1 /*
2 * Copyright (c) 1999, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "opto/superword.hpp"
45 #include "opto/vectornode.hpp"
46 #include "utilities/macros.hpp"
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
52 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
53 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
54 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
55 // so disable this for now
56 return nullptr;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
63 n->in(1) == region->as_CountedLoop()->phi()) {
64 return nullptr;
65 }
66
67 if (cannot_split_division(n, region)) {
68 return nullptr;
69 }
70
71 int wins = 0;
72 assert(!n->is_CFG(), "");
73 assert(region->is_Region(), "");
74
75 const Type* type = n->bottom_type();
76 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
77 Node* phi;
78 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
79 int iid = t_oop->instance_id();
80 int index = C->get_alias_index(t_oop);
81 int offset = t_oop->offset();
82 phi = new PhiNode(region, type, nullptr, iid, index, offset);
83 } else {
84 phi = PhiNode::make_blank(region, n);
85 }
86 uint old_unique = C->unique();
87 for (uint i = 1; i < region->req(); i++) {
88 Node* x;
89 Node* the_clone = nullptr;
90 if (region->in(i) == C->top()) {
91 x = C->top(); // Dead path? Use a dead data op
92 } else {
93 x = n->clone(); // Else clone up the data op
94 the_clone = x; // Remember for possible deletion.
95 // Alter data node to use pre-phi inputs
96 if (n->in(0) == region)
97 x->set_req( 0, region->in(i) );
98 for (uint j = 1; j < n->req(); j++) {
99 Node* in = n->in(j);
100 if (in->is_Phi() && in->in(0) == region)
101 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
102 }
103 }
104 // Check for a 'win' on some paths
105 const Type* t = x->Value(&_igvn);
106
107 bool singleton = t->singleton();
108
109 // A TOP singleton indicates that there are no possible values incoming
110 // along a particular edge. In most cases, this is OK, and the Phi will
111 // be eliminated later in an Ideal call. However, we can't allow this to
112 // happen if the singleton occurs on loop entry, as the elimination of
113 // the PhiNode may cause the resulting node to migrate back to a previous
114 // loop iteration.
115 if (singleton && t == Type::TOP) {
116 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
117 // irreducible loop may not be indicated by an affirmative is_Loop());
118 // therefore, the only top we can split thru a phi is on a backedge of
119 // a loop.
120 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
121 }
122
123 if (singleton) {
124 wins++;
125 x = ((PhaseGVN&)_igvn).makecon(t);
126 } else {
127 // We now call Identity to try to simplify the cloned node.
128 // Note that some Identity methods call phase->type(this).
129 // Make sure that the type array is big enough for
130 // our new node, even though we may throw the node away.
131 // (Note: This tweaking with igvn only works because x is a new node.)
132 _igvn.set_type(x, t);
133 // If x is a TypeNode, capture any more-precise type permanently into Node
134 // otherwise it will be not updated during igvn->transform since
135 // igvn->type(x) is set to x->Value() already.
136 x->raise_bottom_type(t);
137 Node* y = x->Identity(&_igvn);
138 if (y != x) {
139 wins++;
140 x = y;
141 } else {
142 y = _igvn.hash_find(x);
143 if (y == nullptr) {
144 y = similar_subtype_check(x, region->in(i));
145 }
146 if (y) {
147 wins++;
148 x = y;
149 } else {
150 // Else x is a new node we are keeping
151 // We do not need register_new_node_with_optimizer
152 // because set_type has already been called.
153 _igvn._worklist.push(x);
154 }
155 }
156 }
157
158 phi->set_req( i, x );
159
160 if (the_clone == nullptr) {
161 continue;
162 }
163
164 if (the_clone != x) {
165 _igvn.remove_dead_node(the_clone);
166 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
167 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
168 // it is not a win if 'x' moved from an outer to an inner loop
169 // this edge case can only happen for Load nodes
170 wins = 0;
171 break;
172 }
173 }
174 // Too few wins?
175 if (wins <= policy) {
176 _igvn.remove_dead_node(phi);
177 return nullptr;
178 }
179
180 // Record Phi
181 register_new_node( phi, region );
182
183 for (uint i2 = 1; i2 < phi->req(); i2++) {
184 Node *x = phi->in(i2);
185 // If we commoned up the cloned 'x' with another existing Node,
186 // the existing Node picks up a new use. We need to make the
187 // existing Node occur higher up so it dominates its uses.
188 Node *old_ctrl;
189 IdealLoopTree *old_loop;
190
191 if (x->is_Con()) {
192 // Constant's control is always root.
193 set_ctrl(x, C->root());
194 continue;
195 }
196 // The occasional new node
197 if (x->_idx >= old_unique) { // Found a new, unplaced node?
198 old_ctrl = nullptr;
199 old_loop = nullptr; // Not in any prior loop
200 } else {
201 old_ctrl = get_ctrl(x);
202 old_loop = get_loop(old_ctrl); // Get prior loop
203 }
204 // New late point must dominate new use
205 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
206 if (new_ctrl == old_ctrl) // Nothing is changed
207 continue;
208
209 IdealLoopTree *new_loop = get_loop(new_ctrl);
210
211 // Don't move x into a loop if its uses are
212 // outside of loop. Otherwise x will be cloned
213 // for each use outside of this loop.
214 IdealLoopTree *use_loop = get_loop(region);
215 if (!new_loop->is_member(use_loop) &&
216 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
217 // Take early control, later control will be recalculated
218 // during next iteration of loop optimizations.
219 new_ctrl = get_early_ctrl(x);
220 new_loop = get_loop(new_ctrl);
221 }
222 // Set new location
223 set_ctrl(x, new_ctrl);
224 // If changing loop bodies, see if we need to collect into new body
225 if (old_loop != new_loop) {
226 if (old_loop && !old_loop->_child)
227 old_loop->_body.yank(x);
228 if (!new_loop->_child)
229 new_loop->_body.push(x); // Collect body info
230 }
231 }
232
233 return phi;
234 }
235
236 // Test whether node 'x' can move into an inner loop relative to node 'n'.
237 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
238 // BUT it can also return true and 'x' is in the outer loop
239 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
240 IdealLoopTree* n_loop_tree = get_loop(n_loop);
241 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
242 // x_loop_tree should be outer or same loop as n_loop_tree
243 return !x_loop_tree->is_member(n_loop_tree);
244 }
245
246 // Subtype checks that carry profile data don't common so look for a replacement by following edges
247 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
248 if (x->is_SubTypeCheck()) {
249 Node* in1 = x->in(1);
250 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
251 Node* u = in1->fast_out(i);
252 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
253 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
254 Node* bol = u->fast_out(j);
255 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
256 Node* iff = bol->fast_out(k);
257 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
258 // unrelated profile
259 if (iff->is_If() && is_dominator(iff, r_in)) {
260 return u;
261 }
262 }
263 }
264 }
265 }
266 }
267 return nullptr;
268 }
269
270 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
271 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
272 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
273 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
274 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
275 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
276 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
277 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
278 const Type* zero;
279 switch (n->Opcode()) {
280 case Op_DivI:
281 case Op_ModI:
282 zero = TypeInt::ZERO;
283 break;
284 case Op_DivL:
285 case Op_ModL:
286 zero = TypeLong::ZERO;
287 break;
288 default:
289 return false;
290 }
291
292 if (n->in(0) != nullptr) {
293 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
294 return true;
295 }
296
297 Node* divisor = n->in(2);
298 return is_divisor_loop_phi(divisor, region) &&
299 loop_phi_backedge_type_contains_zero(divisor, zero);
300 }
301
302 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
303 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
304 }
305
306 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
307 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
308 }
309
310 //------------------------------dominated_by------------------------------------
311 // Replace the dominated test with an obvious true or false. Place it on the
312 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
313 // live path up to the dominating control.
314 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
315 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
316
317 // prevdom is the dominating projection of the dominating test.
318 assert(iff->Opcode() == Op_If ||
319 iff->Opcode() == Op_CountedLoopEnd ||
320 iff->Opcode() == Op_LongCountedLoopEnd ||
321 iff->Opcode() == Op_RangeCheck ||
322 iff->Opcode() == Op_ParsePredicate,
323 "Check this code when new subtype is added");
324
325 int pop = prevdom->Opcode();
326 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
327 if (flip) {
328 if (pop == Op_IfTrue)
329 pop = Op_IfFalse;
330 else
331 pop = Op_IfTrue;
332 }
333 // 'con' is set to true or false to kill the dominated test.
334 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
335 set_ctrl(con, C->root()); // Constant gets a new use
336 // Hack the dominated test
337 _igvn.replace_input_of(iff, 1, con);
338
339 // If I don't have a reachable TRUE and FALSE path following the IfNode then
340 // I can assume this path reaches an infinite loop. In this case it's not
341 // important to optimize the data Nodes - either the whole compilation will
342 // be tossed or this path (and all data Nodes) will go dead.
343 if (iff->outcnt() != 2) {
344 return;
345 }
346
347 // Make control-dependent data Nodes on the live path (path that will remain
348 // once the dominated IF is removed) become control-dependent on the
349 // dominating projection.
350 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
351
352 if (dp == nullptr) {
353 return;
354 }
355
356 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
357 }
358
359 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
360 IdealLoopTree* old_loop = get_loop(source);
361
362 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
363 Node* out = source->fast_out(i); // Control-dependent node
364 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
365 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
366 assert(out->in(0) == source, "must be control dependent on source");
367 _igvn.replace_input_of(out, 0, dominator);
368 if (pin_array_access_nodes) {
369 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
370 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
371 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
372 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
373 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
374 // dominating check.
375 Node* clone = out->pin_array_access_node();
376 if (clone != nullptr) {
377 clone = _igvn.register_new_node_with_optimizer(clone, out);
378 _igvn.replace_node(out, clone);
379 out = clone;
380 }
381 }
382 set_early_ctrl(out, false);
383 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
384 if (old_loop != new_loop) {
385 if (!old_loop->_child) {
386 old_loop->_body.yank(out);
387 }
388 if (!new_loop->_child) {
389 new_loop->_body.push(out);
390 }
391 }
392 --i;
393 --imax;
394 }
395 }
396 }
397
398 //------------------------------has_local_phi_input----------------------------
399 // Return TRUE if 'n' has Phi inputs from its local block and no other
400 // block-local inputs (all non-local-phi inputs come from earlier blocks)
401 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
402 Node *n_ctrl = get_ctrl(n);
403 // See if some inputs come from a Phi in this block, or from before
404 // this block.
405 uint i;
406 for( i = 1; i < n->req(); i++ ) {
407 Node *phi = n->in(i);
408 if( phi->is_Phi() && phi->in(0) == n_ctrl )
409 break;
410 }
411 if( i >= n->req() )
412 return nullptr; // No Phi inputs; nowhere to clone thru
413
414 // Check for inputs created between 'n' and the Phi input. These
415 // must split as well; they have already been given the chance
416 // (courtesy of a post-order visit) and since they did not we must
417 // recover the 'cost' of splitting them by being very profitable
418 // when splitting 'n'. Since this is unlikely we simply give up.
419 for( i = 1; i < n->req(); i++ ) {
420 Node *m = n->in(i);
421 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
422 // We allow the special case of AddP's with no local inputs.
423 // This allows us to split-up address expressions.
424 if (m->is_AddP() &&
425 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
426 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
427 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
428 // Move the AddP up to the dominating point. That's fine because control of m's inputs
429 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
430 Node* c = find_non_split_ctrl(idom(n_ctrl));
431 if (c->is_OuterStripMinedLoop()) {
432 c->as_Loop()->verify_strip_mined(1);
433 c = c->in(LoopNode::EntryControl);
434 }
435 set_ctrl_and_loop(m, c);
436 continue;
437 }
438 return nullptr;
439 }
440 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
441 }
442
443 return n_ctrl;
444 }
445
446 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
447 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
448 assert(bt == T_INT || bt == T_LONG, "only for integers");
449 int n_op = n->Opcode();
450
451 if (n_op == Op_LShift(bt)) {
452 // Scale is loop invariant
453 Node* scale = n->in(2);
454 Node* scale_ctrl = get_ctrl(scale);
455 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
456 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
457 return nullptr;
458 }
459 const TypeInt* scale_t = scale->bottom_type()->isa_int();
460 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
461 return nullptr; // Dont bother with byte/short masking
462 }
463 // Add must vary with loop (else shift would be loop-invariant)
464 Node* add = n->in(1);
465 Node* add_ctrl = get_ctrl(add);
466 IdealLoopTree* add_loop = get_loop(add_ctrl);
467 if (n_loop != add_loop) {
468 return nullptr; // happens w/ evil ZKM loops
469 }
470
471 // Convert I-V into I+ (0-V); same for V-I
472 if (add->Opcode() == Op_Sub(bt) &&
473 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
474 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
475 Node* zero = _igvn.integercon(0, bt);
476 set_ctrl(zero, C->root());
477 Node* neg = SubNode::make(zero, add->in(2), bt);
478 register_new_node_with_ctrl_of(neg, add->in(2));
479 add = AddNode::make(add->in(1), neg, bt);
480 register_new_node(add, add_ctrl);
481 }
482 if (add->Opcode() != Op_Add(bt)) return nullptr;
483 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
484 // See if one add input is loop invariant
485 Node* add_var = add->in(1);
486 Node* add_var_ctrl = get_ctrl(add_var);
487 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
488 Node* add_invar = add->in(2);
489 Node* add_invar_ctrl = get_ctrl(add_invar);
490 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
491 if (add_invar_loop == n_loop) {
492 // Swap to find the invariant part
493 add_invar = add_var;
494 add_invar_ctrl = add_var_ctrl;
495 add_invar_loop = add_var_loop;
496 add_var = add->in(2);
497 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
498 return nullptr;
499 }
500 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
501 return nullptr; // No invariant part of the add?
502 }
503
504 // Yes! Reshape address expression!
505 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
506 Node* inv_scale_ctrl =
507 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
508 add_invar_ctrl : scale_ctrl;
509 register_new_node(inv_scale, inv_scale_ctrl);
510 Node* var_scale = LShiftNode::make(add_var, scale, bt);
511 register_new_node(var_scale, n_ctrl);
512 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
513 register_new_node(var_add, n_ctrl);
514 _igvn.replace_node(n, var_add);
515 return var_add;
516 }
517 return nullptr;
518 }
519
520 //------------------------------remix_address_expressions----------------------
521 // Rework addressing expressions to get the most loop-invariant stuff
522 // moved out. We'd like to do all associative operators, but it's especially
523 // important (common) to do address expressions.
524 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
525 if (!has_ctrl(n)) return nullptr;
526 Node* n_ctrl = get_ctrl(n);
527 IdealLoopTree* n_loop = get_loop(n_ctrl);
528
529 // See if 'n' mixes loop-varying and loop-invariant inputs and
530 // itself is loop-varying.
531
532 // Only interested in binary ops (and AddP)
533 if (n->req() < 3 || n->req() > 4) return nullptr;
534
535 Node* n1_ctrl = get_ctrl(n->in( 1));
536 Node* n2_ctrl = get_ctrl(n->in( 2));
537 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
538 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
539 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
540 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
541
542 // Does one of my inputs spin in a tighter loop than self?
543 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
544 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
545 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
546 return nullptr; // Leave well enough alone
547 }
548
549 // Is at least one of my inputs loop-invariant?
550 if (n1_loop == n_loop &&
551 n2_loop == n_loop &&
552 n3_loop == n_loop) {
553 return nullptr; // No loop-invariant inputs
554 }
555
556 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
557 if (res != nullptr) {
558 return res;
559 }
560 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
561 if (res != nullptr) {
562 return res;
563 }
564
565 int n_op = n->Opcode();
566 // Replace (I+V) with (V+I)
567 if (n_op == Op_AddI ||
568 n_op == Op_AddL ||
569 n_op == Op_AddF ||
570 n_op == Op_AddD ||
571 n_op == Op_MulI ||
572 n_op == Op_MulL ||
573 n_op == Op_MulF ||
574 n_op == Op_MulD) {
575 if (n2_loop == n_loop) {
576 assert(n1_loop != n_loop, "");
577 n->swap_edges(1, 2);
578 }
579 }
580
581 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
582 // but not if I2 is a constant. Skip for irreducible loops.
583 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
584 if (n2_loop == n_loop && n3_loop != n_loop) {
585 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
586 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
587 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
588 IdealLoopTree* n22loop = get_loop(n22_ctrl);
589 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
590 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
591 n23_loop == n_loop) {
592 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
593 // Stuff new AddP in the loop preheader
594 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
595 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
596 register_new_node(add2, n_ctrl);
597 _igvn.replace_node(n, add2);
598 return add2;
599 }
600 }
601 }
602
603 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
604 if (n2_loop != n_loop && n3_loop == n_loop) {
605 if (n->in(3)->Opcode() == Op_AddX) {
606 Node* V = n->in(3)->in(1);
607 Node* I = n->in(3)->in(2);
608 if (is_member(n_loop,get_ctrl(V))) {
609 } else {
610 Node *tmp = V; V = I; I = tmp;
611 }
612 if (!is_member(n_loop,get_ctrl(I))) {
613 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
614 // Stuff new AddP in the loop preheader
615 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
616 Node* add2 = new AddPNode(n->in(1), add1, V);
617 register_new_node(add2, n_ctrl);
618 _igvn.replace_node(n, add2);
619 return add2;
620 }
621 }
622 }
623 }
624
625 return nullptr;
626 }
627
628 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
629 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
630 assert(n->Opcode() == Op_AddI, "sanity");
631 Node * nn = nullptr;
632 Node * in1 = n->in(1);
633 Node * in2 = n->in(2);
634 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
635 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
636 if (loop_n->is_counted() &&
637 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
638 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
639 Matcher::match_rule_supported(Op_MulAddS2I)) {
640 Node* mul_in1 = in1->in(1);
641 Node* mul_in2 = in1->in(2);
642 Node* mul_in3 = in2->in(1);
643 Node* mul_in4 = in2->in(2);
644 if (mul_in1->Opcode() == Op_LoadS &&
645 mul_in2->Opcode() == Op_LoadS &&
646 mul_in3->Opcode() == Op_LoadS &&
647 mul_in4->Opcode() == Op_LoadS) {
648 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
649 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
650 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
651 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
652 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
653 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
654 // All nodes should be in the same counted loop.
655 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
656 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
657 Node* adr1 = mul_in1->in(MemNode::Address);
658 Node* adr2 = mul_in2->in(MemNode::Address);
659 Node* adr3 = mul_in3->in(MemNode::Address);
660 Node* adr4 = mul_in4->in(MemNode::Address);
661 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
662 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
663 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
664 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
665 register_new_node_with_ctrl_of(nn, n);
666 _igvn.replace_node(n, nn);
667 return nn;
668 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
669 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
670 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
671 register_new_node_with_ctrl_of(nn, n);
672 _igvn.replace_node(n, nn);
673 return nn;
674 }
675 }
676 }
677 }
678 }
679 }
680 return nn;
681 }
682
683 //------------------------------conditional_move-------------------------------
684 // Attempt to replace a Phi with a conditional move. We have some pretty
685 // strict profitability requirements. All Phis at the merge point must
686 // be converted, so we can remove the control flow. We need to limit the
687 // number of c-moves to a small handful. All code that was in the side-arms
688 // of the CFG diamond is now speculatively executed. This code has to be
689 // "cheap enough". We are pretty much limited to CFG diamonds that merge
690 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
691 Node *PhaseIdealLoop::conditional_move( Node *region ) {
692
693 assert(region->is_Region(), "sanity check");
694 if (region->req() != 3) return nullptr;
695
696 // Check for CFG diamond
697 Node *lp = region->in(1);
698 Node *rp = region->in(2);
699 if (!lp || !rp) return nullptr;
700 Node *lp_c = lp->in(0);
701 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
702 IfNode *iff = lp_c->as_If();
703
704 // Check for ops pinned in an arm of the diamond.
705 // Can't remove the control flow in this case
706 if (lp->outcnt() > 1) return nullptr;
707 if (rp->outcnt() > 1) return nullptr;
708
709 IdealLoopTree* r_loop = get_loop(region);
710 assert(r_loop == get_loop(iff), "sanity");
711 // Always convert to CMOVE if all results are used only outside this loop.
712 bool used_inside_loop = (r_loop == _ltree_root);
713
714 // Check profitability
715 int cost = 0;
716 int phis = 0;
717 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
718 Node *out = region->fast_out(i);
719 if (!out->is_Phi()) continue; // Ignore other control edges, etc
720 phis++;
721 PhiNode* phi = out->as_Phi();
722 BasicType bt = phi->type()->basic_type();
723 switch (bt) {
724 case T_DOUBLE:
725 case T_FLOAT:
726 if (C->use_cmove()) {
727 continue; //TODO: maybe we want to add some cost
728 }
729 cost += Matcher::float_cmove_cost(); // Could be very expensive
730 break;
731 case T_LONG: {
732 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
733 }
734 case T_INT: // These all CMOV fine
735 case T_ADDRESS: { // (RawPtr)
736 cost++;
737 break;
738 }
739 case T_NARROWOOP: // Fall through
740 case T_OBJECT: { // Base oops are OK, but not derived oops
741 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
742 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
743 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
744 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
745 // have a Phi for the base here that we convert to a CMOVE all is well
746 // and good. But if the base is dead, we'll not make a CMOVE. Later
747 // the allocator will have to produce a base by creating a CMOVE of the
748 // relevant bases. This puts the allocator in the business of
749 // manufacturing expensive instructions, generally a bad plan.
750 // Just Say No to Conditionally-Moved Derived Pointers.
751 if (tp && tp->offset() != 0)
752 return nullptr;
753 cost++;
754 break;
755 }
756 default:
757 return nullptr; // In particular, can't do memory or I/O
758 }
759 // Add in cost any speculative ops
760 for (uint j = 1; j < region->req(); j++) {
761 Node *proj = region->in(j);
762 Node *inp = phi->in(j);
763 if (get_ctrl(inp) == proj) { // Found local op
764 cost++;
765 // Check for a chain of dependent ops; these will all become
766 // speculative in a CMOV.
767 for (uint k = 1; k < inp->req(); k++)
768 if (get_ctrl(inp->in(k)) == proj)
769 cost += ConditionalMoveLimit; // Too much speculative goo
770 }
771 }
772 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
773 // This will likely Split-If, a higher-payoff operation.
774 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
775 Node* use = phi->fast_out(k);
776 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
777 cost += ConditionalMoveLimit;
778 // Is there a use inside the loop?
779 // Note: check only basic types since CMoveP is pinned.
780 if (!used_inside_loop && is_java_primitive(bt)) {
781 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
782 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
783 used_inside_loop = true;
784 }
785 }
786 }
787 }//for
788 Node* bol = iff->in(1);
789 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
790 if (bol->is_OpaqueTemplateAssertionPredicate()) {
791 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
792 assert(assertion_predicate_has_loop_opaque_node(iff), "must find OpaqueLoop* nodes");
793 return nullptr;
794 }
795 assert(bol->Opcode() == Op_Bool, "Unexpected node");
796 int cmp_op = bol->in(1)->Opcode();
797 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
798 return nullptr;
799 }
800 // It is expensive to generate flags from a float compare.
801 // Avoid duplicated float compare.
802 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
803
804 float infrequent_prob = PROB_UNLIKELY_MAG(3);
805 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
806 if (used_inside_loop) {
807 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
808
809 // BlockLayoutByFrequency optimization moves infrequent branch
810 // from hot path. No point in CMOV'ing in such case (110 is used
811 // instead of 100 to take into account not exactness of float value).
812 if (BlockLayoutByFrequency) {
813 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
814 }
815 }
816 // Check for highly predictable branch. No point in CMOV'ing if
817 // we are going to predict accurately all the time.
818 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
819 //keep going
820 } else if (iff->_prob < infrequent_prob ||
821 iff->_prob > (1.0f - infrequent_prob))
822 return nullptr;
823
824 // --------------
825 // Now replace all Phis with CMOV's
826 Node *cmov_ctrl = iff->in(0);
827 uint flip = (lp->Opcode() == Op_IfTrue);
828 Node_List wq;
829 while (1) {
830 PhiNode* phi = nullptr;
831 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
832 Node *out = region->fast_out(i);
833 if (out->is_Phi()) {
834 phi = out->as_Phi();
835 break;
836 }
837 }
838 if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
839 break;
840 }
841 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
842 // Move speculative ops
843 wq.push(phi);
844 while (wq.size() > 0) {
845 Node *n = wq.pop();
846 for (uint j = 1; j < n->req(); j++) {
847 Node* m = n->in(j);
848 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
849 #ifndef PRODUCT
850 if (PrintOpto && VerifyLoopOptimizations) {
851 tty->print(" speculate: ");
852 m->dump();
853 }
854 #endif
855 set_ctrl(m, cmov_ctrl);
856 wq.push(m);
857 }
858 }
859 }
860 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
861 register_new_node( cmov, cmov_ctrl );
862 _igvn.replace_node( phi, cmov );
863 #ifndef PRODUCT
864 if (TraceLoopOpts) {
865 tty->print("CMOV ");
866 r_loop->dump_head();
867 if (Verbose) {
868 bol->in(1)->dump(1);
869 cmov->dump(1);
870 }
871 }
872 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
873 #endif
874 }
875
876 // The useless CFG diamond will fold up later; see the optimization in
877 // RegionNode::Ideal.
878 _igvn._worklist.push(region);
879
880 return iff->in(1);
881 }
882
883 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
884 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
885 Node* u = m->fast_out(i);
886 if (u->is_CFG()) {
887 if (u->is_NeverBranch()) {
888 u = u->as_NeverBranch()->proj_out(0);
889 enqueue_cfg_uses(u, wq);
890 } else {
891 wq.push(u);
892 }
893 }
894 }
895 }
896
897 // Try moving a store out of a loop, right before the loop
898 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
899 // Store has to be first in the loop body
900 IdealLoopTree *n_loop = get_loop(n_ctrl);
901 if (n->is_Store() && n_loop != _ltree_root &&
902 n_loop->is_loop() && n_loop->_head->is_Loop() &&
903 n->in(0) != nullptr) {
904 Node* address = n->in(MemNode::Address);
905 Node* value = n->in(MemNode::ValueIn);
906 Node* mem = n->in(MemNode::Memory);
907 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
908 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
909
910 // - address and value must be loop invariant
911 // - memory must be a memory Phi for the loop
912 // - Store must be the only store on this memory slice in the
913 // loop: if there's another store following this one then value
914 // written at iteration i by the second store could be overwritten
915 // at iteration i+n by the first store: it's not safe to move the
916 // first store out of the loop
917 // - nothing must observe the memory Phi: it guarantees no read
918 // before the store, we are also guaranteed the store post
919 // dominates the loop head (ignoring a possible early
920 // exit). Otherwise there would be extra Phi involved between the
921 // loop's Phi and the store.
922 // - there must be no early exit from the loop before the Store
923 // (such an exit most of the time would be an extra use of the
924 // memory Phi but sometimes is a bottom memory Phi that takes the
925 // store as input).
926
927 if (!n_loop->is_member(address_loop) &&
928 !n_loop->is_member(value_loop) &&
929 mem->is_Phi() && mem->in(0) == n_loop->_head &&
930 mem->outcnt() == 1 &&
931 mem->in(LoopNode::LoopBackControl) == n) {
932
933 assert(n_loop->_tail != nullptr, "need a tail");
934 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
935
936 // Verify that there's no early exit of the loop before the store.
937 bool ctrl_ok = false;
938 {
939 // Follow control from loop head until n, we exit the loop or
940 // we reach the tail
941 ResourceMark rm;
942 Unique_Node_List wq;
943 wq.push(n_loop->_head);
944
945 for (uint next = 0; next < wq.size(); ++next) {
946 Node *m = wq.at(next);
947 if (m == n->in(0)) {
948 ctrl_ok = true;
949 continue;
950 }
951 assert(!has_ctrl(m), "should be CFG");
952 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
953 ctrl_ok = false;
954 break;
955 }
956 enqueue_cfg_uses(m, wq);
957 if (wq.size() > 10) {
958 ctrl_ok = false;
959 break;
960 }
961 }
962 }
963 if (ctrl_ok) {
964 // move the Store
965 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
966 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
967 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
968 // Disconnect the phi now. An empty phi can confuse other
969 // optimizations in this pass of loop opts.
970 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
971 n_loop->_body.yank(mem);
972
973 set_ctrl_and_loop(n, n->in(0));
974
975 return n;
976 }
977 }
978 }
979 return nullptr;
980 }
981
982 // Try moving a store out of a loop, right after the loop
983 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
984 if (n->is_Store() && n->in(0) != nullptr) {
985 Node *n_ctrl = get_ctrl(n);
986 IdealLoopTree *n_loop = get_loop(n_ctrl);
987 // Store must be in a loop
988 if (n_loop != _ltree_root && !n_loop->_irreducible) {
989 Node* address = n->in(MemNode::Address);
990 Node* value = n->in(MemNode::ValueIn);
991 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
992 // address must be loop invariant
993 if (!n_loop->is_member(address_loop)) {
994 // Store must be last on this memory slice in the loop and
995 // nothing in the loop must observe it
996 Node* phi = nullptr;
997 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
998 Node* u = n->fast_out(i);
999 if (has_ctrl(u)) { // control use?
1000 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1001 if (!n_loop->is_member(u_loop)) {
1002 continue;
1003 }
1004 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1005 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1006 // multiple phis on the same slice are possible
1007 if (phi != nullptr) {
1008 return;
1009 }
1010 phi = u;
1011 continue;
1012 }
1013 }
1014 return;
1015 }
1016 if (phi != nullptr) {
1017 // Nothing in the loop before the store (next iteration)
1018 // must observe the stored value
1019 bool mem_ok = true;
1020 {
1021 ResourceMark rm;
1022 Unique_Node_List wq;
1023 wq.push(phi);
1024 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1025 Node *m = wq.at(next);
1026 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1027 Node* u = m->fast_out(i);
1028 if (u->is_Store() || u->is_Phi()) {
1029 if (u != n) {
1030 wq.push(u);
1031 mem_ok = (wq.size() <= 10);
1032 }
1033 } else {
1034 mem_ok = false;
1035 break;
1036 }
1037 }
1038 }
1039 }
1040 if (mem_ok) {
1041 // Move the store out of the loop if the LCA of all
1042 // users (except for the phi) is outside the loop.
1043 Node* hook = new Node(1);
1044 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1045 _igvn.rehash_node_delayed(phi);
1046 int count = phi->replace_edge(n, hook, &_igvn);
1047 assert(count > 0, "inconsistent phi");
1048
1049 // Compute latest point this store can go
1050 Node* lca = get_late_ctrl(n, get_ctrl(n));
1051 if (lca->is_OuterStripMinedLoop()) {
1052 lca = lca->in(LoopNode::EntryControl);
1053 }
1054 if (n_loop->is_member(get_loop(lca))) {
1055 // LCA is in the loop - bail out
1056 _igvn.replace_node(hook, n);
1057 return;
1058 }
1059 #ifdef ASSERT
1060 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1061 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1062 n_loop->_head->as_Loop()->verify_strip_mined(1);
1063 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1064 IdealLoopTree* outer_loop = get_loop(outer);
1065 assert(n_loop->_parent == outer_loop, "broken loop tree");
1066 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1067 }
1068 #endif
1069 lca = place_outside_loop(lca, n_loop);
1070 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1071 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");
1072
1073 // Move store out of the loop
1074 _igvn.replace_node(hook, n->in(MemNode::Memory));
1075 _igvn.replace_input_of(n, 0, lca);
1076 set_ctrl_and_loop(n, lca);
1077
1078 // Disconnect the phi now. An empty phi can confuse other
1079 // optimizations in this pass of loop opts..
1080 if (phi->in(LoopNode::LoopBackControl) == phi) {
1081 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1082 n_loop->_body.yank(phi);
1083 }
1084 }
1085 }
1086 }
1087 }
1088 }
1089 }
1090
1091 // Split some nodes that take a counted loop phi as input at a counted
1092 // loop can cause vectorization of some expressions to fail
1093 bool PhaseIdealLoop::split_thru_phi_could_prevent_vectorization(Node* n, Node* n_blk) {
1094 if (!n_blk->is_CountedLoop()) {
1095 return false;
1096 }
1097
1098 int opcode = n->Opcode();
1099
1100 if (opcode != Op_AndI &&
1101 opcode != Op_MulI &&
1102 opcode != Op_RotateRight &&
1103 opcode != Op_RShiftI) {
1104 return false;
1105 }
1106
1107 return n->in(1) == n_blk->as_BaseCountedLoop()->phi();
1108 }
1109
1110 //------------------------------split_if_with_blocks_pre-----------------------
1111 // Do the real work in a non-recursive function. Data nodes want to be
1112 // cloned in the pre-order so they can feed each other nicely.
1113 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1114 // Cloning these guys is unlikely to win
1115 int n_op = n->Opcode();
1116 if (n_op == Op_MergeMem) {
1117 return n;
1118 }
1119 if (n->is_Proj()) {
1120 return n;
1121 }
1122 // Do not clone-up CmpFXXX variations, as these are always
1123 // followed by a CmpI
1124 if (n->is_Cmp()) {
1125 return n;
1126 }
1127 // Attempt to use a conditional move instead of a phi/branch
1128 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1129 Node *cmov = conditional_move( n );
1130 if (cmov) {
1131 return cmov;
1132 }
1133 }
1134 if (n->is_CFG() || n->is_LoadStore()) {
1135 return n;
1136 }
1137 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1138 if (!C->major_progress()) { // If chance of no more loop opts...
1139 _igvn._worklist.push(n); // maybe we'll remove them
1140 }
1141 return n;
1142 }
1143
1144 if (n->is_Con()) {
1145 return n; // No cloning for Con nodes
1146 }
1147
1148 Node *n_ctrl = get_ctrl(n);
1149 if (!n_ctrl) {
1150 return n; // Dead node
1151 }
1152
1153 Node* res = try_move_store_before_loop(n, n_ctrl);
1154 if (res != nullptr) {
1155 return n;
1156 }
1157
1158 // Attempt to remix address expressions for loop invariants
1159 Node *m = remix_address_expressions( n );
1160 if( m ) return m;
1161
1162 if (n_op == Op_AddI) {
1163 Node *nn = convert_add_to_muladd( n );
1164 if ( nn ) return nn;
1165 }
1166
1167 if (n->is_ConstraintCast()) {
1168 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1169 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1170 // Node control inputs don't necessarily agree with loop control info (due to
1171 // transformations happened in between), thus additional dominance check is needed
1172 // to keep loop info valid.
1173 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1174 _igvn.replace_node(n, dom_cast);
1175 return dom_cast;
1176 }
1177 }
1178
1179 // Determine if the Node has inputs from some local Phi.
1180 // Returns the block to clone thru.
1181 Node *n_blk = has_local_phi_input( n );
1182 if( !n_blk ) return n;
1183
1184 // Do not clone the trip counter through on a CountedLoop
1185 // (messes up the canonical shape).
1186 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1187 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1188 return n;
1189 }
1190 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1191 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1192 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1193 return n;
1194 }
1195
1196 if (split_thru_phi_could_prevent_vectorization(n, n_blk)) {
1197 return n;
1198 }
1199
1200 // Check for having no control input; not pinned. Allow
1201 // dominating control.
1202 if (n->in(0)) {
1203 Node *dom = idom(n_blk);
1204 if (dom_lca(n->in(0), dom) != n->in(0)) {
1205 return n;
1206 }
1207 }
1208 // Policy: when is it profitable. You must get more wins than
1209 // policy before it is considered profitable. Policy is usually 0,
1210 // so 1 win is considered profitable. Big merges will require big
1211 // cloning, so get a larger policy.
1212 int policy = n_blk->req() >> 2;
1213
1214 // If the loop is a candidate for range check elimination,
1215 // delay splitting through it's phi until a later loop optimization
1216 if (n_blk->is_BaseCountedLoop()) {
1217 IdealLoopTree *lp = get_loop(n_blk);
1218 if (lp && lp->_rce_candidate) {
1219 return n;
1220 }
1221 }
1222
1223 if (must_throttle_split_if()) return n;
1224
1225 // Split 'n' through the merge point if it is profitable
1226 Node *phi = split_thru_phi( n, n_blk, policy );
1227 if (!phi) return n;
1228
1229 // Found a Phi to split thru!
1230 // Replace 'n' with the new phi
1231 _igvn.replace_node( n, phi );
1232 // Moved a load around the loop, 'en-registering' something.
1233 if (n_blk->is_Loop() && n->is_Load() &&
1234 !phi->in(LoopNode::LoopBackControl)->is_Load())
1235 C->set_major_progress();
1236
1237 return phi;
1238 }
1239
1240 static bool merge_point_too_heavy(Compile* C, Node* region) {
1241 // Bail out if the region and its phis have too many users.
1242 int weight = 0;
1243 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1244 weight += region->fast_out(i)->outcnt();
1245 }
1246 int nodes_left = C->max_node_limit() - C->live_nodes();
1247 if (weight * 8 > nodes_left) {
1248 if (PrintOpto) {
1249 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1250 }
1251 return true;
1252 } else {
1253 return false;
1254 }
1255 }
1256
1257 static bool merge_point_safe(Node* region) {
1258 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1259 // having a PhiNode input. This sidesteps the dangerous case where the split
1260 // ConvI2LNode may become TOP if the input Value() does not
1261 // overlap the ConvI2L range, leaving a node which may not dominate its
1262 // uses.
1263 // A better fix for this problem can be found in the BugTraq entry, but
1264 // expediency for Mantis demands this hack.
1265 #ifdef _LP64
1266 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1267 Node* n = region->fast_out(i);
1268 if (n->is_Phi()) {
1269 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1270 Node* m = n->fast_out(j);
1271 if (m->Opcode() == Op_ConvI2L)
1272 return false;
1273 if (m->is_CastII()) {
1274 return false;
1275 }
1276 }
1277 }
1278 }
1279 #endif
1280 return true;
1281 }
1282
1283
1284 //------------------------------place_outside_loop---------------------------------
1285 // Place some computation outside of this loop on the path to the use passed as argument
1286 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1287 Node* head = loop->_head;
1288 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1289 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1290 loop = loop->_parent;
1291 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1292 }
1293
1294 // Pick control right outside the loop
1295 for (;;) {
1296 Node* dom = idom(useblock);
1297 if (loop->is_member(get_loop(dom))) {
1298 break;
1299 }
1300 useblock = dom;
1301 }
1302 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1303 return useblock;
1304 }
1305
1306
1307 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1308 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1309 return false;
1310 }
1311 if (!n->in(0)->is_Region()) {
1312 return false;
1313 }
1314
1315 Node* region = n->in(0);
1316 Node* dom = idom(region);
1317 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1318 return false;
1319 }
1320 IfNode* dom_if = dom->as_If();
1321 Node* proj_true = dom_if->proj_out(1);
1322 Node* proj_false = dom_if->proj_out(0);
1323
1324 for (uint i = 1; i < region->req(); i++) {
1325 if (is_dominator(proj_true, region->in(i))) {
1326 continue;
1327 }
1328 if (is_dominator(proj_false, region->in(i))) {
1329 continue;
1330 }
1331 return false;
1332 }
1333
1334 return true;
1335 }
1336
1337
1338 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1339 if (must_throttle_split_if()) {
1340 return false;
1341 }
1342
1343 // Do not do 'split-if' if irreducible loops are present.
1344 if (_has_irreducible_loops) {
1345 return false;
1346 }
1347
1348 if (merge_point_too_heavy(C, n_ctrl)) {
1349 return false;
1350 }
1351
1352 // Do not do 'split-if' if some paths are dead. First do dead code
1353 // elimination and then see if its still profitable.
1354 for (uint i = 1; i < n_ctrl->req(); i++) {
1355 if (n_ctrl->in(i) == C->top()) {
1356 return false;
1357 }
1358 }
1359
1360 // If trying to do a 'Split-If' at the loop head, it is only
1361 // profitable if the cmp folds up on BOTH paths. Otherwise we
1362 // risk peeling a loop forever.
1363
1364 // CNC - Disabled for now. Requires careful handling of loop
1365 // body selection for the cloned code. Also, make sure we check
1366 // for any input path not being in the same loop as n_ctrl. For
1367 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1368 // because the alternative loop entry points won't be converted
1369 // into LoopNodes.
1370 IdealLoopTree *n_loop = get_loop(n_ctrl);
1371 for (uint j = 1; j < n_ctrl->req(); j++) {
1372 if (get_loop(n_ctrl->in(j)) != n_loop) {
1373 return false;
1374 }
1375 }
1376
1377 // Check for safety of the merge point.
1378 if (!merge_point_safe(n_ctrl)) {
1379 return false;
1380 }
1381
1382 return true;
1383 }
1384
1385 // Detect if the node is the inner strip-mined loop
1386 // Return: null if it's not the case, or the exit of outer strip-mined loop
1387 static Node* is_inner_of_stripmined_loop(const Node* out) {
1388 Node* out_le = nullptr;
1389
1390 if (out->is_CountedLoopEnd()) {
1391 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1392
1393 if (loop != nullptr && loop->is_strip_mined()) {
1394 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1395 }
1396 }
1397
1398 return out_le;
1399 }
1400
1401 //------------------------------split_if_with_blocks_post----------------------
1402 // Do the real work in a non-recursive function. CFG hackery wants to be
1403 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1404 // info.
1405 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1406
1407 // Cloning Cmp through Phi's involves the split-if transform.
1408 // FastLock is not used by an If
1409 if (n->is_Cmp() && !n->is_FastLock()) {
1410 Node *n_ctrl = get_ctrl(n);
1411 // Determine if the Node has inputs from some local Phi.
1412 // Returns the block to clone thru.
1413 Node *n_blk = has_local_phi_input(n);
1414 if (n_blk != n_ctrl) {
1415 return;
1416 }
1417
1418 if (!can_split_if(n_ctrl)) {
1419 return;
1420 }
1421
1422 if (n->outcnt() != 1) {
1423 return; // Multiple bool's from 1 compare?
1424 }
1425 Node *bol = n->unique_out();
1426 assert(bol->is_Bool(), "expect a bool here");
1427 if (bol->outcnt() != 1) {
1428 return;// Multiple branches from 1 compare?
1429 }
1430 Node *iff = bol->unique_out();
1431
1432 // Check some safety conditions
1433 if (iff->is_If()) { // Classic split-if?
1434 if (iff->in(0) != n_ctrl) {
1435 return; // Compare must be in same blk as if
1436 }
1437 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1438 // Can't split CMove with different control.
1439 if (get_ctrl(iff) != n_ctrl) {
1440 return;
1441 }
1442 if (get_ctrl(iff->in(2)) == n_ctrl ||
1443 get_ctrl(iff->in(3)) == n_ctrl) {
1444 return; // Inputs not yet split-up
1445 }
1446 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1447 return; // Loop-invar test gates loop-varying CMOVE
1448 }
1449 } else {
1450 return; // some other kind of node, such as an Allocate
1451 }
1452
1453 // When is split-if profitable? Every 'win' on means some control flow
1454 // goes dead, so it's almost always a win.
1455 int policy = 0;
1456 // Split compare 'n' through the merge point if it is profitable
1457 Node *phi = split_thru_phi( n, n_ctrl, policy);
1458 if (!phi) {
1459 return;
1460 }
1461
1462 // Found a Phi to split thru!
1463 // Replace 'n' with the new phi
1464 _igvn.replace_node(n, phi);
1465
1466 // Now split the bool up thru the phi
1467 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1468 guarantee(bolphi != nullptr, "null boolean phi node");
1469
1470 _igvn.replace_node(bol, bolphi);
1471 assert(iff->in(1) == bolphi, "");
1472
1473 if (bolphi->Value(&_igvn)->singleton()) {
1474 return;
1475 }
1476
1477 // Conditional-move? Must split up now
1478 if (!iff->is_If()) {
1479 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1480 _igvn.replace_node(iff, cmovphi);
1481 return;
1482 }
1483
1484 // Now split the IF
1485 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1486 if ((PrintOpto && VerifyLoopOptimizations) || 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
1553 try_move_store_after_loop(n);
1554 }
1555
1556 // Transform:
1557 //
1558 // if (some_condition) {
1559 // // body 1
1560 // } else {
1561 // // body 2
1562 // }
1563 // if (some_condition) {
1564 // // body 3
1565 // } else {
1566 // // body 4
1567 // }
1568 //
1569 // into:
1570 //
1571 //
1572 // if (some_condition) {
1573 // // body 1
1574 // // body 3
1575 // } else {
1576 // // body 2
1577 // // body 4
1578 // }
1579 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1580 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1581 Node *n_ctrl = n->in(0);
1582 IfNode* dom_if = idom(n_ctrl)->as_If();
1583 if (n->in(1) != dom_if->in(1)) {
1584 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1585 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1586 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1587 _igvn.replace_input_of(n, 1, dom_if->in(1));
1588 }
1589 ProjNode* dom_proj_true = dom_if->proj_out(1);
1590 ProjNode* dom_proj_false = dom_if->proj_out(0);
1591
1592 // Now split the IF
1593 RegionNode* new_false_region;
1594 RegionNode* new_true_region;
1595 do_split_if(n, &new_false_region, &new_true_region);
1596 assert(new_false_region->req() == new_true_region->req(), "");
1597 #ifdef ASSERT
1598 for (uint i = 1; i < new_false_region->req(); ++i) {
1599 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1600 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");
1601 }
1602 #endif
1603 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1604
1605 // We now have:
1606 // if (some_condition) {
1607 // // body 1
1608 // if (some_condition) {
1609 // body3: // new_true_region
1610 // // body3
1611 // } else {
1612 // goto body4;
1613 // }
1614 // } else {
1615 // // body 2
1616 // if (some_condition) {
1617 // goto body3;
1618 // } else {
1619 // body4: // new_false_region
1620 // // body4;
1621 // }
1622 // }
1623 //
1624
1625 // clone pinned nodes thru the resulting regions
1626 push_pinned_nodes_thru_region(dom_if, new_true_region);
1627 push_pinned_nodes_thru_region(dom_if, new_false_region);
1628
1629 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1630 // 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
1631 // unrelated control dependency.
1632 for (uint i = 1; i < new_false_region->req(); i++) {
1633 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1634 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1635 } else {
1636 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1637 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1638 }
1639 }
1640 return true;
1641 }
1642 return false;
1643 }
1644
1645 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1646 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1647 Node* u = region->out(i);
1648 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1649 continue;
1650 }
1651 assert(u->in(0) == region, "not a control dependent node?");
1652 uint j = 1;
1653 for (; j < u->req(); ++j) {
1654 Node* in = u->in(j);
1655 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1656 break;
1657 }
1658 }
1659 if (j == u->req()) {
1660 Node *phi = PhiNode::make_blank(region, u);
1661 for (uint k = 1; k < region->req(); ++k) {
1662 Node* clone = u->clone();
1663 clone->set_req(0, region->in(k));
1664 register_new_node(clone, region->in(k));
1665 phi->init_req(k, clone);
1666 }
1667 register_new_node(phi, region);
1668 _igvn.replace_node(u, phi);
1669 --i;
1670 }
1671 }
1672 }
1673
1674 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1675 if (!dom->is_CountedLoopEnd()) {
1676 return true;
1677 }
1678 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1679 CountedLoopNode* cl = le->loopnode();
1680 if (cl == nullptr) {
1681 return true;
1682 }
1683 if (!cl->is_main_loop()) {
1684 return true;
1685 }
1686 if (cl->is_canonical_loop_entry() == nullptr) {
1687 return true;
1688 }
1689 // Further unrolling is possible so loop exit condition might change
1690 return false;
1691 }
1692
1693 // See if a shared loop-varying computation has no loop-varying uses.
1694 // Happens if something is only used for JVM state in uncommon trap exits,
1695 // like various versions of induction variable+offset. Clone the
1696 // computation per usage to allow it to sink out of the loop.
1697 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1698 if (has_ctrl(n) &&
1699 !n->is_Phi() &&
1700 !n->is_Bool() &&
1701 !n->is_Proj() &&
1702 !n->is_MergeMem() &&
1703 !n->is_CMove() &&
1704 !n->is_OpaqueNotNull() &&
1705 !n->is_OpaqueInitializedAssertionPredicate() &&
1706 !n->is_OpaqueTemplateAssertionPredicate() &&
1707 !n->is_Type()) {
1708 Node *n_ctrl = get_ctrl(n);
1709 IdealLoopTree *n_loop = get_loop(n_ctrl);
1710
1711 if (n->in(0) != nullptr) {
1712 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1713 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1714 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1715 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1716 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1717 Node* maybe_pinned_n = n;
1718 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1719 if (n->depends_only_on_test()) {
1720 Node* pinned_clone = n->pin_array_access_node();
1721 if (pinned_clone != nullptr) {
1722 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1723 // range check for that access. If that condition is replaced by an identical dominating one, then an
1724 // unpinned load would risk floating above its range check.
1725 register_new_node(pinned_clone, n_ctrl);
1726 maybe_pinned_n = pinned_clone;
1727 _igvn.replace_node(n, pinned_clone);
1728 }
1729 }
1730 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1731 }
1732 }
1733 if (n_loop != _ltree_root && n->outcnt() > 1) {
1734 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1735 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1736 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1737 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1738 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1739 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1740 Node* outer_loop_clone = nullptr;
1741 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1742 Node* u = n->last_out(j); // Clone private computation per use
1743 _igvn.rehash_node_delayed(u);
1744 Node* x = nullptr;
1745 if (n->depends_only_on_test()) {
1746 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1747 // range check for that access. If that condition is replaced by an identical dominating one, then an
1748 // unpinned load would risk floating above its range check.
1749 x = n->pin_array_access_node();
1750 }
1751 if (x == nullptr) {
1752 x = n->clone();
1753 }
1754 Node* x_ctrl = nullptr;
1755 if (u->is_Phi()) {
1756 // Replace all uses of normal nodes. Replace Phi uses
1757 // individually, so the separate Nodes can sink down
1758 // different paths.
1759 uint k = 1;
1760 while (u->in(k) != n) k++;
1761 u->set_req(k, x);
1762 // x goes next to Phi input path
1763 x_ctrl = u->in(0)->in(k);
1764 // Find control for 'x' next to use but not inside inner loops.
1765 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1766 --j;
1767 } else { // Normal use
1768 if (has_ctrl(u)) {
1769 x_ctrl = get_ctrl(u);
1770 } else {
1771 x_ctrl = u->in(0);
1772 }
1773 // Find control for 'x' next to use but not inside inner loops.
1774 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1775 // Replace all uses
1776 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1777 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1778 // anymore now that we're going to pin n as well
1779 _igvn.replace_node(u, x);
1780 --j;
1781 } else {
1782 int nb = u->replace_edge(n, x, &_igvn);
1783 j -= nb;
1784 }
1785 }
1786
1787 if (n->is_Load()) {
1788 // For loads, add a control edge to a CFG node outside of the loop
1789 // to force them to not combine and return back inside the loop
1790 // during GVN optimization (4641526).
1791 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1792
1793 IdealLoopTree* x_loop = get_loop(x_ctrl);
1794 Node* x_head = x_loop->_head;
1795 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1796 // Do not add duplicate LoadNodes to the outer strip mined loop
1797 if (outer_loop_clone != nullptr) {
1798 _igvn.replace_node(x, outer_loop_clone);
1799 continue;
1800 }
1801 outer_loop_clone = x;
1802 }
1803 x->set_req(0, x_ctrl);
1804 } else if (n->in(0) != nullptr){
1805 x->set_req(0, x_ctrl);
1806 }
1807 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1808 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1809 register_new_node(x, x_ctrl);
1810
1811 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1812 // All AddP nodes must keep the same base after sinking so:
1813 // 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,
1814 // their bases remain the same.
1815 // (see 2- below)
1816 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1817 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1818 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1819 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1820 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1821 assert(!x->is_Load(), "load should be pinned");
1822 // Use a cast node to pin clone out of loop
1823 Node* cast = nullptr;
1824 for (uint k = 0; k < x->req(); k++) {
1825 Node* in = x->in(k);
1826 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1827 const Type* in_t = _igvn.type(in);
1828 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1829 ConstraintCastNode::UnconditionalDependency, nullptr);
1830 }
1831 if (cast != nullptr) {
1832 Node* prev = _igvn.hash_find_insert(cast);
1833 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1834 cast->destruct(&_igvn);
1835 cast = prev;
1836 } else {
1837 register_new_node(cast, x_ctrl);
1838 }
1839 x->replace_edge(in, cast);
1840 // Chain of AddP nodes:
1841 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1842 if (x->is_AddP() && k == AddPNode::Base) {
1843 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1844 }
1845 break;
1846 }
1847 }
1848 assert(cast != nullptr, "must have added a cast to pin the node");
1849 }
1850 }
1851 _igvn.remove_dead_node(n);
1852 }
1853 _dom_lca_tags_round = 0;
1854 }
1855 }
1856 }
1857
1858 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1859 ResourceMark rm;
1860 Node_List wq;
1861 wq.push(x);
1862 while (wq.size() != 0) {
1863 Node* n = wq.pop();
1864 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1865 Node* u = n->fast_out(i);
1866 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1867 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1868 wq.push(u);
1869 }
1870 }
1871 }
1872 }
1873
1874 // Compute the early control of a node by following its inputs until we reach
1875 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1876 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1877 Node* early_ctrl = nullptr;
1878 ResourceMark rm;
1879 Unique_Node_List wq;
1880 wq.push(n);
1881 for (uint i = 0; i < wq.size(); i++) {
1882 Node* m = wq.at(i);
1883 Node* c = nullptr;
1884 if (m->is_CFG()) {
1885 c = m;
1886 } else if (m->pinned()) {
1887 c = m->in(0);
1888 } else {
1889 for (uint j = 0; j < m->req(); j++) {
1890 Node* in = m->in(j);
1891 if (in != nullptr) {
1892 wq.push(in);
1893 }
1894 }
1895 }
1896 if (c != nullptr) {
1897 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1898 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1899 early_ctrl = c;
1900 }
1901 }
1902 }
1903 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1904 return early_ctrl;
1905 }
1906
1907 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1908 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1909 Node* u = n->fast_out(i);
1910 if (u->is_Opaque1()) {
1911 return false; // Found loop limit, bugfix for 4677003
1912 }
1913 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1914 // get_late_ctrl_with_anti_dep() use their own tag
1915 _dom_lca_tags_round++;
1916 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1917
1918 if (u->is_Phi()) {
1919 for (uint j = 1; j < u->req(); ++j) {
1920 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1921 return false;
1922 }
1923 }
1924 } else {
1925 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1926 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1927 return false;
1928 }
1929 }
1930 }
1931 return true;
1932 }
1933
1934 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1935 if (n->is_Load()) {
1936 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1937 }
1938 IdealLoopTree *u_loop = get_loop(ctrl);
1939 if (u_loop == n_loop) {
1940 return false; // Found loop-varying use
1941 }
1942 if (n_loop->is_member(u_loop)) {
1943 return false; // Found use in inner loop
1944 }
1945 // 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
1946 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1947 // test of the pre loop above the point in the graph where it's pinned.
1948 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1949 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1950 if (is_dominator(pre_loop->loopexit(), ctrl)) {
1951 return false;
1952 }
1953 }
1954 return true;
1955 }
1956
1957 //------------------------------split_if_with_blocks---------------------------
1958 // Check for aggressive application of 'split-if' optimization,
1959 // using basic block level info.
1960 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1961 Node* root = C->root();
1962 visited.set(root->_idx); // first, mark root as visited
1963 // Do pre-visit work for root
1964 Node* n = split_if_with_blocks_pre(root);
1965 uint cnt = n->outcnt();
1966 uint i = 0;
1967
1968 while (true) {
1969 // Visit all children
1970 if (i < cnt) {
1971 Node* use = n->raw_out(i);
1972 ++i;
1973 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1974 // Now do pre-visit work for this use
1975 use = split_if_with_blocks_pre(use);
1976 nstack.push(n, i); // Save parent and next use's index.
1977 n = use; // Process all children of current use.
1978 cnt = use->outcnt();
1979 i = 0;
1980 }
1981 }
1982 else {
1983 // All of n's children have been processed, complete post-processing.
1984 if (cnt != 0 && !n->is_Con()) {
1985 assert(has_node(n), "no dead nodes");
1986 split_if_with_blocks_post(n);
1987 }
1988 if (must_throttle_split_if()) {
1989 nstack.clear();
1990 }
1991 if (nstack.is_empty()) {
1992 // Finished all nodes on stack.
1993 break;
1994 }
1995 // Get saved parent node and next use's index. Visit the rest of uses.
1996 n = nstack.node();
1997 cnt = n->outcnt();
1998 i = nstack.index();
1999 nstack.pop();
2000 }
2001 }
2002 }
2003
2004
2005 //=============================================================================
2006 //
2007 // C L O N E A L O O P B O D Y
2008 //
2009
2010 //------------------------------clone_iff--------------------------------------
2011 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2012 // "Nearly" because all Nodes have been cloned from the original in the loop,
2013 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2014 // through the Phi recursively, and return a Bool.
2015 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2016
2017 // Convert this Phi into a Phi merging Bools
2018 uint i;
2019 for (i = 1; i < phi->req(); i++) {
2020 Node* b = phi->in(i);
2021 if (b->is_Phi()) {
2022 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2023 } else {
2024 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2025 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2026 }
2027 }
2028 Node* n = phi->in(1);
2029 Node* sample_opaque = nullptr;
2030 Node *sample_bool = nullptr;
2031 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2032 sample_opaque = n;
2033 sample_bool = n->in(1);
2034 assert(sample_bool->is_Bool(), "wrong type");
2035 } else {
2036 sample_bool = n;
2037 }
2038 Node *sample_cmp = sample_bool->in(1);
2039
2040 // Make Phis to merge the Cmp's inputs.
2041 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2042 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2043 for (i = 1; i < phi->req(); i++) {
2044 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2045 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2046 phi1->set_req(i, n1);
2047 phi2->set_req(i, n2);
2048 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2049 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2050 }
2051 // See if these Phis have been made before.
2052 // Register with optimizer
2053 Node *hit1 = _igvn.hash_find_insert(phi1);
2054 if (hit1) { // Hit, toss just made Phi
2055 _igvn.remove_dead_node(phi1); // Remove new phi
2056 assert(hit1->is_Phi(), "" );
2057 phi1 = (PhiNode*)hit1; // Use existing phi
2058 } else { // Miss
2059 _igvn.register_new_node_with_optimizer(phi1);
2060 }
2061 Node *hit2 = _igvn.hash_find_insert(phi2);
2062 if (hit2) { // Hit, toss just made Phi
2063 _igvn.remove_dead_node(phi2); // Remove new phi
2064 assert(hit2->is_Phi(), "" );
2065 phi2 = (PhiNode*)hit2; // Use existing phi
2066 } else { // Miss
2067 _igvn.register_new_node_with_optimizer(phi2);
2068 }
2069 // Register Phis with loop/block info
2070 set_ctrl(phi1, phi->in(0));
2071 set_ctrl(phi2, phi->in(0));
2072 // Make a new Cmp
2073 Node *cmp = sample_cmp->clone();
2074 cmp->set_req(1, phi1);
2075 cmp->set_req(2, phi2);
2076 _igvn.register_new_node_with_optimizer(cmp);
2077 set_ctrl(cmp, phi->in(0));
2078
2079 // Make a new Bool
2080 Node *b = sample_bool->clone();
2081 b->set_req(1,cmp);
2082 _igvn.register_new_node_with_optimizer(b);
2083 set_ctrl(b, phi->in(0));
2084
2085 if (sample_opaque != nullptr) {
2086 Node* opaque = sample_opaque->clone();
2087 opaque->set_req(1, b);
2088 _igvn.register_new_node_with_optimizer(opaque);
2089 set_ctrl(opaque, phi->in(0));
2090 return opaque;
2091 }
2092
2093 assert(b->is_Bool(), "");
2094 return b;
2095 }
2096
2097 //------------------------------clone_bool-------------------------------------
2098 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2099 // "Nearly" because all Nodes have been cloned from the original in the loop,
2100 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2101 // through the Phi recursively, and return a Bool.
2102 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2103 uint i;
2104 // Convert this Phi into a Phi merging Bools
2105 for( i = 1; i < phi->req(); i++ ) {
2106 Node *b = phi->in(i);
2107 if( b->is_Phi() ) {
2108 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2109 } else {
2110 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2111 }
2112 }
2113
2114 Node *sample_cmp = phi->in(1);
2115
2116 // Make Phis to merge the Cmp's inputs.
2117 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2118 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2119 for( uint j = 1; j < phi->req(); j++ ) {
2120 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2121 Node *n1, *n2;
2122 if( cmp_top->is_Cmp() ) {
2123 n1 = cmp_top->in(1);
2124 n2 = cmp_top->in(2);
2125 } else {
2126 n1 = n2 = cmp_top;
2127 }
2128 phi1->set_req( j, n1 );
2129 phi2->set_req( j, n2 );
2130 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2131 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2132 }
2133
2134 // See if these Phis have been made before.
2135 // Register with optimizer
2136 Node *hit1 = _igvn.hash_find_insert(phi1);
2137 if( hit1 ) { // Hit, toss just made Phi
2138 _igvn.remove_dead_node(phi1); // Remove new phi
2139 assert( hit1->is_Phi(), "" );
2140 phi1 = (PhiNode*)hit1; // Use existing phi
2141 } else { // Miss
2142 _igvn.register_new_node_with_optimizer(phi1);
2143 }
2144 Node *hit2 = _igvn.hash_find_insert(phi2);
2145 if( hit2 ) { // Hit, toss just made Phi
2146 _igvn.remove_dead_node(phi2); // Remove new phi
2147 assert( hit2->is_Phi(), "" );
2148 phi2 = (PhiNode*)hit2; // Use existing phi
2149 } else { // Miss
2150 _igvn.register_new_node_with_optimizer(phi2);
2151 }
2152 // Register Phis with loop/block info
2153 set_ctrl(phi1, phi->in(0));
2154 set_ctrl(phi2, phi->in(0));
2155 // Make a new Cmp
2156 Node *cmp = sample_cmp->clone();
2157 cmp->set_req( 1, phi1 );
2158 cmp->set_req( 2, phi2 );
2159 _igvn.register_new_node_with_optimizer(cmp);
2160 set_ctrl(cmp, phi->in(0));
2161
2162 assert( cmp->is_Cmp(), "" );
2163 return (CmpNode*)cmp;
2164 }
2165
2166 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2167 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2168 Node_List*& split_if_set, Node_List*& split_bool_set,
2169 Node_List*& split_cex_set, Node_List& worklist,
2170 uint new_counter, CloneLoopMode mode) {
2171 Node* nnn = old_new[old->_idx];
2172 // Copy uses to a worklist, so I can munge the def-use info
2173 // with impunity.
2174 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2175 worklist.push(old->fast_out(j));
2176
2177 while( worklist.size() ) {
2178 Node *use = worklist.pop();
2179 if (!has_node(use)) continue; // Ignore dead nodes
2180 if (use->in(0) == C->top()) continue;
2181 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2182 // Check for data-use outside of loop - at least one of OLD or USE
2183 // must not be a CFG node.
2184 #ifdef ASSERT
2185 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2186 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2187 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2188 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2189 }
2190 #endif
2191 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2192
2193 // If the Data use is an IF, that means we have an IF outside the
2194 // loop that is switching on a condition that is set inside the
2195 // loop. Happens if people set a loop-exit flag; then test the flag
2196 // in the loop to break the loop, then test is again outside the
2197 // loop to determine which way the loop exited.
2198 //
2199 // For several uses we need to make sure that there is no phi between,
2200 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2201 // to avoid such a phi in between.
2202 // For example, it is unexpected that there is a Phi between an
2203 // AllocateArray node and its ValidLengthTest input that could cause
2204 // split if to break.
2205 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2206 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2207 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2208 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2209 // Since this code is highly unlikely, we lazily build the worklist
2210 // of such Nodes to go split.
2211 if (!split_if_set) {
2212 split_if_set = new Node_List();
2213 }
2214 split_if_set->push(use);
2215 }
2216 if (use->is_Bool()) {
2217 if (!split_bool_set) {
2218 split_bool_set = new Node_List();
2219 }
2220 split_bool_set->push(use);
2221 }
2222 if (use->Opcode() == Op_CreateEx) {
2223 if (!split_cex_set) {
2224 split_cex_set = new Node_List();
2225 }
2226 split_cex_set->push(use);
2227 }
2228
2229
2230 // Get "block" use is in
2231 uint idx = 0;
2232 while( use->in(idx) != old ) idx++;
2233 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2234 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2235 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2236 ? prev->in(2)
2237 : idom(prev);
2238 if( use->is_Phi() ) // Phi use is in prior block
2239 cfg = prev->in(idx); // NOT in block of Phi itself
2240 if (cfg->is_top()) { // Use is dead?
2241 _igvn.replace_input_of(use, idx, C->top());
2242 continue;
2243 }
2244
2245 // If use is referenced through control edge... (idx == 0)
2246 if (mode == IgnoreStripMined && idx == 0) {
2247 LoopNode *head = loop->_head->as_Loop();
2248 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2249 // That node is outside the inner loop, leave it outside the
2250 // outer loop as well to not confuse verification code.
2251 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2252 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2253 continue;
2254 }
2255 }
2256
2257 while(!outer_loop->is_member(get_loop(cfg))) {
2258 prev = cfg;
2259 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2260 }
2261 // If the use occurs after merging several exits from the loop, then
2262 // old value must have dominated all those exits. Since the same old
2263 // value was used on all those exits we did not need a Phi at this
2264 // merge point. NOW we do need a Phi here. Each loop exit value
2265 // is now merged with the peeled body exit; each exit gets its own
2266 // private Phi and those Phis need to be merged here.
2267 Node *phi;
2268 if( prev->is_Region() ) {
2269 if( idx == 0 ) { // Updating control edge?
2270 phi = prev; // Just use existing control
2271 } else { // Else need a new Phi
2272 phi = PhiNode::make( prev, old );
2273 // Now recursively fix up the new uses of old!
2274 for( uint i = 1; i < prev->req(); i++ ) {
2275 worklist.push(phi); // Onto worklist once for each 'old' input
2276 }
2277 }
2278 } else {
2279 // Get new RegionNode merging old and new loop exits
2280 prev = old_new[prev->_idx];
2281 assert( prev, "just made this in step 7" );
2282 if( idx == 0) { // Updating control edge?
2283 phi = prev; // Just use existing control
2284 } else { // Else need a new Phi
2285 // Make a new Phi merging data values properly
2286 phi = PhiNode::make( prev, old );
2287 phi->set_req( 1, nnn );
2288 }
2289 }
2290 // If inserting a new Phi, check for prior hits
2291 if( idx != 0 ) {
2292 Node *hit = _igvn.hash_find_insert(phi);
2293 if( hit == nullptr ) {
2294 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2295 } else { // or
2296 // Remove the new phi from the graph and use the hit
2297 _igvn.remove_dead_node(phi);
2298 phi = hit; // Use existing phi
2299 }
2300 set_ctrl(phi, prev);
2301 }
2302 // Make 'use' use the Phi instead of the old loop body exit value
2303 assert(use->in(idx) == old, "old is still input of use");
2304 // We notify all uses of old, including use, and the indirect uses,
2305 // that may now be optimized because we have replaced old with phi.
2306 _igvn.add_users_to_worklist(old);
2307 if (idx == 0 &&
2308 use->depends_only_on_test()) {
2309 Node* pinned_clone = use->pin_array_access_node();
2310 if (pinned_clone != nullptr) {
2311 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2312 // into the region is left, an array load could become dependent on a condition that's not a range check for
2313 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2314 // floating above its range check.
2315 pinned_clone->set_req(0, phi);
2316 register_new_node_with_ctrl_of(pinned_clone, use);
2317 _igvn.replace_node(use, pinned_clone);
2318 continue;
2319 }
2320 }
2321 _igvn.replace_input_of(use, idx, phi);
2322 if( use->_idx >= new_counter ) { // If updating new phis
2323 // Not needed for correctness, but prevents a weak assert
2324 // in AddPNode from tripping (when we end up with different
2325 // base & derived Phis that will become the same after
2326 // IGVN does CSE).
2327 Node *hit = _igvn.hash_find_insert(use);
2328 if( hit ) // Go ahead and re-hash for hits.
2329 _igvn.replace_node( use, hit );
2330 }
2331 }
2332 }
2333 }
2334
2335 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2336 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2337 bool check_old_new) {
2338 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2339 Node* u = n->fast_out(j);
2340 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2341 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2342 Node* c = phase->get_ctrl(u);
2343 IdealLoopTree* u_loop = phase->get_loop(c);
2344 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2345 if (!loop->is_member(u_loop)) {
2346 if (outer_loop->is_member(u_loop)) {
2347 wq.push(u);
2348 } else {
2349 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2350 // the outer loop too
2351 Node* u_c = u->in(0);
2352 if (u_c != nullptr) {
2353 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2354 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2355 wq.push(u);
2356 }
2357 }
2358 }
2359 }
2360 }
2361 }
2362 }
2363
2364 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2365 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2366 Node_List& extra_data_nodes) {
2367 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2368 CountedLoopNode* cl = head->as_CountedLoop();
2369 Node* l = cl->outer_loop();
2370 Node* tail = cl->outer_loop_tail();
2371 IfNode* le = cl->outer_loop_end();
2372 Node* sfpt = cl->outer_safepoint();
2373 CountedLoopEndNode* cle = cl->loopexit();
2374 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2375 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2376 Node* cle_out = cle->proj_out(false);
2377
2378 Node* new_sfpt = nullptr;
2379 Node* new_cle_out = cle_out->clone();
2380 old_new.map(cle_out->_idx, new_cle_out);
2381 if (mode == CloneIncludesStripMined) {
2382 // clone outer loop body
2383 Node* new_l = l->clone();
2384 Node* new_tail = tail->clone();
2385 IfNode* new_le = le->clone()->as_If();
2386 new_sfpt = sfpt->clone();
2387
2388 set_loop(new_l, outer_loop->_parent);
2389 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2390 set_loop(new_cle_out, outer_loop->_parent);
2391 set_idom(new_cle_out, new_cle, dd);
2392 set_loop(new_sfpt, outer_loop->_parent);
2393 set_idom(new_sfpt, new_cle_out, dd);
2394 set_loop(new_le, outer_loop->_parent);
2395 set_idom(new_le, new_sfpt, dd);
2396 set_loop(new_tail, outer_loop->_parent);
2397 set_idom(new_tail, new_le, dd);
2398 set_idom(new_cl, new_l, dd);
2399
2400 old_new.map(l->_idx, new_l);
2401 old_new.map(tail->_idx, new_tail);
2402 old_new.map(le->_idx, new_le);
2403 old_new.map(sfpt->_idx, new_sfpt);
2404
2405 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2406 new_l->set_req(0, new_l);
2407 new_tail->set_req(0, new_le);
2408 new_le->set_req(0, new_sfpt);
2409 new_sfpt->set_req(0, new_cle_out);
2410 new_cle_out->set_req(0, new_cle);
2411 new_cl->set_req(LoopNode::EntryControl, new_l);
2412
2413 _igvn.register_new_node_with_optimizer(new_l);
2414 _igvn.register_new_node_with_optimizer(new_tail);
2415 _igvn.register_new_node_with_optimizer(new_le);
2416 } else {
2417 Node *newhead = old_new[loop->_head->_idx];
2418 newhead->as_Loop()->clear_strip_mined();
2419 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2420 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2421 }
2422 // Look at data node that were assigned a control in the outer
2423 // loop: they are kept in the outer loop by the safepoint so start
2424 // from the safepoint node's inputs.
2425 IdealLoopTree* outer_loop = get_loop(l);
2426 Node_Stack stack(2);
2427 stack.push(sfpt, 1);
2428 uint new_counter = C->unique();
2429 while (stack.size() > 0) {
2430 Node* n = stack.node();
2431 uint i = stack.index();
2432 while (i < n->req() &&
2433 (n->in(i) == nullptr ||
2434 !has_ctrl(n->in(i)) ||
2435 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2436 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2437 i++;
2438 }
2439 if (i < n->req()) {
2440 stack.set_index(i+1);
2441 stack.push(n->in(i), 0);
2442 } else {
2443 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2444 Node* m = n == sfpt ? new_sfpt : n->clone();
2445 if (m != nullptr) {
2446 for (uint i = 0; i < n->req(); i++) {
2447 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2448 m->set_req(i, old_new[m->in(i)->_idx]);
2449 }
2450 }
2451 } else {
2452 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2453 }
2454 if (n != sfpt) {
2455 extra_data_nodes.push(n);
2456 _igvn.register_new_node_with_optimizer(m);
2457 assert(get_ctrl(n) == cle_out, "what other control?");
2458 set_ctrl(m, new_cle_out);
2459 old_new.map(n->_idx, m);
2460 }
2461 stack.pop();
2462 }
2463 }
2464 if (mode == CloneIncludesStripMined) {
2465 _igvn.register_new_node_with_optimizer(new_sfpt);
2466 _igvn.register_new_node_with_optimizer(new_cle_out);
2467 }
2468 // Some other transformation may have pessimistically assigned some
2469 // data nodes to the outer loop. Set their control so they are out
2470 // of the outer loop.
2471 ResourceMark rm;
2472 Unique_Node_List wq;
2473 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2474 Node* old = extra_data_nodes.at(i);
2475 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2476 }
2477
2478 for (uint i = 0; i < loop->_body.size(); i++) {
2479 Node* old = loop->_body.at(i);
2480 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2481 }
2482
2483 Node* inner_out = sfpt->in(0);
2484 if (inner_out->outcnt() > 1) {
2485 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2486 }
2487
2488 Node* new_ctrl = cl->outer_loop_exit();
2489 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2490 for (uint i = 0; i < wq.size(); i++) {
2491 Node* n = wq.at(i);
2492 set_ctrl(n, new_ctrl);
2493 if (n->in(0) != nullptr) {
2494 _igvn.replace_input_of(n, 0, new_ctrl);
2495 }
2496 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2497 }
2498 } else {
2499 Node *newhead = old_new[loop->_head->_idx];
2500 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2501 }
2502 }
2503
2504 //------------------------------clone_loop-------------------------------------
2505 //
2506 // C L O N E A L O O P B O D Y
2507 //
2508 // This is the basic building block of the loop optimizations. It clones an
2509 // entire loop body. It makes an old_new loop body mapping; with this mapping
2510 // you can find the new-loop equivalent to an old-loop node. All new-loop
2511 // nodes are exactly equal to their old-loop counterparts, all edges are the
2512 // same. All exits from the old-loop now have a RegionNode that merges the
2513 // equivalent new-loop path. This is true even for the normal "loop-exit"
2514 // condition. All uses of loop-invariant old-loop values now come from (one
2515 // or more) Phis that merge their new-loop equivalents.
2516 //
2517 // This operation leaves the graph in an illegal state: there are two valid
2518 // control edges coming from the loop pre-header to both loop bodies. I'll
2519 // definitely have to hack the graph after running this transform.
2520 //
2521 // From this building block I will further edit edges to perform loop peeling
2522 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2523 //
2524 // Parameter side_by_size_idom:
2525 // When side_by_size_idom is null, the dominator tree is constructed for
2526 // the clone loop to dominate the original. Used in construction of
2527 // pre-main-post loop sequence.
2528 // When nonnull, the clone and original are side-by-side, both are
2529 // dominated by the side_by_side_idom node. Used in construction of
2530 // unswitched loops.
2531 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2532 CloneLoopMode mode, Node* side_by_side_idom) {
2533
2534 LoopNode* head = loop->_head->as_Loop();
2535 head->verify_strip_mined(1);
2536
2537 if (C->do_vector_loop() && PrintOpto) {
2538 const char* mname = C->method()->name()->as_quoted_ascii();
2539 if (mname != nullptr) {
2540 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2541 }
2542 }
2543
2544 CloneMap& cm = C->clone_map();
2545 if (C->do_vector_loop()) {
2546 cm.set_clone_idx(cm.max_gen()+1);
2547 #ifndef PRODUCT
2548 if (PrintOpto) {
2549 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2550 loop->dump_head();
2551 }
2552 #endif
2553 }
2554
2555 // Step 1: Clone the loop body. Make the old->new mapping.
2556 clone_loop_body(loop->_body, old_new, &cm);
2557
2558 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2559
2560 // Step 2: Fix the edges in the new body. If the old input is outside the
2561 // loop use it. If the old input is INside the loop, use the corresponding
2562 // new node instead.
2563 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2564
2565 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2566 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2567
2568 // Step 3: Now fix control uses. Loop varying control uses have already
2569 // been fixed up (as part of all input edges in Step 2). Loop invariant
2570 // control uses must be either an IfFalse or an IfTrue. Make a merge
2571 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2572 // refer to this.
2573 Node_List worklist;
2574 uint new_counter = C->unique();
2575 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2576
2577 // Step 4: If loop-invariant use is not control, it must be dominated by a
2578 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2579 // there if needed. Make a Phi there merging old and new used values.
2580 Node_List *split_if_set = nullptr;
2581 Node_List *split_bool_set = nullptr;
2582 Node_List *split_cex_set = nullptr;
2583 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2584
2585 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2586 Node* old = extra_data_nodes.at(i);
2587 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2588 split_bool_set, split_cex_set, worklist, new_counter,
2589 mode);
2590 }
2591
2592 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2593 // the loop uses a condition set in the loop. The original IF probably
2594 // takes control from one or more OLD Regions (which in turn get from NEW
2595 // Regions). In any case, there will be a set of Phis for each merge point
2596 // from the IF up to where the original BOOL def exists the loop.
2597 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2598
2599 }
2600
2601 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2602 if (split_if_set) {
2603 while (split_if_set->size()) {
2604 Node *iff = split_if_set->pop();
2605 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2606 if (iff->in(input)->is_Phi()) {
2607 Node *b = clone_iff(iff->in(input)->as_Phi());
2608 _igvn.replace_input_of(iff, input, b);
2609 }
2610 }
2611 }
2612 if (split_bool_set) {
2613 while (split_bool_set->size()) {
2614 Node *b = split_bool_set->pop();
2615 Node *phi = b->in(1);
2616 assert(phi->is_Phi(), "");
2617 CmpNode *cmp = clone_bool((PhiNode*) phi);
2618 _igvn.replace_input_of(b, 1, cmp);
2619 }
2620 }
2621 if (split_cex_set) {
2622 while (split_cex_set->size()) {
2623 Node *b = split_cex_set->pop();
2624 assert(b->in(0)->is_Region(), "");
2625 assert(b->in(1)->is_Phi(), "");
2626 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2627 split_up(b, b->in(0), nullptr);
2628 }
2629 }
2630 }
2631
2632 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2633 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2634 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2635 for(uint i = 0; i < body.size(); i++ ) {
2636 Node* old = body.at(i);
2637 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2638 split_bool_set, split_cex_set, worklist, new_counter,
2639 mode);
2640 }
2641 }
2642
2643 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2644 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2645 LoopNode* head = loop->_head->as_Loop();
2646 for(uint i = 0; i < body.size(); i++ ) {
2647 Node* old = body.at(i);
2648 if( !old->is_CFG() ) continue;
2649
2650 // Copy uses to a worklist, so I can munge the def-use info
2651 // with impunity.
2652 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2653 worklist.push(old->fast_out(j));
2654 }
2655
2656 while (worklist.size()) { // Visit all uses
2657 Node *use = worklist.pop();
2658 if (!has_node(use)) continue; // Ignore dead nodes
2659 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2660 if (!loop->is_member(use_loop) && use->is_CFG()) {
2661 // Both OLD and USE are CFG nodes here.
2662 assert(use->is_Proj(), "" );
2663 Node* nnn = old_new[old->_idx];
2664
2665 Node* newuse = nullptr;
2666 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2667 CountedLoopNode* cl = head->as_CountedLoop();
2668 CountedLoopEndNode* cle = cl->loopexit();
2669 Node* cle_out = cle->proj_out_or_null(false);
2670 if (use == cle_out) {
2671 IfNode* le = cl->outer_loop_end();
2672 use = le->proj_out(false);
2673 use_loop = get_loop(use);
2674 if (mode == CloneIncludesStripMined) {
2675 nnn = old_new[le->_idx];
2676 } else {
2677 newuse = old_new[cle_out->_idx];
2678 }
2679 }
2680 }
2681 if (newuse == nullptr) {
2682 newuse = use->clone();
2683 }
2684
2685 // Clone the loop exit control projection
2686 if (C->do_vector_loop() && cm != nullptr) {
2687 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2688 }
2689 newuse->set_req(0,nnn);
2690 _igvn.register_new_node_with_optimizer(newuse);
2691 set_loop(newuse, use_loop);
2692 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2693
2694 // We need a Region to merge the exit from the peeled body and the
2695 // exit from the old loop body.
2696 RegionNode *r = new RegionNode(3);
2697 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2698 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2699
2700 // The original user of 'use' uses 'r' instead.
2701 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2702 Node* useuse = use->last_out(l);
2703 _igvn.rehash_node_delayed(useuse);
2704 uint uses_found = 0;
2705 if (useuse->in(0) == use) {
2706 useuse->set_req(0, r);
2707 uses_found++;
2708 if (useuse->is_CFG()) {
2709 // This is not a dom_depth > dd_r because when new
2710 // control flow is constructed by a loop opt, a node and
2711 // its dominator can end up at the same dom_depth
2712 assert(dom_depth(useuse) >= dd_r, "");
2713 set_idom(useuse, r, dom_depth(useuse));
2714 }
2715 }
2716 for (uint k = 1; k < useuse->req(); k++) {
2717 if( useuse->in(k) == use ) {
2718 useuse->set_req(k, r);
2719 uses_found++;
2720 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2721 // This is not a dom_depth > dd_r because when new
2722 // control flow is constructed by a loop opt, a node
2723 // and its dominator can end up at the same dom_depth
2724 assert(dom_depth(useuse) >= dd_r , "");
2725 set_idom(useuse, r, dom_depth(useuse));
2726 }
2727 }
2728 }
2729 l -= uses_found; // we deleted 1 or more copies of this edge
2730 }
2731
2732 assert(use->is_Proj(), "loop exit should be projection");
2733 // lazy_replace() below moves all nodes that are:
2734 // - control dependent on the loop exit or
2735 // - have control set to the loop exit
2736 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2737 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2738 // becomes dead and is "replaced" by the region.
2739 Node* use_clone = use->clone();
2740 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2741 // Now finish up 'r'
2742 r->set_req(1, newuse);
2743 r->set_req(2, use_clone);
2744 _igvn.register_new_node_with_optimizer(r);
2745 set_loop(r, use_loop);
2746 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2747 lazy_replace(use, r);
2748 // Map the (cloned) old use to the new merge point
2749 old_new.map(use_clone->_idx, r);
2750 } // End of if a loop-exit test
2751 }
2752 }
2753 }
2754
2755 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2756 IdealLoopTree* parent, bool partial) {
2757 for(uint i = 0; i < body.size(); i++ ) {
2758 Node *old = body.at(i);
2759 Node *nnn = old_new[old->_idx];
2760 // Fix CFG/Loop controlling the new node
2761 if (has_ctrl(old)) {
2762 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2763 } else {
2764 set_loop(nnn, parent);
2765 if (old->outcnt() > 0) {
2766 Node* dom = idom(old);
2767 if (old_new[dom->_idx] != nullptr) {
2768 dom = old_new[dom->_idx];
2769 set_idom(nnn, dom, dd );
2770 }
2771 }
2772 }
2773 // Correct edges to the new node
2774 for (uint j = 0; j < nnn->req(); j++) {
2775 Node *n = nnn->in(j);
2776 if (n != nullptr) {
2777 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2778 if (loop->is_member(old_in_loop)) {
2779 if (old_new[n->_idx] != nullptr) {
2780 nnn->set_req(j, old_new[n->_idx]);
2781 } else {
2782 assert(!body.contains(n), "");
2783 assert(partial, "node not cloned");
2784 }
2785 }
2786 }
2787 }
2788 _igvn.hash_find_insert(nnn);
2789 }
2790 }
2791
2792 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2793 for (uint i = 0; i < body.size(); i++) {
2794 Node* old = body.at(i);
2795 Node* nnn = old->clone();
2796 old_new.map(old->_idx, nnn);
2797 if (C->do_vector_loop() && cm != nullptr) {
2798 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2799 }
2800 _igvn.register_new_node_with_optimizer(nnn);
2801 }
2802 }
2803
2804
2805 //---------------------- stride_of_possible_iv -------------------------------------
2806 // Looks for an iff/bool/comp with one operand of the compare
2807 // being a cycle involving an add and a phi,
2808 // with an optional truncation (left-shift followed by a right-shift)
2809 // of the add. Returns zero if not an iv.
2810 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2811 Node* trunc1 = nullptr;
2812 Node* trunc2 = nullptr;
2813 const TypeInteger* ttype = nullptr;
2814 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2815 return 0;
2816 }
2817 BoolNode* bl = iff->in(1)->as_Bool();
2818 Node* cmp = bl->in(1);
2819 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2820 return 0;
2821 }
2822 // Must have an invariant operand
2823 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2824 return 0;
2825 }
2826 Node* add2 = nullptr;
2827 Node* cmp1 = cmp->in(1);
2828 if (cmp1->is_Phi()) {
2829 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2830 Node* phi = cmp1;
2831 for (uint i = 1; i < phi->req(); i++) {
2832 Node* in = phi->in(i);
2833 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2834 &trunc1, &trunc2, &ttype, T_INT);
2835 if (add && add->in(1) == phi) {
2836 add2 = add->in(2);
2837 break;
2838 }
2839 }
2840 } else {
2841 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2842 Node* addtrunc = cmp1;
2843 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2844 &trunc1, &trunc2, &ttype, T_INT);
2845 if (add && add->in(1)->is_Phi()) {
2846 Node* phi = add->in(1);
2847 for (uint i = 1; i < phi->req(); i++) {
2848 if (phi->in(i) == addtrunc) {
2849 add2 = add->in(2);
2850 break;
2851 }
2852 }
2853 }
2854 }
2855 if (add2 != nullptr) {
2856 const TypeInt* add2t = _igvn.type(add2)->is_int();
2857 if (add2t->is_con()) {
2858 return add2t->get_con();
2859 }
2860 }
2861 return 0;
2862 }
2863
2864
2865 //---------------------- stay_in_loop -------------------------------------
2866 // Return the (unique) control output node that's in the loop (if it exists.)
2867 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2868 Node* unique = nullptr;
2869 if (!n) return nullptr;
2870 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2871 Node* use = n->fast_out(i);
2872 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2873 if (unique != nullptr) {
2874 return nullptr;
2875 }
2876 unique = use;
2877 }
2878 }
2879 return unique;
2880 }
2881
2882 //------------------------------ register_node -------------------------------------
2883 // Utility to register node "n" with PhaseIdealLoop
2884 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2885 _igvn.register_new_node_with_optimizer(n);
2886 loop->_body.push(n);
2887 if (n->is_CFG()) {
2888 set_loop(n, loop);
2889 set_idom(n, pred, ddepth);
2890 } else {
2891 set_ctrl(n, pred);
2892 }
2893 }
2894
2895 //------------------------------ proj_clone -------------------------------------
2896 // Utility to create an if-projection
2897 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2898 ProjNode* c = p->clone()->as_Proj();
2899 c->set_req(0, iff);
2900 return c;
2901 }
2902
2903 //------------------------------ short_circuit_if -------------------------------------
2904 // Force the iff control output to be the live_proj
2905 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2906 guarantee(live_proj != nullptr, "null projection");
2907 int proj_con = live_proj->_con;
2908 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2909 Node *con = _igvn.intcon(proj_con);
2910 set_ctrl(con, C->root());
2911 if (iff) {
2912 iff->set_req(1, con);
2913 }
2914 return con;
2915 }
2916
2917 //------------------------------ insert_if_before_proj -------------------------------------
2918 // Insert a new if before an if projection (* - new node)
2919 //
2920 // before
2921 // if(test)
2922 // / \
2923 // v v
2924 // other-proj proj (arg)
2925 //
2926 // after
2927 // if(test)
2928 // / \
2929 // / v
2930 // | * proj-clone
2931 // v |
2932 // other-proj v
2933 // * new_if(relop(cmp[IU](left,right)))
2934 // / \
2935 // v v
2936 // * new-proj proj
2937 // (returned)
2938 //
2939 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2940 IfNode* iff = proj->in(0)->as_If();
2941 IdealLoopTree *loop = get_loop(proj);
2942 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2943 uint ddepth = dom_depth(proj);
2944
2945 _igvn.rehash_node_delayed(iff);
2946 _igvn.rehash_node_delayed(proj);
2947
2948 proj->set_req(0, nullptr); // temporary disconnect
2949 ProjNode* proj2 = proj_clone(proj, iff);
2950 register_node(proj2, loop, iff, ddepth);
2951
2952 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2953 register_node(cmp, loop, proj2, ddepth);
2954
2955 BoolNode* bol = new BoolNode(cmp, relop);
2956 register_node(bol, loop, proj2, ddepth);
2957
2958 int opcode = iff->Opcode();
2959 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2960 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2961 register_node(new_if, loop, proj2, ddepth);
2962
2963 proj->set_req(0, new_if); // reattach
2964 set_idom(proj, new_if, ddepth);
2965
2966 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2967 guarantee(new_exit != nullptr, "null exit node");
2968 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2969
2970 return new_exit;
2971 }
2972
2973 //------------------------------ insert_region_before_proj -------------------------------------
2974 // Insert a region before an if projection (* - new node)
2975 //
2976 // before
2977 // if(test)
2978 // / |
2979 // v |
2980 // proj v
2981 // other-proj
2982 //
2983 // after
2984 // if(test)
2985 // / |
2986 // v |
2987 // * proj-clone v
2988 // | other-proj
2989 // v
2990 // * new-region
2991 // |
2992 // v
2993 // * dum_if
2994 // / \
2995 // v \
2996 // * dum-proj v
2997 // proj
2998 //
2999 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3000 IfNode* iff = proj->in(0)->as_If();
3001 IdealLoopTree *loop = get_loop(proj);
3002 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3003 uint ddepth = dom_depth(proj);
3004
3005 _igvn.rehash_node_delayed(iff);
3006 _igvn.rehash_node_delayed(proj);
3007
3008 proj->set_req(0, nullptr); // temporary disconnect
3009 ProjNode* proj2 = proj_clone(proj, iff);
3010 register_node(proj2, loop, iff, ddepth);
3011
3012 RegionNode* reg = new RegionNode(2);
3013 reg->set_req(1, proj2);
3014 register_node(reg, loop, iff, ddepth);
3015
3016 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3017 register_node(dum_if, loop, reg, ddepth);
3018
3019 proj->set_req(0, dum_if); // reattach
3020 set_idom(proj, dum_if, ddepth);
3021
3022 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3023 register_node(dum_proj, loop, dum_if, ddepth);
3024
3025 return reg;
3026 }
3027
3028 // Idea
3029 // ----
3030 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3031 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3032 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3033 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3034 // test alone instead:
3035 //
3036 // Before Partial Peeling:
3037 // Loop:
3038 // <peeled section>
3039 // Split off signed loop exit test
3040 // <-- CUT HERE -->
3041 // Unchanged unsigned loop exit test
3042 // <rest of unpeeled section>
3043 // goto Loop
3044 //
3045 // After Partial Peeling:
3046 // <cloned peeled section>
3047 // Cloned split off signed loop exit test
3048 // Loop:
3049 // Unchanged unsigned loop exit test
3050 // <rest of unpeeled section>
3051 // <peeled section>
3052 // Split off signed loop exit test
3053 // goto Loop
3054 //
3055 // Details
3056 // -------
3057 // Before:
3058 // if (i <u limit) Unsigned loop exit condition
3059 // / |
3060 // v v
3061 // exit-proj stay-in-loop-proj
3062 //
3063 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3064 // before the CmpU on the stay-in-loop path and keep both tests:
3065 //
3066 // if (i <u limit) Signed loop exit test
3067 // / |
3068 // / if (i <u limit) Unsigned loop exit test
3069 // / / |
3070 // v v v
3071 // exit-region stay-in-loop-proj
3072 //
3073 // Implementation
3074 // --------------
3075 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3076 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3077 // exit tests is preserved, and their loop nesting is correct.
3078 //
3079 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3080 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3081 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3082 //
3083 // if (stay-in-loop-const) Killed original unsigned loop exit test
3084 // / |
3085 // / v
3086 // / if (i < limit) Split off signed loop exit test
3087 // / / |
3088 // / / v
3089 // / / if (i <u limit) Cloned unsigned loop exit test
3090 // / / / |
3091 // v v v |
3092 // exit-region |
3093 // | |
3094 // dummy-if |
3095 // / | |
3096 // dead | |
3097 // v v
3098 // exit-proj stay-in-loop-proj
3099 //
3100 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3101 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3102 //
3103 // Requirements
3104 // ------------
3105 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3106 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3107 // Otherwise, we need to bail out (see comments in the code below).
3108 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3109 const bool Signed = true;
3110 const bool Unsigned = false;
3111
3112 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3113 if (bol->_test._test != BoolTest::lt) {
3114 return nullptr;
3115 }
3116 CmpNode* cmpu = bol->in(1)->as_Cmp();
3117 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3118
3119 int stride = stride_of_possible_iv(if_cmpu);
3120 if (stride == 0) {
3121 return nullptr;
3122 }
3123
3124 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3125 guarantee(lp_proj != nullptr, "null loop node");
3126
3127 ProjNode* lp_continue = lp_proj->as_Proj();
3128 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3129 if (!lp_exit->is_IfFalse()) {
3130 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3131 // We therefore can't add a single exit condition.
3132 return nullptr;
3133 }
3134 // The unsigned loop exit condition is
3135 // !(i <u limit)
3136 // = i >=u limit
3137 //
3138 // First, we note that for any x for which
3139 // 0 <= x <= INT_MAX
3140 // we can convert x to an unsigned int and still get the same guarantee:
3141 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3142 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3143 //
3144 // With that in mind, if
3145 // limit >= 0 (COND)
3146 // then the unsigned loop exit condition
3147 // i >=u limit (ULE)
3148 // is equivalent to
3149 // i < 0 || i >= limit (SLE-full)
3150 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3151 // (uint) i >=u MAX_INT >= limit >= 0
3152 // or otherwise
3153 // i >= limit >= 0
3154 // holds due to (LEMMA).
3155 //
3156 // For completeness, a counterexample with limit < 0:
3157 // Assume i = -3 and limit = -2:
3158 // i < 0
3159 // -2 < 0
3160 // is true and thus also "i < 0 || i >= limit". But
3161 // i >=u limit
3162 // -3 >=u -2
3163 // is false.
3164 Node* limit = cmpu->in(2);
3165 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3166 if (type_limit->_lo < 0) {
3167 return nullptr;
3168 }
3169
3170 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3171 // stride < 0:
3172 // i < 0 (SLE = SLE-negative)
3173 // stride > 0:
3174 // i >= limit (SLE = SLE-positive)
3175 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3176 //
3177 // Loop:
3178 // <peeled section>
3179 // i >= limit (SLE-positive)
3180 // <-- CUT HERE -->
3181 // i >=u limit (ULE)
3182 // <rest of unpeeled section>
3183 // goto Loop
3184 //
3185 // We exit the loop if:
3186 // (SLE) is true OR (ULE) is true
3187 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3188 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3189 // (SLE) IMPLIES (ULE)
3190 // This indeed holds when (COND) is given:
3191 // - stride > 0:
3192 // i >= limit // (SLE = SLE-positive)
3193 // i >= limit >= 0 // (COND)
3194 // i >=u limit >= 0 // (LEMMA)
3195 // which is the unsigned loop exit condition (ULE).
3196 // - stride < 0:
3197 // i < 0 // (SLE = SLE-negative)
3198 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3199 // MAX_INT >= limit >= 0 // (COND)
3200 // MAX_INT >=u limit >= 0 // (LEMMA)
3201 // and thus from (NEG) and (LEMMA):
3202 // i >=u limit
3203 // which is the unsigned loop exit condition (ULE).
3204 //
3205 //
3206 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3207 // <cloned peeled section>
3208 // i >= limit (SLE-positive)
3209 // Loop:
3210 // i >=u limit (ULE)
3211 // <rest of unpeeled section>
3212 // <peeled section>
3213 // i >= limit (SLE-positive)
3214 // goto Loop
3215 Node* rhs_cmpi;
3216 if (stride > 0) {
3217 rhs_cmpi = limit; // For i >= limit
3218 } else {
3219 rhs_cmpi = _igvn.makecon(TypeInt::ZERO); // For i < 0
3220 set_ctrl(rhs_cmpi, C->root());
3221 }
3222 // Create a new region on the exit path
3223 RegionNode* reg = insert_region_before_proj(lp_exit);
3224 guarantee(reg != nullptr, "null region node");
3225
3226 // Clone the if-cmpu-true-false using a signed compare
3227 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3228 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3229 reg->add_req(cmpi_exit);
3230
3231 // Clone the if-cmpu-true-false
3232 BoolTest::mask rel_u = bol->_test._test;
3233 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3234 reg->add_req(cmpu_exit);
3235
3236 // Force original if to stay in loop.
3237 short_circuit_if(if_cmpu, lp_continue);
3238
3239 return cmpi_exit->in(0)->as_If();
3240 }
3241
3242 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3243 // Remove a previously inserted signed compare loop exit.
3244 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3245 Node* lp_proj = stay_in_loop(if_cmp, loop);
3246 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3247 stay_in_loop(lp_proj, loop)->is_If() &&
3248 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3249 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3250 set_ctrl(con, C->root());
3251 if_cmp->set_req(1, con);
3252 }
3253
3254 //------------------------------ scheduled_nodelist -------------------------------------
3255 // Create a post order schedule of nodes that are in the
3256 // "member" set. The list is returned in "sched".
3257 // The first node in "sched" is the loop head, followed by
3258 // nodes which have no inputs in the "member" set, and then
3259 // followed by the nodes that have an immediate input dependence
3260 // on a node in "sched".
3261 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3262
3263 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3264 VectorSet visited;
3265 Node_Stack nstack(loop->_body.size());
3266
3267 Node* n = loop->_head; // top of stack is cached in "n"
3268 uint idx = 0;
3269 visited.set(n->_idx);
3270
3271 // Initially push all with no inputs from within member set
3272 for(uint i = 0; i < loop->_body.size(); i++ ) {
3273 Node *elt = loop->_body.at(i);
3274 if (member.test(elt->_idx)) {
3275 bool found = false;
3276 for (uint j = 0; j < elt->req(); j++) {
3277 Node* def = elt->in(j);
3278 if (def && member.test(def->_idx) && def != elt) {
3279 found = true;
3280 break;
3281 }
3282 }
3283 if (!found && elt != loop->_head) {
3284 nstack.push(n, idx);
3285 n = elt;
3286 assert(!visited.test(n->_idx), "not seen yet");
3287 visited.set(n->_idx);
3288 }
3289 }
3290 }
3291
3292 // traverse out's that are in the member set
3293 while (true) {
3294 if (idx < n->outcnt()) {
3295 Node* use = n->raw_out(idx);
3296 idx++;
3297 if (!visited.test_set(use->_idx)) {
3298 if (member.test(use->_idx)) {
3299 nstack.push(n, idx);
3300 n = use;
3301 idx = 0;
3302 }
3303 }
3304 } else {
3305 // All outputs processed
3306 sched.push(n);
3307 if (nstack.is_empty()) break;
3308 n = nstack.node();
3309 idx = nstack.index();
3310 nstack.pop();
3311 }
3312 }
3313 }
3314
3315
3316 //------------------------------ has_use_in_set -------------------------------------
3317 // Has a use in the vector set
3318 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3319 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3320 Node* use = n->fast_out(j);
3321 if (vset.test(use->_idx)) {
3322 return true;
3323 }
3324 }
3325 return false;
3326 }
3327
3328
3329 //------------------------------ has_use_internal_to_set -------------------------------------
3330 // Has use internal to the vector set (ie. not in a phi at the loop head)
3331 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3332 Node* head = loop->_head;
3333 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3334 Node* use = n->fast_out(j);
3335 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3336 return true;
3337 }
3338 }
3339 return false;
3340 }
3341
3342
3343 //------------------------------ clone_for_use_outside_loop -------------------------------------
3344 // clone "n" for uses that are outside of loop
3345 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3346 int cloned = 0;
3347 assert(worklist.size() == 0, "should be empty");
3348 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3349 Node* use = n->fast_out(j);
3350 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3351 worklist.push(use);
3352 }
3353 }
3354
3355 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3356 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3357 return -1;
3358 }
3359
3360 while( worklist.size() ) {
3361 Node *use = worklist.pop();
3362 if (!has_node(use) || use->in(0) == C->top()) continue;
3363 uint j;
3364 for (j = 0; j < use->req(); j++) {
3365 if (use->in(j) == n) break;
3366 }
3367 assert(j < use->req(), "must be there");
3368
3369 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3370 Node* n_clone = n->clone();
3371 _igvn.replace_input_of(use, j, n_clone);
3372 cloned++;
3373 Node* use_c;
3374 if (!use->is_Phi()) {
3375 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3376 } else {
3377 // Use in a phi is considered a use in the associated predecessor block
3378 use_c = use->in(0)->in(j);
3379 }
3380 set_ctrl(n_clone, use_c);
3381 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3382 get_loop(use_c)->_body.push(n_clone);
3383 _igvn.register_new_node_with_optimizer(n_clone);
3384 #ifndef PRODUCT
3385 if (TracePartialPeeling) {
3386 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3387 }
3388 #endif
3389 }
3390 return cloned;
3391 }
3392
3393
3394 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3395 // clone "n" for special uses that are in the not_peeled region.
3396 // If these def-uses occur in separate blocks, the code generator
3397 // marks the method as not compilable. For example, if a "BoolNode"
3398 // is in a different basic block than the "IfNode" that uses it, then
3399 // the compilation is aborted in the code generator.
3400 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3401 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3402 if (n->is_Phi() || n->is_Load()) {
3403 return;
3404 }
3405 assert(worklist.size() == 0, "should be empty");
3406 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3407 Node* use = n->fast_out(j);
3408 if ( not_peel.test(use->_idx) &&
3409 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3410 use->in(1) == n) {
3411 worklist.push(use);
3412 }
3413 }
3414 if (worklist.size() > 0) {
3415 // clone "n" and insert it between inputs of "n" and the use
3416 Node* n_clone = n->clone();
3417 loop->_body.push(n_clone);
3418 _igvn.register_new_node_with_optimizer(n_clone);
3419 set_ctrl(n_clone, get_ctrl(n));
3420 sink_list.push(n_clone);
3421 not_peel.set(n_clone->_idx);
3422 #ifndef PRODUCT
3423 if (TracePartialPeeling) {
3424 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3425 }
3426 #endif
3427 while( worklist.size() ) {
3428 Node *use = worklist.pop();
3429 _igvn.rehash_node_delayed(use);
3430 for (uint j = 1; j < use->req(); j++) {
3431 if (use->in(j) == n) {
3432 use->set_req(j, n_clone);
3433 }
3434 }
3435 }
3436 }
3437 }
3438
3439
3440 //------------------------------ insert_phi_for_loop -------------------------------------
3441 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3442 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3443 Node *phi = PhiNode::make(lp, back_edge_val);
3444 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3445 // Use existing phi if it already exists
3446 Node *hit = _igvn.hash_find_insert(phi);
3447 if( hit == nullptr ) {
3448 _igvn.register_new_node_with_optimizer(phi);
3449 set_ctrl(phi, lp);
3450 } else {
3451 // Remove the new phi from the graph and use the hit
3452 _igvn.remove_dead_node(phi);
3453 phi = hit;
3454 }
3455 _igvn.replace_input_of(use, idx, phi);
3456 }
3457
3458 #ifdef ASSERT
3459 //------------------------------ is_valid_loop_partition -------------------------------------
3460 // Validate the loop partition sets: peel and not_peel
3461 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3462 VectorSet& not_peel ) {
3463 uint i;
3464 // Check that peel_list entries are in the peel set
3465 for (i = 0; i < peel_list.size(); i++) {
3466 if (!peel.test(peel_list.at(i)->_idx)) {
3467 return false;
3468 }
3469 }
3470 // Check at loop members are in one of peel set or not_peel set
3471 for (i = 0; i < loop->_body.size(); i++ ) {
3472 Node *def = loop->_body.at(i);
3473 uint di = def->_idx;
3474 // Check that peel set elements are in peel_list
3475 if (peel.test(di)) {
3476 if (not_peel.test(di)) {
3477 return false;
3478 }
3479 // Must be in peel_list also
3480 bool found = false;
3481 for (uint j = 0; j < peel_list.size(); j++) {
3482 if (peel_list.at(j)->_idx == di) {
3483 found = true;
3484 break;
3485 }
3486 }
3487 if (!found) {
3488 return false;
3489 }
3490 } else if (not_peel.test(di)) {
3491 if (peel.test(di)) {
3492 return false;
3493 }
3494 } else {
3495 return false;
3496 }
3497 }
3498 return true;
3499 }
3500
3501 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3502 // Ensure a use outside of loop is of the right form
3503 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3504 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3505 return (use->is_Phi() &&
3506 use_c->is_Region() && use_c->req() == 3 &&
3507 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3508 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3509 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3510 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3511 }
3512
3513 //------------------------------ is_valid_clone_loop_form -------------------------------------
3514 // Ensure that all uses outside of loop are of the right form
3515 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3516 uint orig_exit_idx, uint clone_exit_idx) {
3517 uint len = peel_list.size();
3518 for (uint i = 0; i < len; i++) {
3519 Node *def = peel_list.at(i);
3520
3521 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3522 Node *use = def->fast_out(j);
3523 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3524 if (!loop->is_member(get_loop(use_c))) {
3525 // use is not in the loop, check for correct structure
3526 if (use->in(0) == def) {
3527 // Okay
3528 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3529 return false;
3530 }
3531 }
3532 }
3533 }
3534 return true;
3535 }
3536 #endif
3537
3538 //------------------------------ partial_peel -------------------------------------
3539 // Partially peel (aka loop rotation) the top portion of a loop (called
3540 // the peel section below) by cloning it and placing one copy just before
3541 // the new loop head and the other copy at the bottom of the new loop.
3542 //
3543 // before after where it came from
3544 //
3545 // stmt1 stmt1
3546 // loop: stmt2 clone
3547 // stmt2 if condA goto exitA clone
3548 // if condA goto exitA new_loop: new
3549 // stmt3 stmt3 clone
3550 // if !condB goto loop if condB goto exitB clone
3551 // exitB: stmt2 orig
3552 // stmt4 if !condA goto new_loop orig
3553 // exitA: goto exitA
3554 // exitB:
3555 // stmt4
3556 // exitA:
3557 //
3558 // Step 1: find the cut point: an exit test on probable
3559 // induction variable.
3560 // Step 2: schedule (with cloning) operations in the peel
3561 // section that can be executed after the cut into
3562 // the section that is not peeled. This may need
3563 // to clone operations into exit blocks. For
3564 // instance, a reference to A[i] in the not-peel
3565 // section and a reference to B[i] in an exit block
3566 // may cause a left-shift of i by 2 to be placed
3567 // in the peel block. This step will clone the left
3568 // shift into the exit block and sink the left shift
3569 // from the peel to the not-peel section.
3570 // Step 3: clone the loop, retarget the control, and insert
3571 // phis for values that are live across the new loop
3572 // head. This is very dependent on the graph structure
3573 // from clone_loop. It creates region nodes for
3574 // exit control and associated phi nodes for values
3575 // flow out of the loop through that exit. The region
3576 // node is dominated by the clone's control projection.
3577 // So the clone's peel section is placed before the
3578 // new loop head, and the clone's not-peel section is
3579 // forms the top part of the new loop. The original
3580 // peel section forms the tail of the new loop.
3581 // Step 4: update the dominator tree and recompute the
3582 // dominator depth.
3583 //
3584 // orig
3585 //
3586 // stmt1
3587 // |
3588 // v
3589 // predicates
3590 // |
3591 // v
3592 // loop<----+
3593 // | |
3594 // stmt2 |
3595 // | |
3596 // v |
3597 // ifA |
3598 // / | |
3599 // v v |
3600 // false true ^ <-- last_peel
3601 // / | |
3602 // / ===|==cut |
3603 // / stmt3 | <-- first_not_peel
3604 // / | |
3605 // | v |
3606 // v ifB |
3607 // exitA: / \ |
3608 // / \ |
3609 // v v |
3610 // false true |
3611 // / \ |
3612 // / ----+
3613 // |
3614 // v
3615 // exitB:
3616 // stmt4
3617 //
3618 //
3619 // after clone loop
3620 //
3621 // stmt1
3622 // |
3623 // v
3624 // predicates
3625 // / \
3626 // clone / \ orig
3627 // / \
3628 // / \
3629 // v v
3630 // +---->loop loop<----+
3631 // | | | |
3632 // | stmt2 stmt2 |
3633 // | | | |
3634 // | v v |
3635 // | ifA ifA |
3636 // | | \ / | |
3637 // | v v v v |
3638 // ^ true false false true ^ <-- last_peel
3639 // | | ^ \ / | |
3640 // | cut==|== \ \ / ===|==cut |
3641 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3642 // | | dom | | | |
3643 // | v \ 1v v2 v |
3644 // | ifB regionA ifB |
3645 // | / \ | / \ |
3646 // | / \ v / \ |
3647 // | v v exitA: v v |
3648 // | true false false true |
3649 // | / ^ \ / \ |
3650 // +---- \ \ / ----+
3651 // dom \ /
3652 // \ 1v v2
3653 // regionB
3654 // |
3655 // v
3656 // exitB:
3657 // stmt4
3658 //
3659 //
3660 // after partial peel
3661 //
3662 // stmt1
3663 // |
3664 // v
3665 // predicates
3666 // /
3667 // clone / orig
3668 // / TOP
3669 // / \
3670 // v v
3671 // TOP->loop loop----+
3672 // | | |
3673 // stmt2 stmt2 |
3674 // | | |
3675 // v v |
3676 // ifA ifA |
3677 // | \ / | |
3678 // v v v v |
3679 // true false false true | <-- last_peel
3680 // | ^ \ / +------|---+
3681 // +->newloop \ \ / === ==cut | |
3682 // | stmt3 \ \ / TOP | |
3683 // | | dom | | stmt3 | | <-- first_not_peel
3684 // | v \ 1v v2 v | |
3685 // | ifB regionA ifB ^ v
3686 // | / \ | / \ | |
3687 // | / \ v / \ | |
3688 // | v v exitA: v v | |
3689 // | true false false true | |
3690 // | / ^ \ / \ | |
3691 // | | \ \ / v | |
3692 // | | dom \ / TOP | |
3693 // | | \ 1v v2 | |
3694 // ^ v regionB | |
3695 // | | | | |
3696 // | | v ^ v
3697 // | | exitB: | |
3698 // | | stmt4 | |
3699 // | +------------>-----------------+ |
3700 // | |
3701 // +-----------------<---------------------+
3702 //
3703 //
3704 // final graph
3705 //
3706 // stmt1
3707 // |
3708 // v
3709 // predicates
3710 // |
3711 // v
3712 // stmt2 clone
3713 // |
3714 // v
3715 // ........> ifA clone
3716 // : / |
3717 // dom / |
3718 // : v v
3719 // : false true
3720 // : | |
3721 // : | v
3722 // : | newloop<-----+
3723 // : | | |
3724 // : | stmt3 clone |
3725 // : | | |
3726 // : | v |
3727 // : | ifB |
3728 // : | / \ |
3729 // : | v v |
3730 // : | false true |
3731 // : | | | |
3732 // : | v stmt2 |
3733 // : | exitB: | |
3734 // : | stmt4 v |
3735 // : | ifA orig |
3736 // : | / \ |
3737 // : | / \ |
3738 // : | v v |
3739 // : | false true |
3740 // : | / \ |
3741 // : v v -----+
3742 // RegionA
3743 // |
3744 // v
3745 // exitA
3746 //
3747 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3748
3749 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3750 if (!loop->_head->is_Loop()) {
3751 return false;
3752 }
3753 LoopNode *head = loop->_head->as_Loop();
3754
3755 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3756 return false;
3757 }
3758
3759 // Check for complex exit control
3760 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3761 Node *n = loop->_body.at(ii);
3762 int opc = n->Opcode();
3763 if (n->is_Call() ||
3764 opc == Op_Catch ||
3765 opc == Op_CatchProj ||
3766 opc == Op_Jump ||
3767 opc == Op_JumpProj) {
3768 #ifndef PRODUCT
3769 if (TracePartialPeeling) {
3770 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3771 }
3772 #endif
3773 return false;
3774 }
3775 }
3776
3777 int dd = dom_depth(head);
3778
3779 // Step 1: find cut point
3780
3781 // Walk up dominators to loop head looking for first loop exit
3782 // which is executed on every path thru loop.
3783 IfNode *peel_if = nullptr;
3784 IfNode *peel_if_cmpu = nullptr;
3785
3786 Node *iff = loop->tail();
3787 while (iff != head) {
3788 if (iff->is_If()) {
3789 Node *ctrl = get_ctrl(iff->in(1));
3790 if (ctrl->is_top()) return false; // Dead test on live IF.
3791 // If loop-varying exit-test, check for induction variable
3792 if (loop->is_member(get_loop(ctrl)) &&
3793 loop->is_loop_exit(iff) &&
3794 is_possible_iv_test(iff)) {
3795 Node* cmp = iff->in(1)->in(1);
3796 if (cmp->Opcode() == Op_CmpI) {
3797 peel_if = iff->as_If();
3798 } else {
3799 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3800 peel_if_cmpu = iff->as_If();
3801 }
3802 }
3803 }
3804 iff = idom(iff);
3805 }
3806
3807 // Prefer signed compare over unsigned compare.
3808 IfNode* new_peel_if = nullptr;
3809 if (peel_if == nullptr) {
3810 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3811 return false; // No peel point found
3812 }
3813 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3814 if (new_peel_if == nullptr) {
3815 return false; // No peel point found
3816 }
3817 peel_if = new_peel_if;
3818 }
3819 Node* last_peel = stay_in_loop(peel_if, loop);
3820 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3821 if (first_not_peeled == nullptr || first_not_peeled == head) {
3822 return false;
3823 }
3824
3825 #ifndef PRODUCT
3826 if (TraceLoopOpts) {
3827 tty->print("PartialPeel ");
3828 loop->dump_head();
3829 }
3830
3831 if (TracePartialPeeling) {
3832 tty->print_cr("before partial peel one iteration");
3833 Node_List wl;
3834 Node* t = head->in(2);
3835 while (true) {
3836 wl.push(t);
3837 if (t == head) break;
3838 t = idom(t);
3839 }
3840 while (wl.size() > 0) {
3841 Node* tt = wl.pop();
3842 tt->dump();
3843 if (tt == last_peel) tty->print_cr("-- cut --");
3844 }
3845 }
3846 #endif
3847
3848 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3849
3850 VectorSet peel;
3851 VectorSet not_peel;
3852 Node_List peel_list;
3853 Node_List worklist;
3854 Node_List sink_list;
3855
3856 uint estimate = loop->est_loop_clone_sz(1);
3857 if (exceeding_node_budget(estimate)) {
3858 return false;
3859 }
3860
3861 // Set of cfg nodes to peel are those that are executable from
3862 // the head through last_peel.
3863 assert(worklist.size() == 0, "should be empty");
3864 worklist.push(head);
3865 peel.set(head->_idx);
3866 while (worklist.size() > 0) {
3867 Node *n = worklist.pop();
3868 if (n != last_peel) {
3869 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3870 Node* use = n->fast_out(j);
3871 if (use->is_CFG() &&
3872 loop->is_member(get_loop(use)) &&
3873 !peel.test_set(use->_idx)) {
3874 worklist.push(use);
3875 }
3876 }
3877 }
3878 }
3879
3880 // Set of non-cfg nodes to peel are those that are control
3881 // dependent on the cfg nodes.
3882 for (uint i = 0; i < loop->_body.size(); i++) {
3883 Node *n = loop->_body.at(i);
3884 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3885 if (peel.test(n_c->_idx)) {
3886 peel.set(n->_idx);
3887 } else {
3888 not_peel.set(n->_idx);
3889 }
3890 }
3891
3892 // Step 2: move operations from the peeled section down into the
3893 // not-peeled section
3894
3895 // Get a post order schedule of nodes in the peel region
3896 // Result in right-most operand.
3897 scheduled_nodelist(loop, peel, peel_list);
3898
3899 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3900
3901 // For future check for too many new phis
3902 uint old_phi_cnt = 0;
3903 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3904 Node* use = head->fast_out(j);
3905 if (use->is_Phi()) old_phi_cnt++;
3906 }
3907
3908 #ifndef PRODUCT
3909 if (TracePartialPeeling) {
3910 tty->print_cr("\npeeled list");
3911 }
3912 #endif
3913
3914 // Evacuate nodes in peel region into the not_peeled region if possible
3915 bool too_many_clones = false;
3916 uint new_phi_cnt = 0;
3917 uint cloned_for_outside_use = 0;
3918 for (uint i = 0; i < peel_list.size();) {
3919 Node* n = peel_list.at(i);
3920 #ifndef PRODUCT
3921 if (TracePartialPeeling) n->dump();
3922 #endif
3923 bool incr = true;
3924 if (!n->is_CFG()) {
3925 if (has_use_in_set(n, not_peel)) {
3926 // If not used internal to the peeled region,
3927 // move "n" from peeled to not_peeled region.
3928 if (!has_use_internal_to_set(n, peel, loop)) {
3929 // if not pinned and not a load (which maybe anti-dependent on a store)
3930 // and not a CMove (Matcher expects only bool->cmove).
3931 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3932 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3933 if (C->failing()) return false;
3934 if (new_clones == -1) {
3935 too_many_clones = true;
3936 break;
3937 }
3938 cloned_for_outside_use += new_clones;
3939 sink_list.push(n);
3940 peel.remove(n->_idx);
3941 not_peel.set(n->_idx);
3942 peel_list.remove(i);
3943 incr = false;
3944 #ifndef PRODUCT
3945 if (TracePartialPeeling) {
3946 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3947 n->_idx, get_ctrl(n)->_idx);
3948 }
3949 #endif
3950 }
3951 } else {
3952 // Otherwise check for special def-use cases that span
3953 // the peel/not_peel boundary such as bool->if
3954 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3955 new_phi_cnt++;
3956 }
3957 }
3958 }
3959 if (incr) i++;
3960 }
3961
3962 estimate += cloned_for_outside_use + new_phi_cnt;
3963 bool exceed_node_budget = !may_require_nodes(estimate);
3964 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3965
3966 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3967 #ifndef PRODUCT
3968 if (TracePartialPeeling && exceed_phi_limit) {
3969 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3970 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3971 }
3972 #endif
3973 if (new_peel_if != nullptr) {
3974 remove_cmpi_loop_exit(new_peel_if, loop);
3975 }
3976 // Inhibit more partial peeling on this loop
3977 assert(!head->is_partial_peel_loop(), "not partial peeled");
3978 head->mark_partial_peel_failed();
3979 if (cloned_for_outside_use > 0) {
3980 // Terminate this round of loop opts because
3981 // the graph outside this loop was changed.
3982 C->set_major_progress();
3983 return true;
3984 }
3985 return false;
3986 }
3987
3988 // Step 3: clone loop, retarget control, and insert new phis
3989
3990 // Create new loop head for new phis and to hang
3991 // the nodes being moved (sinked) from the peel region.
3992 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3993 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3994 _igvn.register_new_node_with_optimizer(new_head);
3995 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3996 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3997 set_loop(new_head, loop);
3998 loop->_body.push(new_head);
3999 not_peel.set(new_head->_idx);
4000 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4001 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4002
4003 while (sink_list.size() > 0) {
4004 Node* n = sink_list.pop();
4005 set_ctrl(n, new_head);
4006 }
4007
4008 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4009
4010 clone_loop(loop, old_new, dd, IgnoreStripMined);
4011
4012 const uint clone_exit_idx = 1;
4013 const uint orig_exit_idx = 2;
4014 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4015
4016 Node* head_clone = old_new[head->_idx];
4017 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4018 Node* orig_tail_clone = head_clone->in(2);
4019
4020 // Add phi if "def" node is in peel set and "use" is not
4021
4022 for (uint i = 0; i < peel_list.size(); i++) {
4023 Node *def = peel_list.at(i);
4024 if (!def->is_CFG()) {
4025 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4026 Node *use = def->fast_out(j);
4027 if (has_node(use) && use->in(0) != C->top() &&
4028 (!peel.test(use->_idx) ||
4029 (use->is_Phi() && use->in(0) == head)) ) {
4030 worklist.push(use);
4031 }
4032 }
4033 while( worklist.size() ) {
4034 Node *use = worklist.pop();
4035 for (uint j = 1; j < use->req(); j++) {
4036 Node* n = use->in(j);
4037 if (n == def) {
4038
4039 // "def" is in peel set, "use" is not in peel set
4040 // or "use" is in the entry boundary (a phi) of the peel set
4041
4042 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4043
4044 if ( loop->is_member(get_loop( use_c )) ) {
4045 // use is in loop
4046 if (old_new[use->_idx] != nullptr) { // null for dead code
4047 Node* use_clone = old_new[use->_idx];
4048 _igvn.replace_input_of(use, j, C->top());
4049 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4050 }
4051 } else {
4052 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4053 // use is not in the loop, check if the live range includes the cut
4054 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4055 if (not_peel.test(lp_if->_idx)) {
4056 assert(j == orig_exit_idx, "use from original loop");
4057 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4058 }
4059 }
4060 }
4061 }
4062 }
4063 }
4064 }
4065
4066 // Step 3b: retarget control
4067
4068 // Redirect control to the new loop head if a cloned node in
4069 // the not_peeled region has control that points into the peeled region.
4070 // This necessary because the cloned peeled region will be outside
4071 // the loop.
4072 // from to
4073 // cloned-peeled <---+
4074 // new_head_clone: | <--+
4075 // cloned-not_peeled in(0) in(0)
4076 // orig-peeled
4077
4078 for (uint i = 0; i < loop->_body.size(); i++) {
4079 Node *n = loop->_body.at(i);
4080 if (!n->is_CFG() && n->in(0) != nullptr &&
4081 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4082 Node* n_clone = old_new[n->_idx];
4083 if (n_clone->depends_only_on_test()) {
4084 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4085 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4086 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4087 // floating above its range check.
4088 Node* pinned_clone = n_clone->pin_array_access_node();
4089 if (pinned_clone != nullptr) {
4090 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4091 old_new.map(n->_idx, pinned_clone);
4092 _igvn.replace_node(n_clone, pinned_clone);
4093 n_clone = pinned_clone;
4094 }
4095 }
4096 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4097 }
4098 }
4099
4100 // Backedge of the surviving new_head (the clone) is original last_peel
4101 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4102
4103 // Cut first node in original not_peel set
4104 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4105 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4106 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4107
4108 // Copy head_clone back-branch info to original head
4109 // and remove original head's loop entry and
4110 // clone head's back-branch
4111 _igvn.rehash_node_delayed(head); // Multiple edge updates
4112 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4113 head->set_req(LoopNode::LoopBackControl, C->top());
4114 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4115
4116 // Similarly modify the phis
4117 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4118 Node* use = head->fast_out(k);
4119 if (use->is_Phi() && use->outcnt() > 0) {
4120 Node* use_clone = old_new[use->_idx];
4121 _igvn.rehash_node_delayed(use); // Multiple edge updates
4122 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4123 use->set_req(LoopNode::LoopBackControl, C->top());
4124 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4125 }
4126 }
4127
4128 // Step 4: update dominator tree and dominator depth
4129
4130 set_idom(head, orig_tail_clone, dd);
4131 recompute_dom_depth();
4132
4133 // Inhibit more partial peeling on this loop
4134 new_head_clone->set_partial_peel_loop();
4135 C->set_major_progress();
4136 loop->record_for_igvn();
4137
4138 #ifndef PRODUCT
4139 if (TracePartialPeeling) {
4140 tty->print_cr("\nafter partial peel one iteration");
4141 Node_List wl;
4142 Node* t = last_peel;
4143 while (true) {
4144 wl.push(t);
4145 if (t == head_clone) break;
4146 t = idom(t);
4147 }
4148 while (wl.size() > 0) {
4149 Node* tt = wl.pop();
4150 if (tt == head) tty->print_cr("orig head");
4151 else if (tt == new_head_clone) tty->print_cr("new head");
4152 else if (tt == head_clone) tty->print_cr("clone head");
4153 tt->dump();
4154 }
4155 }
4156 #endif
4157
4158 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4159
4160 return true;
4161 }
4162
4163 // Transform:
4164 //
4165 // loop<-----------------+
4166 // | |
4167 // stmt1 stmt2 .. stmtn |
4168 // | | | |
4169 // \ | / |
4170 // v v v |
4171 // region |
4172 // | |
4173 // shared_stmt |
4174 // | |
4175 // v |
4176 // if |
4177 // / \ |
4178 // | -----------+
4179 // v
4180 //
4181 // into:
4182 //
4183 // loop<-------------------+
4184 // | |
4185 // v |
4186 // +->loop |
4187 // | | |
4188 // | stmt1 stmt2 .. stmtn |
4189 // | | | | |
4190 // | | \ / |
4191 // | | v v |
4192 // | | region1 |
4193 // | | | |
4194 // | shared_stmt shared_stmt |
4195 // | | | |
4196 // | v v |
4197 // | if if |
4198 // | /\ / \ |
4199 // +-- | | -------+
4200 // \ /
4201 // v v
4202 // region2
4203 //
4204 // (region2 is shown to merge mirrored projections of the loop exit
4205 // ifs to make the diagram clearer but they really merge the same
4206 // projection)
4207 //
4208 // Conditions for this transformation to trigger:
4209 // - the path through stmt1 is frequent enough
4210 // - the inner loop will be turned into a counted loop after transformation
4211 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4212 if (!DuplicateBackedge) {
4213 return false;
4214 }
4215 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4216 if (!loop->_head->is_Loop()) {
4217 return false;
4218 }
4219
4220 uint estimate = loop->est_loop_clone_sz(1);
4221 if (exceeding_node_budget(estimate)) {
4222 return false;
4223 }
4224
4225 LoopNode *head = loop->_head->as_Loop();
4226
4227 Node* region = nullptr;
4228 IfNode* exit_test = nullptr;
4229 uint inner;
4230 float f;
4231 if (StressDuplicateBackedge) {
4232 if (head->is_strip_mined()) {
4233 return false;
4234 }
4235 Node* c = head->in(LoopNode::LoopBackControl);
4236
4237 while (c != head) {
4238 if (c->is_Region()) {
4239 region = c;
4240 }
4241 c = idom(c);
4242 }
4243
4244 if (region == nullptr) {
4245 return false;
4246 }
4247
4248 inner = 1;
4249 } else {
4250 // Is the shape of the loop that of a counted loop...
4251 Node* back_control = loop_exit_control(head, loop);
4252 if (back_control == nullptr) {
4253 return false;
4254 }
4255
4256 BoolTest::mask bt = BoolTest::illegal;
4257 float cl_prob = 0;
4258 Node* incr = nullptr;
4259 Node* limit = nullptr;
4260 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4261 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4262 return false;
4263 }
4264
4265 // With an extra phi for the candidate iv?
4266 // Or the region node is the loop head
4267 if (!incr->is_Phi() || incr->in(0) == head) {
4268 return false;
4269 }
4270
4271 PathFrequency pf(head, this);
4272 region = incr->in(0);
4273
4274 // Go over all paths for the extra phi's region and see if that
4275 // path is frequent enough and would match the expected iv shape
4276 // if the extra phi is removed
4277 inner = 0;
4278 for (uint i = 1; i < incr->req(); ++i) {
4279 Node* in = incr->in(i);
4280 Node* trunc1 = nullptr;
4281 Node* trunc2 = nullptr;
4282 const TypeInteger* iv_trunc_t = nullptr;
4283 Node* orig_in = in;
4284 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4285 continue;
4286 }
4287 assert(in->Opcode() == Op_AddI, "wrong increment code");
4288 Node* xphi = nullptr;
4289 Node* stride = loop_iv_stride(in, loop, xphi);
4290
4291 if (stride == nullptr) {
4292 continue;
4293 }
4294
4295 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4296 if (phi == nullptr ||
4297 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4298 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4299 return false;
4300 }
4301
4302 f = pf.to(region->in(i));
4303 if (f > 0.5) {
4304 inner = i;
4305 break;
4306 }
4307 }
4308
4309 if (inner == 0) {
4310 return false;
4311 }
4312
4313 exit_test = back_control->in(0)->as_If();
4314 }
4315
4316 if (idom(region)->is_Catch()) {
4317 return false;
4318 }
4319
4320 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4321 Unique_Node_List wq;
4322 wq.push(head->in(LoopNode::LoopBackControl));
4323 for (uint i = 0; i < wq.size(); i++) {
4324 Node* c = wq.at(i);
4325 assert(get_loop(c) == loop, "not in the right loop?");
4326 if (c->is_Region()) {
4327 if (c != region) {
4328 for (uint j = 1; j < c->req(); ++j) {
4329 wq.push(c->in(j));
4330 }
4331 }
4332 } else {
4333 wq.push(c->in(0));
4334 }
4335 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4336 }
4337
4338 Node* region_dom = idom(region);
4339
4340 // Can't do the transformation if this would cause a membar pair to
4341 // be split
4342 for (uint i = 0; i < wq.size(); i++) {
4343 Node* c = wq.at(i);
4344 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4345 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4346 if (!wq.member(c->as_MemBar()->leading_membar())) {
4347 return false;
4348 }
4349 }
4350 }
4351
4352 // Collect data nodes that need to be clones as well
4353 int dd = dom_depth(head);
4354
4355 for (uint i = 0; i < loop->_body.size(); ++i) {
4356 Node* n = loop->_body.at(i);
4357 if (has_ctrl(n)) {
4358 Node* c = get_ctrl(n);
4359 if (wq.member(c)) {
4360 wq.push(n);
4361 }
4362 } else {
4363 set_idom(n, idom(n), dd);
4364 }
4365 }
4366
4367 // clone shared_stmt
4368 clone_loop_body(wq, old_new, nullptr);
4369
4370 Node* region_clone = old_new[region->_idx];
4371 region_clone->set_req(inner, C->top());
4372 set_idom(region, region->in(inner), dd);
4373
4374 // Prepare the outer loop
4375 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4376 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4377 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4378 set_idom(head, outer_head, dd);
4379
4380 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4381
4382 // Make one of the shared_stmt copies only reachable from stmt1, the
4383 // other only from stmt2..stmtn.
4384 Node* dom = nullptr;
4385 for (uint i = 1; i < region->req(); ++i) {
4386 if (i != inner) {
4387 _igvn.replace_input_of(region, i, C->top());
4388 }
4389 Node* in = region_clone->in(i);
4390 if (in->is_top()) {
4391 continue;
4392 }
4393 if (dom == nullptr) {
4394 dom = in;
4395 } else {
4396 dom = dom_lca(dom, in);
4397 }
4398 }
4399
4400 set_idom(region_clone, dom, dd);
4401
4402 // Set up the outer loop
4403 for (uint i = 0; i < head->outcnt(); i++) {
4404 Node* u = head->raw_out(i);
4405 if (u->is_Phi()) {
4406 Node* outer_phi = u->clone();
4407 outer_phi->set_req(0, outer_head);
4408 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4409 if (backedge == nullptr) {
4410 backedge = u->in(LoopNode::LoopBackControl);
4411 }
4412 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4413 register_new_node(outer_phi, outer_head);
4414 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4415 }
4416 }
4417
4418 // create control and data nodes for out of loop uses (including region2)
4419 Node_List worklist;
4420 uint new_counter = C->unique();
4421 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4422
4423 Node_List *split_if_set = nullptr;
4424 Node_List *split_bool_set = nullptr;
4425 Node_List *split_cex_set = nullptr;
4426 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4427 split_if_set, split_bool_set, split_cex_set);
4428
4429 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4430
4431 if (exit_test != nullptr) {
4432 float cnt = exit_test->_fcnt;
4433 if (cnt != COUNT_UNKNOWN) {
4434 exit_test->_fcnt = cnt * f;
4435 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4436 }
4437 }
4438
4439 C->set_major_progress();
4440
4441 return true;
4442 }
4443
4444 // AutoVectorize the loop: replace scalar ops with vector ops.
4445 PhaseIdealLoop::AutoVectorizeStatus
4446 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4447 // Counted loop only
4448 if (!lpt->is_counted()) {
4449 return AutoVectorizeStatus::Impossible;
4450 }
4451
4452 // Main-loop only
4453 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4454 if (!cl->is_main_loop()) {
4455 return AutoVectorizeStatus::Impossible;
4456 }
4457
4458 VLoop vloop(lpt, false);
4459 if (!vloop.check_preconditions()) {
4460 return AutoVectorizeStatus::TriedAndFailed;
4461 }
4462
4463 // Ensure the shared data is cleared before each use
4464 vshared.clear();
4465
4466 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4467 if (!vloop_analyzer.success()) {
4468 return AutoVectorizeStatus::TriedAndFailed;
4469 }
4470
4471 SuperWord sw(vloop_analyzer);
4472 if (!sw.transform_loop()) {
4473 return AutoVectorizeStatus::TriedAndFailed;
4474 }
4475
4476 return AutoVectorizeStatus::Success;
4477 }
4478
4479 // Returns true if the Reduction node is unordered.
4480 static bool is_unordered_reduction(Node* n) {
4481 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4482 }
4483
4484 // Having ReductionNodes in the loop is expensive. They need to recursively
4485 // fold together the vector values, for every vectorized loop iteration. If
4486 // we encounter the following pattern, we can vector accumulate the values
4487 // inside the loop, and only have a single UnorderedReduction after the loop.
4488 //
4489 // Note: UnorderedReduction represents a ReductionNode which does not require
4490 // calculating in strict order.
4491 //
4492 // CountedLoop init
4493 // | |
4494 // +------+ | +-----------------------+
4495 // | | | |
4496 // PhiNode (s) |
4497 // | |
4498 // | Vector |
4499 // | | |
4500 // UnorderedReduction (first_ur) |
4501 // | |
4502 // ... Vector |
4503 // | | |
4504 // UnorderedReduction (last_ur) |
4505 // | |
4506 // +---------------------+
4507 //
4508 // We patch the graph to look like this:
4509 //
4510 // CountedLoop identity_vector
4511 // | |
4512 // +-------+ | +---------------+
4513 // | | | |
4514 // PhiNode (v) |
4515 // | |
4516 // | Vector |
4517 // | | |
4518 // VectorAccumulator |
4519 // | |
4520 // ... Vector |
4521 // | | |
4522 // init VectorAccumulator |
4523 // | | | |
4524 // UnorderedReduction +-----------+
4525 //
4526 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4527 // use vector_accumulators, which do the same reductions, but only element
4528 // wise. This is a single operation per vector_accumulator, rather than many
4529 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4530 // after the loop, and also reduce the init value into it.
4531 //
4532 // We can not do this with all reductions. Some reductions do not allow the
4533 // reordering of operations (for example float addition/multiplication require
4534 // strict order).
4535 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4536 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4537
4538 // Find all Phi nodes with an unordered Reduction on backedge.
4539 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4540 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4541 Node* phi = cl->fast_out(j);
4542 // We have a phi with a single use, and an unordered Reduction on the backedge.
4543 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4544 continue;
4545 }
4546
4547 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4548 assert(!last_ur->requires_strict_order(), "must be");
4549
4550 // Determine types
4551 const TypeVect* vec_t = last_ur->vect_type();
4552 uint vector_length = vec_t->length();
4553 BasicType bt = vec_t->element_basic_type();
4554
4555 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4556 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4557 const int vopc = VectorNode::opcode(sopc, bt);
4558 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4559 DEBUG_ONLY( last_ur->dump(); )
4560 assert(false, "do not have normal vector op for this reduction");
4561 continue; // not implemented -> fails
4562 }
4563
4564 // Traverse up the chain of unordered Reductions, checking that it loops back to
4565 // the phi. Check that all unordered Reductions only have a single use, except for
4566 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4567 // are outside the loop.
4568 ReductionNode* current = last_ur;
4569 ReductionNode* first_ur = nullptr;
4570 while (true) {
4571 assert(!current->requires_strict_order(), "sanity");
4572
4573 // Expect no ctrl and a vector_input from within the loop.
4574 Node* ctrl = current->in(0);
4575 Node* vector_input = current->in(2);
4576 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4577 DEBUG_ONLY( current->dump(1); )
4578 assert(false, "reduction has ctrl or bad vector_input");
4579 break; // Chain traversal fails.
4580 }
4581
4582 assert(current->vect_type() != nullptr, "must have vector type");
4583 if (current->vect_type() != last_ur->vect_type()) {
4584 // Reductions do not have the same vector type (length and element type).
4585 break; // Chain traversal fails.
4586 }
4587
4588 // Expect single use of an unordered Reduction, except for last_ur.
4589 if (current == last_ur) {
4590 // Expect all uses to be outside the loop, except phi.
4591 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4592 Node* use = current->fast_out(k);
4593 if (use != phi && ctrl_or_self(use) == cl) {
4594 DEBUG_ONLY( current->dump(-1); )
4595 assert(false, "reduction has use inside loop");
4596 // Should not be allowed by SuperWord::mark_reductions
4597 return; // bail out of optimization
4598 }
4599 }
4600 } else {
4601 if (current->outcnt() != 1) {
4602 break; // Chain traversal fails.
4603 }
4604 }
4605
4606 // Expect another unordered Reduction or phi as the scalar input.
4607 Node* scalar_input = current->in(1);
4608 if (is_unordered_reduction(scalar_input) &&
4609 scalar_input->Opcode() == current->Opcode()) {
4610 // Move up the unordered Reduction chain.
4611 current = scalar_input->as_Reduction();
4612 assert(!current->requires_strict_order(), "must be");
4613 } else if (scalar_input == phi) {
4614 // Chain terminates at phi.
4615 first_ur = current;
4616 current = nullptr;
4617 break; // Success.
4618 } else {
4619 // scalar_input is neither phi nor a matching reduction
4620 // Can for example be scalar reduction when we have
4621 // partial vectorization.
4622 break; // Chain traversal fails.
4623 }
4624 }
4625 if (current != nullptr) {
4626 // Chain traversal was not successful.
4627 continue;
4628 }
4629 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4630
4631 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4632 set_ctrl(identity_scalar, C->root());
4633 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4634 register_new_node(identity_vector, C->root());
4635 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4636 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4637
4638 // Turn the scalar phi into a vector phi.
4639 _igvn.rehash_node_delayed(phi);
4640 Node* init = phi->in(1); // Remember init before replacing it.
4641 phi->set_req_X(1, identity_vector, &_igvn);
4642 phi->as_Type()->set_type(vec_t);
4643 _igvn.set_type(phi, vec_t);
4644
4645 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4646 current = first_ur;
4647 while (true) {
4648 // Create vector_accumulator to replace current.
4649 Node* last_vector_accumulator = current->in(1);
4650 Node* vector_input = current->in(2);
4651 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4652 register_new_node(vector_accumulator, cl);
4653 _igvn.replace_node(current, vector_accumulator);
4654 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4655 if (current == last_ur) {
4656 break;
4657 }
4658 current = vector_accumulator->unique_out()->as_Reduction();
4659 assert(!current->requires_strict_order(), "must be");
4660 }
4661
4662 // Create post-loop reduction.
4663 Node* last_accumulator = phi->in(2);
4664 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4665
4666 // Take over uses of last_accumulator that are not in the loop.
4667 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4668 Node* use = last_accumulator->out(i);
4669 if (use != phi && use != post_loop_reduction) {
4670 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4671 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4672 --i;
4673 }
4674 }
4675 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4676 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4677
4678 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4679 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4680 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4681 }
4682 }
4683
4684 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4685 for (uint i = 0; i < _data_nodes.size(); i++) {
4686 clone(_data_nodes[i], new_ctrl);
4687 }
4688 }
4689
4690 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4691 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4692 Node* clone = node->clone();
4693 _phase->igvn().register_new_node_with_optimizer(clone);
4694 _orig_to_new.put(node, clone);
4695 _phase->set_ctrl(clone, new_ctrl);
4696 if (node->is_CastII()) {
4697 clone->set_req(0, new_ctrl);
4698 }
4699 }
4700
4701 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4702 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4703 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4704 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4705 for (uint i = 1; i < node->req(); i++) {
4706 Node** cloned_input = _orig_to_new.get(node->in(i));
4707 if (cloned_input != nullptr) {
4708 // Input was also cloned -> rewire clone to the cloned input.
4709 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4710 }
4711 }
4712 });
4713 }
4714
4715 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4716 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4717 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4718 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4719 Node* new_ctrl) {
4720 for (uint i = 0; i < _data_nodes.size(); i++) {
4721 Node* data_node = _data_nodes[i];
4722 if (data_node->is_Opaque1()) {
4723 transform_opaque_node(transform_strategy, data_node);
4724 } else {
4725 clone(data_node, new_ctrl);
4726 }
4727 }
4728 }
4729
4730 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4731 Node* transformed_node;
4732 if (node->is_OpaqueLoopInit()) {
4733 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4734 } else {
4735 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4736 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4737 }
4738 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4739 _orig_to_new.put(node, transformed_node);
4740 }