1166 if (!C->major_progress()) { // If chance of no more loop opts...
1167 _igvn._worklist.push(n); // maybe we'll remove them
1168 }
1169 return n;
1170 }
1171
1172 if (n->is_Con()) {
1173 return n; // No cloning for Con nodes
1174 }
1175
1176 Node *n_ctrl = get_ctrl(n);
1177 if (!n_ctrl) {
1178 return n; // Dead node
1179 }
1180
1181 Node* res = try_move_store_before_loop(n, n_ctrl);
1182 if (res != nullptr) {
1183 return n;
1184 }
1185
1186 // Attempt to remix address expressions for loop invariants
1187 Node *m = remix_address_expressions( n );
1188 if( m ) return m;
1189
1190 if (n_op == Op_AddI) {
1191 Node *nn = convert_add_to_muladd( n );
1192 if ( nn ) return nn;
1193 }
1194
1195 if (n->is_ConstraintCast()) {
1196 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1197 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1198 // Node control inputs don't necessarily agree with loop control info (due to
1199 // transformations happened in between), thus additional dominance check is needed
1200 // to keep loop info valid.
1201 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1202 _igvn.replace_node(n, dom_cast);
1203 return dom_cast;
1204 }
1205 }
1206
1207 // Determine if the Node has inputs from some local Phi.
1208 // Returns the block to clone thru.
1209 Node *n_blk = has_local_phi_input( n );
1210 if( !n_blk ) return n;
1211
1212 // Do not clone the trip counter through on a CountedLoop
1213 // (messes up the canonical shape).
1214 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1215 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1216 return n;
1217 }
1218 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1219 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1220 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1221 return n;
1222 }
1223
1224 // Check for having no control input; not pinned. Allow
1225 // dominating control.
1226 if (n->in(0)) {
1227 Node *dom = idom(n_blk);
1228 if (dom_lca(n->in(0), dom) != n->in(0)) {
1229 return n;
1230 }
1231 }
1232 // Policy: when is it profitable. You must get more wins than
1233 // policy before it is considered profitable. Policy is usually 0,
1234 // so 1 win is considered profitable. Big merges will require big
1235 // cloning, so get a larger policy.
1236 int policy = n_blk->req() >> 2;
1237
1238 // If the loop is a candidate for range check elimination,
1239 // delay splitting through it's phi until a later loop optimization
1240 if (n_blk->is_BaseCountedLoop()) {
1241 IdealLoopTree *lp = get_loop(n_blk);
1242 if (lp && lp->_rce_candidate) {
1243 return n;
1244 }
1245 }
1246
1247 if (must_throttle_split_if()) return n;
1248
1249 // Split 'n' through the merge point if it is profitable
1250 Node *phi = split_thru_phi( n, n_blk, policy );
1251 if (!phi) return n;
1252
1253 // Found a Phi to split thru!
1254 // Replace 'n' with the new phi
1255 _igvn.replace_node( n, phi );
1256 // Moved a load around the loop, 'en-registering' something.
1257 if (n_blk->is_Loop() && n->is_Load() &&
1258 !phi->in(LoopNode::LoopBackControl)->is_Load())
1259 C->set_major_progress();
1260
1261 return phi;
1262 }
1263
1264 static bool merge_point_too_heavy(Compile* C, Node* region) {
1265 // Bail out if the region and its phis have too many users.
1266 int weight = 0;
1267 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1268 weight += region->fast_out(i)->outcnt();
1269 }
1270 int nodes_left = C->max_node_limit() - C->live_nodes();
1271 if (weight * 8 > nodes_left) {
1272 if (PrintOpto) {
1273 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1274 }
1275 return true;
1276 } else {
1277 return false;
1278 }
1279 }
1280
1281 static bool merge_point_safe(Node* region) {
1282 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1283 // having a PhiNode input. This sidesteps the dangerous case where the split
1284 // ConvI2LNode may become TOP if the input Value() does not
1285 // overlap the ConvI2L range, leaving a node which may not dominate its
1286 // uses.
1287 // A better fix for this problem can be found in the BugTraq entry, but
1288 // expediency for Mantis demands this hack.
1289 #ifdef _LP64
1290 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1291 Node* n = region->fast_out(i);
1292 if (n->is_Phi()) {
1293 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1294 Node* m = n->fast_out(j);
1295 if (m->Opcode() == Op_ConvI2L)
1296 return false;
1297 if (m->is_CastII()) {
1298 return false;
1299 }
1300 }
1301 }
1302 }
1303 #endif
1304 return true;
1305 }
1306
1307
1308 //------------------------------place_outside_loop---------------------------------
1309 // Place some computation outside of this loop on the path to the use passed as argument
1310 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1311 Node* head = loop->_head;
1312 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1313 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1314 loop = loop->_parent;
1315 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1316 }
1317
1318 // Pick control right outside the loop
1319 for (;;) {
1320 Node* dom = idom(useblock);
1321 if (loop->is_member(get_loop(dom)) ||
1322 // NeverBranch nodes are not assigned to the loop when constructed
1323 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) {
1324 break;
1325 }
1326 useblock = dom;
1327 }
1328 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1329 return useblock;
1330 }
1331
1332
1333 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1334 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1335 return false;
1336 }
1337 if (!n->in(0)->is_Region()) {
1338 return false;
1339 }
1340
1341 Node* region = n->in(0);
1342 Node* dom = idom(region);
1343 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1344 return false;
1345 }
1346 IfNode* dom_if = dom->as_If();
1347 Node* proj_true = dom_if->proj_out(1);
1348 Node* proj_false = dom_if->proj_out(0);
1349
1350 for (uint i = 1; i < region->req(); i++) {
1351 if (is_dominator(proj_true, region->in(i))) {
1352 continue;
1353 }
1354 if (is_dominator(proj_false, region->in(i))) {
1355 continue;
1356 }
1357 return false;
1358 }
1359
1360 return true;
1361 }
1362
1363
1364 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1365 if (must_throttle_split_if()) {
1366 return false;
1367 }
1368
1369 // Do not do 'split-if' if irreducible loops are present.
1370 if (_has_irreducible_loops) {
1371 return false;
1372 }
1373
1374 if (merge_point_too_heavy(C, n_ctrl)) {
1375 return false;
1376 }
1377
1378 // Do not do 'split-if' if some paths are dead. First do dead code
1379 // elimination and then see if its still profitable.
1380 for (uint i = 1; i < n_ctrl->req(); i++) {
1381 if (n_ctrl->in(i) == C->top()) {
1382 return false;
1383 }
1384 }
1385
1386 // If trying to do a 'Split-If' at the loop head, it is only
1387 // profitable if the cmp folds up on BOTH paths. Otherwise we
1388 // risk peeling a loop forever.
1389
1390 // CNC - Disabled for now. Requires careful handling of loop
1391 // body selection for the cloned code. Also, make sure we check
1392 // for any input path not being in the same loop as n_ctrl. For
1393 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1394 // because the alternative loop entry points won't be converted
1395 // into LoopNodes.
1396 IdealLoopTree *n_loop = get_loop(n_ctrl);
1397 for (uint j = 1; j < n_ctrl->req(); j++) {
1398 if (get_loop(n_ctrl->in(j)) != n_loop) {
1399 return false;
1400 }
1401 }
1402
1403 // Check for safety of the merge point.
1404 if (!merge_point_safe(n_ctrl)) {
1405 return false;
1406 }
1407
1408 return true;
1409 }
1410
1411 // Detect if the node is the inner strip-mined loop
1412 // Return: null if it's not the case, or the exit of outer strip-mined loop
1413 static Node* is_inner_of_stripmined_loop(const Node* out) {
1414 Node* out_le = nullptr;
1415
1416 if (out->is_CountedLoopEnd()) {
1417 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1418
1419 if (loop != nullptr && loop->is_strip_mined()) {
1420 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1421 }
1422 }
1423
1424 return out_le;
1425 }
1426
1427 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) {
1428 // If the CmpP is a subtype check for a value that has just been
1429 // loaded from an array, the subtype check guarantees the value
1430 // can't be stored in a flat array and the load of the value
1431 // happens with a flat array check then: push the type check
1432 // through the phi of the flat array check. This needs special
1433 // logic because the subtype check's input is not a phi but a
1434 // LoadKlass that must first be cloned through the phi.
1435 if (n->Opcode() != Op_CmpP) {
1436 return false;
1437 }
1438
1439 Node* klassptr = n->in(1);
1440 Node* klasscon = n->in(2);
1441
1442 if (klassptr->is_DecodeNarrowPtr()) {
1443 klassptr = klassptr->in(1);
1444 }
1445
1446 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1447 return false;
1448 }
1449
1450 if (!klasscon->is_Con()) {
1451 return false;
1452 }
1453
1454 Node* addr = klassptr->in(MemNode::Address);
1455
1456 if (!addr->is_AddP()) {
1457 return false;
1458 }
1459
1460 intptr_t offset;
1461 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1462
1463 if (obj == nullptr) {
1464 return false;
1465 }
1466
1467 assert(obj != nullptr && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?");
1468 if (obj->Opcode() == Op_CastPP) {
1469 obj = obj->in(1);
1470 }
1471
1472 if (!obj->is_Phi()) {
1473 return false;
1474 }
1475
1476 Node* region = obj->in(0);
1477
1478 Node* phi = PhiNode::make_blank(region, n->in(1));
1479 for (uint i = 1; i < region->req(); i++) {
1480 Node* in = obj->in(i);
1481 Node* ctrl = region->in(i);
1482 if (addr->in(AddPNode::Base) != obj) {
1483 Node* cast = addr->in(AddPNode::Base);
1484 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph");
1485 Node* cast_clone = cast->clone();
1486 cast_clone->set_req(0, ctrl);
1487 cast_clone->set_req(1, in);
1488 register_new_node(cast_clone, ctrl);
1489 const Type* tcast = cast_clone->Value(&_igvn);
1490 _igvn.set_type(cast_clone, tcast);
1491 cast_clone->as_Type()->set_type(tcast);
1492 in = cast_clone;
1493 }
1494 Node* addr_clone = addr->clone();
1495 addr_clone->set_req(AddPNode::Base, in);
1496 addr_clone->set_req(AddPNode::Address, in);
1497 register_new_node(addr_clone, ctrl);
1498 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1499 Node* klassptr_clone = klassptr->clone();
1500 klassptr_clone->set_req(2, addr_clone);
1501 register_new_node(klassptr_clone, ctrl);
1502 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1503 if (klassptr != n->in(1)) {
1504 Node* decode = n->in(1);
1505 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1506 Node* decode_clone = decode->clone();
1507 decode_clone->set_req(1, klassptr_clone);
1508 register_new_node(decode_clone, ctrl);
1509 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1510 klassptr_clone = decode_clone;
1511 }
1512 phi->set_req(i, klassptr_clone);
1513 }
1514 register_new_node(phi, region);
1515 Node* orig = n->in(1);
1516 _igvn.replace_input_of(n, 1, phi);
1517 split_if_with_blocks_post(n);
1518 if (n->outcnt() != 0) {
1519 _igvn.replace_input_of(n, 1, orig);
1520 _igvn.remove_dead_node(phi);
1521 }
1522 return true;
1523 }
1524
1525 //------------------------------split_if_with_blocks_post----------------------
1526 // Do the real work in a non-recursive function. CFG hackery wants to be
1527 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1528 // info.
1529 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1530
1531 if (flat_array_element_type_check(n)) {
1532 return;
1533 }
1534
1535 // Cloning Cmp through Phi's involves the split-if transform.
1536 // FastLock is not used by an If
1537 if (n->is_Cmp() && !n->is_FastLock()) {
1538 Node *n_ctrl = get_ctrl(n);
1539 // Determine if the Node has inputs from some local Phi.
1540 // Returns the block to clone thru.
1541 Node *n_blk = has_local_phi_input(n);
1542 if (n_blk != n_ctrl) {
1543 return;
1544 }
1545
1546 if (!can_split_if(n_ctrl)) {
1547 return;
1548 }
1549
1550 if (n->outcnt() != 1) {
1551 return; // Multiple bool's from 1 compare?
1552 }
1553 Node *bol = n->unique_out();
1554 assert(bol->is_Bool(), "expect a bool here");
1555 if (bol->outcnt() != 1) {
1556 return;// Multiple branches from 1 compare?
1557 }
1558 Node *iff = bol->unique_out();
1559
1560 // Check some safety conditions
1561 if (iff->is_If()) { // Classic split-if?
1562 if (iff->in(0) != n_ctrl) {
1563 return; // Compare must be in same blk as if
1564 }
1565 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1566 // Can't split CMove with different control.
1567 if (get_ctrl(iff) != n_ctrl) {
1568 return;
1569 }
1570 if (get_ctrl(iff->in(2)) == n_ctrl ||
1571 get_ctrl(iff->in(3)) == n_ctrl) {
1572 return; // Inputs not yet split-up
1573 }
1574 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1575 return; // Loop-invar test gates loop-varying CMOVE
1576 }
1577 } else {
1578 return; // some other kind of node, such as an Allocate
1579 }
1580
1581 // When is split-if profitable? Every 'win' on means some control flow
1582 // goes dead, so it's almost always a win.
1583 int policy = 0;
1584 // Split compare 'n' through the merge point if it is profitable
1585 Node *phi = split_thru_phi( n, n_ctrl, policy);
1586 if (!phi) {
1587 return;
1588 }
1589
1590 // Found a Phi to split thru!
1591 // Replace 'n' with the new phi
1592 _igvn.replace_node(n, phi);
1593
1594 // Now split the bool up thru the phi
1595 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1596 guarantee(bolphi != nullptr, "null boolean phi node");
1597
1598 _igvn.replace_node(bol, bolphi);
1599 assert(iff->in(1) == bolphi, "");
1600
1601 if (bolphi->Value(&_igvn)->singleton()) {
1602 return;
1603 }
1604
1605 // Conditional-move? Must split up now
1606 if (!iff->is_If()) {
1607 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1608 _igvn.replace_node(iff, cmovphi);
1609 return;
1610 }
1611
1612 // Now split the IF
1613 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1614 if ((PrintOpto && VerifyLoopOptimizations) || TraceLoopOpts) {
1615 tty->print_cr("Split-If");
1616 }
1617 do_split_if(iff);
1618 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1619 return;
1620 }
1621
1622 // Two identical ifs back to back can be merged
1623 if (try_merge_identical_ifs(n)) {
1624 return;
1625 }
1626
1627 // Check for an IF ready to split; one that has its
1628 // condition codes input coming from a Phi at the block start.
1629 int n_op = n->Opcode();
1630
1631 // Check for an IF being dominated by another IF same test
1632 if (n_op == Op_If ||
1633 n_op == Op_RangeCheck) {
1634 Node *bol = n->in(1);
1635 uint max = bol->outcnt();
1636 // Check for same test used more than once?
1637 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1638 // Search up IDOMs to see if this IF is dominated.
1639 Node* cmp = bol->in(1);
1640 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1641
1642 // Now search up IDOMs till cutoff, looking for a dominating test
1643 Node *prevdom = n;
1644 Node *dom = idom(prevdom);
1645 while (dom != cutoff) {
1646 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1647 safe_for_if_replacement(dom)) {
1648 // It's invalid to move control dependent data nodes in the inner
1649 // strip-mined loop, because:
1650 // 1) break validation of LoopNode::verify_strip_mined()
1651 // 2) move code with side-effect in strip-mined loop
1652 // Move to the exit of outer strip-mined loop in that case.
1653 Node* out_le = is_inner_of_stripmined_loop(dom);
1654 if (out_le != nullptr) {
1655 prevdom = out_le;
1656 }
1657 // Replace the dominated test with an obvious true or false.
1658 // Place it on the IGVN worklist for later cleanup.
1659 C->set_major_progress();
1660 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1661 // to prevent an array load from floating above its range check. There are three cases:
1662 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1663 // all its array accesses at that point.
1664 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1665 // accesses would start to float, since we don't pin at that point.
1666 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1667 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1668 prevdom->in(0)->Opcode() != Op_RangeCheck;
1669 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1670 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1671 return;
1672 }
1673 prevdom = dom;
1674 dom = idom(prevdom);
1675 }
1676 }
1677 }
1678
1679 try_sink_out_of_loop(n);
1680
1681 try_move_store_after_loop(n);
1682
1683 // Remove multiple allocations of the same inline type
1684 if (n->is_InlineType()) {
1685 n->as_InlineType()->remove_redundant_allocations(this);
1686 }
1687 }
1688
1689 // Transform:
1690 //
1691 // if (some_condition) {
1692 // // body 1
1693 // } else {
1694 // // body 2
1695 // }
1696 // if (some_condition) {
1697 // // body 3
1698 // } else {
1699 // // body 4
1700 // }
1701 //
1702 // into:
1703 //
1704 //
1705 // if (some_condition) {
1706 // // body 1
1707 // // body 3
1708 // } else {
1709 // // body 2
1710 // // body 4
1711 // }
1712 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1713 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1714 Node *n_ctrl = n->in(0);
1715 IfNode* dom_if = idom(n_ctrl)->as_If();
1716 if (n->in(1) != dom_if->in(1)) {
1717 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1718 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1719 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1720 _igvn.replace_input_of(n, 1, dom_if->in(1));
1721 }
1722 ProjNode* dom_proj_true = dom_if->proj_out(1);
1723 ProjNode* dom_proj_false = dom_if->proj_out(0);
1724
1725 // Now split the IF
1726 RegionNode* new_false_region;
1727 RegionNode* new_true_region;
1728 do_split_if(n, &new_false_region, &new_true_region);
1729 assert(new_false_region->req() == new_true_region->req(), "");
1730 #ifdef ASSERT
1731 for (uint i = 1; i < new_false_region->req(); ++i) {
1732 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1733 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");
1734 }
1735 #endif
1736 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1737
1738 // We now have:
1739 // if (some_condition) {
1740 // // body 1
1741 // if (some_condition) {
1742 // body3: // new_true_region
1743 // // body3
1744 // } else {
1745 // goto body4;
1746 // }
1747 // } else {
1748 // // body 2
1749 // if (some_condition) {
1750 // goto body3;
1751 // } else {
1752 // body4: // new_false_region
1753 // // body4;
1754 // }
1755 // }
1756 //
1757
1758 // clone pinned nodes thru the resulting regions
1759 push_pinned_nodes_thru_region(dom_if, new_true_region);
1760 push_pinned_nodes_thru_region(dom_if, new_false_region);
1761
1762 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1763 // 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
1764 // unrelated control dependency.
1765 for (uint i = 1; i < new_false_region->req(); i++) {
1766 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1767 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1768 } else {
1769 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1770 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1771 }
1772 }
1773 return true;
1774 }
1775 return false;
1776 }
1777
1778 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1779 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1780 Node* u = region->out(i);
1781 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1782 continue;
1783 }
1784 assert(u->in(0) == region, "not a control dependent node?");
1785 uint j = 1;
1786 for (; j < u->req(); ++j) {
1787 Node* in = u->in(j);
1788 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1789 break;
1790 }
1791 }
1792 if (j == u->req()) {
1793 Node *phi = PhiNode::make_blank(region, u);
1794 for (uint k = 1; k < region->req(); ++k) {
1795 Node* clone = u->clone();
1796 clone->set_req(0, region->in(k));
1797 register_new_node(clone, region->in(k));
1798 phi->init_req(k, clone);
1799 }
1800 register_new_node(phi, region);
1801 _igvn.replace_node(u, phi);
1802 --i;
1803 }
1804 }
1805 }
1806
1807 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1808 if (!dom->is_CountedLoopEnd()) {
1809 return true;
1810 }
1811 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1812 CountedLoopNode* cl = le->loopnode();
1813 if (cl == nullptr) {
1814 return true;
1815 }
1816 if (!cl->is_main_loop()) {
1817 return true;
1818 }
1819 if (cl->is_canonical_loop_entry() == nullptr) {
1820 return true;
1821 }
1822 // Further unrolling is possible so loop exit condition might change
1823 return false;
1824 }
1825
1826 // See if a shared loop-varying computation has no loop-varying uses.
1827 // Happens if something is only used for JVM state in uncommon trap exits,
1828 // like various versions of induction variable+offset. Clone the
1829 // computation per usage to allow it to sink out of the loop.
1830 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1831 if (has_ctrl(n) &&
1832 !n->is_Phi() &&
1833 !n->is_Bool() &&
1834 !n->is_Proj() &&
1835 !n->is_MergeMem() &&
1836 !n->is_CMove() &&
1837 n->Opcode() != Op_Opaque4 &&
1838 !n->is_Type()) {
1839 Node *n_ctrl = get_ctrl(n);
1840 IdealLoopTree *n_loop = get_loop(n_ctrl);
1841
1842 if (n->in(0) != nullptr) {
1843 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1844 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1845 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1846 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1847 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1848 Node* maybe_pinned_n = n;
1849 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1850 if (n->depends_only_on_test()) {
1851 Node* pinned_clone = n->pin_array_access_node();
1852 if (pinned_clone != nullptr) {
1853 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1854 // range check for that access. If that condition is replaced by an identical dominating one, then an
1855 // unpinned load would risk floating above its range check.
1856 register_new_node(pinned_clone, n_ctrl);
1857 maybe_pinned_n = pinned_clone;
1858 _igvn.replace_node(n, pinned_clone);
1859 }
1860 }
1861 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1862 }
1863 }
1864 if (n_loop != _ltree_root && n->outcnt() > 1) {
1865 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1866 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1867 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1868 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1869 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1870 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1871 Node* outer_loop_clone = nullptr;
1872 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1873 Node* u = n->last_out(j); // Clone private computation per use
1874 _igvn.rehash_node_delayed(u);
1875 Node* x = nullptr;
1876 if (n->depends_only_on_test()) {
1877 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1878 // range check for that access. If that condition is replaced by an identical dominating one, then an
1879 // unpinned load would risk floating above its range check.
1880 x = n->pin_array_access_node();
1881 }
1882 if (x == nullptr) {
1883 x = n->clone();
1884 }
1885 Node* x_ctrl = nullptr;
1886 if (u->is_Phi()) {
1887 // Replace all uses of normal nodes. Replace Phi uses
1888 // individually, so the separate Nodes can sink down
1889 // different paths.
1890 uint k = 1;
1891 while (u->in(k) != n) k++;
1892 u->set_req(k, x);
1893 // x goes next to Phi input path
1894 x_ctrl = u->in(0)->in(k);
1895 // Find control for 'x' next to use but not inside inner loops.
1896 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1897 --j;
1898 } else { // Normal use
1899 if (has_ctrl(u)) {
1900 x_ctrl = get_ctrl(u);
1901 } else {
1902 x_ctrl = u->in(0);
1903 }
1904 // Find control for 'x' next to use but not inside inner loops.
1905 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1906 // Replace all uses
1907 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1908 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1909 // anymore now that we're going to pin n as well
1910 _igvn.replace_node(u, x);
1911 --j;
1912 } else {
1913 int nb = u->replace_edge(n, x, &_igvn);
1914 j -= nb;
1915 }
1916 }
1917
1918 if (n->is_Load()) {
1919 // For loads, add a control edge to a CFG node outside of the loop
1920 // to force them to not combine and return back inside the loop
1921 // during GVN optimization (4641526).
1922 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1923
1924 IdealLoopTree* x_loop = get_loop(x_ctrl);
1925 Node* x_head = x_loop->_head;
1926 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1927 // Do not add duplicate LoadNodes to the outer strip mined loop
1928 if (outer_loop_clone != nullptr) {
1929 _igvn.replace_node(x, outer_loop_clone);
1930 continue;
1931 }
1932 outer_loop_clone = x;
1933 }
1934 x->set_req(0, x_ctrl);
1935 } else if (n->in(0) != nullptr){
1936 x->set_req(0, x_ctrl);
1937 }
1938 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1939 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1940 register_new_node(x, x_ctrl);
1941
1942 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1943 // All AddP nodes must keep the same base after sinking so:
1944 // 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,
1945 // their bases remain the same.
1946 // (see 2- below)
1947 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1948 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1949 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1950 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1951 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1952 assert(!x->is_Load(), "load should be pinned");
1953 // Use a cast node to pin clone out of loop
1954 Node* cast = nullptr;
1955 for (uint k = 0; k < x->req(); k++) {
1956 Node* in = x->in(k);
1957 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1958 const Type* in_t = _igvn.type(in);
1959 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1960 ConstraintCastNode::UnconditionalDependency, nullptr);
1961 }
1962 if (cast != nullptr) {
1963 Node* prev = _igvn.hash_find_insert(cast);
1964 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1965 cast->destruct(&_igvn);
1966 cast = prev;
1967 } else {
1968 register_new_node(cast, x_ctrl);
1969 }
1970 x->replace_edge(in, cast);
1971 // Chain of AddP nodes:
1972 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1973 if (x->is_AddP() && k == AddPNode::Base) {
1974 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1975 }
1976 break;
1977 }
1978 }
1979 assert(cast != nullptr, "must have added a cast to pin the node");
1980 }
1981 }
1982 _igvn.remove_dead_node(n);
1983 }
1984 _dom_lca_tags_round = 0;
1985 }
1986 }
1987 }
1988
1989 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1990 ResourceMark rm;
1991 Node_List wq;
1992 wq.push(x);
1993 while (wq.size() != 0) {
1994 Node* n = wq.pop();
1995 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1996 Node* u = n->fast_out(i);
1997 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1998 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1999 wq.push(u);
2000 }
2001 }
2002 }
2003 }
2004
2005 // Compute the early control of a node by following its inputs until we reach
2006 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
2007 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
2008 Node* early_ctrl = nullptr;
2009 ResourceMark rm;
2010 Unique_Node_List wq;
2011 wq.push(n);
2012 for (uint i = 0; i < wq.size(); i++) {
2013 Node* m = wq.at(i);
2014 Node* c = nullptr;
2015 if (m->is_CFG()) {
2016 c = m;
2017 } else if (m->pinned()) {
2018 c = m->in(0);
2019 } else {
2020 for (uint j = 0; j < m->req(); j++) {
2021 Node* in = m->in(j);
2022 if (in != nullptr) {
2023 wq.push(in);
2024 }
2025 }
2026 }
2027 if (c != nullptr) {
2028 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
2029 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
2030 early_ctrl = c;
2031 }
2032 }
2033 }
2034 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
2035 return early_ctrl;
2036 }
2037
2038 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
2039 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2040 Node* u = n->fast_out(i);
2041 if (u->is_Opaque1()) {
2042 return false; // Found loop limit, bugfix for 4677003
2043 }
2044 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
2045 // get_late_ctrl_with_anti_dep() use their own tag
2046 _dom_lca_tags_round++;
2047 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
2048
2049 if (u->is_Phi()) {
2050 for (uint j = 1; j < u->req(); ++j) {
2051 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
2052 return false;
2053 }
2054 }
2055 } else {
2056 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
2057 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
2058 return false;
2059 }
2060 }
2061 }
2062 return true;
2063 }
2064
2065 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
2066 if (n->is_Load()) {
2067 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
2068 }
2069 IdealLoopTree *u_loop = get_loop(ctrl);
2070 if (u_loop == n_loop) {
2071 return false; // Found loop-varying use
2072 }
2073 if (n_loop->is_member(u_loop)) {
2074 return false; // Found use in inner loop
2075 }
2076 // 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
2077 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
2078 // test of the pre loop above the point in the graph where it's pinned.
2079 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop() &&
2080 u_loop->_head->is_CountedLoop() && u_loop->_head->as_CountedLoop()->is_main_loop() &&
2081 n_loop->_next == get_loop(u_loop->_head->as_CountedLoop()->skip_strip_mined())) {
2082 return false;
2083 }
2084 return true;
2085 }
2086
2087 //------------------------------split_if_with_blocks---------------------------
2088 // Check for aggressive application of 'split-if' optimization,
2089 // using basic block level info.
2090 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
2091 Node* root = C->root();
2092 visited.set(root->_idx); // first, mark root as visited
2093 // Do pre-visit work for root
2094 Node* n = split_if_with_blocks_pre(root);
2095 uint cnt = n->outcnt();
2096 uint i = 0;
2097
2098 while (true) {
2099 // Visit all children
2100 if (i < cnt) {
2101 Node* use = n->raw_out(i);
2102 ++i;
2103 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
2104 // Now do pre-visit work for this use
2105 use = split_if_with_blocks_pre(use);
2106 nstack.push(n, i); // Save parent and next use's index.
2107 n = use; // Process all children of current use.
2108 cnt = use->outcnt();
2109 i = 0;
2110 }
2111 }
2112 else {
2113 // All of n's children have been processed, complete post-processing.
2114 if (cnt != 0 && !n->is_Con()) {
2115 assert(has_node(n), "no dead nodes");
2116 split_if_with_blocks_post(n);
2117 }
2118 if (must_throttle_split_if()) {
2119 nstack.clear();
2120 }
2121 if (nstack.is_empty()) {
2122 // Finished all nodes on stack.
2123 break;
2124 }
2125 // Get saved parent node and next use's index. Visit the rest of uses.
2126 n = nstack.node();
2127 cnt = n->outcnt();
2128 i = nstack.index();
2129 nstack.pop();
2130 }
2131 }
2132 }
2133
2134
2135 //=============================================================================
2136 //
2137 // C L O N E A L O O P B O D Y
2138 //
2139
2140 //------------------------------clone_iff--------------------------------------
2141 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2142 // "Nearly" because all Nodes have been cloned from the original in the loop,
2143 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2144 // through the Phi recursively, and return a Bool.
2145 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2146
2147 // Convert this Phi into a Phi merging Bools
2148 uint i;
2149 for (i = 1; i < phi->req(); i++) {
2150 Node *b = phi->in(i);
2151 if (b->is_Phi()) {
2152 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2153 } else {
2154 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
2155 }
2156 }
2157
2158 Node* n = phi->in(1);
2159 Node* sample_opaque = nullptr;
2160 Node *sample_bool = nullptr;
2161 if (n->Opcode() == Op_Opaque4) {
2162 sample_opaque = n;
2163 sample_bool = n->in(1);
2164 assert(sample_bool->is_Bool(), "wrong type");
2165 } else {
2166 sample_bool = n;
2167 }
2168 Node* sample_cmp = sample_bool->in(1);
2169 const Type* t = Type::TOP;
2170 const TypePtr* at = nullptr;
2171 if (sample_cmp->is_FlatArrayCheck()) {
2172 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2173 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2174 t = Type::MEMORY;
2175 at = TypeRawPtr::BOTTOM;
2176 }
2177
2178 // Make Phis to merge the Cmp's inputs.
2179 PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2180 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2181 for (i = 1; i < phi->req(); i++) {
2182 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2183 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2184 phi1->set_req(i, n1);
2185 phi2->set_req(i, n2);
2186 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2187 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2188 }
2189 // See if these Phis have been made before.
2190 // Register with optimizer
2191 Node *hit1 = _igvn.hash_find_insert(phi1);
2192 if (hit1) { // Hit, toss just made Phi
2193 _igvn.remove_dead_node(phi1); // Remove new phi
2194 assert(hit1->is_Phi(), "" );
2195 phi1 = (PhiNode*)hit1; // Use existing phi
2196 } else { // Miss
2197 _igvn.register_new_node_with_optimizer(phi1);
2198 }
2199 Node *hit2 = _igvn.hash_find_insert(phi2);
2200 if (hit2) { // Hit, toss just made Phi
2201 _igvn.remove_dead_node(phi2); // Remove new phi
2202 assert(hit2->is_Phi(), "" );
2203 phi2 = (PhiNode*)hit2; // Use existing phi
2204 } else { // Miss
2205 _igvn.register_new_node_with_optimizer(phi2);
2206 }
2207 // Register Phis with loop/block info
2208 set_ctrl(phi1, phi->in(0));
2209 set_ctrl(phi2, phi->in(0));
2210 // Make a new Cmp
2211 Node *cmp = sample_cmp->clone();
2212 cmp->set_req(1, phi1);
2213 cmp->set_req(2, phi2);
2214 _igvn.register_new_node_with_optimizer(cmp);
2215 set_ctrl(cmp, phi->in(0));
2216
2217 // Make a new Bool
2218 Node *b = sample_bool->clone();
2219 b->set_req(1,cmp);
2220 _igvn.register_new_node_with_optimizer(b);
2221 set_ctrl(b, phi->in(0));
2222
2223 if (sample_opaque != nullptr) {
2224 Node* opaque = sample_opaque->clone();
2225 opaque->set_req(1, b);
2226 _igvn.register_new_node_with_optimizer(opaque);
2227 set_ctrl(opaque, phi->in(0));
2228 return opaque;
2229 }
2230
2231 assert(b->is_Bool(), "");
2232 return b;
2233 }
2234
2235 //------------------------------clone_bool-------------------------------------
2236 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2237 // "Nearly" because all Nodes have been cloned from the original in the loop,
2238 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2239 // through the Phi recursively, and return a Bool.
2240 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2241 uint i;
2242 // Convert this Phi into a Phi merging Bools
2243 for( i = 1; i < phi->req(); i++ ) {
2244 Node *b = phi->in(i);
2245 if( b->is_Phi() ) {
2246 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2247 } else {
2248 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2249 }
2250 }
2251
2252 Node *sample_cmp = phi->in(1);
2253
2254 // Make Phis to merge the Cmp's inputs.
2255 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2256 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2257 for( uint j = 1; j < phi->req(); j++ ) {
2258 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2259 Node *n1, *n2;
2260 if( cmp_top->is_Cmp() ) {
2261 n1 = cmp_top->in(1);
2262 n2 = cmp_top->in(2);
2263 } else {
2264 n1 = n2 = cmp_top;
2265 }
2266 phi1->set_req( j, n1 );
2267 phi2->set_req( j, n2 );
2268 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2269 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2270 }
2271
2272 // See if these Phis have been made before.
2273 // Register with optimizer
2274 Node *hit1 = _igvn.hash_find_insert(phi1);
2275 if( hit1 ) { // Hit, toss just made Phi
2276 _igvn.remove_dead_node(phi1); // Remove new phi
2277 assert( hit1->is_Phi(), "" );
2278 phi1 = (PhiNode*)hit1; // Use existing phi
2279 } else { // Miss
2280 _igvn.register_new_node_with_optimizer(phi1);
2281 }
2282 Node *hit2 = _igvn.hash_find_insert(phi2);
2283 if( hit2 ) { // Hit, toss just made Phi
2284 _igvn.remove_dead_node(phi2); // Remove new phi
2285 assert( hit2->is_Phi(), "" );
2286 phi2 = (PhiNode*)hit2; // Use existing phi
2287 } else { // Miss
2288 _igvn.register_new_node_with_optimizer(phi2);
2289 }
2290 // Register Phis with loop/block info
2291 set_ctrl(phi1, phi->in(0));
2292 set_ctrl(phi2, phi->in(0));
2293 // Make a new Cmp
2294 Node *cmp = sample_cmp->clone();
2295 cmp->set_req( 1, phi1 );
2296 cmp->set_req( 2, phi2 );
2297 _igvn.register_new_node_with_optimizer(cmp);
2298 set_ctrl(cmp, phi->in(0));
2299
2300 assert( cmp->is_Cmp(), "" );
2301 return (CmpNode*)cmp;
2302 }
2303
2304 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2305 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2306 Node_List*& split_if_set, Node_List*& split_bool_set,
2307 Node_List*& split_cex_set, Node_List& worklist,
2308 uint new_counter, CloneLoopMode mode) {
2309 Node* nnn = old_new[old->_idx];
2310 // Copy uses to a worklist, so I can munge the def-use info
2311 // with impunity.
2312 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2313 worklist.push(old->fast_out(j));
2314
2315 while( worklist.size() ) {
2316 Node *use = worklist.pop();
2317 if (!has_node(use)) continue; // Ignore dead nodes
2318 if (use->in(0) == C->top()) continue;
2319 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2320 // Check for data-use outside of loop - at least one of OLD or USE
2321 // must not be a CFG node.
2322 #ifdef ASSERT
2323 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2324 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2325 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2326 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2327 }
2328 #endif
2329 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2330
2331 // If the Data use is an IF, that means we have an IF outside of the
2332 // loop that is switching on a condition that is set inside of the
2333 // loop. Happens if people set a loop-exit flag; then test the flag
2334 // in the loop to break the loop, then test is again outside of the
2335 // loop to determine which way the loop exited.
2336 // Loop predicate If node connects to Bool node through Opaque1 node.
2337 //
2338 // If the use is an AllocateArray through its ValidLengthTest input,
2339 // make sure the Bool/Cmp input is cloned down to avoid a Phi between
2340 // the AllocateArray node and its ValidLengthTest input that could cause
2341 // split if to break.
2342 if (use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque4 ||
2343 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2344 // Since this code is highly unlikely, we lazily build the worklist
2345 // of such Nodes to go split.
2346 if (!split_if_set) {
2347 split_if_set = new Node_List();
2348 }
2349 split_if_set->push(use);
2350 }
2351 if (use->is_Bool()) {
2352 if (!split_bool_set) {
2353 split_bool_set = new Node_List();
2354 }
2355 split_bool_set->push(use);
2356 }
2357 if (use->Opcode() == Op_CreateEx) {
2358 if (!split_cex_set) {
2359 split_cex_set = new Node_List();
2360 }
2361 split_cex_set->push(use);
2362 }
2363
2364
2365 // Get "block" use is in
2366 uint idx = 0;
2367 while( use->in(idx) != old ) idx++;
2368 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2369 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2370 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2371 ? prev->in(2)
2372 : idom(prev);
2373 if( use->is_Phi() ) // Phi use is in prior block
2374 cfg = prev->in(idx); // NOT in block of Phi itself
2375 if (cfg->is_top()) { // Use is dead?
2376 _igvn.replace_input_of(use, idx, C->top());
2377 continue;
2378 }
2379
2380 // If use is referenced through control edge... (idx == 0)
2381 if (mode == IgnoreStripMined && idx == 0) {
2382 LoopNode *head = loop->_head->as_Loop();
2383 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2384 // That node is outside the inner loop, leave it outside the
2385 // outer loop as well to not confuse verification code.
2386 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2387 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2388 continue;
2389 }
2390 }
2391
2392 while(!outer_loop->is_member(get_loop(cfg))) {
2393 prev = cfg;
2394 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2395 }
2396 // If the use occurs after merging several exits from the loop, then
2397 // old value must have dominated all those exits. Since the same old
2398 // value was used on all those exits we did not need a Phi at this
2399 // merge point. NOW we do need a Phi here. Each loop exit value
2400 // is now merged with the peeled body exit; each exit gets its own
2401 // private Phi and those Phis need to be merged here.
2402 Node *phi;
2403 if( prev->is_Region() ) {
2404 if( idx == 0 ) { // Updating control edge?
2405 phi = prev; // Just use existing control
2406 } else { // Else need a new Phi
2407 phi = PhiNode::make( prev, old );
2408 // Now recursively fix up the new uses of old!
2409 for( uint i = 1; i < prev->req(); i++ ) {
2410 worklist.push(phi); // Onto worklist once for each 'old' input
2411 }
2412 }
2413 } else {
2414 // Get new RegionNode merging old and new loop exits
2415 prev = old_new[prev->_idx];
2416 assert( prev, "just made this in step 7" );
2417 if( idx == 0) { // Updating control edge?
2418 phi = prev; // Just use existing control
2419 } else { // Else need a new Phi
2420 // Make a new Phi merging data values properly
2421 phi = PhiNode::make( prev, old );
2422 phi->set_req( 1, nnn );
2423 }
2424 }
2425 // If inserting a new Phi, check for prior hits
2426 if( idx != 0 ) {
2427 Node *hit = _igvn.hash_find_insert(phi);
2428 if( hit == nullptr ) {
2429 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2430 } else { // or
2431 // Remove the new phi from the graph and use the hit
2432 _igvn.remove_dead_node(phi);
2433 phi = hit; // Use existing phi
2434 }
2435 set_ctrl(phi, prev);
2436 }
2437 // Make 'use' use the Phi instead of the old loop body exit value
2438 assert(use->in(idx) == old, "old is still input of use");
2439 // We notify all uses of old, including use, and the indirect uses,
2440 // that may now be optimized because we have replaced old with phi.
2441 _igvn.add_users_to_worklist(old);
2442 if (idx == 0 &&
2443 use->depends_only_on_test()) {
2444 Node* pinned_clone = use->pin_array_access_node();
2445 if (pinned_clone != nullptr) {
2446 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2447 // into the region is left, an array load could become dependent on a condition that's not a range check for
2448 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2449 // floating above its range check.
2450 pinned_clone->set_req(0, phi);
2451 register_new_node(pinned_clone, get_ctrl(use));
2452 _igvn.replace_node(use, pinned_clone);
2453 continue;
2454 }
2455 }
2456 _igvn.replace_input_of(use, idx, phi);
2457 if( use->_idx >= new_counter ) { // If updating new phis
2458 // Not needed for correctness, but prevents a weak assert
2459 // in AddPNode from tripping (when we end up with different
2460 // base & derived Phis that will become the same after
2461 // IGVN does CSE).
2462 Node *hit = _igvn.hash_find_insert(use);
2463 if( hit ) // Go ahead and re-hash for hits.
2464 _igvn.replace_node( use, hit );
2465 }
2466 }
2467 }
2468 }
2469
2470 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2471 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2472 bool check_old_new) {
2473 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2474 Node* u = n->fast_out(j);
2475 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2476 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2477 Node* c = phase->get_ctrl(u);
2478 IdealLoopTree* u_loop = phase->get_loop(c);
2479 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2480 if (!loop->is_member(u_loop)) {
2481 if (outer_loop->is_member(u_loop)) {
2482 wq.push(u);
2483 } else {
2484 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2485 // the outer loop too
2486 Node* u_c = u->in(0);
2487 if (u_c != nullptr) {
2488 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2489 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2490 wq.push(u);
2491 }
2492 }
2493 }
2494 }
2495 }
2496 }
2497 }
2498
2499 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2500 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2501 Node_List& extra_data_nodes) {
2502 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2503 CountedLoopNode* cl = head->as_CountedLoop();
2504 Node* l = cl->outer_loop();
2505 Node* tail = cl->outer_loop_tail();
2506 IfNode* le = cl->outer_loop_end();
2507 Node* sfpt = cl->outer_safepoint();
2508 CountedLoopEndNode* cle = cl->loopexit();
2509 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2510 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2511 Node* cle_out = cle->proj_out(false);
2512
2513 Node* new_sfpt = nullptr;
2514 Node* new_cle_out = cle_out->clone();
2515 old_new.map(cle_out->_idx, new_cle_out);
2516 if (mode == CloneIncludesStripMined) {
2517 // clone outer loop body
2518 Node* new_l = l->clone();
2519 Node* new_tail = tail->clone();
2520 IfNode* new_le = le->clone()->as_If();
2521 new_sfpt = sfpt->clone();
2522
2523 set_loop(new_l, outer_loop->_parent);
2524 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2525 set_loop(new_cle_out, outer_loop->_parent);
2526 set_idom(new_cle_out, new_cle, dd);
2527 set_loop(new_sfpt, outer_loop->_parent);
2528 set_idom(new_sfpt, new_cle_out, dd);
2529 set_loop(new_le, outer_loop->_parent);
2530 set_idom(new_le, new_sfpt, dd);
2531 set_loop(new_tail, outer_loop->_parent);
2532 set_idom(new_tail, new_le, dd);
2533 set_idom(new_cl, new_l, dd);
2534
2535 old_new.map(l->_idx, new_l);
2536 old_new.map(tail->_idx, new_tail);
2537 old_new.map(le->_idx, new_le);
2538 old_new.map(sfpt->_idx, new_sfpt);
2539
2540 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2541 new_l->set_req(0, new_l);
2542 new_tail->set_req(0, new_le);
2543 new_le->set_req(0, new_sfpt);
2544 new_sfpt->set_req(0, new_cle_out);
2545 new_cle_out->set_req(0, new_cle);
2546 new_cl->set_req(LoopNode::EntryControl, new_l);
2547
2548 _igvn.register_new_node_with_optimizer(new_l);
2549 _igvn.register_new_node_with_optimizer(new_tail);
2550 _igvn.register_new_node_with_optimizer(new_le);
2551 } else {
2552 Node *newhead = old_new[loop->_head->_idx];
2553 newhead->as_Loop()->clear_strip_mined();
2554 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2555 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2556 }
2557 // Look at data node that were assigned a control in the outer
2558 // loop: they are kept in the outer loop by the safepoint so start
2559 // from the safepoint node's inputs.
2560 IdealLoopTree* outer_loop = get_loop(l);
2561 Node_Stack stack(2);
2562 stack.push(sfpt, 1);
2563 uint new_counter = C->unique();
2564 while (stack.size() > 0) {
2565 Node* n = stack.node();
2566 uint i = stack.index();
2567 while (i < n->req() &&
2568 (n->in(i) == nullptr ||
2569 !has_ctrl(n->in(i)) ||
2570 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2571 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2572 i++;
2573 }
2574 if (i < n->req()) {
2575 stack.set_index(i+1);
2576 stack.push(n->in(i), 0);
2577 } else {
2578 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2579 Node* m = n == sfpt ? new_sfpt : n->clone();
2580 if (m != nullptr) {
2581 for (uint i = 0; i < n->req(); i++) {
2582 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2583 m->set_req(i, old_new[m->in(i)->_idx]);
2584 }
2585 }
2586 } else {
2587 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2588 }
2589 if (n != sfpt) {
2590 extra_data_nodes.push(n);
2591 _igvn.register_new_node_with_optimizer(m);
2592 assert(get_ctrl(n) == cle_out, "what other control?");
2593 set_ctrl(m, new_cle_out);
2594 old_new.map(n->_idx, m);
2595 }
2596 stack.pop();
2597 }
2598 }
2599 if (mode == CloneIncludesStripMined) {
2600 _igvn.register_new_node_with_optimizer(new_sfpt);
2601 _igvn.register_new_node_with_optimizer(new_cle_out);
2602 }
2603 // Some other transformation may have pessimistically assigned some
2604 // data nodes to the outer loop. Set their control so they are out
2605 // of the outer loop.
2606 ResourceMark rm;
2607 Unique_Node_List wq;
2608 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2609 Node* old = extra_data_nodes.at(i);
2610 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2611 }
2612
2613 for (uint i = 0; i < loop->_body.size(); i++) {
2614 Node* old = loop->_body.at(i);
2615 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2616 }
2617
2618 Node* inner_out = sfpt->in(0);
2619 if (inner_out->outcnt() > 1) {
2620 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2621 }
2622
2623 Node* new_ctrl = cl->outer_loop_exit();
2624 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2625 for (uint i = 0; i < wq.size(); i++) {
2626 Node* n = wq.at(i);
2627 set_ctrl(n, new_ctrl);
2628 if (n->in(0) != nullptr) {
2629 _igvn.replace_input_of(n, 0, new_ctrl);
2630 }
2631 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2632 }
2633 } else {
2634 Node *newhead = old_new[loop->_head->_idx];
2635 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2636 }
2637 }
2638
2639 //------------------------------clone_loop-------------------------------------
2640 //
2641 // C L O N E A L O O P B O D Y
2642 //
2643 // This is the basic building block of the loop optimizations. It clones an
2644 // entire loop body. It makes an old_new loop body mapping; with this mapping
2645 // you can find the new-loop equivalent to an old-loop node. All new-loop
2646 // nodes are exactly equal to their old-loop counterparts, all edges are the
2647 // same. All exits from the old-loop now have a RegionNode that merges the
2648 // equivalent new-loop path. This is true even for the normal "loop-exit"
2649 // condition. All uses of loop-invariant old-loop values now come from (one
2650 // or more) Phis that merge their new-loop equivalents.
2651 //
2652 // This operation leaves the graph in an illegal state: there are two valid
2653 // control edges coming from the loop pre-header to both loop bodies. I'll
2654 // definitely have to hack the graph after running this transform.
2655 //
2656 // From this building block I will further edit edges to perform loop peeling
2657 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2658 //
2659 // Parameter side_by_size_idom:
2660 // When side_by_size_idom is null, the dominator tree is constructed for
2661 // the clone loop to dominate the original. Used in construction of
2662 // pre-main-post loop sequence.
2663 // When nonnull, the clone and original are side-by-side, both are
2664 // dominated by the side_by_side_idom node. Used in construction of
2665 // unswitched loops.
2666 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2667 CloneLoopMode mode, Node* side_by_side_idom) {
2668
2669 LoopNode* head = loop->_head->as_Loop();
2670 head->verify_strip_mined(1);
2671
2672 if (C->do_vector_loop() && PrintOpto) {
2673 const char* mname = C->method()->name()->as_quoted_ascii();
2674 if (mname != nullptr) {
2675 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2676 }
2677 }
2678
2679 CloneMap& cm = C->clone_map();
2680 if (C->do_vector_loop()) {
2681 cm.set_clone_idx(cm.max_gen()+1);
2682 #ifndef PRODUCT
2683 if (PrintOpto) {
2684 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2685 loop->dump_head();
2686 }
2687 #endif
2688 }
2689
2690 // Step 1: Clone the loop body. Make the old->new mapping.
2691 clone_loop_body(loop->_body, old_new, &cm);
2692
2693 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2694
2695 // Step 2: Fix the edges in the new body. If the old input is outside the
2696 // loop use it. If the old input is INside the loop, use the corresponding
2697 // new node instead.
2698 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2699
2700 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2701 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2702
2703 // Step 3: Now fix control uses. Loop varying control uses have already
2704 // been fixed up (as part of all input edges in Step 2). Loop invariant
2705 // control uses must be either an IfFalse or an IfTrue. Make a merge
2706 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2707 // refer to this.
2708 Node_List worklist;
2709 uint new_counter = C->unique();
2710 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2711
2712 // Step 4: If loop-invariant use is not control, it must be dominated by a
2713 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2714 // there if needed. Make a Phi there merging old and new used values.
2715 Node_List *split_if_set = nullptr;
2716 Node_List *split_bool_set = nullptr;
2717 Node_List *split_cex_set = nullptr;
2718 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2719
2720 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2721 Node* old = extra_data_nodes.at(i);
2722 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2723 split_bool_set, split_cex_set, worklist, new_counter,
2724 mode);
2725 }
2726
2727 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2728 // the loop uses a condition set in the loop. The original IF probably
2729 // takes control from one or more OLD Regions (which in turn get from NEW
2730 // Regions). In any case, there will be a set of Phis for each merge point
2731 // from the IF up to where the original BOOL def exists the loop.
2732 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2733
2734 }
2735
2736 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2737 if (split_if_set) {
2738 while (split_if_set->size()) {
2739 Node *iff = split_if_set->pop();
2740 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2741 if (iff->in(input)->is_Phi()) {
2742 Node *b = clone_iff(iff->in(input)->as_Phi());
2743 _igvn.replace_input_of(iff, input, b);
2744 }
2745 }
2746 }
2747 if (split_bool_set) {
2748 while (split_bool_set->size()) {
2749 Node *b = split_bool_set->pop();
2750 Node *phi = b->in(1);
2751 assert(phi->is_Phi(), "");
2752 CmpNode *cmp = clone_bool((PhiNode*) phi);
2753 _igvn.replace_input_of(b, 1, cmp);
2754 }
2755 }
2756 if (split_cex_set) {
2757 while (split_cex_set->size()) {
2758 Node *b = split_cex_set->pop();
2759 assert(b->in(0)->is_Region(), "");
2760 assert(b->in(1)->is_Phi(), "");
2761 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2762 split_up(b, b->in(0), nullptr);
2763 }
2764 }
2765 }
2766
2767 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2768 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2769 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2770 for(uint i = 0; i < body.size(); i++ ) {
2771 Node* old = body.at(i);
2772 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2773 split_bool_set, split_cex_set, worklist, new_counter,
2774 mode);
2775 }
2776 }
2777
2778 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2779 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2780 LoopNode* head = loop->_head->as_Loop();
2781 for(uint i = 0; i < body.size(); i++ ) {
2782 Node* old = body.at(i);
2783 if( !old->is_CFG() ) continue;
2784
2785 // Copy uses to a worklist, so I can munge the def-use info
2786 // with impunity.
2787 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2788 worklist.push(old->fast_out(j));
2789 }
2790
2791 while (worklist.size()) { // Visit all uses
2792 Node *use = worklist.pop();
2793 if (!has_node(use)) continue; // Ignore dead nodes
2794 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2795 if (!loop->is_member(use_loop) && use->is_CFG()) {
2796 // Both OLD and USE are CFG nodes here.
2797 assert(use->is_Proj(), "" );
2798 Node* nnn = old_new[old->_idx];
2799
2800 Node* newuse = nullptr;
2801 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2802 CountedLoopNode* cl = head->as_CountedLoop();
2803 CountedLoopEndNode* cle = cl->loopexit();
2804 Node* cle_out = cle->proj_out_or_null(false);
2805 if (use == cle_out) {
2806 IfNode* le = cl->outer_loop_end();
2807 use = le->proj_out(false);
2808 use_loop = get_loop(use);
2809 if (mode == CloneIncludesStripMined) {
2810 nnn = old_new[le->_idx];
2811 } else {
2812 newuse = old_new[cle_out->_idx];
2813 }
2814 }
2815 }
2816 if (newuse == nullptr) {
2817 newuse = use->clone();
2818 }
2819
2820 // Clone the loop exit control projection
2821 if (C->do_vector_loop() && cm != nullptr) {
2822 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2823 }
2824 newuse->set_req(0,nnn);
2825 _igvn.register_new_node_with_optimizer(newuse);
2826 set_loop(newuse, use_loop);
2827 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2828
2829 // We need a Region to merge the exit from the peeled body and the
2830 // exit from the old loop body.
2831 RegionNode *r = new RegionNode(3);
2832 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2833 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2834
2835 // The original user of 'use' uses 'r' instead.
2836 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2837 Node* useuse = use->last_out(l);
2838 _igvn.rehash_node_delayed(useuse);
2839 uint uses_found = 0;
2840 if (useuse->in(0) == use) {
2841 useuse->set_req(0, r);
2842 uses_found++;
2843 if (useuse->is_CFG()) {
2844 // This is not a dom_depth > dd_r because when new
2845 // control flow is constructed by a loop opt, a node and
2846 // its dominator can end up at the same dom_depth
2847 assert(dom_depth(useuse) >= dd_r, "");
2848 set_idom(useuse, r, dom_depth(useuse));
2849 }
2850 }
2851 for (uint k = 1; k < useuse->req(); k++) {
2852 if( useuse->in(k) == use ) {
2853 useuse->set_req(k, r);
2854 uses_found++;
2855 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2856 // This is not a dom_depth > dd_r because when new
2857 // control flow is constructed by a loop opt, a node
2858 // and its dominator can end up at the same dom_depth
2859 assert(dom_depth(useuse) >= dd_r , "");
2860 set_idom(useuse, r, dom_depth(useuse));
2861 }
2862 }
2863 }
2864 l -= uses_found; // we deleted 1 or more copies of this edge
2865 }
2866
2867 assert(use->is_Proj(), "loop exit should be projection");
2868 // lazy_replace() below moves all nodes that are:
2869 // - control dependent on the loop exit or
2870 // - have control set to the loop exit
2871 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2872 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2873 // becomes dead and is "replaced" by the region.
2874 Node* use_clone = use->clone();
2875 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2876 // Now finish up 'r'
2877 r->set_req(1, newuse);
2878 r->set_req(2, use_clone);
2879 _igvn.register_new_node_with_optimizer(r);
2880 set_loop(r, use_loop);
2881 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2882 lazy_replace(use, r);
2883 // Map the (cloned) old use to the new merge point
2884 old_new.map(use_clone->_idx, r);
2885 } // End of if a loop-exit test
2886 }
2887 }
2888 }
2889
2890 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2891 IdealLoopTree* parent, bool partial) {
2892 for(uint i = 0; i < body.size(); i++ ) {
2893 Node *old = body.at(i);
2894 Node *nnn = old_new[old->_idx];
2895 // Fix CFG/Loop controlling the new node
2896 if (has_ctrl(old)) {
2897 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2898 } else {
2899 set_loop(nnn, parent);
2900 if (old->outcnt() > 0) {
2901 Node* dom = idom(old);
2902 if (old_new[dom->_idx] != nullptr) {
2903 dom = old_new[dom->_idx];
2904 set_idom(nnn, dom, dd );
2905 }
2906 }
2907 }
2908 // Correct edges to the new node
2909 for (uint j = 0; j < nnn->req(); j++) {
2910 Node *n = nnn->in(j);
2911 if (n != nullptr) {
2912 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2913 if (loop->is_member(old_in_loop)) {
2914 if (old_new[n->_idx] != nullptr) {
2915 nnn->set_req(j, old_new[n->_idx]);
2916 } else {
2917 assert(!body.contains(n), "");
2918 assert(partial, "node not cloned");
2919 }
2920 }
2921 }
2922 }
2923 _igvn.hash_find_insert(nnn);
2924 }
2925 }
2926
2927 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2928 for (uint i = 0; i < body.size(); i++) {
2929 Node* old = body.at(i);
2930 Node* nnn = old->clone();
2931 old_new.map(old->_idx, nnn);
2932 if (C->do_vector_loop() && cm != nullptr) {
2933 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2934 }
2935 _igvn.register_new_node_with_optimizer(nnn);
2936 }
2937 }
2938
2939
2940 //---------------------- stride_of_possible_iv -------------------------------------
2941 // Looks for an iff/bool/comp with one operand of the compare
2942 // being a cycle involving an add and a phi,
2943 // with an optional truncation (left-shift followed by a right-shift)
2944 // of the add. Returns zero if not an iv.
2945 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2946 Node* trunc1 = nullptr;
2947 Node* trunc2 = nullptr;
2948 const TypeInteger* ttype = nullptr;
2949 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2950 return 0;
2951 }
2952 BoolNode* bl = iff->in(1)->as_Bool();
2953 Node* cmp = bl->in(1);
2954 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2955 return 0;
2956 }
2957 // Must have an invariant operand
2958 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2959 return 0;
2960 }
2961 Node* add2 = nullptr;
2962 Node* cmp1 = cmp->in(1);
2963 if (cmp1->is_Phi()) {
2964 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2965 Node* phi = cmp1;
2966 for (uint i = 1; i < phi->req(); i++) {
2967 Node* in = phi->in(i);
2968 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2969 &trunc1, &trunc2, &ttype, T_INT);
2970 if (add && add->in(1) == phi) {
2971 add2 = add->in(2);
2972 break;
2973 }
2974 }
2975 } else {
2976 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2977 Node* addtrunc = cmp1;
2978 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2979 &trunc1, &trunc2, &ttype, T_INT);
2980 if (add && add->in(1)->is_Phi()) {
2981 Node* phi = add->in(1);
2982 for (uint i = 1; i < phi->req(); i++) {
2983 if (phi->in(i) == addtrunc) {
2984 add2 = add->in(2);
2985 break;
2986 }
2987 }
2988 }
2989 }
2990 if (add2 != nullptr) {
2991 const TypeInt* add2t = _igvn.type(add2)->is_int();
2992 if (add2t->is_con()) {
2993 return add2t->get_con();
2994 }
2995 }
2996 return 0;
2997 }
2998
2999
3000 //---------------------- stay_in_loop -------------------------------------
3001 // Return the (unique) control output node that's in the loop (if it exists.)
3002 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
3003 Node* unique = nullptr;
3004 if (!n) return nullptr;
3005 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3006 Node* use = n->fast_out(i);
3007 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
3008 if (unique != nullptr) {
3009 return nullptr;
3010 }
3011 unique = use;
3012 }
3013 }
3014 return unique;
3015 }
3016
3017 //------------------------------ register_node -------------------------------------
3018 // Utility to register node "n" with PhaseIdealLoop
3019 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
3020 _igvn.register_new_node_with_optimizer(n);
3021 loop->_body.push(n);
3022 if (n->is_CFG()) {
3023 set_loop(n, loop);
3024 set_idom(n, pred, ddepth);
3025 } else {
3026 set_ctrl(n, pred);
3027 }
3028 }
3029
3030 //------------------------------ proj_clone -------------------------------------
3031 // Utility to create an if-projection
3032 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
3033 ProjNode* c = p->clone()->as_Proj();
3034 c->set_req(0, iff);
3035 return c;
3036 }
3037
3038 //------------------------------ short_circuit_if -------------------------------------
3039 // Force the iff control output to be the live_proj
3040 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
3041 guarantee(live_proj != nullptr, "null projection");
3042 int proj_con = live_proj->_con;
3043 assert(proj_con == 0 || proj_con == 1, "false or true projection");
3044 Node *con = _igvn.intcon(proj_con);
3045 set_ctrl(con, C->root());
3046 if (iff) {
3047 iff->set_req(1, con);
3048 }
3049 return con;
3050 }
3051
3052 //------------------------------ insert_if_before_proj -------------------------------------
3053 // Insert a new if before an if projection (* - new node)
3054 //
3055 // before
3056 // if(test)
3057 // / \
3058 // v v
3059 // other-proj proj (arg)
3060 //
3061 // after
3062 // if(test)
3063 // / \
3064 // / v
3065 // | * proj-clone
3066 // v |
3067 // other-proj v
3068 // * new_if(relop(cmp[IU](left,right)))
3069 // / \
3070 // v v
3071 // * new-proj proj
3072 // (returned)
3073 //
3074 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
3075 IfNode* iff = proj->in(0)->as_If();
3076 IdealLoopTree *loop = get_loop(proj);
3077 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3078 uint ddepth = dom_depth(proj);
3079
3080 _igvn.rehash_node_delayed(iff);
3081 _igvn.rehash_node_delayed(proj);
3082
3083 proj->set_req(0, nullptr); // temporary disconnect
3084 ProjNode* proj2 = proj_clone(proj, iff);
3085 register_node(proj2, loop, iff, ddepth);
3086
3087 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
3088 register_node(cmp, loop, proj2, ddepth);
3089
3090 BoolNode* bol = new BoolNode(cmp, relop);
3091 register_node(bol, loop, proj2, ddepth);
3092
3093 int opcode = iff->Opcode();
3094 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
3095 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
3096 register_node(new_if, loop, proj2, ddepth);
3097
3098 proj->set_req(0, new_if); // reattach
3099 set_idom(proj, new_if, ddepth);
3100
3101 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
3102 guarantee(new_exit != nullptr, "null exit node");
3103 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
3104
3105 return new_exit;
3106 }
3107
3108 //------------------------------ insert_region_before_proj -------------------------------------
3109 // Insert a region before an if projection (* - new node)
3110 //
3111 // before
3112 // if(test)
3113 // / |
3114 // v |
3115 // proj v
3116 // other-proj
3117 //
3118 // after
3119 // if(test)
3120 // / |
3121 // v |
3122 // * proj-clone v
3123 // | other-proj
3124 // v
3125 // * new-region
3126 // |
3127 // v
3128 // * dum_if
3129 // / \
3130 // v \
3131 // * dum-proj v
3132 // proj
3133 //
3134 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3135 IfNode* iff = proj->in(0)->as_If();
3136 IdealLoopTree *loop = get_loop(proj);
3137 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3138 uint ddepth = dom_depth(proj);
3139
3140 _igvn.rehash_node_delayed(iff);
3141 _igvn.rehash_node_delayed(proj);
3142
3143 proj->set_req(0, nullptr); // temporary disconnect
3144 ProjNode* proj2 = proj_clone(proj, iff);
3145 register_node(proj2, loop, iff, ddepth);
3146
3147 RegionNode* reg = new RegionNode(2);
3148 reg->set_req(1, proj2);
3149 register_node(reg, loop, iff, ddepth);
3150
3151 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3152 register_node(dum_if, loop, reg, ddepth);
3153
3154 proj->set_req(0, dum_if); // reattach
3155 set_idom(proj, dum_if, ddepth);
3156
3157 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3158 register_node(dum_proj, loop, dum_if, ddepth);
3159
3160 return reg;
3161 }
3162
3163 //------------------------------ insert_cmpi_loop_exit -------------------------------------
3164 // Clone a signed compare loop exit from an unsigned compare and
3165 // insert it before the unsigned cmp on the stay-in-loop path.
3166 // All new nodes inserted in the dominator tree between the original
3167 // if and it's projections. The original if test is replaced with
3168 // a constant to force the stay-in-loop path.
3169 //
3170 // This is done to make sure that the original if and it's projections
3171 // still dominate the same set of control nodes, that the ctrl() relation
3172 // from data nodes to them is preserved, and that their loop nesting is
3173 // preserved.
3174 //
3175 // before
3176 // if(i <u limit) unsigned compare loop exit
3177 // / |
3178 // v v
3179 // exit-proj stay-in-loop-proj
3180 //
3181 // after
3182 // if(stay-in-loop-const) original if
3183 // / |
3184 // / v
3185 // / if(i < limit) new signed test
3186 // / / |
3187 // / / v
3188 // / / if(i <u limit) new cloned unsigned test
3189 // / / / |
3190 // v v v |
3191 // region |
3192 // | |
3193 // dum-if |
3194 // / | |
3195 // ether | |
3196 // v v
3197 // exit-proj stay-in-loop-proj
3198 //
3199 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
3200 const bool Signed = true;
3201 const bool Unsigned = false;
3202
3203 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3204 if (bol->_test._test != BoolTest::lt) return nullptr;
3205 CmpNode* cmpu = bol->in(1)->as_Cmp();
3206 if (cmpu->Opcode() != Op_CmpU) return nullptr;
3207 int stride = stride_of_possible_iv(if_cmpu);
3208 if (stride == 0) return nullptr;
3209
3210 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3211 guarantee(lp_proj != nullptr, "null loop node");
3212
3213 ProjNode* lp_continue = lp_proj->as_Proj();
3214 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3215 if (!lp_exit->is_IfFalse()) {
3216 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3217 // We therefore can't add a single exit condition.
3218 return nullptr;
3219 }
3220 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
3221 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
3222 Node* limit = nullptr;
3223 if (stride > 0) {
3224 limit = cmpu->in(2);
3225 } else {
3226 limit = _igvn.makecon(TypeInt::ZERO);
3227 set_ctrl(limit, C->root());
3228 }
3229 // Create a new region on the exit path
3230 RegionNode* reg = insert_region_before_proj(lp_exit);
3231 guarantee(reg != nullptr, "null region node");
3232
3233 // Clone the if-cmpu-true-false using a signed compare
3234 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3235 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
3236 reg->add_req(cmpi_exit);
3237
3238 // Clone the if-cmpu-true-false
3239 BoolTest::mask rel_u = bol->_test._test;
3240 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3241 reg->add_req(cmpu_exit);
3242
3243 // Force original if to stay in loop.
3244 short_circuit_if(if_cmpu, lp_continue);
3245
3246 return cmpi_exit->in(0)->as_If();
3247 }
3248
3249 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3250 // Remove a previously inserted signed compare loop exit.
3251 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3252 Node* lp_proj = stay_in_loop(if_cmp, loop);
3253 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3254 stay_in_loop(lp_proj, loop)->is_If() &&
3255 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3256 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3257 set_ctrl(con, C->root());
3258 if_cmp->set_req(1, con);
3259 }
3260
3261 //------------------------------ scheduled_nodelist -------------------------------------
3262 // Create a post order schedule of nodes that are in the
3263 // "member" set. The list is returned in "sched".
3264 // The first node in "sched" is the loop head, followed by
3265 // nodes which have no inputs in the "member" set, and then
3266 // followed by the nodes that have an immediate input dependence
3267 // on a node in "sched".
3268 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3269
3270 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3271 VectorSet visited;
3272 Node_Stack nstack(loop->_body.size());
3273
3274 Node* n = loop->_head; // top of stack is cached in "n"
3275 uint idx = 0;
3276 visited.set(n->_idx);
3277
3278 // Initially push all with no inputs from within member set
3279 for(uint i = 0; i < loop->_body.size(); i++ ) {
3280 Node *elt = loop->_body.at(i);
3281 if (member.test(elt->_idx)) {
3282 bool found = false;
3283 for (uint j = 0; j < elt->req(); j++) {
3284 Node* def = elt->in(j);
3285 if (def && member.test(def->_idx) && def != elt) {
3286 found = true;
3287 break;
3288 }
3289 }
3290 if (!found && elt != loop->_head) {
3291 nstack.push(n, idx);
3292 n = elt;
3293 assert(!visited.test(n->_idx), "not seen yet");
3294 visited.set(n->_idx);
3295 }
3296 }
3297 }
3298
3299 // traverse out's that are in the member set
3300 while (true) {
3301 if (idx < n->outcnt()) {
3302 Node* use = n->raw_out(idx);
3303 idx++;
3304 if (!visited.test_set(use->_idx)) {
3305 if (member.test(use->_idx)) {
3306 nstack.push(n, idx);
3307 n = use;
3308 idx = 0;
3309 }
3310 }
3311 } else {
3312 // All outputs processed
3313 sched.push(n);
3314 if (nstack.is_empty()) break;
3315 n = nstack.node();
3316 idx = nstack.index();
3317 nstack.pop();
3318 }
3319 }
3320 }
3321
3322
3323 //------------------------------ has_use_in_set -------------------------------------
3324 // Has a use in the vector set
3325 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3326 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3327 Node* use = n->fast_out(j);
3328 if (vset.test(use->_idx)) {
3329 return true;
3330 }
3331 }
3332 return false;
3333 }
3334
3335
3336 //------------------------------ has_use_internal_to_set -------------------------------------
3337 // Has use internal to the vector set (ie. not in a phi at the loop head)
3338 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3339 Node* head = loop->_head;
3340 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3341 Node* use = n->fast_out(j);
3342 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3343 return true;
3344 }
3345 }
3346 return false;
3347 }
3348
3349
3350 //------------------------------ clone_for_use_outside_loop -------------------------------------
3351 // clone "n" for uses that are outside of loop
3352 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3353 int cloned = 0;
3354 assert(worklist.size() == 0, "should be empty");
3355 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3356 Node* use = n->fast_out(j);
3357 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3358 worklist.push(use);
3359 }
3360 }
3361
3362 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3363 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3364 return -1;
3365 }
3366
3367 while( worklist.size() ) {
3368 Node *use = worklist.pop();
3369 if (!has_node(use) || use->in(0) == C->top()) continue;
3370 uint j;
3371 for (j = 0; j < use->req(); j++) {
3372 if (use->in(j) == n) break;
3373 }
3374 assert(j < use->req(), "must be there");
3375
3376 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3377 Node* n_clone = n->clone();
3378 _igvn.replace_input_of(use, j, n_clone);
3379 cloned++;
3380 Node* use_c;
3381 if (!use->is_Phi()) {
3382 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3383 } else {
3384 // Use in a phi is considered a use in the associated predecessor block
3385 use_c = use->in(0)->in(j);
3386 }
3387 set_ctrl(n_clone, use_c);
3388 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3389 get_loop(use_c)->_body.push(n_clone);
3390 _igvn.register_new_node_with_optimizer(n_clone);
3391 #ifndef PRODUCT
3392 if (TracePartialPeeling) {
3393 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3394 }
3395 #endif
3396 }
3397 return cloned;
3398 }
3399
3400
3401 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3402 // clone "n" for special uses that are in the not_peeled region.
3403 // If these def-uses occur in separate blocks, the code generator
3404 // marks the method as not compilable. For example, if a "BoolNode"
3405 // is in a different basic block than the "IfNode" that uses it, then
3406 // the compilation is aborted in the code generator.
3407 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3408 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3409 if (n->is_Phi() || n->is_Load()) {
3410 return;
3411 }
3412 assert(worklist.size() == 0, "should be empty");
3413 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3414 Node* use = n->fast_out(j);
3415 if ( not_peel.test(use->_idx) &&
3416 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3417 use->in(1) == n) {
3418 worklist.push(use);
3419 }
3420 }
3421 if (worklist.size() > 0) {
3422 // clone "n" and insert it between inputs of "n" and the use
3423 Node* n_clone = n->clone();
3424 loop->_body.push(n_clone);
3425 _igvn.register_new_node_with_optimizer(n_clone);
3426 set_ctrl(n_clone, get_ctrl(n));
3427 sink_list.push(n_clone);
3428 not_peel.set(n_clone->_idx);
3429 #ifndef PRODUCT
3430 if (TracePartialPeeling) {
3431 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3432 }
3433 #endif
3434 while( worklist.size() ) {
3435 Node *use = worklist.pop();
3436 _igvn.rehash_node_delayed(use);
3437 for (uint j = 1; j < use->req(); j++) {
3438 if (use->in(j) == n) {
3439 use->set_req(j, n_clone);
3440 }
3441 }
3442 }
3443 }
3444 }
3445
3446
3447 //------------------------------ insert_phi_for_loop -------------------------------------
3448 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3449 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3450 Node *phi = PhiNode::make(lp, back_edge_val);
3451 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3452 // Use existing phi if it already exists
3453 Node *hit = _igvn.hash_find_insert(phi);
3454 if( hit == nullptr ) {
3455 _igvn.register_new_node_with_optimizer(phi);
3456 set_ctrl(phi, lp);
3457 } else {
3458 // Remove the new phi from the graph and use the hit
3459 _igvn.remove_dead_node(phi);
3460 phi = hit;
3461 }
3462 _igvn.replace_input_of(use, idx, phi);
3463 }
3464
3465 #ifdef ASSERT
3466 //------------------------------ is_valid_loop_partition -------------------------------------
3467 // Validate the loop partition sets: peel and not_peel
3468 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3469 VectorSet& not_peel ) {
3470 uint i;
3471 // Check that peel_list entries are in the peel set
3472 for (i = 0; i < peel_list.size(); i++) {
3473 if (!peel.test(peel_list.at(i)->_idx)) {
3474 return false;
3475 }
3476 }
3477 // Check at loop members are in one of peel set or not_peel set
3478 for (i = 0; i < loop->_body.size(); i++ ) {
3479 Node *def = loop->_body.at(i);
3480 uint di = def->_idx;
3481 // Check that peel set elements are in peel_list
3482 if (peel.test(di)) {
3483 if (not_peel.test(di)) {
3484 return false;
3485 }
3486 // Must be in peel_list also
3487 bool found = false;
3488 for (uint j = 0; j < peel_list.size(); j++) {
3489 if (peel_list.at(j)->_idx == di) {
3490 found = true;
3491 break;
3492 }
3493 }
3494 if (!found) {
3495 return false;
3496 }
3497 } else if (not_peel.test(di)) {
3498 if (peel.test(di)) {
3499 return false;
3500 }
3501 } else {
3502 return false;
3503 }
3504 }
3505 return true;
3506 }
3507
3508 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3509 // Ensure a use outside of loop is of the right form
3510 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3511 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3512 return (use->is_Phi() &&
3513 use_c->is_Region() && use_c->req() == 3 &&
3514 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3515 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3516 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3517 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3518 }
3519
3520 //------------------------------ is_valid_clone_loop_form -------------------------------------
3521 // Ensure that all uses outside of loop are of the right form
3522 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3523 uint orig_exit_idx, uint clone_exit_idx) {
3524 uint len = peel_list.size();
3525 for (uint i = 0; i < len; i++) {
3526 Node *def = peel_list.at(i);
3527
3528 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3529 Node *use = def->fast_out(j);
3530 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3531 if (!loop->is_member(get_loop(use_c))) {
3532 // use is not in the loop, check for correct structure
3533 if (use->in(0) == def) {
3534 // Okay
3535 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3536 return false;
3537 }
3538 }
3539 }
3540 }
3541 return true;
3542 }
3543 #endif
3544
3545 //------------------------------ partial_peel -------------------------------------
3546 // Partially peel (aka loop rotation) the top portion of a loop (called
3547 // the peel section below) by cloning it and placing one copy just before
3548 // the new loop head and the other copy at the bottom of the new loop.
3549 //
3550 // before after where it came from
3551 //
3552 // stmt1 stmt1
3553 // loop: stmt2 clone
3554 // stmt2 if condA goto exitA clone
3555 // if condA goto exitA new_loop: new
3556 // stmt3 stmt3 clone
3557 // if !condB goto loop if condB goto exitB clone
3558 // exitB: stmt2 orig
3559 // stmt4 if !condA goto new_loop orig
3560 // exitA: goto exitA
3561 // exitB:
3562 // stmt4
3563 // exitA:
3564 //
3565 // Step 1: find the cut point: an exit test on probable
3566 // induction variable.
3567 // Step 2: schedule (with cloning) operations in the peel
3568 // section that can be executed after the cut into
3569 // the section that is not peeled. This may need
3570 // to clone operations into exit blocks. For
3571 // instance, a reference to A[i] in the not-peel
3572 // section and a reference to B[i] in an exit block
3573 // may cause a left-shift of i by 2 to be placed
3574 // in the peel block. This step will clone the left
3575 // shift into the exit block and sink the left shift
3576 // from the peel to the not-peel section.
3577 // Step 3: clone the loop, retarget the control, and insert
3578 // phis for values that are live across the new loop
3579 // head. This is very dependent on the graph structure
3580 // from clone_loop. It creates region nodes for
3581 // exit control and associated phi nodes for values
3582 // flow out of the loop through that exit. The region
3583 // node is dominated by the clone's control projection.
3584 // So the clone's peel section is placed before the
3585 // new loop head, and the clone's not-peel section is
3586 // forms the top part of the new loop. The original
3587 // peel section forms the tail of the new loop.
3588 // Step 4: update the dominator tree and recompute the
3589 // dominator depth.
3590 //
3591 // orig
3592 //
3593 // stmt1
3594 // |
3595 // v
3596 // predicates
3597 // |
3598 // v
3599 // loop<----+
3600 // | |
3601 // stmt2 |
3602 // | |
3603 // v |
3604 // ifA |
3605 // / | |
3606 // v v |
3607 // false true ^ <-- last_peel
3608 // / | |
3609 // / ===|==cut |
3610 // / stmt3 | <-- first_not_peel
3611 // / | |
3612 // | v |
3613 // v ifB |
3614 // exitA: / \ |
3615 // / \ |
3616 // v v |
3617 // false true |
3618 // / \ |
3619 // / ----+
3620 // |
3621 // v
3622 // exitB:
3623 // stmt4
3624 //
3625 //
3626 // after clone loop
3627 //
3628 // stmt1
3629 // |
3630 // v
3631 // predicates
3632 // / \
3633 // clone / \ orig
3634 // / \
3635 // / \
3636 // v v
3637 // +---->loop loop<----+
3638 // | | | |
3639 // | stmt2 stmt2 |
3640 // | | | |
3641 // | v v |
3642 // | ifA ifA |
3643 // | | \ / | |
3644 // | v v v v |
3645 // ^ true false false true ^ <-- last_peel
3646 // | | ^ \ / | |
3647 // | cut==|== \ \ / ===|==cut |
3648 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3649 // | | dom | | | |
3650 // | v \ 1v v2 v |
3651 // | ifB regionA ifB |
3652 // | / \ | / \ |
3653 // | / \ v / \ |
3654 // | v v exitA: v v |
3655 // | true false false true |
3656 // | / ^ \ / \ |
3657 // +---- \ \ / ----+
3658 // dom \ /
3659 // \ 1v v2
3660 // regionB
3661 // |
3662 // v
3663 // exitB:
3664 // stmt4
3665 //
3666 //
3667 // after partial peel
3668 //
3669 // stmt1
3670 // |
3671 // v
3672 // predicates
3673 // /
3674 // clone / orig
3675 // / TOP
3676 // / \
3677 // v v
3678 // TOP->loop loop----+
3679 // | | |
3680 // stmt2 stmt2 |
3681 // | | |
3682 // v v |
3683 // ifA ifA |
3684 // | \ / | |
3685 // v v v v |
3686 // true false false true | <-- last_peel
3687 // | ^ \ / +------|---+
3688 // +->newloop \ \ / === ==cut | |
3689 // | stmt3 \ \ / TOP | |
3690 // | | dom | | stmt3 | | <-- first_not_peel
3691 // | v \ 1v v2 v | |
3692 // | ifB regionA ifB ^ v
3693 // | / \ | / \ | |
3694 // | / \ v / \ | |
3695 // | v v exitA: v v | |
3696 // | true false false true | |
3697 // | / ^ \ / \ | |
3698 // | | \ \ / v | |
3699 // | | dom \ / TOP | |
3700 // | | \ 1v v2 | |
3701 // ^ v regionB | |
3702 // | | | | |
3703 // | | v ^ v
3704 // | | exitB: | |
3705 // | | stmt4 | |
3706 // | +------------>-----------------+ |
3707 // | |
3708 // +-----------------<---------------------+
3709 //
3710 //
3711 // final graph
3712 //
3713 // stmt1
3714 // |
3715 // v
3716 // predicates
3717 // |
3718 // v
3719 // stmt2 clone
3720 // |
3721 // v
3722 // ........> ifA clone
3723 // : / |
3724 // dom / |
3725 // : v v
3726 // : false true
3727 // : | |
3728 // : | v
3729 // : | newloop<-----+
3730 // : | | |
3731 // : | stmt3 clone |
3732 // : | | |
3733 // : | v |
3734 // : | ifB |
3735 // : | / \ |
3736 // : | v v |
3737 // : | false true |
3738 // : | | | |
3739 // : | v stmt2 |
3740 // : | exitB: | |
3741 // : | stmt4 v |
3742 // : | ifA orig |
3743 // : | / \ |
3744 // : | / \ |
3745 // : | v v |
3746 // : | false true |
3747 // : | / \ |
3748 // : v v -----+
3749 // RegionA
3750 // |
3751 // v
3752 // exitA
3753 //
3754 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3755
3756 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3757 if (!loop->_head->is_Loop()) {
3758 return false;
3759 }
3760 LoopNode *head = loop->_head->as_Loop();
3761
3762 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3763 return false;
3764 }
3765
3766 // Check for complex exit control
3767 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3768 Node *n = loop->_body.at(ii);
3769 int opc = n->Opcode();
3770 if (n->is_Call() ||
3771 opc == Op_Catch ||
3772 opc == Op_CatchProj ||
3773 opc == Op_Jump ||
3774 opc == Op_JumpProj) {
3775 #ifndef PRODUCT
3776 if (TracePartialPeeling) {
3777 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3778 }
3779 #endif
3780 return false;
3781 }
3782 }
3783
3784 int dd = dom_depth(head);
3785
3786 // Step 1: find cut point
3787
3788 // Walk up dominators to loop head looking for first loop exit
3789 // which is executed on every path thru loop.
3790 IfNode *peel_if = nullptr;
3791 IfNode *peel_if_cmpu = nullptr;
3792
3793 Node *iff = loop->tail();
3794 while (iff != head) {
3795 if (iff->is_If()) {
3796 Node *ctrl = get_ctrl(iff->in(1));
3797 if (ctrl->is_top()) return false; // Dead test on live IF.
3798 // If loop-varying exit-test, check for induction variable
3799 if (loop->is_member(get_loop(ctrl)) &&
3800 loop->is_loop_exit(iff) &&
3801 is_possible_iv_test(iff)) {
3802 Node* cmp = iff->in(1)->in(1);
3803 if (cmp->Opcode() == Op_CmpI) {
3804 peel_if = iff->as_If();
3805 } else {
3806 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3807 peel_if_cmpu = iff->as_If();
3808 }
3809 }
3810 }
3811 iff = idom(iff);
3812 }
3813
3814 // Prefer signed compare over unsigned compare.
3815 IfNode* new_peel_if = nullptr;
3816 if (peel_if == nullptr) {
3817 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3818 return false; // No peel point found
3819 }
3820 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3821 if (new_peel_if == nullptr) {
3822 return false; // No peel point found
3823 }
3824 peel_if = new_peel_if;
3825 }
3826 Node* last_peel = stay_in_loop(peel_if, loop);
3827 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3828 if (first_not_peeled == nullptr || first_not_peeled == head) {
3829 return false;
3830 }
3831
3832 #ifndef PRODUCT
3833 if (TraceLoopOpts) {
3834 tty->print("PartialPeel ");
3835 loop->dump_head();
3836 }
3837
3838 if (TracePartialPeeling) {
3839 tty->print_cr("before partial peel one iteration");
3840 Node_List wl;
3841 Node* t = head->in(2);
3842 while (true) {
3843 wl.push(t);
3844 if (t == head) break;
3845 t = idom(t);
3846 }
3847 while (wl.size() > 0) {
3848 Node* tt = wl.pop();
3849 tt->dump();
3850 if (tt == last_peel) tty->print_cr("-- cut --");
3851 }
3852 }
3853 #endif
3854
3855 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3856
3857 VectorSet peel;
3858 VectorSet not_peel;
3859 Node_List peel_list;
3860 Node_List worklist;
3861 Node_List sink_list;
3862
3863 uint estimate = loop->est_loop_clone_sz(1);
3864 if (exceeding_node_budget(estimate)) {
3865 return false;
3866 }
3867
3868 // Set of cfg nodes to peel are those that are executable from
3869 // the head through last_peel.
3870 assert(worklist.size() == 0, "should be empty");
3871 worklist.push(head);
3872 peel.set(head->_idx);
3873 while (worklist.size() > 0) {
3874 Node *n = worklist.pop();
3875 if (n != last_peel) {
3876 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3877 Node* use = n->fast_out(j);
3878 if (use->is_CFG() &&
3879 loop->is_member(get_loop(use)) &&
3880 !peel.test_set(use->_idx)) {
3881 worklist.push(use);
3882 }
3883 }
3884 }
3885 }
3886
3887 // Set of non-cfg nodes to peel are those that are control
3888 // dependent on the cfg nodes.
3889 for (uint i = 0; i < loop->_body.size(); i++) {
3890 Node *n = loop->_body.at(i);
3891 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3892 if (peel.test(n_c->_idx)) {
3893 peel.set(n->_idx);
3894 } else {
3895 not_peel.set(n->_idx);
3896 }
3897 }
3898
3899 // Step 2: move operations from the peeled section down into the
3900 // not-peeled section
3901
3902 // Get a post order schedule of nodes in the peel region
3903 // Result in right-most operand.
3904 scheduled_nodelist(loop, peel, peel_list);
3905
3906 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3907
3908 // For future check for too many new phis
3909 uint old_phi_cnt = 0;
3910 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3911 Node* use = head->fast_out(j);
3912 if (use->is_Phi()) old_phi_cnt++;
3913 }
3914
3915 #ifndef PRODUCT
3916 if (TracePartialPeeling) {
3917 tty->print_cr("\npeeled list");
3918 }
3919 #endif
3920
3921 // Evacuate nodes in peel region into the not_peeled region if possible
3922 bool too_many_clones = false;
3923 uint new_phi_cnt = 0;
3924 uint cloned_for_outside_use = 0;
3925 for (uint i = 0; i < peel_list.size();) {
3926 Node* n = peel_list.at(i);
3927 #ifndef PRODUCT
3928 if (TracePartialPeeling) n->dump();
3929 #endif
3930 bool incr = true;
3931 if (!n->is_CFG()) {
3932 if (has_use_in_set(n, not_peel)) {
3933 // If not used internal to the peeled region,
3934 // move "n" from peeled to not_peeled region.
3935 if (!has_use_internal_to_set(n, peel, loop)) {
3936 // if not pinned and not a load (which maybe anti-dependent on a store)
3937 // and not a CMove (Matcher expects only bool->cmove).
3938 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3939 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3940 if (C->failing()) return false;
3941 if (new_clones == -1) {
3942 too_many_clones = true;
3943 break;
3944 }
3945 cloned_for_outside_use += new_clones;
3946 sink_list.push(n);
3947 peel.remove(n->_idx);
3948 not_peel.set(n->_idx);
3949 peel_list.remove(i);
3950 incr = false;
3951 #ifndef PRODUCT
3952 if (TracePartialPeeling) {
3953 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3954 n->_idx, get_ctrl(n)->_idx);
3955 }
3956 #endif
3957 }
3958 } else {
3959 // Otherwise check for special def-use cases that span
3960 // the peel/not_peel boundary such as bool->if
3961 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3962 new_phi_cnt++;
3963 }
3964 }
3965 }
3966 if (incr) i++;
3967 }
3968
3969 estimate += cloned_for_outside_use + new_phi_cnt;
3970 bool exceed_node_budget = !may_require_nodes(estimate);
3971 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3972
3973 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3974 #ifndef PRODUCT
3975 if (TracePartialPeeling && exceed_phi_limit) {
3976 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3977 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3978 }
3979 #endif
3980 if (new_peel_if != nullptr) {
3981 remove_cmpi_loop_exit(new_peel_if, loop);
3982 }
3983 // Inhibit more partial peeling on this loop
3984 assert(!head->is_partial_peel_loop(), "not partial peeled");
3985 head->mark_partial_peel_failed();
3986 if (cloned_for_outside_use > 0) {
3987 // Terminate this round of loop opts because
3988 // the graph outside this loop was changed.
3989 C->set_major_progress();
3990 return true;
3991 }
3992 return false;
3993 }
3994
3995 // Step 3: clone loop, retarget control, and insert new phis
3996
3997 // Create new loop head for new phis and to hang
3998 // the nodes being moved (sinked) from the peel region.
3999 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4000 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4001 _igvn.register_new_node_with_optimizer(new_head);
4002 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4003 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4004 set_loop(new_head, loop);
4005 loop->_body.push(new_head);
4006 not_peel.set(new_head->_idx);
4007 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4008 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4009
4010 while (sink_list.size() > 0) {
4011 Node* n = sink_list.pop();
4012 set_ctrl(n, new_head);
4013 }
4014
4015 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4016
4017 clone_loop(loop, old_new, dd, IgnoreStripMined);
4018
4019 const uint clone_exit_idx = 1;
4020 const uint orig_exit_idx = 2;
4021 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4022
4023 Node* head_clone = old_new[head->_idx];
4024 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4025 Node* orig_tail_clone = head_clone->in(2);
4026
4027 // Add phi if "def" node is in peel set and "use" is not
4028
4029 for (uint i = 0; i < peel_list.size(); i++) {
4030 Node *def = peel_list.at(i);
4031 if (!def->is_CFG()) {
4032 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4033 Node *use = def->fast_out(j);
4034 if (has_node(use) && use->in(0) != C->top() &&
4035 (!peel.test(use->_idx) ||
4036 (use->is_Phi() && use->in(0) == head)) ) {
4037 worklist.push(use);
4038 }
4039 }
4040 while( worklist.size() ) {
4041 Node *use = worklist.pop();
4042 for (uint j = 1; j < use->req(); j++) {
4043 Node* n = use->in(j);
4044 if (n == def) {
4045
4046 // "def" is in peel set, "use" is not in peel set
4047 // or "use" is in the entry boundary (a phi) of the peel set
4048
4049 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4050
4051 if ( loop->is_member(get_loop( use_c )) ) {
4052 // use is in loop
4053 if (old_new[use->_idx] != nullptr) { // null for dead code
4054 Node* use_clone = old_new[use->_idx];
4055 _igvn.replace_input_of(use, j, C->top());
4056 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4057 }
4058 } else {
4059 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4060 // use is not in the loop, check if the live range includes the cut
4061 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4062 if (not_peel.test(lp_if->_idx)) {
4063 assert(j == orig_exit_idx, "use from original loop");
4064 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4065 }
4066 }
4067 }
4068 }
4069 }
4070 }
4071 }
4072
4073 // Step 3b: retarget control
4074
4075 // Redirect control to the new loop head if a cloned node in
4076 // the not_peeled region has control that points into the peeled region.
4077 // This necessary because the cloned peeled region will be outside
4078 // the loop.
4079 // from to
4080 // cloned-peeled <---+
4081 // new_head_clone: | <--+
4082 // cloned-not_peeled in(0) in(0)
4083 // orig-peeled
4084
4085 for (uint i = 0; i < loop->_body.size(); i++) {
4086 Node *n = loop->_body.at(i);
4087 if (!n->is_CFG() && n->in(0) != nullptr &&
4088 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4089 Node* n_clone = old_new[n->_idx];
4090 if (n_clone->depends_only_on_test()) {
4091 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4092 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4093 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4094 // floating above its range check.
4095 Node* pinned_clone = n_clone->pin_array_access_node();
4096 if (pinned_clone != nullptr) {
4097 register_new_node(pinned_clone, get_ctrl(n_clone));
4098 old_new.map(n->_idx, pinned_clone);
4099 _igvn.replace_node(n_clone, pinned_clone);
4100 n_clone = pinned_clone;
4101 }
4102 }
4103 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4104 }
4105 }
4106
4107 // Backedge of the surviving new_head (the clone) is original last_peel
4108 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4109
4110 // Cut first node in original not_peel set
4111 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4112 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4113 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4114
4115 // Copy head_clone back-branch info to original head
4116 // and remove original head's loop entry and
4117 // clone head's back-branch
4118 _igvn.rehash_node_delayed(head); // Multiple edge updates
4119 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4120 head->set_req(LoopNode::LoopBackControl, C->top());
4121 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4122
4123 // Similarly modify the phis
4124 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4125 Node* use = head->fast_out(k);
4126 if (use->is_Phi() && use->outcnt() > 0) {
4127 Node* use_clone = old_new[use->_idx];
4128 _igvn.rehash_node_delayed(use); // Multiple edge updates
4129 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4130 use->set_req(LoopNode::LoopBackControl, C->top());
4131 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4132 }
4133 }
4134
4135 // Step 4: update dominator tree and dominator depth
4136
4137 set_idom(head, orig_tail_clone, dd);
4138 recompute_dom_depth();
4139
4140 // Inhibit more partial peeling on this loop
4141 new_head_clone->set_partial_peel_loop();
4142 C->set_major_progress();
4143 loop->record_for_igvn();
4144
4145 #ifndef PRODUCT
4146 if (TracePartialPeeling) {
4147 tty->print_cr("\nafter partial peel one iteration");
4148 Node_List wl;
4149 Node* t = last_peel;
4150 while (true) {
4151 wl.push(t);
4152 if (t == head_clone) break;
4153 t = idom(t);
4154 }
4155 while (wl.size() > 0) {
4156 Node* tt = wl.pop();
4157 if (tt == head) tty->print_cr("orig head");
4158 else if (tt == new_head_clone) tty->print_cr("new head");
4159 else if (tt == head_clone) tty->print_cr("clone head");
4160 tt->dump();
4161 }
4162 }
4163 #endif
4164
4165 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4166
4167 return true;
4168 }
4169
4170 // Transform:
4171 //
4172 // loop<-----------------+
4173 // | |
4174 // stmt1 stmt2 .. stmtn |
4175 // | | | |
4176 // \ | / |
4177 // v v v |
4178 // region |
4179 // | |
4180 // shared_stmt |
4181 // | |
4182 // v |
4183 // if |
4184 // / \ |
4185 // | -----------+
4186 // v
4187 //
4188 // into:
4189 //
4190 // loop<-------------------+
4191 // | |
4192 // v |
4193 // +->loop |
4194 // | | |
4195 // | stmt1 stmt2 .. stmtn |
4196 // | | | | |
4197 // | | \ / |
4198 // | | v v |
4199 // | | region1 |
4200 // | | | |
4201 // | shared_stmt shared_stmt |
4202 // | | | |
4203 // | v v |
4204 // | if if |
4205 // | /\ / \ |
4206 // +-- | | -------+
4207 // \ /
4208 // v v
4209 // region2
4210 //
4211 // (region2 is shown to merge mirrored projections of the loop exit
4212 // ifs to make the diagram clearer but they really merge the same
4213 // projection)
4214 //
4215 // Conditions for this transformation to trigger:
4216 // - the path through stmt1 is frequent enough
4217 // - the inner loop will be turned into a counted loop after transformation
4218 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4219 if (!DuplicateBackedge) {
4220 return false;
4221 }
4222 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4223 if (!loop->_head->is_Loop()) {
4224 return false;
4225 }
4226
4227 uint estimate = loop->est_loop_clone_sz(1);
4228 if (exceeding_node_budget(estimate)) {
4229 return false;
4230 }
4231
4232 LoopNode *head = loop->_head->as_Loop();
4233
4234 Node* region = nullptr;
4235 IfNode* exit_test = nullptr;
4236 uint inner;
4237 float f;
4238 if (StressDuplicateBackedge) {
4239 if (head->is_strip_mined()) {
4240 return false;
4241 }
4242 Node* c = head->in(LoopNode::LoopBackControl);
4243
4244 while (c != head) {
4245 if (c->is_Region()) {
4246 region = c;
4247 }
4248 c = idom(c);
4249 }
4250
4251 if (region == nullptr) {
4252 return false;
4253 }
4254
4255 inner = 1;
4256 } else {
4257 // Is the shape of the loop that of a counted loop...
4258 Node* back_control = loop_exit_control(head, loop);
4259 if (back_control == nullptr) {
4260 return false;
4261 }
4262
4263 BoolTest::mask bt = BoolTest::illegal;
4264 float cl_prob = 0;
4265 Node* incr = nullptr;
4266 Node* limit = nullptr;
4267 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4268 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4269 return false;
4270 }
4271
4272 // With an extra phi for the candidate iv?
4273 // Or the region node is the loop head
4274 if (!incr->is_Phi() || incr->in(0) == head) {
4275 return false;
4276 }
4277
4278 PathFrequency pf(head, this);
4279 region = incr->in(0);
4280
4281 // Go over all paths for the extra phi's region and see if that
4282 // path is frequent enough and would match the expected iv shape
4283 // if the extra phi is removed
4284 inner = 0;
4285 for (uint i = 1; i < incr->req(); ++i) {
4286 Node* in = incr->in(i);
4287 Node* trunc1 = nullptr;
4288 Node* trunc2 = nullptr;
4289 const TypeInteger* iv_trunc_t = nullptr;
4290 Node* orig_in = in;
4291 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4292 continue;
4293 }
4294 assert(in->Opcode() == Op_AddI, "wrong increment code");
4295 Node* xphi = nullptr;
4296 Node* stride = loop_iv_stride(in, loop, xphi);
4297
4298 if (stride == nullptr) {
4299 continue;
4300 }
4301
4302 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4303 if (phi == nullptr ||
4304 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4305 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4306 return false;
4307 }
4308
4309 f = pf.to(region->in(i));
4310 if (f > 0.5) {
4311 inner = i;
4312 break;
4313 }
4314 }
4315
4316 if (inner == 0) {
4317 return false;
4318 }
4319
4320 exit_test = back_control->in(0)->as_If();
4321 }
4322
4323 if (idom(region)->is_Catch()) {
4324 return false;
4325 }
4326
4327 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4328 Unique_Node_List wq;
4329 wq.push(head->in(LoopNode::LoopBackControl));
4330 for (uint i = 0; i < wq.size(); i++) {
4331 Node* c = wq.at(i);
4332 assert(get_loop(c) == loop, "not in the right loop?");
4333 if (c->is_Region()) {
4334 if (c != region) {
4335 for (uint j = 1; j < c->req(); ++j) {
4336 wq.push(c->in(j));
4337 }
4338 }
4339 } else {
4340 wq.push(c->in(0));
4341 }
4342 assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
4343 }
4344
4345 Node* region_dom = idom(region);
4346
4347 // Can't do the transformation if this would cause a membar pair to
4348 // be split
4349 for (uint i = 0; i < wq.size(); i++) {
4350 Node* c = wq.at(i);
4351 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4352 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4353 if (!wq.member(c->as_MemBar()->leading_membar())) {
4354 return false;
4355 }
4356 }
4357 }
4358
4359 // Collect data nodes that need to be clones as well
4360 int dd = dom_depth(head);
4361
4362 for (uint i = 0; i < loop->_body.size(); ++i) {
4363 Node* n = loop->_body.at(i);
4364 if (has_ctrl(n)) {
4365 Node* c = get_ctrl(n);
4366 if (wq.member(c)) {
4367 wq.push(n);
4368 }
4369 } else {
4370 set_idom(n, idom(n), dd);
4371 }
4372 }
4373
4374 // clone shared_stmt
4375 clone_loop_body(wq, old_new, nullptr);
4376
4377 Node* region_clone = old_new[region->_idx];
4378 region_clone->set_req(inner, C->top());
4379 set_idom(region, region->in(inner), dd);
4380
4381 // Prepare the outer loop
4382 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4383 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4384 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4385 set_idom(head, outer_head, dd);
4386
4387 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4388
4389 // Make one of the shared_stmt copies only reachable from stmt1, the
4390 // other only from stmt2..stmtn.
4391 Node* dom = nullptr;
4392 for (uint i = 1; i < region->req(); ++i) {
4393 if (i != inner) {
4394 _igvn.replace_input_of(region, i, C->top());
4395 }
4396 Node* in = region_clone->in(i);
4397 if (in->is_top()) {
4398 continue;
4399 }
4400 if (dom == nullptr) {
4401 dom = in;
4402 } else {
4403 dom = dom_lca(dom, in);
4404 }
4405 }
4406
4407 set_idom(region_clone, dom, dd);
4408
4409 // Set up the outer loop
4410 for (uint i = 0; i < head->outcnt(); i++) {
4411 Node* u = head->raw_out(i);
4412 if (u->is_Phi()) {
4413 Node* outer_phi = u->clone();
4414 outer_phi->set_req(0, outer_head);
4415 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4416 if (backedge == nullptr) {
4417 backedge = u->in(LoopNode::LoopBackControl);
4418 }
4419 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4420 register_new_node(outer_phi, outer_head);
4421 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4422 }
4423 }
4424
4425 // create control and data nodes for out of loop uses (including region2)
4426 Node_List worklist;
4427 uint new_counter = C->unique();
4428 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4429
4430 Node_List *split_if_set = nullptr;
4431 Node_List *split_bool_set = nullptr;
4432 Node_List *split_cex_set = nullptr;
4433 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4434 split_if_set, split_bool_set, split_cex_set);
4435
4436 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4437
4438 if (exit_test != nullptr) {
4439 float cnt = exit_test->_fcnt;
4440 if (cnt != COUNT_UNKNOWN) {
4441 exit_test->_fcnt = cnt * f;
4442 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4443 }
4444 }
4445
4446 C->set_major_progress();
4447
4448 return true;
4449 }
4450
4451 // AutoVectorize the loop: replace scalar ops with vector ops.
4452 PhaseIdealLoop::AutoVectorizeStatus
4453 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4454 // Counted loop only
4455 if (!lpt->is_counted()) {
4456 return AutoVectorizeStatus::Impossible;
4457 }
4458
4459 // Main-loop only
4460 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4461 if (!cl->is_main_loop()) {
4462 return AutoVectorizeStatus::Impossible;
4463 }
4464
4465 VLoop vloop(lpt, false);
4466 if (!vloop.check_preconditions()) {
4467 return AutoVectorizeStatus::TriedAndFailed;
4468 }
4469
4470 // Ensure the shared data is cleared before each use
4471 vshared.clear();
4472
4473 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4474 if (!vloop_analyzer.success()) {
4475 return AutoVectorizeStatus::TriedAndFailed;
4476 }
4477
4478 SuperWord sw(vloop_analyzer);
4479 if (!sw.transform_loop()) {
4480 return AutoVectorizeStatus::TriedAndFailed;
4481 }
4482
4483 return AutoVectorizeStatus::Success;
4484 }
4485
4486 // Having ReductionNodes in the loop is expensive. They need to recursively
4487 // fold together the vector values, for every vectorized loop iteration. If
4488 // we encounter the following pattern, we can vector accumulate the values
4489 // inside the loop, and only have a single UnorderedReduction after the loop.
4490 //
4491 // CountedLoop init
4492 // | |
4493 // +------+ | +-----------------------+
4494 // | | | |
4495 // PhiNode (s) |
4496 // | |
4497 // | Vector |
4498 // | | |
4499 // UnorderedReduction (first_ur) |
4500 // | |
4501 // ... Vector |
4502 // | | |
4503 // UnorderedReduction (last_ur) |
4504 // | |
4505 // +---------------------+
4506 //
4507 // We patch the graph to look like this:
4508 //
4509 // CountedLoop identity_vector
4510 // | |
4511 // +-------+ | +---------------+
4512 // | | | |
4513 // PhiNode (v) |
4514 // | |
4515 // | Vector |
4516 // | | |
4517 // VectorAccumulator |
4518 // | |
4519 // ... Vector |
4520 // | | |
4521 // init VectorAccumulator |
4522 // | | | |
4523 // UnorderedReduction +-----------+
4524 //
4525 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4526 // use vector_accumulators, which do the same reductions, but only element
4527 // wise. This is a single operation per vector_accumulator, rather than many
4528 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4529 // after the loop, and also reduce the init value into it.
4530 // We can not do this with all reductions. Some reductions do not allow the
4531 // reordering of operations (for example float addition).
4532 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4533 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4534
4535 // Find all Phi nodes with UnorderedReduction on backedge.
4536 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4537 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4538 Node* phi = cl->fast_out(j);
4539 // We have a phi with a single use, and a UnorderedReduction on the backedge.
4540 if (!phi->is_Phi() || phi->outcnt() != 1 || !phi->in(2)->is_UnorderedReduction()) {
4541 continue;
4542 }
4543
4544 UnorderedReductionNode* last_ur = phi->in(2)->as_UnorderedReduction();
4545
4546 // Determine types
4547 const TypeVect* vec_t = last_ur->vect_type();
4548 uint vector_length = vec_t->length();
4549 BasicType bt = vec_t->element_basic_type();
4550 const Type* bt_t = Type::get_const_basic_type(bt);
4551
4552 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4553 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4554 const int vopc = VectorNode::opcode(sopc, bt);
4555 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4556 DEBUG_ONLY( last_ur->dump(); )
4557 assert(false, "do not have normal vector op for this reduction");
4558 continue; // not implemented -> fails
4559 }
4560
4561 // Traverse up the chain of UnorderedReductions, checking that it loops back to
4562 // the phi. Check that all UnorderedReductions only have a single use, except for
4563 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4564 // are outside the loop.
4565 UnorderedReductionNode* current = last_ur;
4566 UnorderedReductionNode* first_ur = nullptr;
4567 while (true) {
4568 assert(current->is_UnorderedReduction(), "sanity");
4569
4570 // Expect no ctrl and a vector_input from within the loop.
4571 Node* ctrl = current->in(0);
4572 Node* vector_input = current->in(2);
4573 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4574 DEBUG_ONLY( current->dump(1); )
4575 assert(false, "reduction has ctrl or bad vector_input");
4576 break; // Chain traversal fails.
4577 }
4578
4579 assert(current->vect_type() != nullptr, "must have vector type");
4580 if (current->vect_type() != last_ur->vect_type()) {
4581 // Reductions do not have the same vector type (length and element type).
4582 break; // Chain traversal fails.
4583 }
4584
4585 // Expect single use of UnorderedReduction, except for last_ur.
4586 if (current == last_ur) {
4587 // Expect all uses to be outside the loop, except phi.
4588 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4589 Node* use = current->fast_out(k);
4590 if (use != phi && ctrl_or_self(use) == cl) {
4591 DEBUG_ONLY( current->dump(-1); )
4592 assert(false, "reduction has use inside loop");
4593 // Should not be allowed by SuperWord::mark_reductions
4594 return; // bail out of optimization
4595 }
4596 }
4597 } else {
4598 if (current->outcnt() != 1) {
4599 break; // Chain traversal fails.
4600 }
4601 }
4602
4603 // Expect another UnorderedReduction or phi as the scalar input.
4604 Node* scalar_input = current->in(1);
4605 if (scalar_input->is_UnorderedReduction() &&
4606 scalar_input->Opcode() == current->Opcode()) {
4607 // Move up the UnorderedReduction chain.
4608 current = scalar_input->as_UnorderedReduction();
4609 } else if (scalar_input == phi) {
4610 // Chain terminates at phi.
4611 first_ur = current;
4612 current = nullptr;
4613 break; // Success.
4614 } else {
4615 // scalar_input is neither phi nor a matching reduction
4616 // Can for example be scalar reduction when we have
4617 // partial vectorization.
4618 break; // Chain traversal fails.
4619 }
4620 }
4621 if (current != nullptr) {
4622 // Chain traversal was not successful.
4623 continue;
4624 }
4625 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4626
4627 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4628 set_ctrl(identity_scalar, C->root());
4629 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t);
4630 register_new_node(identity_vector, C->root());
4631 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4632 VectorNode::trace_new_vector(identity_vector, "UnorderedReduction");
4633
4634 // Turn the scalar phi into a vector phi.
4635 _igvn.rehash_node_delayed(phi);
4636 Node* init = phi->in(1); // Remember init before replacing it.
4637 phi->set_req_X(1, identity_vector, &_igvn);
4638 phi->as_Type()->set_type(vec_t);
4639 _igvn.set_type(phi, vec_t);
4640
4641 // Traverse down the chain of UnorderedReductions, and replace them with vector_accumulators.
4642 current = first_ur;
4643 while (true) {
4644 // Create vector_accumulator to replace current.
4645 Node* last_vector_accumulator = current->in(1);
4646 Node* vector_input = current->in(2);
4647 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4648 register_new_node(vector_accumulator, cl);
4649 _igvn.replace_node(current, vector_accumulator);
4650 VectorNode::trace_new_vector(vector_accumulator, "UnorderedReduction");
4651 if (current == last_ur) {
4652 break;
4653 }
4654 current = vector_accumulator->unique_out()->as_UnorderedReduction();
4655 }
4656
4657 // Create post-loop reduction.
4658 Node* last_accumulator = phi->in(2);
4659 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4660
4661 // Take over uses of last_accumulator that are not in the loop.
4662 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4663 Node* use = last_accumulator->out(i);
4664 if (use != phi && use != post_loop_reduction) {
4665 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4666 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4667 --i;
4668 }
4669 }
4670 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4671 VectorNode::trace_new_vector(post_loop_reduction, "UnorderedReduction");
4672
4673 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4674 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4675 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4676 }
4677 }
4678
4679 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4680 for (uint i = 0; i < _data_nodes.size(); i++) {
4681 clone(_data_nodes[i], new_ctrl);
4682 }
4683 }
4684
4685 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4686 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4687 Node* clone = node->clone();
4688 _phase->igvn().register_new_node_with_optimizer(clone);
4689 _orig_to_new.put(node, clone);
4690 _phase->set_ctrl(clone, new_ctrl);
4691 }
4692
4693 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4694 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4695 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4696 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4697 for (uint i = 1; i < node->req(); i++) {
4698 Node** cloned_input = _orig_to_new.get(node->in(i));
4699 if (cloned_input != nullptr) {
4700 // Input was also cloned -> rewire clone to the cloned input.
4701 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4702 }
4703 }
4704 });
4705 }
4706
4707 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4708 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4709 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4710 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4711 Node* new_ctrl) {
4712 for (uint i = 0; i < _data_nodes.size(); i++) {
4713 Node* data_node = _data_nodes[i];
4714 if (data_node->is_Opaque1()) {
4715 transform_opaque_node(transform_strategy, data_node);
4716 } else {
4717 clone(data_node, new_ctrl);
4718 }
4719 }
4720 }
4721
4722 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4723 Node* transformed_node;
4724 if (node->is_OpaqueLoopInit()) {
4725 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4726 } else {
4727 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4728 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4729 }
4730 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4731 _orig_to_new.put(node, transformed_node);
4732 }