1 /*
2 * Copyright (c) 1999, 2025, 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 "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/subtypenode.hpp"
42 #include "opto/superword.hpp"
43 #include "opto/vectornode.hpp"
44 #include "utilities/checkedCast.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
52 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
53 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
54 // so disable this for now
55 return nullptr;
56 }
57
58 // Splitting range check CastIIs through a loop induction Phi can
59 // cause new Phis to be created that are left unrelated to the loop
60 // induction Phi and prevent optimizations (vectorization)
61 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
62 n->in(1) == region->as_CountedLoop()->phi()) {
63 return nullptr;
64 }
65
66 if (cannot_split_division(n, region)) {
67 return nullptr;
68 }
69
70 SplitThruPhiWins wins(region);
71 assert(!n->is_CFG(), "");
72 assert(region->is_Region(), "");
73
74 const Type* type = n->bottom_type();
75 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
76 Node* phi;
77 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
78 int iid = t_oop->instance_id();
79 int index = C->get_alias_index(t_oop);
80 int offset = t_oop->offset();
81 phi = new PhiNode(region, type, nullptr, iid, index, offset);
82 } else {
83 phi = PhiNode::make_blank(region, n);
84 }
85 uint old_unique = C->unique();
86 for (uint i = 1; i < region->req(); i++) {
87 Node* x;
88 Node* the_clone = nullptr;
89 if (region->in(i) == C->top()) {
90 x = C->top(); // Dead path? Use a dead data op
91 } else {
92 x = n->clone(); // Else clone up the data op
93 the_clone = x; // Remember for possible deletion.
94 // Alter data node to use pre-phi inputs
95 if (n->in(0) == region)
96 x->set_req( 0, region->in(i) );
97 for (uint j = 1; j < n->req(); j++) {
98 Node* in = n->in(j);
99 if (in->is_Phi() && in->in(0) == region)
100 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
101 }
102 }
103 // Check for a 'win' on some paths
104 const Type* t = x->Value(&_igvn);
105
106 bool singleton = t->singleton();
107
108 // A TOP singleton indicates that there are no possible values incoming
109 // along a particular edge. In most cases, this is OK, and the Phi will
110 // be eliminated later in an Ideal call. However, we can't allow this to
111 // happen if the singleton occurs on loop entry, as the elimination of
112 // the PhiNode may cause the resulting node to migrate back to a previous
113 // loop iteration.
114 if (singleton && t == Type::TOP) {
115 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
116 // irreducible loop may not be indicated by an affirmative is_Loop());
117 // therefore, the only top we can split thru a phi is on a backedge of
118 // a loop.
119 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
120 }
121
122 if (singleton) {
123 wins.add_win(i);
124 x = makecon(t);
125 } else {
126 // We now call Identity to try to simplify the cloned node.
127 // Note that some Identity methods call phase->type(this).
128 // Make sure that the type array is big enough for
129 // our new node, even though we may throw the node away.
130 // (Note: This tweaking with igvn only works because x is a new node.)
131 _igvn.set_type(x, t);
132 // If x is a TypeNode, capture any more-precise type permanently into Node
133 // otherwise it will be not updated during igvn->transform since
134 // igvn->type(x) is set to x->Value() already.
135 x->raise_bottom_type(t);
136 Node* y = x->Identity(&_igvn);
137 if (y != x) {
138 wins.add_win(i);
139 x = y;
140 } else {
141 y = _igvn.hash_find(x);
142 if (y == nullptr) {
143 y = similar_subtype_check(x, region->in(i));
144 }
145 if (y) {
146 wins.add_win(i);
147 x = y;
148 } else {
149 // Else x is a new node we are keeping
150 // We do not need register_new_node_with_optimizer
151 // because set_type has already been called.
152 _igvn._worklist.push(x);
153 }
154 }
155 }
156
157 phi->set_req( i, x );
158
159 if (the_clone == nullptr) {
160 continue;
161 }
162
163 if (the_clone != x) {
164 _igvn.remove_dead_node(the_clone);
165 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
166 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
167 // it is not a win if 'x' moved from an outer to an inner loop
168 // this edge case can only happen for Load nodes
169 wins.reset();
170 break;
171 }
172 }
173 // Too few wins?
174 if (!wins.profitable(policy)) {
175 _igvn.remove_dead_node(phi);
176 return nullptr;
177 }
178
179 // Record Phi
180 register_new_node( phi, region );
181
182 for (uint i2 = 1; i2 < phi->req(); i2++) {
183 Node *x = phi->in(i2);
184 // If we commoned up the cloned 'x' with another existing Node,
185 // the existing Node picks up a new use. We need to make the
186 // existing Node occur higher up so it dominates its uses.
187 Node *old_ctrl;
188 IdealLoopTree *old_loop;
189
190 if (x->is_Con()) {
191 assert(get_ctrl(x) == C->root(), "constant control is not root");
192 continue;
193 }
194 // The occasional new node
195 if (x->_idx >= old_unique) { // Found a new, unplaced node?
196 old_ctrl = nullptr;
197 old_loop = nullptr; // Not in any prior loop
198 } else {
199 old_ctrl = get_ctrl(x);
200 old_loop = get_loop(old_ctrl); // Get prior loop
201 }
202 // New late point must dominate new use
203 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
204 if (new_ctrl == old_ctrl) // Nothing is changed
205 continue;
206
207 IdealLoopTree *new_loop = get_loop(new_ctrl);
208
209 // Don't move x into a loop if its uses are
210 // outside of loop. Otherwise x will be cloned
211 // for each use outside of this loop.
212 IdealLoopTree *use_loop = get_loop(region);
213 if (!new_loop->is_member(use_loop) &&
214 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
215 // Take early control, later control will be recalculated
216 // during next iteration of loop optimizations.
217 new_ctrl = get_early_ctrl(x);
218 new_loop = get_loop(new_ctrl);
219 }
220 // Set new location
221 set_ctrl(x, new_ctrl);
222 // If changing loop bodies, see if we need to collect into new body
223 if (old_loop != new_loop) {
224 if (old_loop && !old_loop->_child)
225 old_loop->_body.yank(x);
226 if (!new_loop->_child)
227 new_loop->_body.push(x); // Collect body info
228 }
229 }
230
231 split_thru_phi_yank_old_nodes(n, region);
232 _igvn.replace_node(n, phi);
233
234 #ifndef PRODUCT
235 if (TraceLoopOpts) {
236 tty->print_cr("Split %d %s through %d Phi in %d %s",
237 n->_idx, n->Name(), phi->_idx, region->_idx, region->Name());
238 }
239 #endif // !PRODUCT
240
241 return phi;
242 }
243
244 // If the region is a Loop, we are removing the old n,
245 // and need to yank it from the _body. If any phi we
246 // just split through now has no use any more, it also
247 // has to be removed.
248 void PhaseIdealLoop::split_thru_phi_yank_old_nodes(Node* n, Node* region) {
249 IdealLoopTree* region_loop = get_loop(region);
250 if (region->is_Loop() && region_loop->is_innermost()) {
251 region_loop->_body.yank(n);
252 for (uint j = 1; j < n->req(); j++) {
253 PhiNode* phi = n->in(j)->isa_Phi();
254 // Check that phi belongs to the region and only has n as a use.
255 if (phi != nullptr &&
256 phi->in(0) == region &&
257 phi->unique_multiple_edges_out_or_null() == n) {
258 assert(get_ctrl(phi) == region, "sanity");
259 assert(get_ctrl(n) == region, "sanity");
260 region_loop->_body.yank(phi);
261 }
262 }
263 }
264 }
265
266 // Test whether node 'x' can move into an inner loop relative to node 'n'.
267 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
268 // BUT it can also return true and 'x' is in the outer loop
269 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
270 IdealLoopTree* n_loop_tree = get_loop(n_loop);
271 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
272 // x_loop_tree should be outer or same loop as n_loop_tree
273 return !x_loop_tree->is_member(n_loop_tree);
274 }
275
276 // Subtype checks that carry profile data don't common so look for a replacement by following edges
277 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
278 if (x->is_SubTypeCheck()) {
279 Node* in1 = x->in(1);
280 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
281 Node* u = in1->fast_out(i);
282 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
283 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
284 Node* bol = u->fast_out(j);
285 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
286 Node* iff = bol->fast_out(k);
287 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
288 // unrelated profile
289 if (iff->is_If() && is_dominator(iff, r_in)) {
290 return u;
291 }
292 }
293 }
294 }
295 }
296 }
297 return nullptr;
298 }
299
300 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
301 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
302 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
303 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
304 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
305 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
306 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
307 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
308 const Type* zero;
309 switch (n->Opcode()) {
310 case Op_DivI:
311 case Op_ModI:
312 case Op_UDivI:
313 case Op_UModI:
314 zero = TypeInt::ZERO;
315 break;
316 case Op_DivL:
317 case Op_ModL:
318 case Op_UDivL:
319 case Op_UModL:
320 zero = TypeLong::ZERO;
321 break;
322 default:
323 return false;
324 }
325
326 if (n->in(0) != nullptr) {
327 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
328 return true;
329 }
330
331 Node* divisor = n->in(2);
332 return is_divisor_loop_phi(divisor, region) &&
333 loop_phi_backedge_type_contains_zero(divisor, zero);
334 }
335
336 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
337 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
338 }
339
340 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
341 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
342 }
343
344 //------------------------------dominated_by------------------------------------
345 // Replace the dominated test with an obvious true or false. Place it on the
346 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
347 // live path up to the dominating control.
348 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
349 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
350
351 // prevdom is the dominating projection of the dominating test.
352 assert(iff->Opcode() == Op_If ||
353 iff->Opcode() == Op_CountedLoopEnd ||
354 iff->Opcode() == Op_LongCountedLoopEnd ||
355 iff->Opcode() == Op_RangeCheck ||
356 iff->Opcode() == Op_ParsePredicate,
357 "Check this code when new subtype is added");
358
359 int pop = prevdom->Opcode();
360 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
361 if (flip) {
362 if (pop == Op_IfTrue)
363 pop = Op_IfFalse;
364 else
365 pop = Op_IfTrue;
366 }
367 // 'con' is set to true or false to kill the dominated test.
368 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
369 // Hack the dominated test
370 _igvn.replace_input_of(iff, 1, con);
371
372 // If I don't have a reachable TRUE and FALSE path following the IfNode then
373 // I can assume this path reaches an infinite loop. In this case it's not
374 // important to optimize the data Nodes - either the whole compilation will
375 // be tossed or this path (and all data Nodes) will go dead.
376 if (iff->outcnt() != 2) {
377 return;
378 }
379
380 // Make control-dependent data Nodes on the live path (path that will remain
381 // once the dominated IF is removed) become control-dependent on the
382 // dominating projection.
383 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
384
385 if (dp == nullptr) {
386 return;
387 }
388
389 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
390 }
391
392 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
393 IdealLoopTree* old_loop = get_loop(source);
394
395 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
396 Node* out = source->fast_out(i); // Control-dependent node
397 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
398 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
399 assert(out->in(0) == source, "must be control dependent on source");
400 _igvn.replace_input_of(out, 0, dominator);
401 if (pin_array_access_nodes) {
402 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
403 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
404 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
405 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
406 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
407 // dominating check.
408 Node* clone = out->pin_array_access_node();
409 if (clone != nullptr) {
410 clone = _igvn.register_new_node_with_optimizer(clone, out);
411 _igvn.replace_node(out, clone);
412 out = clone;
413 }
414 }
415 set_early_ctrl(out, false);
416 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
417 if (old_loop != new_loop) {
418 if (!old_loop->_child) {
419 old_loop->_body.yank(out);
420 }
421 if (!new_loop->_child) {
422 new_loop->_body.push(out);
423 }
424 }
425 --i;
426 --imax;
427 }
428 }
429 }
430
431 //------------------------------has_local_phi_input----------------------------
432 // Return TRUE if 'n' has Phi inputs from its local block and no other
433 // block-local inputs (all non-local-phi inputs come from earlier blocks)
434 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
435 Node *n_ctrl = get_ctrl(n);
436 // See if some inputs come from a Phi in this block, or from before
437 // this block.
438 uint i;
439 for( i = 1; i < n->req(); i++ ) {
440 Node *phi = n->in(i);
441 if( phi->is_Phi() && phi->in(0) == n_ctrl )
442 break;
443 }
444 if( i >= n->req() )
445 return nullptr; // No Phi inputs; nowhere to clone thru
446
447 // Check for inputs created between 'n' and the Phi input. These
448 // must split as well; they have already been given the chance
449 // (courtesy of a post-order visit) and since they did not we must
450 // recover the 'cost' of splitting them by being very profitable
451 // when splitting 'n'. Since this is unlikely we simply give up.
452 for( i = 1; i < n->req(); i++ ) {
453 Node *m = n->in(i);
454 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
455 // We allow the special case of AddP's with no local inputs.
456 // This allows us to split-up address expressions.
457 if (m->is_AddP() &&
458 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
459 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
460 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
461 // Move the AddP up to the dominating point. That's fine because control of m's inputs
462 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
463 Node* c = find_non_split_ctrl(idom(n_ctrl));
464 if (c->is_OuterStripMinedLoop()) {
465 c->as_Loop()->verify_strip_mined(1);
466 c = c->in(LoopNode::EntryControl);
467 }
468 set_ctrl_and_loop(m, c);
469 continue;
470 }
471 return nullptr;
472 }
473 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
474 }
475
476 return n_ctrl;
477 }
478
479 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
480 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
481 assert(bt == T_INT || bt == T_LONG, "only for integers");
482 int n_op = n->Opcode();
483
484 if (n_op == Op_LShift(bt)) {
485 // Scale is loop invariant
486 Node* scale = n->in(2);
487 Node* scale_ctrl = get_ctrl(scale);
488 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
489 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
490 return nullptr;
491 }
492 const TypeInt* scale_t = scale->bottom_type()->isa_int();
493 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
494 return nullptr; // Dont bother with byte/short masking
495 }
496 // Add must vary with loop (else shift would be loop-invariant)
497 Node* add = n->in(1);
498 Node* add_ctrl = get_ctrl(add);
499 IdealLoopTree* add_loop = get_loop(add_ctrl);
500 if (n_loop != add_loop) {
501 return nullptr; // happens w/ evil ZKM loops
502 }
503
504 // Convert I-V into I+ (0-V); same for V-I
505 if (add->Opcode() == Op_Sub(bt) &&
506 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
507 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
508 Node* zero = integercon(0, bt);
509 Node* neg = SubNode::make(zero, add->in(2), bt);
510 register_new_node_with_ctrl_of(neg, add->in(2));
511 add = AddNode::make(add->in(1), neg, bt);
512 register_new_node(add, add_ctrl);
513 }
514 if (add->Opcode() != Op_Add(bt)) return nullptr;
515 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
516 // See if one add input is loop invariant
517 Node* add_var = add->in(1);
518 Node* add_var_ctrl = get_ctrl(add_var);
519 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
520 Node* add_invar = add->in(2);
521 Node* add_invar_ctrl = get_ctrl(add_invar);
522 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
523 if (add_invar_loop == n_loop) {
524 // Swap to find the invariant part
525 add_invar = add_var;
526 add_invar_ctrl = add_var_ctrl;
527 add_invar_loop = add_var_loop;
528 add_var = add->in(2);
529 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
530 return nullptr;
531 }
532 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
533 return nullptr; // No invariant part of the add?
534 }
535
536 // Yes! Reshape address expression!
537 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
538 Node* inv_scale_ctrl =
539 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
540 add_invar_ctrl : scale_ctrl;
541 register_new_node(inv_scale, inv_scale_ctrl);
542 Node* var_scale = LShiftNode::make(add_var, scale, bt);
543 register_new_node(var_scale, n_ctrl);
544 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
545 register_new_node(var_add, n_ctrl);
546 _igvn.replace_node(n, var_add);
547 return var_add;
548 }
549 return nullptr;
550 }
551
552 //------------------------------remix_address_expressions----------------------
553 // Rework addressing expressions to get the most loop-invariant stuff
554 // moved out. We'd like to do all associative operators, but it's especially
555 // important (common) to do address expressions.
556 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
557 if (!has_ctrl(n)) return nullptr;
558 Node* n_ctrl = get_ctrl(n);
559 IdealLoopTree* n_loop = get_loop(n_ctrl);
560
561 // See if 'n' mixes loop-varying and loop-invariant inputs and
562 // itself is loop-varying.
563
564 // Only interested in binary ops (and AddP)
565 if (n->req() < 3 || n->req() > 4) return nullptr;
566
567 Node* n1_ctrl = get_ctrl(n->in( 1));
568 Node* n2_ctrl = get_ctrl(n->in( 2));
569 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
570 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
571 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
572 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
573
574 // Does one of my inputs spin in a tighter loop than self?
575 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
576 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
577 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
578 return nullptr; // Leave well enough alone
579 }
580
581 // Is at least one of my inputs loop-invariant?
582 if (n1_loop == n_loop &&
583 n2_loop == n_loop &&
584 n3_loop == n_loop) {
585 return nullptr; // No loop-invariant inputs
586 }
587
588 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
589 if (res != nullptr) {
590 return res;
591 }
592 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
593 if (res != nullptr) {
594 return res;
595 }
596
597 int n_op = n->Opcode();
598 // Replace (I+V) with (V+I)
599 if (n_op == Op_AddI ||
600 n_op == Op_AddL ||
601 n_op == Op_AddF ||
602 n_op == Op_AddD ||
603 n_op == Op_MulI ||
604 n_op == Op_MulL ||
605 n_op == Op_MulF ||
606 n_op == Op_MulD) {
607 if (n2_loop == n_loop) {
608 assert(n1_loop != n_loop, "");
609 n->swap_edges(1, 2);
610 }
611 }
612
613 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
614 // but not if I2 is a constant. Skip for irreducible loops.
615 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
616 if (n2_loop == n_loop && n3_loop != n_loop) {
617 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
618 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
619 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
620 IdealLoopTree* n22loop = get_loop(n22_ctrl);
621 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
622 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
623 n23_loop == n_loop) {
624 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
625 // Stuff new AddP in the loop preheader
626 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
627 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
628 register_new_node(add2, n_ctrl);
629 _igvn.replace_node(n, add2);
630 return add2;
631 }
632 }
633 }
634
635 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
636 if (n2_loop != n_loop && n3_loop == n_loop) {
637 if (n->in(3)->Opcode() == Op_AddX) {
638 Node* V = n->in(3)->in(1);
639 Node* I = n->in(3)->in(2);
640 if (is_member(n_loop,get_ctrl(V))) {
641 } else {
642 Node *tmp = V; V = I; I = tmp;
643 }
644 if (!is_member(n_loop,get_ctrl(I))) {
645 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
646 // Stuff new AddP in the loop preheader
647 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
648 Node* add2 = new AddPNode(n->in(1), add1, V);
649 register_new_node(add2, n_ctrl);
650 _igvn.replace_node(n, add2);
651 return add2;
652 }
653 }
654 }
655 }
656
657 return nullptr;
658 }
659
660 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
661 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
662 assert(n->Opcode() == Op_AddI, "sanity");
663 Node * nn = nullptr;
664 Node * in1 = n->in(1);
665 Node * in2 = n->in(2);
666 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
667 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
668 if (loop_n->is_counted() &&
669 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
670 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
671 Matcher::match_rule_supported(Op_MulAddS2I)) {
672 Node* mul_in1 = in1->in(1);
673 Node* mul_in2 = in1->in(2);
674 Node* mul_in3 = in2->in(1);
675 Node* mul_in4 = in2->in(2);
676 if (mul_in1->Opcode() == Op_LoadS &&
677 mul_in2->Opcode() == Op_LoadS &&
678 mul_in3->Opcode() == Op_LoadS &&
679 mul_in4->Opcode() == Op_LoadS) {
680 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
681 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
682 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
683 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
684 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
685 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
686 // All nodes should be in the same counted loop.
687 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
688 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
689 Node* adr1 = mul_in1->in(MemNode::Address);
690 Node* adr2 = mul_in2->in(MemNode::Address);
691 Node* adr3 = mul_in3->in(MemNode::Address);
692 Node* adr4 = mul_in4->in(MemNode::Address);
693 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
694 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
695 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
696 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
697 register_new_node_with_ctrl_of(nn, n);
698 _igvn.replace_node(n, nn);
699 return nn;
700 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
701 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
702 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
703 register_new_node_with_ctrl_of(nn, n);
704 _igvn.replace_node(n, nn);
705 return nn;
706 }
707 }
708 }
709 }
710 }
711 }
712 return nn;
713 }
714
715 //------------------------------conditional_move-------------------------------
716 // Attempt to replace a Phi with a conditional move. We have some pretty
717 // strict profitability requirements. All Phis at the merge point must
718 // be converted, so we can remove the control flow. We need to limit the
719 // number of c-moves to a small handful. All code that was in the side-arms
720 // of the CFG diamond is now speculatively executed. This code has to be
721 // "cheap enough". We are pretty much limited to CFG diamonds that merge
722 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
723 Node *PhaseIdealLoop::conditional_move( Node *region ) {
724
725 assert(region->is_Region(), "sanity check");
726 if (region->req() != 3) return nullptr;
727
728 // Check for CFG diamond
729 Node *lp = region->in(1);
730 Node *rp = region->in(2);
731 if (!lp || !rp) return nullptr;
732 Node *lp_c = lp->in(0);
733 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
734 IfNode *iff = lp_c->as_If();
735
736 // Check for ops pinned in an arm of the diamond.
737 // Can't remove the control flow in this case
738 if (lp->outcnt() > 1) return nullptr;
739 if (rp->outcnt() > 1) return nullptr;
740
741 IdealLoopTree* r_loop = get_loop(region);
742 assert(r_loop == get_loop(iff), "sanity");
743 // Always convert to CMOVE if all results are used only outside this loop.
744 bool used_inside_loop = (r_loop == _ltree_root);
745
746 // Check profitability
747 int cost = 0;
748 int phis = 0;
749 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
750 Node *out = region->fast_out(i);
751 if (!out->is_Phi()) continue; // Ignore other control edges, etc
752 phis++;
753 PhiNode* phi = out->as_Phi();
754 BasicType bt = phi->type()->basic_type();
755 switch (bt) {
756 case T_DOUBLE:
757 case T_FLOAT:
758 if (C->use_cmove()) {
759 continue; //TODO: maybe we want to add some cost
760 }
761 cost += Matcher::float_cmove_cost(); // Could be very expensive
762 break;
763 case T_LONG: {
764 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
765 }
766 case T_INT: // These all CMOV fine
767 case T_ADDRESS: { // (RawPtr)
768 cost++;
769 break;
770 }
771 case T_NARROWOOP: // Fall through
772 case T_OBJECT: { // Base oops are OK, but not derived oops
773 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
774 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
775 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
776 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
777 // have a Phi for the base here that we convert to a CMOVE all is well
778 // and good. But if the base is dead, we'll not make a CMOVE. Later
779 // the allocator will have to produce a base by creating a CMOVE of the
780 // relevant bases. This puts the allocator in the business of
781 // manufacturing expensive instructions, generally a bad plan.
782 // Just Say No to Conditionally-Moved Derived Pointers.
783 if (tp && tp->offset() != 0)
784 return nullptr;
785 cost++;
786 break;
787 }
788 default:
789 return nullptr; // In particular, can't do memory or I/O
790 }
791 // Add in cost any speculative ops
792 for (uint j = 1; j < region->req(); j++) {
793 Node *proj = region->in(j);
794 Node *inp = phi->in(j);
795 if (get_ctrl(inp) == proj) { // Found local op
796 cost++;
797 // Check for a chain of dependent ops; these will all become
798 // speculative in a CMOV.
799 for (uint k = 1; k < inp->req(); k++)
800 if (get_ctrl(inp->in(k)) == proj)
801 cost += ConditionalMoveLimit; // Too much speculative goo
802 }
803 }
804 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
805 // This will likely Split-If, a higher-payoff operation.
806 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
807 Node* use = phi->fast_out(k);
808 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
809 cost += ConditionalMoveLimit;
810 // Is there a use inside the loop?
811 // Note: check only basic types since CMoveP is pinned.
812 if (!used_inside_loop && is_java_primitive(bt)) {
813 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
814 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
815 used_inside_loop = true;
816 }
817 }
818 }
819 }//for
820 Node* bol = iff->in(1);
821 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
822 if (bol->is_OpaqueTemplateAssertionPredicate()) {
823 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
824 return nullptr;
825 }
826 if (bol->is_OpaqueMultiversioning()) {
827 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
828 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
829 // and will make the multiversion_if constant fold in the next IGVN round.
830 return nullptr;
831 }
832 if (!bol->is_Bool()) {
833 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
834 return nullptr;
835 }
836 int cmp_op = bol->in(1)->Opcode();
837 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
838 return nullptr;
839 }
840 // It is expensive to generate flags from a float compare.
841 // Avoid duplicated float compare.
842 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
843
844 float infrequent_prob = PROB_UNLIKELY_MAG(3);
845 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
846 if (used_inside_loop) {
847 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
848
849 // BlockLayoutByFrequency optimization moves infrequent branch
850 // from hot path. No point in CMOV'ing in such case (110 is used
851 // instead of 100 to take into account not exactness of float value).
852 if (BlockLayoutByFrequency) {
853 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
854 }
855 }
856 // Check for highly predictable branch. No point in CMOV'ing if
857 // we are going to predict accurately all the time.
858 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
859 //keep going
860 } else if (iff->_prob < infrequent_prob ||
861 iff->_prob > (1.0f - infrequent_prob))
862 return nullptr;
863
864 // --------------
865 // Now replace all Phis with CMOV's
866 Node *cmov_ctrl = iff->in(0);
867 uint flip = (lp->Opcode() == Op_IfTrue);
868 Node_List wq;
869 while (1) {
870 PhiNode* phi = nullptr;
871 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
872 Node *out = region->fast_out(i);
873 if (out->is_Phi()) {
874 phi = out->as_Phi();
875 break;
876 }
877 }
878 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
879 break;
880 }
881 // Move speculative ops
882 wq.push(phi);
883 while (wq.size() > 0) {
884 Node *n = wq.pop();
885 for (uint j = 1; j < n->req(); j++) {
886 Node* m = n->in(j);
887 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
888 set_ctrl(m, cmov_ctrl);
889 wq.push(m);
890 }
891 }
892 }
893 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
894 register_new_node(cmov, cmov_ctrl);
895 _igvn.replace_node(phi, cmov);
896 #ifndef PRODUCT
897 if (TraceLoopOpts) {
898 tty->print("CMOV ");
899 r_loop->dump_head();
900 if (Verbose) {
901 bol->in(1)->dump(1);
902 cmov->dump(1);
903 }
904 }
905 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
906 #endif
907 }
908
909 // The useless CFG diamond will fold up later; see the optimization in
910 // RegionNode::Ideal.
911 _igvn._worklist.push(region);
912
913 return iff->in(1);
914 }
915
916 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
917 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
918 Node* u = m->fast_out(i);
919 if (u->is_CFG()) {
920 if (u->is_NeverBranch()) {
921 u = u->as_NeverBranch()->proj_out(0);
922 enqueue_cfg_uses(u, wq);
923 } else {
924 wq.push(u);
925 }
926 }
927 }
928 }
929
930 // Try moving a store out of a loop, right before the loop
931 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
932 // Store has to be first in the loop body
933 IdealLoopTree *n_loop = get_loop(n_ctrl);
934 if (n->is_Store() && n_loop != _ltree_root &&
935 n_loop->is_loop() && n_loop->_head->is_Loop() &&
936 n->in(0) != nullptr) {
937 Node* address = n->in(MemNode::Address);
938 Node* value = n->in(MemNode::ValueIn);
939 Node* mem = n->in(MemNode::Memory);
940 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
941 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
942
943 // - address and value must be loop invariant
944 // - memory must be a memory Phi for the loop
945 // - Store must be the only store on this memory slice in the
946 // loop: if there's another store following this one then value
947 // written at iteration i by the second store could be overwritten
948 // at iteration i+n by the first store: it's not safe to move the
949 // first store out of the loop
950 // - nothing must observe the memory Phi: it guarantees no read
951 // before the store, we are also guaranteed the store post
952 // dominates the loop head (ignoring a possible early
953 // exit). Otherwise there would be extra Phi involved between the
954 // loop's Phi and the store.
955 // - there must be no early exit from the loop before the Store
956 // (such an exit most of the time would be an extra use of the
957 // memory Phi but sometimes is a bottom memory Phi that takes the
958 // store as input).
959
960 if (!n_loop->is_member(address_loop) &&
961 !n_loop->is_member(value_loop) &&
962 mem->is_Phi() && mem->in(0) == n_loop->_head &&
963 mem->outcnt() == 1 &&
964 mem->in(LoopNode::LoopBackControl) == n) {
965
966 assert(n_loop->_tail != nullptr, "need a tail");
967 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
968
969 // Verify that there's no early exit of the loop before the store.
970 bool ctrl_ok = false;
971 {
972 // Follow control from loop head until n, we exit the loop or
973 // we reach the tail
974 ResourceMark rm;
975 Unique_Node_List wq;
976 wq.push(n_loop->_head);
977
978 for (uint next = 0; next < wq.size(); ++next) {
979 Node *m = wq.at(next);
980 if (m == n->in(0)) {
981 ctrl_ok = true;
982 continue;
983 }
984 assert(!has_ctrl(m), "should be CFG");
985 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
986 ctrl_ok = false;
987 break;
988 }
989 enqueue_cfg_uses(m, wq);
990 if (wq.size() > 10) {
991 ctrl_ok = false;
992 break;
993 }
994 }
995 }
996 if (ctrl_ok) {
997 // move the Store
998 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
999 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
1000 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
1001 // Disconnect the phi now. An empty phi can confuse other
1002 // optimizations in this pass of loop opts.
1003 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
1004 n_loop->_body.yank(mem);
1005
1006 set_ctrl_and_loop(n, n->in(0));
1007
1008 return n;
1009 }
1010 }
1011 }
1012 return nullptr;
1013 }
1014
1015 // Try moving a store out of a loop, right after the loop
1016 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
1017 if (n->is_Store() && n->in(0) != nullptr) {
1018 Node *n_ctrl = get_ctrl(n);
1019 IdealLoopTree *n_loop = get_loop(n_ctrl);
1020 // Store must be in a loop
1021 if (n_loop != _ltree_root && !n_loop->_irreducible) {
1022 Node* address = n->in(MemNode::Address);
1023 Node* value = n->in(MemNode::ValueIn);
1024 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
1025 // address must be loop invariant
1026 if (!n_loop->is_member(address_loop)) {
1027 // Store must be last on this memory slice in the loop and
1028 // nothing in the loop must observe it
1029 Node* phi = nullptr;
1030 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1031 Node* u = n->fast_out(i);
1032 if (has_ctrl(u)) { // control use?
1033 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1034 if (!n_loop->is_member(u_loop)) {
1035 continue;
1036 }
1037 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1038 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1039 // multiple phis on the same slice are possible
1040 if (phi != nullptr) {
1041 return;
1042 }
1043 phi = u;
1044 continue;
1045 }
1046 }
1047 return;
1048 }
1049 if (phi != nullptr) {
1050 // Nothing in the loop before the store (next iteration)
1051 // must observe the stored value
1052 bool mem_ok = true;
1053 {
1054 ResourceMark rm;
1055 Unique_Node_List wq;
1056 wq.push(phi);
1057 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1058 Node *m = wq.at(next);
1059 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1060 Node* u = m->fast_out(i);
1061 if (u->is_Store() || u->is_Phi()) {
1062 if (u != n) {
1063 wq.push(u);
1064 mem_ok = (wq.size() <= 10);
1065 }
1066 } else {
1067 mem_ok = false;
1068 break;
1069 }
1070 }
1071 }
1072 }
1073 if (mem_ok) {
1074 // Move the store out of the loop if the LCA of all
1075 // users (except for the phi) is outside the loop.
1076 Node* hook = new Node(1);
1077 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1078 _igvn.rehash_node_delayed(phi);
1079 int count = phi->replace_edge(n, hook, &_igvn);
1080 assert(count > 0, "inconsistent phi");
1081
1082 // Compute latest point this store can go
1083 Node* lca = get_late_ctrl(n, get_ctrl(n));
1084 if (lca->is_OuterStripMinedLoop()) {
1085 lca = lca->in(LoopNode::EntryControl);
1086 }
1087 if (n_loop->is_member(get_loop(lca))) {
1088 // LCA is in the loop - bail out
1089 _igvn.replace_node(hook, n);
1090 return;
1091 }
1092 #ifdef ASSERT
1093 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1094 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1095 n_loop->_head->as_Loop()->verify_strip_mined(1);
1096 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1097 IdealLoopTree* outer_loop = get_loop(outer);
1098 assert(n_loop->_parent == outer_loop, "broken loop tree");
1099 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1100 }
1101 #endif
1102 lca = place_outside_loop(lca, n_loop);
1103 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1104 assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1105
1106 // Move store out of the loop
1107 _igvn.replace_node(hook, n->in(MemNode::Memory));
1108 _igvn.replace_input_of(n, 0, lca);
1109 set_ctrl_and_loop(n, lca);
1110
1111 // Disconnect the phi now. An empty phi can confuse other
1112 // optimizations in this pass of loop opts..
1113 if (phi->in(LoopNode::LoopBackControl) == phi) {
1114 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1115 n_loop->_body.yank(phi);
1116 }
1117 }
1118 }
1119 }
1120 }
1121 }
1122 }
1123
1124 //------------------------------split_if_with_blocks_pre-----------------------
1125 // Do the real work in a non-recursive function. Data nodes want to be
1126 // cloned in the pre-order so they can feed each other nicely.
1127 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1128 // Cloning these guys is unlikely to win
1129 int n_op = n->Opcode();
1130 if (n_op == Op_MergeMem) {
1131 return n;
1132 }
1133 if (n->is_Proj()) {
1134 return n;
1135 }
1136 // Do not clone-up CmpFXXX variations, as these are always
1137 // followed by a CmpI
1138 if (n->is_Cmp()) {
1139 return n;
1140 }
1141 // Attempt to use a conditional move instead of a phi/branch
1142 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1143 Node *cmov = conditional_move( n );
1144 if (cmov) {
1145 return cmov;
1146 }
1147 }
1148 if (n->is_CFG() || n->is_LoadStore()) {
1149 return n;
1150 }
1151 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1152 if (!C->major_progress()) { // If chance of no more loop opts...
1153 _igvn._worklist.push(n); // maybe we'll remove them
1154 }
1155 return n;
1156 }
1157
1158 if (n->is_Con()) {
1159 return n; // No cloning for Con nodes
1160 }
1161
1162 Node *n_ctrl = get_ctrl(n);
1163 if (!n_ctrl) {
1164 return n; // Dead node
1165 }
1166
1167 Node* res = try_move_store_before_loop(n, n_ctrl);
1168 if (res != nullptr) {
1169 return n;
1170 }
1171
1172 // Attempt to remix address expressions for loop invariants
1173 Node *m = remix_address_expressions( n );
1174 if( m ) return m;
1175
1176 if (n_op == Op_AddI) {
1177 Node *nn = convert_add_to_muladd( n );
1178 if ( nn ) return nn;
1179 }
1180
1181 if (n->is_ConstraintCast()) {
1182 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1183 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1184 // Node control inputs don't necessarily agree with loop control info (due to
1185 // transformations happened in between), thus additional dominance check is needed
1186 // to keep loop info valid.
1187 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1188 _igvn.replace_node(n, dom_cast);
1189 return dom_cast;
1190 }
1191 }
1192
1193 // Determine if the Node has inputs from some local Phi.
1194 // Returns the block to clone thru.
1195 Node *n_blk = has_local_phi_input( n );
1196 if( !n_blk ) return n;
1197
1198 // Do not clone the trip counter through on a CountedLoop
1199 // (messes up the canonical shape).
1200 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1201 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1202 return n;
1203 }
1204 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1205 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1206 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1207 return n;
1208 }
1209
1210 // Check for having no control input; not pinned. Allow
1211 // dominating control.
1212 if (n->in(0)) {
1213 Node *dom = idom(n_blk);
1214 if (dom_lca(n->in(0), dom) != n->in(0)) {
1215 return n;
1216 }
1217 }
1218 // Policy: when is it profitable. You must get more wins than
1219 // policy before it is considered profitable. Policy is usually 0,
1220 // so 1 win is considered profitable. Big merges will require big
1221 // cloning, so get a larger policy.
1222 int policy = n_blk->req() >> 2;
1223
1224 // If the loop is a candidate for range check elimination,
1225 // delay splitting through it's phi until a later loop optimization
1226 if (n_blk->is_BaseCountedLoop()) {
1227 IdealLoopTree *lp = get_loop(n_blk);
1228 if (lp && lp->_rce_candidate) {
1229 return n;
1230 }
1231 }
1232
1233 if (must_throttle_split_if()) return n;
1234
1235 // Split 'n' through the merge point if it is profitable, replacing it with a new phi.
1236 Node* phi = split_thru_phi(n, n_blk, policy);
1237 if (phi == nullptr) { return n; }
1238
1239 // Moved a load around the loop, 'en-registering' something.
1240 if (n_blk->is_Loop() && n->is_Load() &&
1241 !phi->in(LoopNode::LoopBackControl)->is_Load())
1242 C->set_major_progress();
1243
1244 return phi;
1245 }
1246
1247 static bool merge_point_too_heavy(Compile* C, Node* region) {
1248 // Bail out if the region and its phis have too many users.
1249 int weight = 0;
1250 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1251 weight += region->fast_out(i)->outcnt();
1252 }
1253 int nodes_left = C->max_node_limit() - C->live_nodes();
1254 if (weight * 8 > nodes_left) {
1255 if (PrintOpto) {
1256 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1257 }
1258 return true;
1259 } else {
1260 return false;
1261 }
1262 }
1263
1264 static bool merge_point_safe(Node* region) {
1265 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1266 // having a PhiNode input. This sidesteps the dangerous case where the split
1267 // ConvI2LNode may become TOP if the input Value() does not
1268 // overlap the ConvI2L range, leaving a node which may not dominate its
1269 // uses.
1270 // A better fix for this problem can be found in the BugTraq entry, but
1271 // expediency for Mantis demands this hack.
1272 #ifdef _LP64
1273 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1274 Node* n = region->fast_out(i);
1275 if (n->is_Phi()) {
1276 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1277 Node* m = n->fast_out(j);
1278 if (m->Opcode() == Op_ConvI2L)
1279 return false;
1280 if (m->is_CastII()) {
1281 return false;
1282 }
1283 }
1284 }
1285 }
1286 #endif
1287 return true;
1288 }
1289
1290
1291 //------------------------------place_outside_loop---------------------------------
1292 // Place some computation outside of this loop on the path to the use passed as argument
1293 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1294 Node* head = loop->_head;
1295 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1296 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1297 loop = loop->_parent;
1298 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1299 }
1300
1301 // Pick control right outside the loop
1302 for (;;) {
1303 Node* dom = idom(useblock);
1304 if (loop->is_member(get_loop(dom))) {
1305 break;
1306 }
1307 useblock = dom;
1308 }
1309 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1310 return useblock;
1311 }
1312
1313
1314 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1315 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1316 return false;
1317 }
1318 if (!n->in(0)->is_Region()) {
1319 return false;
1320 }
1321
1322 Node* region = n->in(0);
1323 Node* dom = idom(region);
1324 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1325 return false;
1326 }
1327 IfNode* dom_if = dom->as_If();
1328 Node* proj_true = dom_if->proj_out(1);
1329 Node* proj_false = dom_if->proj_out(0);
1330
1331 for (uint i = 1; i < region->req(); i++) {
1332 if (is_dominator(proj_true, region->in(i))) {
1333 continue;
1334 }
1335 if (is_dominator(proj_false, region->in(i))) {
1336 continue;
1337 }
1338 return false;
1339 }
1340
1341 return true;
1342 }
1343
1344
1345 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1346 if (must_throttle_split_if()) {
1347 return false;
1348 }
1349
1350 // Do not do 'split-if' if irreducible loops are present.
1351 if (_has_irreducible_loops) {
1352 return false;
1353 }
1354
1355 if (merge_point_too_heavy(C, n_ctrl)) {
1356 return false;
1357 }
1358
1359 // Do not do 'split-if' if some paths are dead. First do dead code
1360 // elimination and then see if its still profitable.
1361 for (uint i = 1; i < n_ctrl->req(); i++) {
1362 if (n_ctrl->in(i) == C->top()) {
1363 return false;
1364 }
1365 }
1366
1367 // If trying to do a 'Split-If' at the loop head, it is only
1368 // profitable if the cmp folds up on BOTH paths. Otherwise we
1369 // risk peeling a loop forever.
1370
1371 // CNC - Disabled for now. Requires careful handling of loop
1372 // body selection for the cloned code. Also, make sure we check
1373 // for any input path not being in the same loop as n_ctrl. For
1374 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1375 // because the alternative loop entry points won't be converted
1376 // into LoopNodes.
1377 IdealLoopTree *n_loop = get_loop(n_ctrl);
1378 for (uint j = 1; j < n_ctrl->req(); j++) {
1379 if (get_loop(n_ctrl->in(j)) != n_loop) {
1380 return false;
1381 }
1382 }
1383
1384 // Check for safety of the merge point.
1385 if (!merge_point_safe(n_ctrl)) {
1386 return false;
1387 }
1388
1389 return true;
1390 }
1391
1392 // Detect if the node is the inner strip-mined loop
1393 // Return: null if it's not the case, or the exit of outer strip-mined loop
1394 static Node* is_inner_of_stripmined_loop(const Node* out) {
1395 Node* out_le = nullptr;
1396
1397 if (out->is_CountedLoopEnd()) {
1398 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1399
1400 if (loop != nullptr && loop->is_strip_mined()) {
1401 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1402 }
1403 }
1404
1405 return out_le;
1406 }
1407
1408 //------------------------------split_if_with_blocks_post----------------------
1409 // Do the real work in a non-recursive function. CFG hackery wants to be
1410 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1411 // info.
1412 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1413
1414 // Cloning Cmp through Phi's involves the split-if transform.
1415 // FastLock is not used by an If
1416 if (n->is_Cmp() && !n->is_FastLock()) {
1417 Node *n_ctrl = get_ctrl(n);
1418 // Determine if the Node has inputs from some local Phi.
1419 // Returns the block to clone thru.
1420 Node *n_blk = has_local_phi_input(n);
1421 if (n_blk != n_ctrl) {
1422 return;
1423 }
1424
1425 if (!can_split_if(n_ctrl)) {
1426 return;
1427 }
1428
1429 if (n->outcnt() != 1) {
1430 return; // Multiple bool's from 1 compare?
1431 }
1432 Node *bol = n->unique_out();
1433 assert(bol->is_Bool(), "expect a bool here");
1434 if (bol->outcnt() != 1) {
1435 return;// Multiple branches from 1 compare?
1436 }
1437 Node *iff = bol->unique_out();
1438
1439 // Check some safety conditions
1440 if (iff->is_If()) { // Classic split-if?
1441 if (iff->in(0) != n_ctrl) {
1442 return; // Compare must be in same blk as if
1443 }
1444 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1445 // Can't split CMove with different control.
1446 if (get_ctrl(iff) != n_ctrl) {
1447 return;
1448 }
1449 if (get_ctrl(iff->in(2)) == n_ctrl ||
1450 get_ctrl(iff->in(3)) == n_ctrl) {
1451 return; // Inputs not yet split-up
1452 }
1453 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1454 return; // Loop-invar test gates loop-varying CMOVE
1455 }
1456 } else {
1457 return; // some other kind of node, such as an Allocate
1458 }
1459
1460 // When is split-if profitable? Every 'win' on means some control flow
1461 // goes dead, so it's almost always a win.
1462 int policy = 0;
1463 // Split compare 'n' through the merge point if it is profitable
1464 Node *phi = split_thru_phi( n, n_ctrl, policy);
1465 if (!phi) {
1466 return;
1467 }
1468
1469 // Now split the bool up thru the phi
1470 Node* bolphi = split_thru_phi(bol, n_ctrl, -1);
1471 guarantee(bolphi != nullptr, "null boolean phi node");
1472 assert(iff->in(1) == bolphi, "");
1473
1474 if (bolphi->Value(&_igvn)->singleton()) {
1475 return;
1476 }
1477
1478 // Conditional-move? Must split up now
1479 if (!iff->is_If()) {
1480 Node* cmovphi = split_thru_phi(iff, n_ctrl, -1);
1481 return;
1482 }
1483
1484 // Now split the IF
1485 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1486 if (TraceLoopOpts) {
1487 tty->print_cr("Split-If");
1488 }
1489 do_split_if(iff);
1490 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1491 return;
1492 }
1493
1494 // Two identical ifs back to back can be merged
1495 if (try_merge_identical_ifs(n)) {
1496 return;
1497 }
1498
1499 // Check for an IF ready to split; one that has its
1500 // condition codes input coming from a Phi at the block start.
1501 int n_op = n->Opcode();
1502
1503 // Check for an IF being dominated by another IF same test
1504 if (n_op == Op_If ||
1505 n_op == Op_RangeCheck) {
1506 Node *bol = n->in(1);
1507 uint max = bol->outcnt();
1508 // Check for same test used more than once?
1509 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1510 // Search up IDOMs to see if this IF is dominated.
1511 Node* cmp = bol->in(1);
1512 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1513
1514 // Now search up IDOMs till cutoff, looking for a dominating test
1515 Node *prevdom = n;
1516 Node *dom = idom(prevdom);
1517 while (dom != cutoff) {
1518 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1519 safe_for_if_replacement(dom)) {
1520 // It's invalid to move control dependent data nodes in the inner
1521 // strip-mined loop, because:
1522 // 1) break validation of LoopNode::verify_strip_mined()
1523 // 2) move code with side-effect in strip-mined loop
1524 // Move to the exit of outer strip-mined loop in that case.
1525 Node* out_le = is_inner_of_stripmined_loop(dom);
1526 if (out_le != nullptr) {
1527 prevdom = out_le;
1528 }
1529 // Replace the dominated test with an obvious true or false.
1530 // Place it on the IGVN worklist for later cleanup.
1531 C->set_major_progress();
1532 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1533 // to prevent an array load from floating above its range check. There are three cases:
1534 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1535 // all its array accesses at that point.
1536 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1537 // accesses would start to float, since we don't pin at that point.
1538 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1539 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1540 prevdom->in(0)->Opcode() != Op_RangeCheck;
1541 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1542 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1543 return;
1544 }
1545 prevdom = dom;
1546 dom = idom(prevdom);
1547 }
1548 }
1549 }
1550
1551 try_sink_out_of_loop(n);
1552 if (C->failing()) {
1553 return;
1554 }
1555
1556 try_move_store_after_loop(n);
1557 }
1558
1559 // Transform:
1560 //
1561 // if (some_condition) {
1562 // // body 1
1563 // } else {
1564 // // body 2
1565 // }
1566 // if (some_condition) {
1567 // // body 3
1568 // } else {
1569 // // body 4
1570 // }
1571 //
1572 // into:
1573 //
1574 //
1575 // if (some_condition) {
1576 // // body 1
1577 // // body 3
1578 // } else {
1579 // // body 2
1580 // // body 4
1581 // }
1582 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1583 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1584 Node *n_ctrl = n->in(0);
1585 IfNode* dom_if = idom(n_ctrl)->as_If();
1586 if (n->in(1) != dom_if->in(1)) {
1587 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1588 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1589 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1590 _igvn.replace_input_of(n, 1, dom_if->in(1));
1591 }
1592 ProjNode* dom_proj_true = dom_if->proj_out(1);
1593 ProjNode* dom_proj_false = dom_if->proj_out(0);
1594
1595 // Now split the IF
1596 RegionNode* new_false_region;
1597 RegionNode* new_true_region;
1598 do_split_if(n, &new_false_region, &new_true_region);
1599 assert(new_false_region->req() == new_true_region->req(), "");
1600 #ifdef ASSERT
1601 for (uint i = 1; i < new_false_region->req(); ++i) {
1602 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1603 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");
1604 }
1605 #endif
1606 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1607
1608 // We now have:
1609 // if (some_condition) {
1610 // // body 1
1611 // if (some_condition) {
1612 // body3: // new_true_region
1613 // // body3
1614 // } else {
1615 // goto body4;
1616 // }
1617 // } else {
1618 // // body 2
1619 // if (some_condition) {
1620 // goto body3;
1621 // } else {
1622 // body4: // new_false_region
1623 // // body4;
1624 // }
1625 // }
1626 //
1627
1628 // clone pinned nodes thru the resulting regions
1629 push_pinned_nodes_thru_region(dom_if, new_true_region);
1630 push_pinned_nodes_thru_region(dom_if, new_false_region);
1631
1632 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1633 // 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
1634 // unrelated control dependency.
1635 for (uint i = 1; i < new_false_region->req(); i++) {
1636 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1637 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1638 } else {
1639 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1640 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1641 }
1642 }
1643 return true;
1644 }
1645 return false;
1646 }
1647
1648 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1649 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1650 Node* u = region->out(i);
1651 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1652 continue;
1653 }
1654 assert(u->in(0) == region, "not a control dependent node?");
1655 uint j = 1;
1656 for (; j < u->req(); ++j) {
1657 Node* in = u->in(j);
1658 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1659 break;
1660 }
1661 }
1662 if (j == u->req()) {
1663 Node *phi = PhiNode::make_blank(region, u);
1664 for (uint k = 1; k < region->req(); ++k) {
1665 Node* clone = u->clone();
1666 clone->set_req(0, region->in(k));
1667 register_new_node(clone, region->in(k));
1668 phi->init_req(k, clone);
1669 }
1670 register_new_node(phi, region);
1671 _igvn.replace_node(u, phi);
1672 --i;
1673 }
1674 }
1675 }
1676
1677 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1678 if (!dom->is_CountedLoopEnd()) {
1679 return true;
1680 }
1681 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1682 CountedLoopNode* cl = le->loopnode();
1683 if (cl == nullptr) {
1684 return true;
1685 }
1686 if (!cl->is_main_loop()) {
1687 return true;
1688 }
1689 if (cl->is_canonical_loop_entry() == nullptr) {
1690 return true;
1691 }
1692 // Further unrolling is possible so loop exit condition might change
1693 return false;
1694 }
1695
1696 // See if a shared loop-varying computation has no loop-varying uses.
1697 // Happens if something is only used for JVM state in uncommon trap exits,
1698 // like various versions of induction variable+offset. Clone the
1699 // computation per usage to allow it to sink out of the loop.
1700 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1701 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1702 n->in(1)->bottom_type()->isa_rawptr() &&
1703 !n->bottom_type()->isa_rawptr();
1704
1705 if (has_ctrl(n) &&
1706 !n->is_Phi() &&
1707 !n->is_Bool() &&
1708 !n->is_Proj() &&
1709 !n->is_MergeMem() &&
1710 !n->is_CMove() &&
1711 !n->is_OpaqueNotNull() &&
1712 !n->is_OpaqueInitializedAssertionPredicate() &&
1713 !n->is_OpaqueTemplateAssertionPredicate() &&
1714 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1715 (KillPathsReachableByDeadTypeNode || !n->is_Type())
1716 ) {
1717 Node *n_ctrl = get_ctrl(n);
1718 IdealLoopTree *n_loop = get_loop(n_ctrl);
1719
1720 if (n->in(0) != nullptr) {
1721 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1722 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1723 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1724 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1725 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1726 Node* maybe_pinned_n = n;
1727 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1728 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) {
1729 if (n->depends_only_on_test()) {
1730 Node* pinned_clone = n->pin_array_access_node();
1731 if (pinned_clone != nullptr) {
1732 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1733 // range check for that access. If that condition is replaced by an identical dominating one, then an
1734 // unpinned load would risk floating above its range check.
1735 register_new_node(pinned_clone, n_ctrl);
1736 maybe_pinned_n = pinned_clone;
1737 _igvn.replace_node(n, pinned_clone);
1738 }
1739 }
1740 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1741 }
1742 }
1743 }
1744 if (n_loop != _ltree_root && n->outcnt() > 1) {
1745 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1746 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1747 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1748 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1749 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1750 if (n->is_Store() || n->is_LoadStore()) {
1751 assert(false, "no node with a side effect");
1752 C->record_failure("no node with a side effect");
1753 return;
1754 }
1755 Node* outer_loop_clone = nullptr;
1756 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1757 Node* u = n->last_out(j); // Clone private computation per use
1758 _igvn.rehash_node_delayed(u);
1759 Node* x = nullptr;
1760 if (n->depends_only_on_test()) {
1761 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1762 // range check for that access. If that condition is replaced by an identical dominating one, then an
1763 // unpinned load would risk floating above its range check.
1764 x = n->pin_array_access_node();
1765 }
1766 if (x == nullptr) {
1767 x = n->clone();
1768 }
1769 Node* x_ctrl = nullptr;
1770 if (u->is_Phi()) {
1771 // Replace all uses of normal nodes. Replace Phi uses
1772 // individually, so the separate Nodes can sink down
1773 // different paths.
1774 uint k = 1;
1775 while (u->in(k) != n) k++;
1776 u->set_req(k, x);
1777 // x goes next to Phi input path
1778 x_ctrl = u->in(0)->in(k);
1779 // Find control for 'x' next to use but not inside inner loops.
1780 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1781 --j;
1782 } else { // Normal use
1783 if (has_ctrl(u)) {
1784 x_ctrl = get_ctrl(u);
1785 } else {
1786 x_ctrl = u->in(0);
1787 }
1788 // Find control for 'x' next to use but not inside inner loops.
1789 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1790 // Replace all uses
1791 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1792 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1793 // anymore now that we're going to pin n as well
1794 _igvn.replace_node(u, x);
1795 --j;
1796 } else {
1797 int nb = u->replace_edge(n, x, &_igvn);
1798 j -= nb;
1799 }
1800 }
1801
1802 if (n->is_Load()) {
1803 // For loads, add a control edge to a CFG node outside of the loop
1804 // to force them to not combine and return back inside the loop
1805 // during GVN optimization (4641526).
1806 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1807
1808 IdealLoopTree* x_loop = get_loop(x_ctrl);
1809 Node* x_head = x_loop->_head;
1810 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1811 // Do not add duplicate LoadNodes to the outer strip mined loop
1812 if (outer_loop_clone != nullptr) {
1813 _igvn.replace_node(x, outer_loop_clone);
1814 continue;
1815 }
1816 outer_loop_clone = x;
1817 }
1818 x->set_req(0, x_ctrl);
1819 } else if (n->in(0) != nullptr){
1820 x->set_req(0, x_ctrl);
1821 }
1822 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1823 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1824 register_new_node(x, x_ctrl);
1825
1826 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1827 // All AddP nodes must keep the same base after sinking so:
1828 // 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,
1829 // their bases remain the same.
1830 // (see 2- below)
1831 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1832 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1833 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1834 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1835 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1836 assert(!x->is_Load(), "load should be pinned");
1837 // Use a cast node to pin clone out of loop
1838 Node* cast = nullptr;
1839 for (uint k = 0; k < x->req(); k++) {
1840 Node* in = x->in(k);
1841 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1842 const Type* in_t = _igvn.type(in);
1843 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1844 ConstraintCastNode::UnconditionalDependency, nullptr);
1845 }
1846 if (cast != nullptr) {
1847 Node* prev = _igvn.hash_find_insert(cast);
1848 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1849 cast->destruct(&_igvn);
1850 cast = prev;
1851 } else {
1852 register_new_node(cast, x_ctrl);
1853 }
1854 x->replace_edge(in, cast);
1855 // Chain of AddP nodes:
1856 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1857 if (x->is_AddP() && k == AddPNode::Base) {
1858 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1859 }
1860 break;
1861 }
1862 }
1863 assert(cast != nullptr, "must have added a cast to pin the node");
1864 }
1865 }
1866 _igvn.remove_dead_node(n);
1867 }
1868 _dom_lca_tags_round = 0;
1869 }
1870 }
1871 }
1872
1873 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1874 ResourceMark rm;
1875 Node_List wq;
1876 wq.push(x);
1877 while (wq.size() != 0) {
1878 Node* n = wq.pop();
1879 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1880 Node* u = n->fast_out(i);
1881 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1882 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1883 wq.push(u);
1884 }
1885 }
1886 }
1887 }
1888
1889 // Compute the early control of a node by following its inputs until we reach
1890 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1891 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1892 Node* early_ctrl = nullptr;
1893 ResourceMark rm;
1894 Unique_Node_List wq;
1895 wq.push(n);
1896 for (uint i = 0; i < wq.size(); i++) {
1897 Node* m = wq.at(i);
1898 Node* c = nullptr;
1899 if (m->is_CFG()) {
1900 c = m;
1901 } else if (m->pinned()) {
1902 c = m->in(0);
1903 } else {
1904 for (uint j = 0; j < m->req(); j++) {
1905 Node* in = m->in(j);
1906 if (in != nullptr) {
1907 wq.push(in);
1908 }
1909 }
1910 }
1911 if (c != nullptr) {
1912 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1913 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1914 early_ctrl = c;
1915 }
1916 }
1917 }
1918 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1919 return early_ctrl;
1920 }
1921
1922 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1923 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1924 Node* u = n->fast_out(i);
1925 if (u->is_Opaque1()) {
1926 return false; // Found loop limit, bugfix for 4677003
1927 }
1928 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1929 // get_late_ctrl_with_anti_dep() use their own tag
1930 _dom_lca_tags_round++;
1931 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1932
1933 if (u->is_Phi()) {
1934 for (uint j = 1; j < u->req(); ++j) {
1935 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1936 return false;
1937 }
1938 }
1939 } else {
1940 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1941 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1942 return false;
1943 }
1944 }
1945 }
1946 return true;
1947 }
1948
1949 // 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
1950 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1951 // test of the pre loop above the point in the graph where it's pinned. This results in a broken graph. One way to avoid
1952 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code
1953 // is generated for the main loop.
1954 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) {
1955 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1956 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1957 if (is_dominator(pre_loop->loopexit(), ctrl)) {
1958 return true;
1959 }
1960 }
1961 return false;
1962 }
1963
1964 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1965 if (n->is_Load()) {
1966 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1967 }
1968 IdealLoopTree *u_loop = get_loop(ctrl);
1969 if (u_loop == n_loop) {
1970 return false; // Found loop-varying use
1971 }
1972 if (n_loop->is_member(u_loop)) {
1973 return false; // Found use in inner loop
1974 }
1975 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) {
1976 return false;
1977 }
1978 return true;
1979 }
1980
1981 //------------------------------split_if_with_blocks---------------------------
1982 // Check for aggressive application of 'split-if' optimization,
1983 // using basic block level info.
1984 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1985 Node* root = C->root();
1986 visited.set(root->_idx); // first, mark root as visited
1987 // Do pre-visit work for root
1988 Node* n = split_if_with_blocks_pre(root);
1989 uint cnt = n->outcnt();
1990 uint i = 0;
1991
1992 while (true) {
1993 // Visit all children
1994 if (i < cnt) {
1995 Node* use = n->raw_out(i);
1996 ++i;
1997 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1998 // Now do pre-visit work for this use
1999 use = split_if_with_blocks_pre(use);
2000 nstack.push(n, i); // Save parent and next use's index.
2001 n = use; // Process all children of current use.
2002 cnt = use->outcnt();
2003 i = 0;
2004 }
2005 }
2006 else {
2007 // All of n's children have been processed, complete post-processing.
2008 if (cnt != 0 && !n->is_Con()) {
2009 assert(has_node(n), "no dead nodes");
2010 split_if_with_blocks_post(n);
2011 if (C->failing()) {
2012 return;
2013 }
2014 }
2015 if (must_throttle_split_if()) {
2016 nstack.clear();
2017 }
2018 if (nstack.is_empty()) {
2019 // Finished all nodes on stack.
2020 break;
2021 }
2022 // Get saved parent node and next use's index. Visit the rest of uses.
2023 n = nstack.node();
2024 cnt = n->outcnt();
2025 i = nstack.index();
2026 nstack.pop();
2027 }
2028 }
2029 }
2030
2031
2032 //=============================================================================
2033 //
2034 // C L O N E A L O O P B O D Y
2035 //
2036
2037 //------------------------------clone_iff--------------------------------------
2038 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2039 // "Nearly" because all Nodes have been cloned from the original in the loop,
2040 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2041 // through the Phi recursively, and return a Bool.
2042 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2043
2044 // Convert this Phi into a Phi merging Bools
2045 uint i;
2046 for (i = 1; i < phi->req(); i++) {
2047 Node* b = phi->in(i);
2048 if (b->is_Phi()) {
2049 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2050 } else {
2051 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2052 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2053 }
2054 }
2055 Node* n = phi->in(1);
2056 Node* sample_opaque = nullptr;
2057 Node *sample_bool = nullptr;
2058 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2059 sample_opaque = n;
2060 sample_bool = n->in(1);
2061 assert(sample_bool->is_Bool(), "wrong type");
2062 } else {
2063 sample_bool = n;
2064 }
2065 Node *sample_cmp = sample_bool->in(1);
2066
2067 // Make Phis to merge the Cmp's inputs.
2068 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2069 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2070 for (i = 1; i < phi->req(); i++) {
2071 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2072 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2073 phi1->set_req(i, n1);
2074 phi2->set_req(i, n2);
2075 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2076 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2077 }
2078 // See if these Phis have been made before.
2079 // Register with optimizer
2080 Node *hit1 = _igvn.hash_find_insert(phi1);
2081 if (hit1) { // Hit, toss just made Phi
2082 _igvn.remove_dead_node(phi1); // Remove new phi
2083 assert(hit1->is_Phi(), "" );
2084 phi1 = (PhiNode*)hit1; // Use existing phi
2085 } else { // Miss
2086 _igvn.register_new_node_with_optimizer(phi1);
2087 }
2088 Node *hit2 = _igvn.hash_find_insert(phi2);
2089 if (hit2) { // Hit, toss just made Phi
2090 _igvn.remove_dead_node(phi2); // Remove new phi
2091 assert(hit2->is_Phi(), "" );
2092 phi2 = (PhiNode*)hit2; // Use existing phi
2093 } else { // Miss
2094 _igvn.register_new_node_with_optimizer(phi2);
2095 }
2096 // Register Phis with loop/block info
2097 set_ctrl(phi1, phi->in(0));
2098 set_ctrl(phi2, phi->in(0));
2099 // Make a new Cmp
2100 Node *cmp = sample_cmp->clone();
2101 cmp->set_req(1, phi1);
2102 cmp->set_req(2, phi2);
2103 _igvn.register_new_node_with_optimizer(cmp);
2104 set_ctrl(cmp, phi->in(0));
2105
2106 // Make a new Bool
2107 Node *b = sample_bool->clone();
2108 b->set_req(1,cmp);
2109 _igvn.register_new_node_with_optimizer(b);
2110 set_ctrl(b, phi->in(0));
2111
2112 if (sample_opaque != nullptr) {
2113 Node* opaque = sample_opaque->clone();
2114 opaque->set_req(1, b);
2115 _igvn.register_new_node_with_optimizer(opaque);
2116 set_ctrl(opaque, phi->in(0));
2117 return opaque;
2118 }
2119
2120 assert(b->is_Bool(), "");
2121 return b;
2122 }
2123
2124 //------------------------------clone_bool-------------------------------------
2125 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2126 // "Nearly" because all Nodes have been cloned from the original in the loop,
2127 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2128 // through the Phi recursively, and return a Bool.
2129 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2130 uint i;
2131 // Convert this Phi into a Phi merging Bools
2132 for( i = 1; i < phi->req(); i++ ) {
2133 Node *b = phi->in(i);
2134 if( b->is_Phi() ) {
2135 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2136 } else {
2137 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2138 }
2139 }
2140
2141 Node *sample_cmp = phi->in(1);
2142
2143 // Make Phis to merge the Cmp's inputs.
2144 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2145 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2146 for( uint j = 1; j < phi->req(); j++ ) {
2147 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2148 Node *n1, *n2;
2149 if( cmp_top->is_Cmp() ) {
2150 n1 = cmp_top->in(1);
2151 n2 = cmp_top->in(2);
2152 } else {
2153 n1 = n2 = cmp_top;
2154 }
2155 phi1->set_req( j, n1 );
2156 phi2->set_req( j, n2 );
2157 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2158 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2159 }
2160
2161 // See if these Phis have been made before.
2162 // Register with optimizer
2163 Node *hit1 = _igvn.hash_find_insert(phi1);
2164 if( hit1 ) { // Hit, toss just made Phi
2165 _igvn.remove_dead_node(phi1); // Remove new phi
2166 assert( hit1->is_Phi(), "" );
2167 phi1 = (PhiNode*)hit1; // Use existing phi
2168 } else { // Miss
2169 _igvn.register_new_node_with_optimizer(phi1);
2170 }
2171 Node *hit2 = _igvn.hash_find_insert(phi2);
2172 if( hit2 ) { // Hit, toss just made Phi
2173 _igvn.remove_dead_node(phi2); // Remove new phi
2174 assert( hit2->is_Phi(), "" );
2175 phi2 = (PhiNode*)hit2; // Use existing phi
2176 } else { // Miss
2177 _igvn.register_new_node_with_optimizer(phi2);
2178 }
2179 // Register Phis with loop/block info
2180 set_ctrl(phi1, phi->in(0));
2181 set_ctrl(phi2, phi->in(0));
2182 // Make a new Cmp
2183 Node *cmp = sample_cmp->clone();
2184 cmp->set_req( 1, phi1 );
2185 cmp->set_req( 2, phi2 );
2186 _igvn.register_new_node_with_optimizer(cmp);
2187 set_ctrl(cmp, phi->in(0));
2188
2189 assert( cmp->is_Cmp(), "" );
2190 return (CmpNode*)cmp;
2191 }
2192
2193 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2194 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2195 Node_List*& split_if_set, Node_List*& split_bool_set,
2196 Node_List*& split_cex_set, Node_List& worklist,
2197 uint new_counter, CloneLoopMode mode) {
2198 Node* nnn = old_new[old->_idx];
2199 // Copy uses to a worklist, so I can munge the def-use info
2200 // with impunity.
2201 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2202 worklist.push(old->fast_out(j));
2203
2204 while( worklist.size() ) {
2205 Node *use = worklist.pop();
2206 if (!has_node(use)) continue; // Ignore dead nodes
2207 if (use->in(0) == C->top()) continue;
2208 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2209 // Check for data-use outside of loop - at least one of OLD or USE
2210 // must not be a CFG node.
2211 #ifdef ASSERT
2212 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2213 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2214 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2215 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2216 }
2217 #endif
2218 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2219
2220 // If the Data use is an IF, that means we have an IF outside the
2221 // loop that is switching on a condition that is set inside the
2222 // loop. Happens if people set a loop-exit flag; then test the flag
2223 // in the loop to break the loop, then test is again outside the
2224 // loop to determine which way the loop exited.
2225 //
2226 // For several uses we need to make sure that there is no phi between,
2227 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2228 // to avoid such a phi in between.
2229 // For example, it is unexpected that there is a Phi between an
2230 // AllocateArray node and its ValidLengthTest input that could cause
2231 // split if to break.
2232 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2233 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2234 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2235 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2236 // Since this code is highly unlikely, we lazily build the worklist
2237 // of such Nodes to go split.
2238 if (!split_if_set) {
2239 split_if_set = new Node_List();
2240 }
2241 split_if_set->push(use);
2242 }
2243 if (use->is_Bool()) {
2244 if (!split_bool_set) {
2245 split_bool_set = new Node_List();
2246 }
2247 split_bool_set->push(use);
2248 }
2249 if (use->Opcode() == Op_CreateEx) {
2250 if (!split_cex_set) {
2251 split_cex_set = new Node_List();
2252 }
2253 split_cex_set->push(use);
2254 }
2255
2256
2257 // Get "block" use is in
2258 uint idx = 0;
2259 while( use->in(idx) != old ) idx++;
2260 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2261 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2262 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2263 ? prev->in(2)
2264 : idom(prev);
2265 if( use->is_Phi() ) // Phi use is in prior block
2266 cfg = prev->in(idx); // NOT in block of Phi itself
2267 if (cfg->is_top()) { // Use is dead?
2268 _igvn.replace_input_of(use, idx, C->top());
2269 continue;
2270 }
2271
2272 // If use is referenced through control edge... (idx == 0)
2273 if (mode == IgnoreStripMined && idx == 0) {
2274 LoopNode *head = loop->_head->as_Loop();
2275 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2276 // That node is outside the inner loop, leave it outside the
2277 // outer loop as well to not confuse verification code.
2278 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2279 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2280 continue;
2281 }
2282 }
2283
2284 while(!outer_loop->is_member(get_loop(cfg))) {
2285 prev = cfg;
2286 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2287 }
2288 // If the use occurs after merging several exits from the loop, then
2289 // old value must have dominated all those exits. Since the same old
2290 // value was used on all those exits we did not need a Phi at this
2291 // merge point. NOW we do need a Phi here. Each loop exit value
2292 // is now merged with the peeled body exit; each exit gets its own
2293 // private Phi and those Phis need to be merged here.
2294 Node *phi;
2295 if( prev->is_Region() ) {
2296 if( idx == 0 ) { // Updating control edge?
2297 phi = prev; // Just use existing control
2298 } else { // Else need a new Phi
2299 phi = PhiNode::make( prev, old );
2300 // Now recursively fix up the new uses of old!
2301 for( uint i = 1; i < prev->req(); i++ ) {
2302 worklist.push(phi); // Onto worklist once for each 'old' input
2303 }
2304 }
2305 } else {
2306 // Get new RegionNode merging old and new loop exits
2307 prev = old_new[prev->_idx];
2308 assert( prev, "just made this in step 7" );
2309 if( idx == 0) { // Updating control edge?
2310 phi = prev; // Just use existing control
2311 } else { // Else need a new Phi
2312 // Make a new Phi merging data values properly
2313 phi = PhiNode::make( prev, old );
2314 phi->set_req( 1, nnn );
2315 }
2316 }
2317 // If inserting a new Phi, check for prior hits
2318 if( idx != 0 ) {
2319 Node *hit = _igvn.hash_find_insert(phi);
2320 if( hit == nullptr ) {
2321 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2322 } else { // or
2323 // Remove the new phi from the graph and use the hit
2324 _igvn.remove_dead_node(phi);
2325 phi = hit; // Use existing phi
2326 }
2327 set_ctrl(phi, prev);
2328 }
2329 // Make 'use' use the Phi instead of the old loop body exit value
2330 assert(use->in(idx) == old, "old is still input of use");
2331 // We notify all uses of old, including use, and the indirect uses,
2332 // that may now be optimized because we have replaced old with phi.
2333 _igvn.add_users_to_worklist(old);
2334 if (idx == 0 &&
2335 use->depends_only_on_test()) {
2336 Node* pinned_clone = use->pin_array_access_node();
2337 if (pinned_clone != nullptr) {
2338 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2339 // into the region is left, an array load could become dependent on a condition that's not a range check for
2340 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2341 // floating above its range check.
2342 pinned_clone->set_req(0, phi);
2343 register_new_node_with_ctrl_of(pinned_clone, use);
2344 _igvn.replace_node(use, pinned_clone);
2345 continue;
2346 }
2347 }
2348 _igvn.replace_input_of(use, idx, phi);
2349 if( use->_idx >= new_counter ) { // If updating new phis
2350 // Not needed for correctness, but prevents a weak assert
2351 // in AddPNode from tripping (when we end up with different
2352 // base & derived Phis that will become the same after
2353 // IGVN does CSE).
2354 Node *hit = _igvn.hash_find_insert(use);
2355 if( hit ) // Go ahead and re-hash for hits.
2356 _igvn.replace_node( use, hit );
2357 }
2358 }
2359 }
2360 }
2361
2362 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2363 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2364 bool check_old_new) {
2365 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2366 Node* u = n->fast_out(j);
2367 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2368 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2369 Node* c = phase->get_ctrl(u);
2370 IdealLoopTree* u_loop = phase->get_loop(c);
2371 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2372 if (!loop->is_member(u_loop)) {
2373 if (outer_loop->is_member(u_loop)) {
2374 wq.push(u);
2375 } else {
2376 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2377 // the outer loop too
2378 Node* u_c = u->in(0);
2379 if (u_c != nullptr) {
2380 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2381 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2382 wq.push(u);
2383 }
2384 }
2385 }
2386 }
2387 }
2388 }
2389 }
2390
2391 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2392 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2393 Node_List& extra_data_nodes) {
2394 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2395 CountedLoopNode* cl = head->as_CountedLoop();
2396 Node* l = cl->outer_loop();
2397 Node* tail = cl->outer_loop_tail();
2398 IfNode* le = cl->outer_loop_end();
2399 Node* sfpt = cl->outer_safepoint();
2400 CountedLoopEndNode* cle = cl->loopexit();
2401 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2402 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2403 Node* cle_out = cle->proj_out(false);
2404
2405 Node* new_sfpt = nullptr;
2406 Node* new_cle_out = cle_out->clone();
2407 old_new.map(cle_out->_idx, new_cle_out);
2408 if (mode == CloneIncludesStripMined) {
2409 // clone outer loop body
2410 Node* new_l = l->clone();
2411 Node* new_tail = tail->clone();
2412 IfNode* new_le = le->clone()->as_If();
2413 new_sfpt = sfpt->clone();
2414
2415 set_loop(new_l, outer_loop->_parent);
2416 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2417 set_loop(new_cle_out, outer_loop->_parent);
2418 set_idom(new_cle_out, new_cle, dd);
2419 set_loop(new_sfpt, outer_loop->_parent);
2420 set_idom(new_sfpt, new_cle_out, dd);
2421 set_loop(new_le, outer_loop->_parent);
2422 set_idom(new_le, new_sfpt, dd);
2423 set_loop(new_tail, outer_loop->_parent);
2424 set_idom(new_tail, new_le, dd);
2425 set_idom(new_cl, new_l, dd);
2426
2427 old_new.map(l->_idx, new_l);
2428 old_new.map(tail->_idx, new_tail);
2429 old_new.map(le->_idx, new_le);
2430 old_new.map(sfpt->_idx, new_sfpt);
2431
2432 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2433 new_l->set_req(0, new_l);
2434 new_tail->set_req(0, new_le);
2435 new_le->set_req(0, new_sfpt);
2436 new_sfpt->set_req(0, new_cle_out);
2437 new_cle_out->set_req(0, new_cle);
2438 new_cl->set_req(LoopNode::EntryControl, new_l);
2439
2440 _igvn.register_new_node_with_optimizer(new_l);
2441 _igvn.register_new_node_with_optimizer(new_tail);
2442 _igvn.register_new_node_with_optimizer(new_le);
2443 } else {
2444 Node *newhead = old_new[loop->_head->_idx];
2445 newhead->as_Loop()->clear_strip_mined();
2446 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2447 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2448 }
2449 // Look at data node that were assigned a control in the outer
2450 // loop: they are kept in the outer loop by the safepoint so start
2451 // from the safepoint node's inputs.
2452 IdealLoopTree* outer_loop = get_loop(l);
2453 Node_Stack stack(2);
2454 stack.push(sfpt, 1);
2455 uint new_counter = C->unique();
2456 while (stack.size() > 0) {
2457 Node* n = stack.node();
2458 uint i = stack.index();
2459 while (i < n->req() &&
2460 (n->in(i) == nullptr ||
2461 !has_ctrl(n->in(i)) ||
2462 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2463 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2464 i++;
2465 }
2466 if (i < n->req()) {
2467 stack.set_index(i+1);
2468 stack.push(n->in(i), 0);
2469 } else {
2470 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2471 Node* m = n == sfpt ? new_sfpt : n->clone();
2472 if (m != nullptr) {
2473 for (uint i = 0; i < n->req(); i++) {
2474 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2475 m->set_req(i, old_new[m->in(i)->_idx]);
2476 }
2477 }
2478 } else {
2479 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2480 }
2481 if (n != sfpt) {
2482 extra_data_nodes.push(n);
2483 _igvn.register_new_node_with_optimizer(m);
2484 assert(get_ctrl(n) == cle_out, "what other control?");
2485 set_ctrl(m, new_cle_out);
2486 old_new.map(n->_idx, m);
2487 }
2488 stack.pop();
2489 }
2490 }
2491 if (mode == CloneIncludesStripMined) {
2492 _igvn.register_new_node_with_optimizer(new_sfpt);
2493 _igvn.register_new_node_with_optimizer(new_cle_out);
2494 }
2495 // Some other transformation may have pessimistically assigned some
2496 // data nodes to the outer loop. Set their control so they are out
2497 // of the outer loop.
2498 ResourceMark rm;
2499 Unique_Node_List wq;
2500 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2501 Node* old = extra_data_nodes.at(i);
2502 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2503 }
2504
2505 for (uint i = 0; i < loop->_body.size(); i++) {
2506 Node* old = loop->_body.at(i);
2507 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2508 }
2509
2510 Node* inner_out = sfpt->in(0);
2511 if (inner_out->outcnt() > 1) {
2512 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2513 }
2514
2515 Node* new_ctrl = cl->outer_loop_exit();
2516 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2517 for (uint i = 0; i < wq.size(); i++) {
2518 Node* n = wq.at(i);
2519 set_ctrl(n, new_ctrl);
2520 if (n->in(0) != nullptr) {
2521 _igvn.replace_input_of(n, 0, new_ctrl);
2522 }
2523 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2524 }
2525 } else {
2526 Node *newhead = old_new[loop->_head->_idx];
2527 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2528 }
2529 }
2530
2531 //------------------------------clone_loop-------------------------------------
2532 //
2533 // C L O N E A L O O P B O D Y
2534 //
2535 // This is the basic building block of the loop optimizations. It clones an
2536 // entire loop body. It makes an old_new loop body mapping; with this mapping
2537 // you can find the new-loop equivalent to an old-loop node. All new-loop
2538 // nodes are exactly equal to their old-loop counterparts, all edges are the
2539 // same. All exits from the old-loop now have a RegionNode that merges the
2540 // equivalent new-loop path. This is true even for the normal "loop-exit"
2541 // condition. All uses of loop-invariant old-loop values now come from (one
2542 // or more) Phis that merge their new-loop equivalents.
2543 //
2544 // This operation leaves the graph in an illegal state: there are two valid
2545 // control edges coming from the loop pre-header to both loop bodies. I'll
2546 // definitely have to hack the graph after running this transform.
2547 //
2548 // From this building block I will further edit edges to perform loop peeling
2549 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2550 //
2551 // Parameter side_by_size_idom:
2552 // When side_by_size_idom is null, the dominator tree is constructed for
2553 // the clone loop to dominate the original. Used in construction of
2554 // pre-main-post loop sequence.
2555 // When nonnull, the clone and original are side-by-side, both are
2556 // dominated by the side_by_side_idom node. Used in construction of
2557 // unswitched loops.
2558 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2559 CloneLoopMode mode, Node* side_by_side_idom) {
2560
2561 LoopNode* head = loop->_head->as_Loop();
2562 head->verify_strip_mined(1);
2563
2564 if (C->do_vector_loop() && PrintOpto) {
2565 const char* mname = C->method()->name()->as_quoted_ascii();
2566 if (mname != nullptr) {
2567 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2568 }
2569 }
2570
2571 CloneMap& cm = C->clone_map();
2572 if (C->do_vector_loop()) {
2573 cm.set_clone_idx(cm.max_gen()+1);
2574 #ifndef PRODUCT
2575 if (PrintOpto) {
2576 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2577 loop->dump_head();
2578 }
2579 #endif
2580 }
2581
2582 // Step 1: Clone the loop body. Make the old->new mapping.
2583 clone_loop_body(loop->_body, old_new, &cm);
2584
2585 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2586
2587 // Step 2: Fix the edges in the new body. If the old input is outside the
2588 // loop use it. If the old input is INside the loop, use the corresponding
2589 // new node instead.
2590 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2591
2592 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2593 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2594
2595 // Step 3: Now fix control uses. Loop varying control uses have already
2596 // been fixed up (as part of all input edges in Step 2). Loop invariant
2597 // control uses must be either an IfFalse or an IfTrue. Make a merge
2598 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2599 // refer to this.
2600 Node_List worklist;
2601 uint new_counter = C->unique();
2602 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2603
2604 // Step 4: If loop-invariant use is not control, it must be dominated by a
2605 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2606 // there if needed. Make a Phi there merging old and new used values.
2607 Node_List *split_if_set = nullptr;
2608 Node_List *split_bool_set = nullptr;
2609 Node_List *split_cex_set = nullptr;
2610 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2611
2612 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2613 Node* old = extra_data_nodes.at(i);
2614 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2615 split_bool_set, split_cex_set, worklist, new_counter,
2616 mode);
2617 }
2618
2619 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2620 // the loop uses a condition set in the loop. The original IF probably
2621 // takes control from one or more OLD Regions (which in turn get from NEW
2622 // Regions). In any case, there will be a set of Phis for each merge point
2623 // from the IF up to where the original BOOL def exists the loop.
2624 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2625
2626 }
2627
2628 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2629 if (split_if_set) {
2630 while (split_if_set->size()) {
2631 Node *iff = split_if_set->pop();
2632 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2633 if (iff->in(input)->is_Phi()) {
2634 Node *b = clone_iff(iff->in(input)->as_Phi());
2635 _igvn.replace_input_of(iff, input, b);
2636 }
2637 }
2638 }
2639 if (split_bool_set) {
2640 while (split_bool_set->size()) {
2641 Node *b = split_bool_set->pop();
2642 Node *phi = b->in(1);
2643 assert(phi->is_Phi(), "");
2644 CmpNode *cmp = clone_bool((PhiNode*) phi);
2645 _igvn.replace_input_of(b, 1, cmp);
2646 }
2647 }
2648 if (split_cex_set) {
2649 while (split_cex_set->size()) {
2650 Node *b = split_cex_set->pop();
2651 assert(b->in(0)->is_Region(), "");
2652 assert(b->in(1)->is_Phi(), "");
2653 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2654 split_up(b, b->in(0), nullptr);
2655 }
2656 }
2657 }
2658
2659 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2660 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2661 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2662 for(uint i = 0; i < body.size(); i++ ) {
2663 Node* old = body.at(i);
2664 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2665 split_bool_set, split_cex_set, worklist, new_counter,
2666 mode);
2667 }
2668 }
2669
2670 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2671 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2672 LoopNode* head = loop->_head->as_Loop();
2673 for(uint i = 0; i < body.size(); i++ ) {
2674 Node* old = body.at(i);
2675 if( !old->is_CFG() ) continue;
2676
2677 // Copy uses to a worklist, so I can munge the def-use info
2678 // with impunity.
2679 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2680 worklist.push(old->fast_out(j));
2681 }
2682
2683 while (worklist.size()) { // Visit all uses
2684 Node *use = worklist.pop();
2685 if (!has_node(use)) continue; // Ignore dead nodes
2686 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2687 if (!loop->is_member(use_loop) && use->is_CFG()) {
2688 // Both OLD and USE are CFG nodes here.
2689 assert(use->is_Proj(), "" );
2690 Node* nnn = old_new[old->_idx];
2691
2692 Node* newuse = nullptr;
2693 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2694 CountedLoopNode* cl = head->as_CountedLoop();
2695 CountedLoopEndNode* cle = cl->loopexit();
2696 // is use the projection that exits the loop from the CountedLoopEndNode?
2697 if (use->in(0) == cle) {
2698 IfFalseNode* cle_out = use->as_IfFalse();
2699 IfNode* le = cl->outer_loop_end();
2700 use = le->proj_out(false);
2701 use_loop = get_loop(use);
2702 if (mode == CloneIncludesStripMined) {
2703 nnn = old_new[le->_idx];
2704 } else {
2705 newuse = old_new[cle_out->_idx];
2706 }
2707 }
2708 }
2709 if (newuse == nullptr) {
2710 newuse = use->clone();
2711 }
2712
2713 // Clone the loop exit control projection
2714 if (C->do_vector_loop() && cm != nullptr) {
2715 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2716 }
2717 newuse->set_req(0,nnn);
2718 _igvn.register_new_node_with_optimizer(newuse);
2719 set_loop(newuse, use_loop);
2720 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2721
2722 // We need a Region to merge the exit from the peeled body and the
2723 // exit from the old loop body.
2724 RegionNode *r = new RegionNode(3);
2725 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2726 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2727
2728 // The original user of 'use' uses 'r' instead.
2729 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2730 Node* useuse = use->last_out(l);
2731 _igvn.rehash_node_delayed(useuse);
2732 uint uses_found = 0;
2733 if (useuse->in(0) == use) {
2734 useuse->set_req(0, r);
2735 uses_found++;
2736 if (useuse->is_CFG()) {
2737 // This is not a dom_depth > dd_r because when new
2738 // control flow is constructed by a loop opt, a node and
2739 // its dominator can end up at the same dom_depth
2740 assert(dom_depth(useuse) >= dd_r, "");
2741 set_idom(useuse, r, dom_depth(useuse));
2742 }
2743 }
2744 for (uint k = 1; k < useuse->req(); k++) {
2745 if( useuse->in(k) == use ) {
2746 useuse->set_req(k, r);
2747 uses_found++;
2748 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2749 // This is not a dom_depth > dd_r because when new
2750 // control flow is constructed by a loop opt, a node
2751 // and its dominator can end up at the same dom_depth
2752 assert(dom_depth(useuse) >= dd_r , "");
2753 set_idom(useuse, r, dom_depth(useuse));
2754 }
2755 }
2756 }
2757 l -= uses_found; // we deleted 1 or more copies of this edge
2758 }
2759
2760 assert(use->is_Proj(), "loop exit should be projection");
2761 // replace_node_and_forward_ctrl() below moves all nodes that are:
2762 // - control dependent on the loop exit or
2763 // - have control set to the loop exit
2764 // below the post-loop merge point.
2765 // replace_node_and_forward_ctrl() takes a dead control as first input.
2766 // To make it possible to use it, the loop exit projection is cloned and becomes the
2767 // new exit projection. The initial one becomes dead and is "replaced" by the region.
2768 Node* use_clone = use->clone();
2769 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2770 // Now finish up 'r'
2771 r->set_req(1, newuse);
2772 r->set_req(2, use_clone);
2773 _igvn.register_new_node_with_optimizer(r);
2774 set_loop(r, use_loop);
2775 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2776 replace_node_and_forward_ctrl(use, r);
2777 // Map the (cloned) old use to the new merge point
2778 old_new.map(use_clone->_idx, r);
2779 } // End of if a loop-exit test
2780 }
2781 }
2782 }
2783
2784 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2785 IdealLoopTree* parent, bool partial) {
2786 for(uint i = 0; i < body.size(); i++ ) {
2787 Node *old = body.at(i);
2788 Node *nnn = old_new[old->_idx];
2789 // Fix CFG/Loop controlling the new node
2790 if (has_ctrl(old)) {
2791 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2792 } else {
2793 set_loop(nnn, parent);
2794 if (old->outcnt() > 0) {
2795 Node* dom = idom(old);
2796 if (old_new[dom->_idx] != nullptr) {
2797 dom = old_new[dom->_idx];
2798 set_idom(nnn, dom, dd );
2799 }
2800 }
2801 }
2802 // Correct edges to the new node
2803 for (uint j = 0; j < nnn->req(); j++) {
2804 Node *n = nnn->in(j);
2805 if (n != nullptr) {
2806 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2807 if (loop->is_member(old_in_loop)) {
2808 if (old_new[n->_idx] != nullptr) {
2809 nnn->set_req(j, old_new[n->_idx]);
2810 } else {
2811 assert(!body.contains(n), "");
2812 assert(partial, "node not cloned");
2813 }
2814 }
2815 }
2816 }
2817 _igvn.hash_find_insert(nnn);
2818 }
2819 }
2820
2821 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2822 for (uint i = 0; i < body.size(); i++) {
2823 Node* old = body.at(i);
2824 Node* nnn = old->clone();
2825 old_new.map(old->_idx, nnn);
2826 if (C->do_vector_loop() && cm != nullptr) {
2827 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2828 }
2829 _igvn.register_new_node_with_optimizer(nnn);
2830 }
2831 }
2832
2833
2834 //---------------------- stride_of_possible_iv -------------------------------------
2835 // Looks for an iff/bool/comp with one operand of the compare
2836 // being a cycle involving an add and a phi,
2837 // with an optional truncation (left-shift followed by a right-shift)
2838 // of the add. Returns zero if not an iv.
2839 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2840 Node* trunc1 = nullptr;
2841 Node* trunc2 = nullptr;
2842 const TypeInteger* ttype = nullptr;
2843 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2844 return 0;
2845 }
2846 BoolNode* bl = iff->in(1)->as_Bool();
2847 Node* cmp = bl->in(1);
2848 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2849 return 0;
2850 }
2851 // Must have an invariant operand
2852 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2853 return 0;
2854 }
2855 Node* add2 = nullptr;
2856 Node* cmp1 = cmp->in(1);
2857 if (cmp1->is_Phi()) {
2858 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2859 Node* phi = cmp1;
2860 for (uint i = 1; i < phi->req(); i++) {
2861 Node* in = phi->in(i);
2862 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2863 &trunc1, &trunc2, &ttype, T_INT);
2864 if (add && add->in(1) == phi) {
2865 add2 = add->in(2);
2866 break;
2867 }
2868 }
2869 } else {
2870 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2871 Node* addtrunc = cmp1;
2872 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2873 &trunc1, &trunc2, &ttype, T_INT);
2874 if (add && add->in(1)->is_Phi()) {
2875 Node* phi = add->in(1);
2876 for (uint i = 1; i < phi->req(); i++) {
2877 if (phi->in(i) == addtrunc) {
2878 add2 = add->in(2);
2879 break;
2880 }
2881 }
2882 }
2883 }
2884 if (add2 != nullptr) {
2885 const TypeInt* add2t = _igvn.type(add2)->is_int();
2886 if (add2t->is_con()) {
2887 return add2t->get_con();
2888 }
2889 }
2890 return 0;
2891 }
2892
2893
2894 //---------------------- stay_in_loop -------------------------------------
2895 // Return the (unique) control output node that's in the loop (if it exists.)
2896 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2897 Node* unique = nullptr;
2898 if (!n) return nullptr;
2899 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2900 Node* use = n->fast_out(i);
2901 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2902 if (unique != nullptr) {
2903 return nullptr;
2904 }
2905 unique = use;
2906 }
2907 }
2908 return unique;
2909 }
2910
2911 //------------------------------ register_node -------------------------------------
2912 // Utility to register node "n" with PhaseIdealLoop
2913 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2914 _igvn.register_new_node_with_optimizer(n);
2915 loop->_body.push(n);
2916 if (n->is_CFG()) {
2917 set_loop(n, loop);
2918 set_idom(n, pred, ddepth);
2919 } else {
2920 set_ctrl(n, pred);
2921 }
2922 }
2923
2924 //------------------------------ proj_clone -------------------------------------
2925 // Utility to create an if-projection
2926 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2927 ProjNode* c = p->clone()->as_Proj();
2928 c->set_req(0, iff);
2929 return c;
2930 }
2931
2932 //------------------------------ short_circuit_if -------------------------------------
2933 // Force the iff control output to be the live_proj
2934 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2935 guarantee(live_proj != nullptr, "null projection");
2936 int proj_con = live_proj->_con;
2937 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2938 Node* con = intcon(proj_con);
2939 if (iff) {
2940 iff->set_req(1, con);
2941 }
2942 return con;
2943 }
2944
2945 //------------------------------ insert_if_before_proj -------------------------------------
2946 // Insert a new if before an if projection (* - new node)
2947 //
2948 // before
2949 // if(test)
2950 // / \
2951 // v v
2952 // other-proj proj (arg)
2953 //
2954 // after
2955 // if(test)
2956 // / \
2957 // / v
2958 // | * proj-clone
2959 // v |
2960 // other-proj v
2961 // * new_if(relop(cmp[IU](left,right)))
2962 // / \
2963 // v v
2964 // * new-proj proj
2965 // (returned)
2966 //
2967 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2968 IfNode* iff = proj->in(0)->as_If();
2969 IdealLoopTree *loop = get_loop(proj);
2970 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2971 uint ddepth = dom_depth(proj);
2972
2973 _igvn.rehash_node_delayed(iff);
2974 _igvn.rehash_node_delayed(proj);
2975
2976 proj->set_req(0, nullptr); // temporary disconnect
2977 ProjNode* proj2 = proj_clone(proj, iff);
2978 register_node(proj2, loop, iff, ddepth);
2979
2980 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2981 register_node(cmp, loop, proj2, ddepth);
2982
2983 BoolNode* bol = new BoolNode(cmp, relop);
2984 register_node(bol, loop, proj2, ddepth);
2985
2986 int opcode = iff->Opcode();
2987 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2988 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2989 register_node(new_if, loop, proj2, ddepth);
2990
2991 proj->set_req(0, new_if); // reattach
2992 set_idom(proj, new_if, ddepth);
2993
2994 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2995 guarantee(new_exit != nullptr, "null exit node");
2996 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2997
2998 return new_exit;
2999 }
3000
3001 //------------------------------ insert_region_before_proj -------------------------------------
3002 // Insert a region before an if projection (* - new node)
3003 //
3004 // before
3005 // if(test)
3006 // / |
3007 // v |
3008 // proj v
3009 // other-proj
3010 //
3011 // after
3012 // if(test)
3013 // / |
3014 // v |
3015 // * proj-clone v
3016 // | other-proj
3017 // v
3018 // * new-region
3019 // |
3020 // v
3021 // * dum_if
3022 // / \
3023 // v \
3024 // * dum-proj v
3025 // proj
3026 //
3027 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3028 IfNode* iff = proj->in(0)->as_If();
3029 IdealLoopTree *loop = get_loop(proj);
3030 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3031 uint ddepth = dom_depth(proj);
3032
3033 _igvn.rehash_node_delayed(iff);
3034 _igvn.rehash_node_delayed(proj);
3035
3036 proj->set_req(0, nullptr); // temporary disconnect
3037 ProjNode* proj2 = proj_clone(proj, iff);
3038 register_node(proj2, loop, iff, ddepth);
3039
3040 RegionNode* reg = new RegionNode(2);
3041 reg->set_req(1, proj2);
3042 register_node(reg, loop, iff, ddepth);
3043
3044 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3045 register_node(dum_if, loop, reg, ddepth);
3046
3047 proj->set_req(0, dum_if); // reattach
3048 set_idom(proj, dum_if, ddepth);
3049
3050 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3051 register_node(dum_proj, loop, dum_if, ddepth);
3052
3053 return reg;
3054 }
3055
3056 // Idea
3057 // ----
3058 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3059 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3060 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3061 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3062 // test alone instead:
3063 //
3064 // Before Partial Peeling:
3065 // Loop:
3066 // <peeled section>
3067 // Split off signed loop exit test
3068 // <-- CUT HERE -->
3069 // Unchanged unsigned loop exit test
3070 // <rest of unpeeled section>
3071 // goto Loop
3072 //
3073 // After Partial Peeling:
3074 // <cloned peeled section>
3075 // Cloned split off signed loop exit test
3076 // Loop:
3077 // Unchanged unsigned loop exit test
3078 // <rest of unpeeled section>
3079 // <peeled section>
3080 // Split off signed loop exit test
3081 // goto Loop
3082 //
3083 // Details
3084 // -------
3085 // Before:
3086 // if (i <u limit) Unsigned loop exit condition
3087 // / |
3088 // v v
3089 // exit-proj stay-in-loop-proj
3090 //
3091 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3092 // before the CmpU on the stay-in-loop path and keep both tests:
3093 //
3094 // if (i <u limit) Signed loop exit test
3095 // / |
3096 // / if (i <u limit) Unsigned loop exit test
3097 // / / |
3098 // v v v
3099 // exit-region stay-in-loop-proj
3100 //
3101 // Implementation
3102 // --------------
3103 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3104 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3105 // exit tests is preserved, and their loop nesting is correct.
3106 //
3107 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3108 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3109 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3110 //
3111 // if (stay-in-loop-const) Killed original unsigned loop exit test
3112 // / |
3113 // / v
3114 // / if (i < limit) Split off signed loop exit test
3115 // / / |
3116 // / / v
3117 // / / if (i <u limit) Cloned unsigned loop exit test
3118 // / / / |
3119 // v v v |
3120 // exit-region |
3121 // | |
3122 // dummy-if |
3123 // / | |
3124 // dead | |
3125 // v v
3126 // exit-proj stay-in-loop-proj
3127 //
3128 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3129 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3130 //
3131 // Requirements
3132 // ------------
3133 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3134 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3135 // Otherwise, we need to bail out (see comments in the code below).
3136 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3137 const bool Signed = true;
3138 const bool Unsigned = false;
3139
3140 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3141 if (bol->_test._test != BoolTest::lt) {
3142 return nullptr;
3143 }
3144 CmpNode* cmpu = bol->in(1)->as_Cmp();
3145 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3146
3147 int stride = stride_of_possible_iv(if_cmpu);
3148 if (stride == 0) {
3149 return nullptr;
3150 }
3151
3152 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3153 guarantee(lp_proj != nullptr, "null loop node");
3154
3155 ProjNode* lp_continue = lp_proj->as_Proj();
3156 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3157 if (!lp_exit->is_IfFalse()) {
3158 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3159 // We therefore can't add a single exit condition.
3160 return nullptr;
3161 }
3162 // The unsigned loop exit condition is
3163 // !(i <u limit)
3164 // = i >=u limit
3165 //
3166 // First, we note that for any x for which
3167 // 0 <= x <= INT_MAX
3168 // we can convert x to an unsigned int and still get the same guarantee:
3169 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3170 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3171 //
3172 // With that in mind, if
3173 // limit >= 0 (COND)
3174 // then the unsigned loop exit condition
3175 // i >=u limit (ULE)
3176 // is equivalent to
3177 // i < 0 || i >= limit (SLE-full)
3178 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3179 // (uint) i >=u MAX_INT >= limit >= 0
3180 // or otherwise
3181 // i >= limit >= 0
3182 // holds due to (LEMMA).
3183 //
3184 // For completeness, a counterexample with limit < 0:
3185 // Assume i = -3 and limit = -2:
3186 // i < 0
3187 // -2 < 0
3188 // is true and thus also "i < 0 || i >= limit". But
3189 // i >=u limit
3190 // -3 >=u -2
3191 // is false.
3192 Node* limit = cmpu->in(2);
3193 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3194 if (type_limit->_lo < 0) {
3195 return nullptr;
3196 }
3197
3198 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3199 // stride < 0:
3200 // i < 0 (SLE = SLE-negative)
3201 // stride > 0:
3202 // i >= limit (SLE = SLE-positive)
3203 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3204 //
3205 // Loop:
3206 // <peeled section>
3207 // i >= limit (SLE-positive)
3208 // <-- CUT HERE -->
3209 // i >=u limit (ULE)
3210 // <rest of unpeeled section>
3211 // goto Loop
3212 //
3213 // We exit the loop if:
3214 // (SLE) is true OR (ULE) is true
3215 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3216 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3217 // (SLE) IMPLIES (ULE)
3218 // This indeed holds when (COND) is given:
3219 // - stride > 0:
3220 // i >= limit // (SLE = SLE-positive)
3221 // i >= limit >= 0 // (COND)
3222 // i >=u limit >= 0 // (LEMMA)
3223 // which is the unsigned loop exit condition (ULE).
3224 // - stride < 0:
3225 // i < 0 // (SLE = SLE-negative)
3226 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3227 // MAX_INT >= limit >= 0 // (COND)
3228 // MAX_INT >=u limit >= 0 // (LEMMA)
3229 // and thus from (NEG) and (LEMMA):
3230 // i >=u limit
3231 // which is the unsigned loop exit condition (ULE).
3232 //
3233 //
3234 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3235 // <cloned peeled section>
3236 // i >= limit (SLE-positive)
3237 // Loop:
3238 // i >=u limit (ULE)
3239 // <rest of unpeeled section>
3240 // <peeled section>
3241 // i >= limit (SLE-positive)
3242 // goto Loop
3243 Node* rhs_cmpi;
3244 if (stride > 0) {
3245 rhs_cmpi = limit; // For i >= limit
3246 } else {
3247 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3248 }
3249 // Create a new region on the exit path
3250 RegionNode* reg = insert_region_before_proj(lp_exit);
3251 guarantee(reg != nullptr, "null region node");
3252
3253 // Clone the if-cmpu-true-false using a signed compare
3254 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3255 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3256 reg->add_req(cmpi_exit);
3257
3258 // Clone the if-cmpu-true-false
3259 BoolTest::mask rel_u = bol->_test._test;
3260 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3261 reg->add_req(cmpu_exit);
3262
3263 // Force original if to stay in loop.
3264 short_circuit_if(if_cmpu, lp_continue);
3265
3266 return cmpi_exit->in(0)->as_If();
3267 }
3268
3269 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3270 // Remove a previously inserted signed compare loop exit.
3271 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3272 Node* lp_proj = stay_in_loop(if_cmp, loop);
3273 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3274 stay_in_loop(lp_proj, loop)->is_If() &&
3275 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3276 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3277 if_cmp->set_req(1, con);
3278 }
3279
3280 //------------------------------ scheduled_nodelist -------------------------------------
3281 // Create a post order schedule of nodes that are in the
3282 // "member" set. The list is returned in "sched".
3283 // The first node in "sched" is the loop head, followed by
3284 // nodes which have no inputs in the "member" set, and then
3285 // followed by the nodes that have an immediate input dependence
3286 // on a node in "sched".
3287 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3288
3289 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3290 VectorSet visited;
3291 Node_Stack nstack(loop->_body.size());
3292
3293 Node* n = loop->_head; // top of stack is cached in "n"
3294 uint idx = 0;
3295 visited.set(n->_idx);
3296
3297 // Initially push all with no inputs from within member set
3298 for(uint i = 0; i < loop->_body.size(); i++ ) {
3299 Node *elt = loop->_body.at(i);
3300 if (member.test(elt->_idx)) {
3301 bool found = false;
3302 for (uint j = 0; j < elt->req(); j++) {
3303 Node* def = elt->in(j);
3304 if (def && member.test(def->_idx) && def != elt) {
3305 found = true;
3306 break;
3307 }
3308 }
3309 if (!found && elt != loop->_head) {
3310 nstack.push(n, idx);
3311 n = elt;
3312 assert(!visited.test(n->_idx), "not seen yet");
3313 visited.set(n->_idx);
3314 }
3315 }
3316 }
3317
3318 // traverse out's that are in the member set
3319 while (true) {
3320 if (idx < n->outcnt()) {
3321 Node* use = n->raw_out(idx);
3322 idx++;
3323 if (!visited.test_set(use->_idx)) {
3324 if (member.test(use->_idx)) {
3325 nstack.push(n, idx);
3326 n = use;
3327 idx = 0;
3328 }
3329 }
3330 } else {
3331 // All outputs processed
3332 sched.push(n);
3333 if (nstack.is_empty()) break;
3334 n = nstack.node();
3335 idx = nstack.index();
3336 nstack.pop();
3337 }
3338 }
3339 }
3340
3341
3342 //------------------------------ has_use_in_set -------------------------------------
3343 // Has a use in the vector set
3344 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3345 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3346 Node* use = n->fast_out(j);
3347 if (vset.test(use->_idx)) {
3348 return true;
3349 }
3350 }
3351 return false;
3352 }
3353
3354
3355 //------------------------------ has_use_internal_to_set -------------------------------------
3356 // Has use internal to the vector set (ie. not in a phi at the loop head)
3357 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3358 Node* head = loop->_head;
3359 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3360 Node* use = n->fast_out(j);
3361 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3362 return true;
3363 }
3364 }
3365 return false;
3366 }
3367
3368
3369 //------------------------------ clone_for_use_outside_loop -------------------------------------
3370 // clone "n" for uses that are outside of loop
3371 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3372 int cloned = 0;
3373 assert(worklist.size() == 0, "should be empty");
3374 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3375 Node* use = n->fast_out(j);
3376 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3377 worklist.push(use);
3378 }
3379 }
3380
3381 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3382 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3383 return -1;
3384 }
3385
3386 while( worklist.size() ) {
3387 Node *use = worklist.pop();
3388 if (!has_node(use) || use->in(0) == C->top()) continue;
3389 uint j;
3390 for (j = 0; j < use->req(); j++) {
3391 if (use->in(j) == n) break;
3392 }
3393 assert(j < use->req(), "must be there");
3394
3395 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3396 Node* n_clone = n->clone();
3397 _igvn.replace_input_of(use, j, n_clone);
3398 cloned++;
3399 Node* use_c;
3400 if (!use->is_Phi()) {
3401 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3402 } else {
3403 // Use in a phi is considered a use in the associated predecessor block
3404 use_c = use->in(0)->in(j);
3405 }
3406 set_ctrl(n_clone, use_c);
3407 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3408 get_loop(use_c)->_body.push(n_clone);
3409 _igvn.register_new_node_with_optimizer(n_clone);
3410 #ifndef PRODUCT
3411 if (TracePartialPeeling) {
3412 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3413 }
3414 #endif
3415 }
3416 return cloned;
3417 }
3418
3419
3420 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3421 // clone "n" for special uses that are in the not_peeled region.
3422 // If these def-uses occur in separate blocks, the code generator
3423 // marks the method as not compilable. For example, if a "BoolNode"
3424 // is in a different basic block than the "IfNode" that uses it, then
3425 // the compilation is aborted in the code generator.
3426 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3427 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3428 if (n->is_Phi() || n->is_Load()) {
3429 return;
3430 }
3431 assert(worklist.size() == 0, "should be empty");
3432 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3433 Node* use = n->fast_out(j);
3434 if ( not_peel.test(use->_idx) &&
3435 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3436 use->in(1) == n) {
3437 worklist.push(use);
3438 }
3439 }
3440 if (worklist.size() > 0) {
3441 // clone "n" and insert it between inputs of "n" and the use
3442 Node* n_clone = n->clone();
3443 loop->_body.push(n_clone);
3444 _igvn.register_new_node_with_optimizer(n_clone);
3445 set_ctrl(n_clone, get_ctrl(n));
3446 sink_list.push(n_clone);
3447 not_peel.set(n_clone->_idx);
3448 #ifndef PRODUCT
3449 if (TracePartialPeeling) {
3450 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3451 }
3452 #endif
3453 while( worklist.size() ) {
3454 Node *use = worklist.pop();
3455 _igvn.rehash_node_delayed(use);
3456 for (uint j = 1; j < use->req(); j++) {
3457 if (use->in(j) == n) {
3458 use->set_req(j, n_clone);
3459 }
3460 }
3461 }
3462 }
3463 }
3464
3465
3466 //------------------------------ insert_phi_for_loop -------------------------------------
3467 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3468 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3469 Node *phi = PhiNode::make(lp, back_edge_val);
3470 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3471 // Use existing phi if it already exists
3472 Node *hit = _igvn.hash_find_insert(phi);
3473 if( hit == nullptr ) {
3474 _igvn.register_new_node_with_optimizer(phi);
3475 set_ctrl(phi, lp);
3476 } else {
3477 // Remove the new phi from the graph and use the hit
3478 _igvn.remove_dead_node(phi);
3479 phi = hit;
3480 }
3481 _igvn.replace_input_of(use, idx, phi);
3482 }
3483
3484 #ifdef ASSERT
3485 //------------------------------ is_valid_loop_partition -------------------------------------
3486 // Validate the loop partition sets: peel and not_peel
3487 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3488 VectorSet& not_peel ) {
3489 uint i;
3490 // Check that peel_list entries are in the peel set
3491 for (i = 0; i < peel_list.size(); i++) {
3492 if (!peel.test(peel_list.at(i)->_idx)) {
3493 return false;
3494 }
3495 }
3496 // Check at loop members are in one of peel set or not_peel set
3497 for (i = 0; i < loop->_body.size(); i++ ) {
3498 Node *def = loop->_body.at(i);
3499 uint di = def->_idx;
3500 // Check that peel set elements are in peel_list
3501 if (peel.test(di)) {
3502 if (not_peel.test(di)) {
3503 return false;
3504 }
3505 // Must be in peel_list also
3506 bool found = false;
3507 for (uint j = 0; j < peel_list.size(); j++) {
3508 if (peel_list.at(j)->_idx == di) {
3509 found = true;
3510 break;
3511 }
3512 }
3513 if (!found) {
3514 return false;
3515 }
3516 } else if (not_peel.test(di)) {
3517 if (peel.test(di)) {
3518 return false;
3519 }
3520 } else {
3521 return false;
3522 }
3523 }
3524 return true;
3525 }
3526
3527 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3528 // Ensure a use outside of loop is of the right form
3529 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3530 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3531 return (use->is_Phi() &&
3532 use_c->is_Region() && use_c->req() == 3 &&
3533 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3534 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3535 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3536 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3537 }
3538
3539 //------------------------------ is_valid_clone_loop_form -------------------------------------
3540 // Ensure that all uses outside of loop are of the right form
3541 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3542 uint orig_exit_idx, uint clone_exit_idx) {
3543 uint len = peel_list.size();
3544 for (uint i = 0; i < len; i++) {
3545 Node *def = peel_list.at(i);
3546
3547 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3548 Node *use = def->fast_out(j);
3549 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3550 if (!loop->is_member(get_loop(use_c))) {
3551 // use is not in the loop, check for correct structure
3552 if (use->in(0) == def) {
3553 // Okay
3554 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3555 return false;
3556 }
3557 }
3558 }
3559 }
3560 return true;
3561 }
3562 #endif
3563
3564 //------------------------------ partial_peel -------------------------------------
3565 // Partially peel (aka loop rotation) the top portion of a loop (called
3566 // the peel section below) by cloning it and placing one copy just before
3567 // the new loop head and the other copy at the bottom of the new loop.
3568 //
3569 // before after where it came from
3570 //
3571 // stmt1 stmt1
3572 // loop: stmt2 clone
3573 // stmt2 if condA goto exitA clone
3574 // if condA goto exitA new_loop: new
3575 // stmt3 stmt3 clone
3576 // if !condB goto loop if condB goto exitB clone
3577 // exitB: stmt2 orig
3578 // stmt4 if !condA goto new_loop orig
3579 // exitA: goto exitA
3580 // exitB:
3581 // stmt4
3582 // exitA:
3583 //
3584 // Step 1: find the cut point: an exit test on probable
3585 // induction variable.
3586 // Step 2: schedule (with cloning) operations in the peel
3587 // section that can be executed after the cut into
3588 // the section that is not peeled. This may need
3589 // to clone operations into exit blocks. For
3590 // instance, a reference to A[i] in the not-peel
3591 // section and a reference to B[i] in an exit block
3592 // may cause a left-shift of i by 2 to be placed
3593 // in the peel block. This step will clone the left
3594 // shift into the exit block and sink the left shift
3595 // from the peel to the not-peel section.
3596 // Step 3: clone the loop, retarget the control, and insert
3597 // phis for values that are live across the new loop
3598 // head. This is very dependent on the graph structure
3599 // from clone_loop. It creates region nodes for
3600 // exit control and associated phi nodes for values
3601 // flow out of the loop through that exit. The region
3602 // node is dominated by the clone's control projection.
3603 // So the clone's peel section is placed before the
3604 // new loop head, and the clone's not-peel section is
3605 // forms the top part of the new loop. The original
3606 // peel section forms the tail of the new loop.
3607 // Step 4: update the dominator tree and recompute the
3608 // dominator depth.
3609 //
3610 // orig
3611 //
3612 // stmt1
3613 // |
3614 // v
3615 // predicates
3616 // |
3617 // v
3618 // loop<----+
3619 // | |
3620 // stmt2 |
3621 // | |
3622 // v |
3623 // ifA |
3624 // / | |
3625 // v v |
3626 // false true ^ <-- last_peel
3627 // / | |
3628 // / ===|==cut |
3629 // / stmt3 | <-- first_not_peel
3630 // / | |
3631 // | v |
3632 // v ifB |
3633 // exitA: / \ |
3634 // / \ |
3635 // v v |
3636 // false true |
3637 // / \ |
3638 // / ----+
3639 // |
3640 // v
3641 // exitB:
3642 // stmt4
3643 //
3644 //
3645 // after clone loop
3646 //
3647 // stmt1
3648 // |
3649 // v
3650 // predicates
3651 // / \
3652 // clone / \ orig
3653 // / \
3654 // / \
3655 // v v
3656 // +---->loop loop<----+
3657 // | | | |
3658 // | stmt2 stmt2 |
3659 // | | | |
3660 // | v v |
3661 // | ifA ifA |
3662 // | | \ / | |
3663 // | v v v v |
3664 // ^ true false false true ^ <-- last_peel
3665 // | | ^ \ / | |
3666 // | cut==|== \ \ / ===|==cut |
3667 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3668 // | | dom | | | |
3669 // | v \ 1v v2 v |
3670 // | ifB regionA ifB |
3671 // | / \ | / \ |
3672 // | / \ v / \ |
3673 // | v v exitA: v v |
3674 // | true false false true |
3675 // | / ^ \ / \ |
3676 // +---- \ \ / ----+
3677 // dom \ /
3678 // \ 1v v2
3679 // regionB
3680 // |
3681 // v
3682 // exitB:
3683 // stmt4
3684 //
3685 //
3686 // after partial peel
3687 //
3688 // stmt1
3689 // |
3690 // v
3691 // predicates
3692 // /
3693 // clone / orig
3694 // / TOP
3695 // / \
3696 // v v
3697 // TOP->loop loop----+
3698 // | | |
3699 // stmt2 stmt2 |
3700 // | | |
3701 // v v |
3702 // ifA ifA |
3703 // | \ / | |
3704 // v v v v |
3705 // true false false true | <-- last_peel
3706 // | ^ \ / +------|---+
3707 // +->newloop \ \ / === ==cut | |
3708 // | stmt3 \ \ / TOP | |
3709 // | | dom | | stmt3 | | <-- first_not_peel
3710 // | v \ 1v v2 v | |
3711 // | ifB regionA ifB ^ v
3712 // | / \ | / \ | |
3713 // | / \ v / \ | |
3714 // | v v exitA: v v | |
3715 // | true false false true | |
3716 // | / ^ \ / \ | |
3717 // | | \ \ / v | |
3718 // | | dom \ / TOP | |
3719 // | | \ 1v v2 | |
3720 // ^ v regionB | |
3721 // | | | | |
3722 // | | v ^ v
3723 // | | exitB: | |
3724 // | | stmt4 | |
3725 // | +------------>-----------------+ |
3726 // | |
3727 // +-----------------<---------------------+
3728 //
3729 //
3730 // final graph
3731 //
3732 // stmt1
3733 // |
3734 // v
3735 // predicates
3736 // |
3737 // v
3738 // stmt2 clone
3739 // |
3740 // v
3741 // ........> ifA clone
3742 // : / |
3743 // dom / |
3744 // : v v
3745 // : false true
3746 // : | |
3747 // : | v
3748 // : | newloop<-----+
3749 // : | | |
3750 // : | stmt3 clone |
3751 // : | | |
3752 // : | v |
3753 // : | ifB |
3754 // : | / \ |
3755 // : | v v |
3756 // : | false true |
3757 // : | | | |
3758 // : | v stmt2 |
3759 // : | exitB: | |
3760 // : | stmt4 v |
3761 // : | ifA orig |
3762 // : | / \ |
3763 // : | / \ |
3764 // : | v v |
3765 // : | false true |
3766 // : | / \ |
3767 // : v v -----+
3768 // RegionA
3769 // |
3770 // v
3771 // exitA
3772 //
3773 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3774
3775 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3776 if (!loop->_head->is_Loop()) {
3777 return false;
3778 }
3779 LoopNode *head = loop->_head->as_Loop();
3780
3781 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3782 return false;
3783 }
3784
3785 // Check for complex exit control
3786 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3787 Node *n = loop->_body.at(ii);
3788 int opc = n->Opcode();
3789 if (n->is_Call() ||
3790 opc == Op_Catch ||
3791 opc == Op_CatchProj ||
3792 opc == Op_Jump ||
3793 opc == Op_JumpProj) {
3794 #ifndef PRODUCT
3795 if (TracePartialPeeling) {
3796 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3797 }
3798 #endif
3799 return false;
3800 }
3801 }
3802
3803 int dd = dom_depth(head);
3804
3805 // Step 1: find cut point
3806
3807 // Walk up dominators to loop head looking for first loop exit
3808 // which is executed on every path thru loop.
3809 IfNode *peel_if = nullptr;
3810 IfNode *peel_if_cmpu = nullptr;
3811
3812 Node *iff = loop->tail();
3813 while (iff != head) {
3814 if (iff->is_If()) {
3815 Node *ctrl = get_ctrl(iff->in(1));
3816 if (ctrl->is_top()) return false; // Dead test on live IF.
3817 // If loop-varying exit-test, check for induction variable
3818 if (loop->is_member(get_loop(ctrl)) &&
3819 loop->is_loop_exit(iff) &&
3820 is_possible_iv_test(iff)) {
3821 Node* cmp = iff->in(1)->in(1);
3822 if (cmp->Opcode() == Op_CmpI) {
3823 peel_if = iff->as_If();
3824 } else {
3825 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3826 peel_if_cmpu = iff->as_If();
3827 }
3828 }
3829 }
3830 iff = idom(iff);
3831 }
3832
3833 // Prefer signed compare over unsigned compare.
3834 IfNode* new_peel_if = nullptr;
3835 if (peel_if == nullptr) {
3836 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3837 return false; // No peel point found
3838 }
3839 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3840 if (new_peel_if == nullptr) {
3841 return false; // No peel point found
3842 }
3843 peel_if = new_peel_if;
3844 }
3845 Node* last_peel = stay_in_loop(peel_if, loop);
3846 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3847 if (first_not_peeled == nullptr || first_not_peeled == head) {
3848 return false;
3849 }
3850
3851 #ifndef PRODUCT
3852 if (TraceLoopOpts) {
3853 tty->print("PartialPeel ");
3854 loop->dump_head();
3855 }
3856
3857 if (TracePartialPeeling) {
3858 tty->print_cr("before partial peel one iteration");
3859 Node_List wl;
3860 Node* t = head->in(2);
3861 while (true) {
3862 wl.push(t);
3863 if (t == head) break;
3864 t = idom(t);
3865 }
3866 while (wl.size() > 0) {
3867 Node* tt = wl.pop();
3868 tt->dump();
3869 if (tt == last_peel) tty->print_cr("-- cut --");
3870 }
3871 }
3872 #endif
3873
3874 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3875
3876 VectorSet peel;
3877 VectorSet not_peel;
3878 Node_List peel_list;
3879 Node_List worklist;
3880 Node_List sink_list;
3881
3882 uint estimate = loop->est_loop_clone_sz(1);
3883 if (exceeding_node_budget(estimate)) {
3884 return false;
3885 }
3886
3887 // Set of cfg nodes to peel are those that are executable from
3888 // the head through last_peel.
3889 assert(worklist.size() == 0, "should be empty");
3890 worklist.push(head);
3891 peel.set(head->_idx);
3892 while (worklist.size() > 0) {
3893 Node *n = worklist.pop();
3894 if (n != last_peel) {
3895 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3896 Node* use = n->fast_out(j);
3897 if (use->is_CFG() &&
3898 loop->is_member(get_loop(use)) &&
3899 !peel.test_set(use->_idx)) {
3900 worklist.push(use);
3901 }
3902 }
3903 }
3904 }
3905
3906 // Set of non-cfg nodes to peel are those that are control
3907 // dependent on the cfg nodes.
3908 for (uint i = 0; i < loop->_body.size(); i++) {
3909 Node *n = loop->_body.at(i);
3910 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3911 if (peel.test(n_c->_idx)) {
3912 peel.set(n->_idx);
3913 } else {
3914 not_peel.set(n->_idx);
3915 }
3916 }
3917
3918 // Step 2: move operations from the peeled section down into the
3919 // not-peeled section
3920
3921 // Get a post order schedule of nodes in the peel region
3922 // Result in right-most operand.
3923 scheduled_nodelist(loop, peel, peel_list);
3924
3925 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3926
3927 // For future check for too many new phis
3928 uint old_phi_cnt = 0;
3929 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3930 Node* use = head->fast_out(j);
3931 if (use->is_Phi()) old_phi_cnt++;
3932 }
3933
3934 #ifndef PRODUCT
3935 if (TracePartialPeeling) {
3936 tty->print_cr("\npeeled list");
3937 }
3938 #endif
3939
3940 // Evacuate nodes in peel region into the not_peeled region if possible
3941 bool too_many_clones = false;
3942 uint new_phi_cnt = 0;
3943 uint cloned_for_outside_use = 0;
3944 for (uint i = 0; i < peel_list.size();) {
3945 Node* n = peel_list.at(i);
3946 #ifndef PRODUCT
3947 if (TracePartialPeeling) n->dump();
3948 #endif
3949 bool incr = true;
3950 if (!n->is_CFG()) {
3951 if (has_use_in_set(n, not_peel)) {
3952 // If not used internal to the peeled region,
3953 // move "n" from peeled to not_peeled region.
3954 if (!has_use_internal_to_set(n, peel, loop)) {
3955 // if not pinned and not a load (which maybe anti-dependent on a store)
3956 // and not a CMove (Matcher expects only bool->cmove).
3957 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3958 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3959 if (C->failing()) return false;
3960 if (new_clones == -1) {
3961 too_many_clones = true;
3962 break;
3963 }
3964 cloned_for_outside_use += new_clones;
3965 sink_list.push(n);
3966 peel.remove(n->_idx);
3967 not_peel.set(n->_idx);
3968 peel_list.remove(i);
3969 incr = false;
3970 #ifndef PRODUCT
3971 if (TracePartialPeeling) {
3972 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3973 n->_idx, get_ctrl(n)->_idx);
3974 }
3975 #endif
3976 }
3977 } else {
3978 // Otherwise check for special def-use cases that span
3979 // the peel/not_peel boundary such as bool->if
3980 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3981 new_phi_cnt++;
3982 }
3983 }
3984 }
3985 if (incr) i++;
3986 }
3987
3988 estimate += cloned_for_outside_use + new_phi_cnt;
3989 bool exceed_node_budget = !may_require_nodes(estimate);
3990 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3991
3992 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3993 #ifndef PRODUCT
3994 if (TracePartialPeeling && exceed_phi_limit) {
3995 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3996 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3997 }
3998 #endif
3999 if (new_peel_if != nullptr) {
4000 remove_cmpi_loop_exit(new_peel_if, loop);
4001 }
4002 // Inhibit more partial peeling on this loop
4003 assert(!head->is_partial_peel_loop(), "not partial peeled");
4004 head->mark_partial_peel_failed();
4005 if (cloned_for_outside_use > 0) {
4006 // Terminate this round of loop opts because
4007 // the graph outside this loop was changed.
4008 C->set_major_progress();
4009 return true;
4010 }
4011 return false;
4012 }
4013
4014 // Step 3: clone loop, retarget control, and insert new phis
4015
4016 // Create new loop head for new phis and to hang
4017 // the nodes being moved (sinked) from the peel region.
4018 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4019 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4020 _igvn.register_new_node_with_optimizer(new_head);
4021 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4022 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4023 set_loop(new_head, loop);
4024 loop->_body.push(new_head);
4025 not_peel.set(new_head->_idx);
4026 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4027 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4028
4029 while (sink_list.size() > 0) {
4030 Node* n = sink_list.pop();
4031 set_ctrl(n, new_head);
4032 }
4033
4034 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4035
4036 clone_loop(loop, old_new, dd, IgnoreStripMined);
4037
4038 const uint clone_exit_idx = 1;
4039 const uint orig_exit_idx = 2;
4040 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4041
4042 Node* head_clone = old_new[head->_idx];
4043 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4044 Node* orig_tail_clone = head_clone->in(2);
4045
4046 // Add phi if "def" node is in peel set and "use" is not
4047
4048 for (uint i = 0; i < peel_list.size(); i++) {
4049 Node *def = peel_list.at(i);
4050 if (!def->is_CFG()) {
4051 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4052 Node *use = def->fast_out(j);
4053 if (has_node(use) && use->in(0) != C->top() &&
4054 (!peel.test(use->_idx) ||
4055 (use->is_Phi() && use->in(0) == head)) ) {
4056 worklist.push(use);
4057 }
4058 }
4059 while( worklist.size() ) {
4060 Node *use = worklist.pop();
4061 for (uint j = 1; j < use->req(); j++) {
4062 Node* n = use->in(j);
4063 if (n == def) {
4064
4065 // "def" is in peel set, "use" is not in peel set
4066 // or "use" is in the entry boundary (a phi) of the peel set
4067
4068 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4069
4070 if ( loop->is_member(get_loop( use_c )) ) {
4071 // use is in loop
4072 if (old_new[use->_idx] != nullptr) { // null for dead code
4073 Node* use_clone = old_new[use->_idx];
4074 _igvn.replace_input_of(use, j, C->top());
4075 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4076 }
4077 } else {
4078 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4079 // use is not in the loop, check if the live range includes the cut
4080 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4081 if (not_peel.test(lp_if->_idx)) {
4082 assert(j == orig_exit_idx, "use from original loop");
4083 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4084 }
4085 }
4086 }
4087 }
4088 }
4089 }
4090 }
4091
4092 // Step 3b: retarget control
4093
4094 // Redirect control to the new loop head if a cloned node in
4095 // the not_peeled region has control that points into the peeled region.
4096 // This necessary because the cloned peeled region will be outside
4097 // the loop.
4098 // from to
4099 // cloned-peeled <---+
4100 // new_head_clone: | <--+
4101 // cloned-not_peeled in(0) in(0)
4102 // orig-peeled
4103
4104 for (uint i = 0; i < loop->_body.size(); i++) {
4105 Node *n = loop->_body.at(i);
4106 if (!n->is_CFG() && n->in(0) != nullptr &&
4107 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4108 Node* n_clone = old_new[n->_idx];
4109 if (n_clone->depends_only_on_test()) {
4110 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4111 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4112 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4113 // floating above its range check.
4114 Node* pinned_clone = n_clone->pin_array_access_node();
4115 if (pinned_clone != nullptr) {
4116 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4117 old_new.map(n->_idx, pinned_clone);
4118 _igvn.replace_node(n_clone, pinned_clone);
4119 n_clone = pinned_clone;
4120 }
4121 }
4122 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4123 }
4124 }
4125
4126 // Backedge of the surviving new_head (the clone) is original last_peel
4127 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4128
4129 // Cut first node in original not_peel set
4130 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4131 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4132 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4133
4134 // Copy head_clone back-branch info to original head
4135 // and remove original head's loop entry and
4136 // clone head's back-branch
4137 _igvn.rehash_node_delayed(head); // Multiple edge updates
4138 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4139 head->set_req(LoopNode::LoopBackControl, C->top());
4140 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4141
4142 // Similarly modify the phis
4143 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4144 Node* use = head->fast_out(k);
4145 if (use->is_Phi() && use->outcnt() > 0) {
4146 Node* use_clone = old_new[use->_idx];
4147 _igvn.rehash_node_delayed(use); // Multiple edge updates
4148 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4149 use->set_req(LoopNode::LoopBackControl, C->top());
4150 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4151 }
4152 }
4153
4154 // Step 4: update dominator tree and dominator depth
4155
4156 set_idom(head, orig_tail_clone, dd);
4157 recompute_dom_depth();
4158
4159 // Inhibit more partial peeling on this loop
4160 new_head_clone->set_partial_peel_loop();
4161 C->set_major_progress();
4162 loop->record_for_igvn();
4163
4164 #ifndef PRODUCT
4165 if (TracePartialPeeling) {
4166 tty->print_cr("\nafter partial peel one iteration");
4167 Node_List wl;
4168 Node* t = last_peel;
4169 while (true) {
4170 wl.push(t);
4171 if (t == head_clone) break;
4172 t = idom(t);
4173 }
4174 while (wl.size() > 0) {
4175 Node* tt = wl.pop();
4176 if (tt == head) tty->print_cr("orig head");
4177 else if (tt == new_head_clone) tty->print_cr("new head");
4178 else if (tt == head_clone) tty->print_cr("clone head");
4179 tt->dump();
4180 }
4181 }
4182 #endif
4183
4184 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4185
4186 return true;
4187 }
4188
4189 // Transform:
4190 //
4191 // loop<-----------------+
4192 // | |
4193 // stmt1 stmt2 .. stmtn |
4194 // | | | |
4195 // \ | / |
4196 // v v v |
4197 // region |
4198 // | |
4199 // shared_stmt |
4200 // | |
4201 // v |
4202 // if |
4203 // / \ |
4204 // | -----------+
4205 // v
4206 //
4207 // into:
4208 //
4209 // loop<-------------------+
4210 // | |
4211 // v |
4212 // +->loop |
4213 // | | |
4214 // | stmt1 stmt2 .. stmtn |
4215 // | | | | |
4216 // | | \ / |
4217 // | | v v |
4218 // | | region1 |
4219 // | | | |
4220 // | shared_stmt shared_stmt |
4221 // | | | |
4222 // | v v |
4223 // | if if |
4224 // | /\ / \ |
4225 // +-- | | -------+
4226 // \ /
4227 // v v
4228 // region2
4229 //
4230 // (region2 is shown to merge mirrored projections of the loop exit
4231 // ifs to make the diagram clearer but they really merge the same
4232 // projection)
4233 //
4234 // Conditions for this transformation to trigger:
4235 // - the path through stmt1 is frequent enough
4236 // - the inner loop will be turned into a counted loop after transformation
4237 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4238 if (!DuplicateBackedge) {
4239 return false;
4240 }
4241 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4242 if (!loop->_head->is_Loop()) {
4243 return false;
4244 }
4245
4246 uint estimate = loop->est_loop_clone_sz(1);
4247 if (exceeding_node_budget(estimate)) {
4248 return false;
4249 }
4250
4251 LoopNode *head = loop->_head->as_Loop();
4252
4253 Node* region = nullptr;
4254 IfNode* exit_test = nullptr;
4255 uint inner;
4256 float f;
4257 if (StressDuplicateBackedge) {
4258 if (head->is_strip_mined()) {
4259 return false;
4260 }
4261 Node* c = head->in(LoopNode::LoopBackControl);
4262
4263 while (c != head) {
4264 if (c->is_Region()) {
4265 region = c;
4266 }
4267 c = idom(c);
4268 }
4269
4270 if (region == nullptr) {
4271 return false;
4272 }
4273
4274 inner = 1;
4275 } else {
4276 // Is the shape of the loop that of a counted loop...
4277 Node* back_control = loop_exit_control(head, loop);
4278 if (back_control == nullptr) {
4279 return false;
4280 }
4281
4282 BoolTest::mask bt = BoolTest::illegal;
4283 float cl_prob = 0;
4284 Node* incr = nullptr;
4285 Node* limit = nullptr;
4286 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4287 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4288 return false;
4289 }
4290
4291 // With an extra phi for the candidate iv?
4292 // Or the region node is the loop head
4293 if (!incr->is_Phi() || incr->in(0) == head) {
4294 return false;
4295 }
4296
4297 PathFrequency pf(head, this);
4298 region = incr->in(0);
4299
4300 // Go over all paths for the extra phi's region and see if that
4301 // path is frequent enough and would match the expected iv shape
4302 // if the extra phi is removed
4303 inner = 0;
4304 for (uint i = 1; i < incr->req(); ++i) {
4305 Node* in = incr->in(i);
4306 Node* trunc1 = nullptr;
4307 Node* trunc2 = nullptr;
4308 const TypeInteger* iv_trunc_t = nullptr;
4309 Node* orig_in = in;
4310 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4311 continue;
4312 }
4313 assert(in->Opcode() == Op_AddI, "wrong increment code");
4314 Node* xphi = nullptr;
4315 Node* stride = loop_iv_stride(in, xphi);
4316
4317 if (stride == nullptr) {
4318 continue;
4319 }
4320
4321 PhiNode* phi = loop_iv_phi(xphi, nullptr, head);
4322 if (phi == nullptr ||
4323 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4324 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4325 return false;
4326 }
4327
4328 f = pf.to(region->in(i));
4329 if (f > 0.5) {
4330 inner = i;
4331 break;
4332 }
4333 }
4334
4335 if (inner == 0) {
4336 return false;
4337 }
4338
4339 exit_test = back_control->in(0)->as_If();
4340 }
4341
4342 if (idom(region)->is_Catch()) {
4343 return false;
4344 }
4345
4346 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4347 Unique_Node_List wq;
4348 wq.push(head->in(LoopNode::LoopBackControl));
4349 for (uint i = 0; i < wq.size(); i++) {
4350 Node* c = wq.at(i);
4351 assert(get_loop(c) == loop, "not in the right loop?");
4352 if (c->is_Region()) {
4353 if (c != region) {
4354 for (uint j = 1; j < c->req(); ++j) {
4355 wq.push(c->in(j));
4356 }
4357 }
4358 } else {
4359 wq.push(c->in(0));
4360 }
4361 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4362 }
4363
4364 Node* region_dom = idom(region);
4365
4366 // Can't do the transformation if this would cause a membar pair to
4367 // be split
4368 for (uint i = 0; i < wq.size(); i++) {
4369 Node* c = wq.at(i);
4370 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4371 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4372 if (!wq.member(c->as_MemBar()->leading_membar())) {
4373 return false;
4374 }
4375 }
4376 }
4377 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head);
4378
4379 // Collect data nodes that need to be clones as well
4380 int dd = dom_depth(head);
4381
4382 for (uint i = 0; i < loop->_body.size(); ++i) {
4383 Node* n = loop->_body.at(i);
4384 if (has_ctrl(n)) {
4385 Node* c = get_ctrl(n);
4386 if (wq.member(c)) {
4387 wq.push(n);
4388 }
4389 } else {
4390 set_idom(n, idom(n), dd);
4391 }
4392 }
4393
4394 // clone shared_stmt
4395 clone_loop_body(wq, old_new, nullptr);
4396
4397 Node* region_clone = old_new[region->_idx];
4398 region_clone->set_req(inner, C->top());
4399 set_idom(region, region->in(inner), dd);
4400
4401 // Prepare the outer loop
4402 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4403 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4404 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4405 set_idom(head, outer_head, dd);
4406
4407 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4408
4409 // Make one of the shared_stmt copies only reachable from stmt1, the
4410 // other only from stmt2..stmtn.
4411 Node* dom = nullptr;
4412 for (uint i = 1; i < region->req(); ++i) {
4413 if (i != inner) {
4414 _igvn.replace_input_of(region, i, C->top());
4415 }
4416 Node* in = region_clone->in(i);
4417 if (in->is_top()) {
4418 continue;
4419 }
4420 if (dom == nullptr) {
4421 dom = in;
4422 } else {
4423 dom = dom_lca(dom, in);
4424 }
4425 }
4426
4427 set_idom(region_clone, dom, dd);
4428
4429 // Set up the outer loop
4430 for (uint i = 0; i < head->outcnt(); i++) {
4431 Node* u = head->raw_out(i);
4432 if (u->is_Phi()) {
4433 Node* outer_phi = u->clone();
4434 outer_phi->set_req(0, outer_head);
4435 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4436 if (backedge == nullptr) {
4437 backedge = u->in(LoopNode::LoopBackControl);
4438 }
4439 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4440 register_new_node(outer_phi, outer_head);
4441 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4442 }
4443 }
4444
4445 // create control and data nodes for out of loop uses (including region2)
4446 Node_List worklist;
4447 uint new_counter = C->unique();
4448 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4449
4450 Node_List *split_if_set = nullptr;
4451 Node_List *split_bool_set = nullptr;
4452 Node_List *split_cex_set = nullptr;
4453 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4454 split_if_set, split_bool_set, split_cex_set);
4455
4456 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4457
4458 if (exit_test != nullptr) {
4459 float cnt = exit_test->_fcnt;
4460 if (cnt != COUNT_UNKNOWN) {
4461 exit_test->_fcnt = cnt * f;
4462 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4463 }
4464 }
4465
4466 C->set_major_progress();
4467
4468 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head);
4469
4470 return true;
4471 }
4472
4473 // AutoVectorize the loop: replace scalar ops with vector ops.
4474 PhaseIdealLoop::AutoVectorizeStatus
4475 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4476 // Counted loop only
4477 if (!lpt->is_counted()) {
4478 return AutoVectorizeStatus::Impossible;
4479 }
4480
4481 // Main-loop only
4482 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4483 if (!cl->is_main_loop()) {
4484 return AutoVectorizeStatus::Impossible;
4485 }
4486
4487 VLoop vloop(lpt, false);
4488 if (!vloop.check_preconditions()) {
4489 return AutoVectorizeStatus::TriedAndFailed;
4490 }
4491
4492 // Ensure the shared data is cleared before each use
4493 vshared.clear();
4494
4495 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4496 if (!vloop_analyzer.success()) {
4497 return AutoVectorizeStatus::TriedAndFailed;
4498 }
4499
4500 SuperWord sw(vloop_analyzer);
4501 if (!sw.transform_loop()) {
4502 return AutoVectorizeStatus::TriedAndFailed;
4503 }
4504
4505 return AutoVectorizeStatus::Success;
4506 }
4507
4508 // Just before insert_pre_post_loops, we can multiversion the loop:
4509 //
4510 // multiversion_if
4511 // | |
4512 // fast_loop slow_loop
4513 //
4514 // In the fast_loop we can make speculative assumptions, and put the
4515 // conditions into the multiversion_if. If the conditions hold at runtime,
4516 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4517 // instead which does not make any of the speculative assumptions.
4518 //
4519 // Note: we only multiversion the loop if the loop does not have any
4520 // auto vectorization check Predicate. If we have that predicate,
4521 // then we can simply add the speculative assumption checks to
4522 // that Predicate. This means we do not need to duplicate the
4523 // loop - we have a smaller graph and save compile time. Should
4524 // the conditions ever fail, then we deopt / trap at the Predicate
4525 // and recompile without that Predicate. At that point we will
4526 // multiversion the loop, so that we can still have speculative
4527 // runtime checks.
4528 //
4529 // We perform the multiversioning when the loop is still in its single
4530 // iteration form, even before we insert pre and post loops. This makes
4531 // the cloning much simpler. However, this means that both the fast
4532 // and the slow loop have to be optimized independently (adding pre
4533 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4534 // we may end up not needing any speculative assumptions in the fast_loop
4535 // and then rejecting the slow_loop by constant folding the multiversion_if.
4536 //
4537 // Therefore, we "delay" the optimization of the slow_loop until we add
4538 // at least one speculative assumption for the fast_loop. If we never
4539 // add such a speculative runtime check, the OpaqueMultiversioningNode
4540 // of the multiversion_if constant folds to true after loop opts, and the
4541 // multiversion_if folds away the "delayed" slow_loop. If we add any
4542 // speculative assumption, then we notify the OpaqueMultiversioningNode
4543 // with "notify_slow_loop_that_it_can_resume_optimizations".
4544 //
4545 // Note: new runtime checks can be added to the multiversion_if with
4546 // PhaseIdealLoop::create_new_if_for_multiversion
4547 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4548 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4549 LoopNode* outer_loop = cl->skip_strip_mined();
4550 Node* entry = outer_loop->in(LoopNode::EntryControl);
4551
4552 // Check we have multiversioning enabled, and are not already multiversioned.
4553 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4554
4555 // Check that we do not have a parse-predicate where we can add the runtime checks
4556 // during auto-vectorization.
4557 const Predicates predicates(entry);
4558 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4559 if (predicate_block->has_parse_predicate()) { return; }
4560
4561 // Check node budget.
4562 uint estimate = lpt->est_loop_clone_sz(2);
4563 if (!may_require_nodes(estimate)) { return; }
4564
4565 do_multiversioning(lpt, old_new);
4566 }
4567
4568 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4569 for (uint i = 0; i < _data_nodes.size(); i++) {
4570 clone(_data_nodes[i], new_ctrl);
4571 }
4572 }
4573
4574 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4575 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4576 Node* clone = node->clone();
4577 _phase->igvn().register_new_node_with_optimizer(clone);
4578 _orig_to_new.put(node, clone);
4579 _phase->set_ctrl(clone, new_ctrl);
4580 if (node->is_CastII()) {
4581 clone->set_req(0, new_ctrl);
4582 }
4583 }
4584
4585 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4586 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4587 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4588 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4589 for (uint i = 1; i < node->req(); i++) {
4590 Node** cloned_input = _orig_to_new.get(node->in(i));
4591 if (cloned_input != nullptr) {
4592 // Input was also cloned -> rewire clone to the cloned input.
4593 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4594 }
4595 }
4596 });
4597 }
4598
4599 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4600 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4601 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4602 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4603 Node* new_ctrl) {
4604 for (uint i = 0; i < _data_nodes.size(); i++) {
4605 Node* data_node = _data_nodes[i];
4606 if (data_node->is_Opaque1()) {
4607 transform_opaque_node(transform_strategy, data_node);
4608 } else {
4609 clone(data_node, new_ctrl);
4610 }
4611 }
4612 }
4613
4614 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4615 Node* transformed_node;
4616 if (node->is_OpaqueLoopInit()) {
4617 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4618 } else {
4619 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4620 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4621 }
4622 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4623 _orig_to_new.put(node, transformed_node);
4624 }