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_Opaque4()) {
791 // Ignore Template Assertion Predicates with Opaque4 nodes.
792 assert(assertion_predicate_has_loop_opaque_node(iff),
793 "must be Template Assertion Predicate, non-null-check with Opaque4 cannot form a diamond with Halt");
794 return nullptr;
795 }
796 assert(bol->Opcode() == Op_Bool, "Unexpected node");
797 int cmp_op = bol->in(1)->Opcode();
798 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
799 return nullptr;
800 }
801 // It is expensive to generate flags from a float compare.
802 // Avoid duplicated float compare.
803 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
804
805 float infrequent_prob = PROB_UNLIKELY_MAG(3);
806 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
807 if (used_inside_loop) {
808 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
809
810 // BlockLayoutByFrequency optimization moves infrequent branch
811 // from hot path. No point in CMOV'ing in such case (110 is used
812 // instead of 100 to take into account not exactness of float value).
813 if (BlockLayoutByFrequency) {
814 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
815 }
816 }
817 // Check for highly predictable branch. No point in CMOV'ing if
818 // we are going to predict accurately all the time.
819 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
820 //keep going
821 } else if (iff->_prob < infrequent_prob ||
822 iff->_prob > (1.0f - infrequent_prob))
823 return nullptr;
824
825 // --------------
826 // Now replace all Phis with CMOV's
827 Node *cmov_ctrl = iff->in(0);
828 uint flip = (lp->Opcode() == Op_IfTrue);
829 Node_List wq;
830 while (1) {
831 PhiNode* phi = nullptr;
832 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
833 Node *out = region->fast_out(i);
834 if (out->is_Phi()) {
835 phi = out->as_Phi();
836 break;
837 }
838 }
839 if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
840 break;
841 }
842 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
843 // Move speculative ops
844 wq.push(phi);
845 while (wq.size() > 0) {
846 Node *n = wq.pop();
847 for (uint j = 1; j < n->req(); j++) {
848 Node* m = n->in(j);
849 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
850 #ifndef PRODUCT
851 if (PrintOpto && VerifyLoopOptimizations) {
852 tty->print(" speculate: ");
853 m->dump();
854 }
855 #endif
856 set_ctrl(m, cmov_ctrl);
857 wq.push(m);
858 }
859 }
860 }
861 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
862 register_new_node( cmov, cmov_ctrl );
863 _igvn.replace_node( phi, cmov );
864 #ifndef PRODUCT
865 if (TraceLoopOpts) {
866 tty->print("CMOV ");
867 r_loop->dump_head();
868 if (Verbose) {
869 bol->in(1)->dump(1);
870 cmov->dump(1);
871 }
872 }
873 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
874 #endif
875 }
876
877 // The useless CFG diamond will fold up later; see the optimization in
878 // RegionNode::Ideal.
879 _igvn._worklist.push(region);
880
881 return iff->in(1);
882 }
883
884 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
885 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
886 Node* u = m->fast_out(i);
887 if (u->is_CFG()) {
888 if (u->is_NeverBranch()) {
889 u = u->as_NeverBranch()->proj_out(0);
890 enqueue_cfg_uses(u, wq);
891 } else {
892 wq.push(u);
893 }
894 }
895 }
896 }
897
898 // Try moving a store out of a loop, right before the loop
899 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
900 // Store has to be first in the loop body
901 IdealLoopTree *n_loop = get_loop(n_ctrl);
902 if (n->is_Store() && n_loop != _ltree_root &&
903 n_loop->is_loop() && n_loop->_head->is_Loop() &&
904 n->in(0) != nullptr) {
905 Node* address = n->in(MemNode::Address);
906 Node* value = n->in(MemNode::ValueIn);
907 Node* mem = n->in(MemNode::Memory);
908 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
909 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
910
911 // - address and value must be loop invariant
912 // - memory must be a memory Phi for the loop
913 // - Store must be the only store on this memory slice in the
914 // loop: if there's another store following this one then value
915 // written at iteration i by the second store could be overwritten
916 // at iteration i+n by the first store: it's not safe to move the
917 // first store out of the loop
918 // - nothing must observe the memory Phi: it guarantees no read
919 // before the store, we are also guaranteed the store post
920 // dominates the loop head (ignoring a possible early
921 // exit). Otherwise there would be extra Phi involved between the
922 // loop's Phi and the store.
923 // - there must be no early exit from the loop before the Store
924 // (such an exit most of the time would be an extra use of the
925 // memory Phi but sometimes is a bottom memory Phi that takes the
926 // store as input).
927
928 if (!n_loop->is_member(address_loop) &&
929 !n_loop->is_member(value_loop) &&
930 mem->is_Phi() && mem->in(0) == n_loop->_head &&
931 mem->outcnt() == 1 &&
932 mem->in(LoopNode::LoopBackControl) == n) {
933
934 assert(n_loop->_tail != nullptr, "need a tail");
935 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
936
937 // Verify that there's no early exit of the loop before the store.
938 bool ctrl_ok = false;
939 {
940 // Follow control from loop head until n, we exit the loop or
941 // we reach the tail
942 ResourceMark rm;
943 Unique_Node_List wq;
944 wq.push(n_loop->_head);
945
946 for (uint next = 0; next < wq.size(); ++next) {
947 Node *m = wq.at(next);
948 if (m == n->in(0)) {
949 ctrl_ok = true;
950 continue;
951 }
952 assert(!has_ctrl(m), "should be CFG");
953 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
954 ctrl_ok = false;
955 break;
956 }
957 enqueue_cfg_uses(m, wq);
958 if (wq.size() > 10) {
959 ctrl_ok = false;
960 break;
961 }
962 }
963 }
964 if (ctrl_ok) {
965 // move the Store
966 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
967 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
968 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
969 // Disconnect the phi now. An empty phi can confuse other
970 // optimizations in this pass of loop opts.
971 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
972 n_loop->_body.yank(mem);
973
974 set_ctrl_and_loop(n, n->in(0));
975
976 return n;
977 }
978 }
979 }
980 return nullptr;
981 }
982
983 // Try moving a store out of a loop, right after the loop
984 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
985 if (n->is_Store() && n->in(0) != nullptr) {
986 Node *n_ctrl = get_ctrl(n);
987 IdealLoopTree *n_loop = get_loop(n_ctrl);
988 // Store must be in a loop
989 if (n_loop != _ltree_root && !n_loop->_irreducible) {
990 Node* address = n->in(MemNode::Address);
991 Node* value = n->in(MemNode::ValueIn);
992 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
993 // address must be loop invariant
994 if (!n_loop->is_member(address_loop)) {
995 // Store must be last on this memory slice in the loop and
996 // nothing in the loop must observe it
997 Node* phi = nullptr;
998 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
999 Node* u = n->fast_out(i);
1000 if (has_ctrl(u)) { // control use?
1001 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1002 if (!n_loop->is_member(u_loop)) {
1003 continue;
1004 }
1005 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1006 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1007 // multiple phis on the same slice are possible
1008 if (phi != nullptr) {
1009 return;
1010 }
1011 phi = u;
1012 continue;
1013 }
1014 }
1015 return;
1016 }
1017 if (phi != nullptr) {
1018 // Nothing in the loop before the store (next iteration)
1019 // must observe the stored value
1020 bool mem_ok = true;
1021 {
1022 ResourceMark rm;
1023 Unique_Node_List wq;
1024 wq.push(phi);
1025 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1026 Node *m = wq.at(next);
1027 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1028 Node* u = m->fast_out(i);
1029 if (u->is_Store() || u->is_Phi()) {
1030 if (u != n) {
1031 wq.push(u);
1032 mem_ok = (wq.size() <= 10);
1033 }
1034 } else {
1035 mem_ok = false;
1036 break;
1037 }
1038 }
1039 }
1040 }
1041 if (mem_ok) {
1042 // Move the store out of the loop if the LCA of all
1043 // users (except for the phi) is outside the loop.
1044 Node* hook = new Node(1);
1045 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1046 _igvn.rehash_node_delayed(phi);
1047 int count = phi->replace_edge(n, hook, &_igvn);
1048 assert(count > 0, "inconsistent phi");
1049
1050 // Compute latest point this store can go
1051 Node* lca = get_late_ctrl(n, get_ctrl(n));
1052 if (lca->is_OuterStripMinedLoop()) {
1053 lca = lca->in(LoopNode::EntryControl);
1054 }
1055 if (n_loop->is_member(get_loop(lca))) {
1056 // LCA is in the loop - bail out
1057 _igvn.replace_node(hook, n);
1058 return;
1059 }
1060 #ifdef ASSERT
1061 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1062 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1063 n_loop->_head->as_Loop()->verify_strip_mined(1);
1064 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1065 IdealLoopTree* outer_loop = get_loop(outer);
1066 assert(n_loop->_parent == outer_loop, "broken loop tree");
1067 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1068 }
1069 #endif
1070 lca = place_outside_loop(lca, n_loop);
1071 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1072 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");
1073
1074 // Move store out of the loop
1075 _igvn.replace_node(hook, n->in(MemNode::Memory));
1076 _igvn.replace_input_of(n, 0, lca);
1077 set_ctrl_and_loop(n, lca);
1078
1079 // Disconnect the phi now. An empty phi can confuse other
1080 // optimizations in this pass of loop opts..
1081 if (phi->in(LoopNode::LoopBackControl) == phi) {
1082 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1083 n_loop->_body.yank(phi);
1084 }
1085 }
1086 }
1087 }
1088 }
1089 }
1090 }
1091
1092 // Split some nodes that take a counted loop phi as input at a counted
1093 // loop can cause vectorization of some expressions to fail
1094 bool PhaseIdealLoop::split_thru_phi_could_prevent_vectorization(Node* n, Node* n_blk) {
1095 if (!n_blk->is_CountedLoop()) {
1096 return false;
1097 }
1098
1099 int opcode = n->Opcode();
1100
1101 if (opcode != Op_AndI &&
1102 opcode != Op_MulI &&
1103 opcode != Op_RotateRight &&
1104 opcode != Op_RShiftI) {
1105 return false;
1106 }
1107
1108 return n->in(1) == n_blk->as_BaseCountedLoop()->phi();
1109 }
1110
1111 //------------------------------split_if_with_blocks_pre-----------------------
1112 // Do the real work in a non-recursive function. Data nodes want to be
1113 // cloned in the pre-order so they can feed each other nicely.
1114 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1115 // Cloning these guys is unlikely to win
1116 int n_op = n->Opcode();
1117 if (n_op == Op_MergeMem) {
1118 return n;
1119 }
1120 if (n->is_Proj()) {
1121 return n;
1122 }
1123 // Do not clone-up CmpFXXX variations, as these are always
1124 // followed by a CmpI
1125 if (n->is_Cmp()) {
1126 return n;
1127 }
1128 // Attempt to use a conditional move instead of a phi/branch
1129 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1130 Node *cmov = conditional_move( n );
1131 if (cmov) {
1132 return cmov;
1133 }
1134 }
1135 if (n->is_CFG() || n->is_LoadStore()) {
1136 return n;
1137 }
1138 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1139 if (!C->major_progress()) { // If chance of no more loop opts...
1140 _igvn._worklist.push(n); // maybe we'll remove them
1141 }
1142 return n;
1143 }
1144
1145 if (n->is_Con()) {
1146 return n; // No cloning for Con nodes
1147 }
1148
1149 Node *n_ctrl = get_ctrl(n);
1150 if (!n_ctrl) {
1151 return n; // Dead node
1152 }
1153
1154 Node* res = try_move_store_before_loop(n, n_ctrl);
1155 if (res != nullptr) {
1156 return n;
1157 }
1158
1159 // Attempt to remix address expressions for loop invariants
1160 Node *m = remix_address_expressions( n );
1161 if( m ) return m;
1162
1163 if (n_op == Op_AddI) {
1164 Node *nn = convert_add_to_muladd( n );
1165 if ( nn ) return nn;
1166 }
1167
1168 if (n->is_ConstraintCast()) {
1169 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1170 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1171 // Node control inputs don't necessarily agree with loop control info (due to
1172 // transformations happened in between), thus additional dominance check is needed
1173 // to keep loop info valid.
1174 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1175 _igvn.replace_node(n, dom_cast);
1176 return dom_cast;
1177 }
1178 }
1179
1180 // Determine if the Node has inputs from some local Phi.
1181 // Returns the block to clone thru.
1182 Node *n_blk = has_local_phi_input( n );
1183 if( !n_blk ) return n;
1184
1185 // Do not clone the trip counter through on a CountedLoop
1186 // (messes up the canonical shape).
1187 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1188 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1189 return n;
1190 }
1191 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1192 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1193 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1194 return n;
1195 }
1196
1197 if (split_thru_phi_could_prevent_vectorization(n, n_blk)) {
1198 return n;
1199 }
1200
1201 // Check for having no control input; not pinned. Allow
1202 // dominating control.
1203 if (n->in(0)) {
1204 Node *dom = idom(n_blk);
1205 if (dom_lca(n->in(0), dom) != n->in(0)) {
1206 return n;
1207 }
1208 }
1209 // Policy: when is it profitable. You must get more wins than
1210 // policy before it is considered profitable. Policy is usually 0,
1211 // so 1 win is considered profitable. Big merges will require big
1212 // cloning, so get a larger policy.
1213 int policy = n_blk->req() >> 2;
1214
1215 // If the loop is a candidate for range check elimination,
1216 // delay splitting through it's phi until a later loop optimization
1217 if (n_blk->is_BaseCountedLoop()) {
1218 IdealLoopTree *lp = get_loop(n_blk);
1219 if (lp && lp->_rce_candidate) {
1220 return n;
1221 }
1222 }
1223
1224 if (must_throttle_split_if()) return n;
1225
1226 // Split 'n' through the merge point if it is profitable
1227 Node *phi = split_thru_phi( n, n_blk, policy );
1228 if (!phi) return n;
1229
1230 // Found a Phi to split thru!
1231 // Replace 'n' with the new phi
1232 _igvn.replace_node( n, phi );
1233 // Moved a load around the loop, 'en-registering' something.
1234 if (n_blk->is_Loop() && n->is_Load() &&
1235 !phi->in(LoopNode::LoopBackControl)->is_Load())
1236 C->set_major_progress();
1237
1238 return phi;
1239 }
1240
1241 static bool merge_point_too_heavy(Compile* C, Node* region) {
1242 // Bail out if the region and its phis have too many users.
1243 int weight = 0;
1244 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1245 weight += region->fast_out(i)->outcnt();
1246 }
1247 int nodes_left = C->max_node_limit() - C->live_nodes();
1248 if (weight * 8 > nodes_left) {
1249 if (PrintOpto) {
1250 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1251 }
1252 return true;
1253 } else {
1254 return false;
1255 }
1256 }
1257
1258 static bool merge_point_safe(Node* region) {
1259 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1260 // having a PhiNode input. This sidesteps the dangerous case where the split
1261 // ConvI2LNode may become TOP if the input Value() does not
1262 // overlap the ConvI2L range, leaving a node which may not dominate its
1263 // uses.
1264 // A better fix for this problem can be found in the BugTraq entry, but
1265 // expediency for Mantis demands this hack.
1266 #ifdef _LP64
1267 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1268 Node* n = region->fast_out(i);
1269 if (n->is_Phi()) {
1270 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1271 Node* m = n->fast_out(j);
1272 if (m->Opcode() == Op_ConvI2L)
1273 return false;
1274 if (m->is_CastII()) {
1275 return false;
1276 }
1277 }
1278 }
1279 }
1280 #endif
1281 return true;
1282 }
1283
1284
1285 //------------------------------place_outside_loop---------------------------------
1286 // Place some computation outside of this loop on the path to the use passed as argument
1287 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1288 Node* head = loop->_head;
1289 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1290 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1291 loop = loop->_parent;
1292 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1293 }
1294
1295 // Pick control right outside the loop
1296 for (;;) {
1297 Node* dom = idom(useblock);
1298 if (loop->is_member(get_loop(dom))) {
1299 break;
1300 }
1301 useblock = dom;
1302 }
1303 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1304 return useblock;
1305 }
1306
1307
1308 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1309 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1310 return false;
1311 }
1312 if (!n->in(0)->is_Region()) {
1313 return false;
1314 }
1315
1316 Node* region = n->in(0);
1317 Node* dom = idom(region);
1318 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1319 return false;
1320 }
1321 IfNode* dom_if = dom->as_If();
1322 Node* proj_true = dom_if->proj_out(1);
1323 Node* proj_false = dom_if->proj_out(0);
1324
1325 for (uint i = 1; i < region->req(); i++) {
1326 if (is_dominator(proj_true, region->in(i))) {
1327 continue;
1328 }
1329 if (is_dominator(proj_false, region->in(i))) {
1330 continue;
1331 }
1332 return false;
1333 }
1334
1335 return true;
1336 }
1337
1338
1339 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1340 if (must_throttle_split_if()) {
1341 return false;
1342 }
1343
1344 // Do not do 'split-if' if irreducible loops are present.
1345 if (_has_irreducible_loops) {
1346 return false;
1347 }
1348
1349 if (merge_point_too_heavy(C, n_ctrl)) {
1350 return false;
1351 }
1352
1353 // Do not do 'split-if' if some paths are dead. First do dead code
1354 // elimination and then see if its still profitable.
1355 for (uint i = 1; i < n_ctrl->req(); i++) {
1356 if (n_ctrl->in(i) == C->top()) {
1357 return false;
1358 }
1359 }
1360
1361 // If trying to do a 'Split-If' at the loop head, it is only
1362 // profitable if the cmp folds up on BOTH paths. Otherwise we
1363 // risk peeling a loop forever.
1364
1365 // CNC - Disabled for now. Requires careful handling of loop
1366 // body selection for the cloned code. Also, make sure we check
1367 // for any input path not being in the same loop as n_ctrl. For
1368 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1369 // because the alternative loop entry points won't be converted
1370 // into LoopNodes.
1371 IdealLoopTree *n_loop = get_loop(n_ctrl);
1372 for (uint j = 1; j < n_ctrl->req(); j++) {
1373 if (get_loop(n_ctrl->in(j)) != n_loop) {
1374 return false;
1375 }
1376 }
1377
1378 // Check for safety of the merge point.
1379 if (!merge_point_safe(n_ctrl)) {
1380 return false;
1381 }
1382
1383 return true;
1384 }
1385
1386 // Detect if the node is the inner strip-mined loop
1387 // Return: null if it's not the case, or the exit of outer strip-mined loop
1388 static Node* is_inner_of_stripmined_loop(const Node* out) {
1389 Node* out_le = nullptr;
1390
1391 if (out->is_CountedLoopEnd()) {
1392 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1393
1394 if (loop != nullptr && loop->is_strip_mined()) {
1395 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1396 }
1397 }
1398
1399 return out_le;
1400 }
1401
1402 //------------------------------split_if_with_blocks_post----------------------
1403 // Do the real work in a non-recursive function. CFG hackery wants to be
1404 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1405 // info.
1406 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1407
1408 // Cloning Cmp through Phi's involves the split-if transform.
1409 // FastLock is not used by an If
1410 if (n->is_Cmp() && !n->is_FastLock()) {
1411 Node *n_ctrl = get_ctrl(n);
1412 // Determine if the Node has inputs from some local Phi.
1413 // Returns the block to clone thru.
1414 Node *n_blk = has_local_phi_input(n);
1415 if (n_blk != n_ctrl) {
1416 return;
1417 }
1418
1419 if (!can_split_if(n_ctrl)) {
1420 return;
1421 }
1422
1423 if (n->outcnt() != 1) {
1424 return; // Multiple bool's from 1 compare?
1425 }
1426 Node *bol = n->unique_out();
1427 assert(bol->is_Bool(), "expect a bool here");
1428 if (bol->outcnt() != 1) {
1429 return;// Multiple branches from 1 compare?
1430 }
1431 Node *iff = bol->unique_out();
1432
1433 // Check some safety conditions
1434 if (iff->is_If()) { // Classic split-if?
1435 if (iff->in(0) != n_ctrl) {
1436 return; // Compare must be in same blk as if
1437 }
1438 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1439 // Can't split CMove with different control.
1440 if (get_ctrl(iff) != n_ctrl) {
1441 return;
1442 }
1443 if (get_ctrl(iff->in(2)) == n_ctrl ||
1444 get_ctrl(iff->in(3)) == n_ctrl) {
1445 return; // Inputs not yet split-up
1446 }
1447 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1448 return; // Loop-invar test gates loop-varying CMOVE
1449 }
1450 } else {
1451 return; // some other kind of node, such as an Allocate
1452 }
1453
1454 // When is split-if profitable? Every 'win' on means some control flow
1455 // goes dead, so it's almost always a win.
1456 int policy = 0;
1457 // Split compare 'n' through the merge point if it is profitable
1458 Node *phi = split_thru_phi( n, n_ctrl, policy);
1459 if (!phi) {
1460 return;
1461 }
1462
1463 // Found a Phi to split thru!
1464 // Replace 'n' with the new phi
1465 _igvn.replace_node(n, phi);
1466
1467 // Now split the bool up thru the phi
1468 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1469 guarantee(bolphi != nullptr, "null boolean phi node");
1470
1471 _igvn.replace_node(bol, bolphi);
1472 assert(iff->in(1) == bolphi, "");
1473
1474 if (bolphi->Value(&_igvn)->singleton()) {
1475 return;
1476 }
1477
1478 // Conditional-move? Must split up now
1479 if (!iff->is_If()) {
1480 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1481 _igvn.replace_node(iff, cmovphi);
1482 return;
1483 }
1484
1485 // Now split the IF
1486 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1487 if ((PrintOpto && VerifyLoopOptimizations) || TraceLoopOpts) {
1488 tty->print_cr("Split-If");
1489 }
1490 do_split_if(iff);
1491 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1492 return;
1493 }
1494
1495 // Two identical ifs back to back can be merged
1496 if (try_merge_identical_ifs(n)) {
1497 return;
1498 }
1499
1500 // Check for an IF ready to split; one that has its
1501 // condition codes input coming from a Phi at the block start.
1502 int n_op = n->Opcode();
1503
1504 // Check for an IF being dominated by another IF same test
1505 if (n_op == Op_If ||
1506 n_op == Op_RangeCheck) {
1507 Node *bol = n->in(1);
1508 uint max = bol->outcnt();
1509 // Check for same test used more than once?
1510 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1511 // Search up IDOMs to see if this IF is dominated.
1512 Node* cmp = bol->in(1);
1513 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1514
1515 // Now search up IDOMs till cutoff, looking for a dominating test
1516 Node *prevdom = n;
1517 Node *dom = idom(prevdom);
1518 while (dom != cutoff) {
1519 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1520 safe_for_if_replacement(dom)) {
1521 // It's invalid to move control dependent data nodes in the inner
1522 // strip-mined loop, because:
1523 // 1) break validation of LoopNode::verify_strip_mined()
1524 // 2) move code with side-effect in strip-mined loop
1525 // Move to the exit of outer strip-mined loop in that case.
1526 Node* out_le = is_inner_of_stripmined_loop(dom);
1527 if (out_le != nullptr) {
1528 prevdom = out_le;
1529 }
1530 // Replace the dominated test with an obvious true or false.
1531 // Place it on the IGVN worklist for later cleanup.
1532 C->set_major_progress();
1533 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1534 // to prevent an array load from floating above its range check. There are three cases:
1535 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1536 // all its array accesses at that point.
1537 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1538 // accesses would start to float, since we don't pin at that point.
1539 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1540 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1541 prevdom->in(0)->Opcode() != Op_RangeCheck;
1542 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1543 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1544 return;
1545 }
1546 prevdom = dom;
1547 dom = idom(prevdom);
1548 }
1549 }
1550 }
1551
1552 try_sink_out_of_loop(n);
1553
1554 try_move_store_after_loop(n);
1555 }
1556
1557 // Transform:
1558 //
1559 // if (some_condition) {
1560 // // body 1
1561 // } else {
1562 // // body 2
1563 // }
1564 // if (some_condition) {
1565 // // body 3
1566 // } else {
1567 // // body 4
1568 // }
1569 //
1570 // into:
1571 //
1572 //
1573 // if (some_condition) {
1574 // // body 1
1575 // // body 3
1576 // } else {
1577 // // body 2
1578 // // body 4
1579 // }
1580 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1581 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1582 Node *n_ctrl = n->in(0);
1583 IfNode* dom_if = idom(n_ctrl)->as_If();
1584 if (n->in(1) != dom_if->in(1)) {
1585 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1586 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1587 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1588 _igvn.replace_input_of(n, 1, dom_if->in(1));
1589 }
1590 ProjNode* dom_proj_true = dom_if->proj_out(1);
1591 ProjNode* dom_proj_false = dom_if->proj_out(0);
1592
1593 // Now split the IF
1594 RegionNode* new_false_region;
1595 RegionNode* new_true_region;
1596 do_split_if(n, &new_false_region, &new_true_region);
1597 assert(new_false_region->req() == new_true_region->req(), "");
1598 #ifdef ASSERT
1599 for (uint i = 1; i < new_false_region->req(); ++i) {
1600 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1601 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");
1602 }
1603 #endif
1604 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1605
1606 // We now have:
1607 // if (some_condition) {
1608 // // body 1
1609 // if (some_condition) {
1610 // body3: // new_true_region
1611 // // body3
1612 // } else {
1613 // goto body4;
1614 // }
1615 // } else {
1616 // // body 2
1617 // if (some_condition) {
1618 // goto body3;
1619 // } else {
1620 // body4: // new_false_region
1621 // // body4;
1622 // }
1623 // }
1624 //
1625
1626 // clone pinned nodes thru the resulting regions
1627 push_pinned_nodes_thru_region(dom_if, new_true_region);
1628 push_pinned_nodes_thru_region(dom_if, new_false_region);
1629
1630 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1631 // 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
1632 // unrelated control dependency.
1633 for (uint i = 1; i < new_false_region->req(); i++) {
1634 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1635 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1636 } else {
1637 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1638 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1639 }
1640 }
1641 return true;
1642 }
1643 return false;
1644 }
1645
1646 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1647 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1648 Node* u = region->out(i);
1649 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1650 continue;
1651 }
1652 assert(u->in(0) == region, "not a control dependent node?");
1653 uint j = 1;
1654 for (; j < u->req(); ++j) {
1655 Node* in = u->in(j);
1656 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1657 break;
1658 }
1659 }
1660 if (j == u->req()) {
1661 Node *phi = PhiNode::make_blank(region, u);
1662 for (uint k = 1; k < region->req(); ++k) {
1663 Node* clone = u->clone();
1664 clone->set_req(0, region->in(k));
1665 register_new_node(clone, region->in(k));
1666 phi->init_req(k, clone);
1667 }
1668 register_new_node(phi, region);
1669 _igvn.replace_node(u, phi);
1670 --i;
1671 }
1672 }
1673 }
1674
1675 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1676 if (!dom->is_CountedLoopEnd()) {
1677 return true;
1678 }
1679 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1680 CountedLoopNode* cl = le->loopnode();
1681 if (cl == nullptr) {
1682 return true;
1683 }
1684 if (!cl->is_main_loop()) {
1685 return true;
1686 }
1687 if (cl->is_canonical_loop_entry() == nullptr) {
1688 return true;
1689 }
1690 // Further unrolling is possible so loop exit condition might change
1691 return false;
1692 }
1693
1694 // See if a shared loop-varying computation has no loop-varying uses.
1695 // Happens if something is only used for JVM state in uncommon trap exits,
1696 // like various versions of induction variable+offset. Clone the
1697 // computation per usage to allow it to sink out of the loop.
1698 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1699 if (has_ctrl(n) &&
1700 !n->is_Phi() &&
1701 !n->is_Bool() &&
1702 !n->is_Proj() &&
1703 !n->is_MergeMem() &&
1704 !n->is_CMove() &&
1705 !n->is_Opaque4() &&
1706 !n->is_OpaqueInitializedAssertionPredicate() &&
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 u_loop->_head->is_CountedLoop() && u_loop->_head->as_CountedLoop()->is_main_loop() &&
1950 n_loop->_next == get_loop(u_loop->_head->as_CountedLoop()->skip_strip_mined())) {
1951 return false;
1952 }
1953 return true;
1954 }
1955
1956 //------------------------------split_if_with_blocks---------------------------
1957 // Check for aggressive application of 'split-if' optimization,
1958 // using basic block level info.
1959 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1960 Node* root = C->root();
1961 visited.set(root->_idx); // first, mark root as visited
1962 // Do pre-visit work for root
1963 Node* n = split_if_with_blocks_pre(root);
1964 uint cnt = n->outcnt();
1965 uint i = 0;
1966
1967 while (true) {
1968 // Visit all children
1969 if (i < cnt) {
1970 Node* use = n->raw_out(i);
1971 ++i;
1972 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1973 // Now do pre-visit work for this use
1974 use = split_if_with_blocks_pre(use);
1975 nstack.push(n, i); // Save parent and next use's index.
1976 n = use; // Process all children of current use.
1977 cnt = use->outcnt();
1978 i = 0;
1979 }
1980 }
1981 else {
1982 // All of n's children have been processed, complete post-processing.
1983 if (cnt != 0 && !n->is_Con()) {
1984 assert(has_node(n), "no dead nodes");
1985 split_if_with_blocks_post(n);
1986 }
1987 if (must_throttle_split_if()) {
1988 nstack.clear();
1989 }
1990 if (nstack.is_empty()) {
1991 // Finished all nodes on stack.
1992 break;
1993 }
1994 // Get saved parent node and next use's index. Visit the rest of uses.
1995 n = nstack.node();
1996 cnt = n->outcnt();
1997 i = nstack.index();
1998 nstack.pop();
1999 }
2000 }
2001 }
2002
2003
2004 //=============================================================================
2005 //
2006 // C L O N E A L O O P B O D Y
2007 //
2008
2009 //------------------------------clone_iff--------------------------------------
2010 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2011 // "Nearly" because all Nodes have been cloned from the original in the loop,
2012 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2013 // through the Phi recursively, and return a Bool.
2014 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2015
2016 // Convert this Phi into a Phi merging Bools
2017 uint i;
2018 for (i = 1; i < phi->req(); i++) {
2019 Node* b = phi->in(i);
2020 if (b->is_Phi()) {
2021 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2022 } else {
2023 assert(b->is_Bool() || b->is_Opaque4() || b->is_OpaqueInitializedAssertionPredicate(),
2024 "bool, non-null check with Opaque4 node or Initialized Assertion Predicate with its Opaque node");
2025 }
2026 }
2027 Node* n = phi->in(1);
2028 Node* sample_opaque = nullptr;
2029 Node *sample_bool = nullptr;
2030 if (n->is_Opaque4() || n->is_OpaqueInitializedAssertionPredicate()) {
2031 sample_opaque = n;
2032 sample_bool = n->in(1);
2033 assert(sample_bool->is_Bool(), "wrong type");
2034 } else {
2035 sample_bool = n;
2036 }
2037 Node *sample_cmp = sample_bool->in(1);
2038
2039 // Make Phis to merge the Cmp's inputs.
2040 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2041 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2042 for (i = 1; i < phi->req(); i++) {
2043 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2044 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2045 phi1->set_req(i, n1);
2046 phi2->set_req(i, n2);
2047 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2048 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2049 }
2050 // See if these Phis have been made before.
2051 // Register with optimizer
2052 Node *hit1 = _igvn.hash_find_insert(phi1);
2053 if (hit1) { // Hit, toss just made Phi
2054 _igvn.remove_dead_node(phi1); // Remove new phi
2055 assert(hit1->is_Phi(), "" );
2056 phi1 = (PhiNode*)hit1; // Use existing phi
2057 } else { // Miss
2058 _igvn.register_new_node_with_optimizer(phi1);
2059 }
2060 Node *hit2 = _igvn.hash_find_insert(phi2);
2061 if (hit2) { // Hit, toss just made Phi
2062 _igvn.remove_dead_node(phi2); // Remove new phi
2063 assert(hit2->is_Phi(), "" );
2064 phi2 = (PhiNode*)hit2; // Use existing phi
2065 } else { // Miss
2066 _igvn.register_new_node_with_optimizer(phi2);
2067 }
2068 // Register Phis with loop/block info
2069 set_ctrl(phi1, phi->in(0));
2070 set_ctrl(phi2, phi->in(0));
2071 // Make a new Cmp
2072 Node *cmp = sample_cmp->clone();
2073 cmp->set_req(1, phi1);
2074 cmp->set_req(2, phi2);
2075 _igvn.register_new_node_with_optimizer(cmp);
2076 set_ctrl(cmp, phi->in(0));
2077
2078 // Make a new Bool
2079 Node *b = sample_bool->clone();
2080 b->set_req(1,cmp);
2081 _igvn.register_new_node_with_optimizer(b);
2082 set_ctrl(b, phi->in(0));
2083
2084 if (sample_opaque != nullptr) {
2085 Node* opaque = sample_opaque->clone();
2086 opaque->set_req(1, b);
2087 _igvn.register_new_node_with_optimizer(opaque);
2088 set_ctrl(opaque, phi->in(0));
2089 return opaque;
2090 }
2091
2092 assert(b->is_Bool(), "");
2093 return b;
2094 }
2095
2096 //------------------------------clone_bool-------------------------------------
2097 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2098 // "Nearly" because all Nodes have been cloned from the original in the loop,
2099 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2100 // through the Phi recursively, and return a Bool.
2101 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2102 uint i;
2103 // Convert this Phi into a Phi merging Bools
2104 for( i = 1; i < phi->req(); i++ ) {
2105 Node *b = phi->in(i);
2106 if( b->is_Phi() ) {
2107 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2108 } else {
2109 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2110 }
2111 }
2112
2113 Node *sample_cmp = phi->in(1);
2114
2115 // Make Phis to merge the Cmp's inputs.
2116 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2117 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2118 for( uint j = 1; j < phi->req(); j++ ) {
2119 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2120 Node *n1, *n2;
2121 if( cmp_top->is_Cmp() ) {
2122 n1 = cmp_top->in(1);
2123 n2 = cmp_top->in(2);
2124 } else {
2125 n1 = n2 = cmp_top;
2126 }
2127 phi1->set_req( j, n1 );
2128 phi2->set_req( j, n2 );
2129 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2130 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2131 }
2132
2133 // See if these Phis have been made before.
2134 // Register with optimizer
2135 Node *hit1 = _igvn.hash_find_insert(phi1);
2136 if( hit1 ) { // Hit, toss just made Phi
2137 _igvn.remove_dead_node(phi1); // Remove new phi
2138 assert( hit1->is_Phi(), "" );
2139 phi1 = (PhiNode*)hit1; // Use existing phi
2140 } else { // Miss
2141 _igvn.register_new_node_with_optimizer(phi1);
2142 }
2143 Node *hit2 = _igvn.hash_find_insert(phi2);
2144 if( hit2 ) { // Hit, toss just made Phi
2145 _igvn.remove_dead_node(phi2); // Remove new phi
2146 assert( hit2->is_Phi(), "" );
2147 phi2 = (PhiNode*)hit2; // Use existing phi
2148 } else { // Miss
2149 _igvn.register_new_node_with_optimizer(phi2);
2150 }
2151 // Register Phis with loop/block info
2152 set_ctrl(phi1, phi->in(0));
2153 set_ctrl(phi2, phi->in(0));
2154 // Make a new Cmp
2155 Node *cmp = sample_cmp->clone();
2156 cmp->set_req( 1, phi1 );
2157 cmp->set_req( 2, phi2 );
2158 _igvn.register_new_node_with_optimizer(cmp);
2159 set_ctrl(cmp, phi->in(0));
2160
2161 assert( cmp->is_Cmp(), "" );
2162 return (CmpNode*)cmp;
2163 }
2164
2165 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2166 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2167 Node_List*& split_if_set, Node_List*& split_bool_set,
2168 Node_List*& split_cex_set, Node_List& worklist,
2169 uint new_counter, CloneLoopMode mode) {
2170 Node* nnn = old_new[old->_idx];
2171 // Copy uses to a worklist, so I can munge the def-use info
2172 // with impunity.
2173 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2174 worklist.push(old->fast_out(j));
2175
2176 while( worklist.size() ) {
2177 Node *use = worklist.pop();
2178 if (!has_node(use)) continue; // Ignore dead nodes
2179 if (use->in(0) == C->top()) continue;
2180 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2181 // Check for data-use outside of loop - at least one of OLD or USE
2182 // must not be a CFG node.
2183 #ifdef ASSERT
2184 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2185 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2186 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2187 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2188 }
2189 #endif
2190 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2191
2192 // If the Data use is an IF, that means we have an IF outside the
2193 // loop that is switching on a condition that is set inside the
2194 // loop. Happens if people set a loop-exit flag; then test the flag
2195 // in the loop to break the loop, then test is again outside the
2196 // loop to determine which way the loop exited.
2197 //
2198 // For several uses we need to make sure that there is no phi between,
2199 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2200 // to avoid such a phi in between.
2201 // For example, it is unexpected that there is a Phi between an
2202 // AllocateArray node and its ValidLengthTest input that could cause
2203 // split if to break.
2204 if (use->is_If() || use->is_CMove() || use->is_Opaque4() || use->is_OpaqueInitializedAssertionPredicate() ||
2205 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2206 // Since this code is highly unlikely, we lazily build the worklist
2207 // of such Nodes to go split.
2208 if (!split_if_set) {
2209 split_if_set = new Node_List();
2210 }
2211 split_if_set->push(use);
2212 }
2213 if (use->is_Bool()) {
2214 if (!split_bool_set) {
2215 split_bool_set = new Node_List();
2216 }
2217 split_bool_set->push(use);
2218 }
2219 if (use->Opcode() == Op_CreateEx) {
2220 if (!split_cex_set) {
2221 split_cex_set = new Node_List();
2222 }
2223 split_cex_set->push(use);
2224 }
2225
2226
2227 // Get "block" use is in
2228 uint idx = 0;
2229 while( use->in(idx) != old ) idx++;
2230 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2231 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2232 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2233 ? prev->in(2)
2234 : idom(prev);
2235 if( use->is_Phi() ) // Phi use is in prior block
2236 cfg = prev->in(idx); // NOT in block of Phi itself
2237 if (cfg->is_top()) { // Use is dead?
2238 _igvn.replace_input_of(use, idx, C->top());
2239 continue;
2240 }
2241
2242 // If use is referenced through control edge... (idx == 0)
2243 if (mode == IgnoreStripMined && idx == 0) {
2244 LoopNode *head = loop->_head->as_Loop();
2245 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2246 // That node is outside the inner loop, leave it outside the
2247 // outer loop as well to not confuse verification code.
2248 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2249 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2250 continue;
2251 }
2252 }
2253
2254 while(!outer_loop->is_member(get_loop(cfg))) {
2255 prev = cfg;
2256 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2257 }
2258 // If the use occurs after merging several exits from the loop, then
2259 // old value must have dominated all those exits. Since the same old
2260 // value was used on all those exits we did not need a Phi at this
2261 // merge point. NOW we do need a Phi here. Each loop exit value
2262 // is now merged with the peeled body exit; each exit gets its own
2263 // private Phi and those Phis need to be merged here.
2264 Node *phi;
2265 if( prev->is_Region() ) {
2266 if( idx == 0 ) { // Updating control edge?
2267 phi = prev; // Just use existing control
2268 } else { // Else need a new Phi
2269 phi = PhiNode::make( prev, old );
2270 // Now recursively fix up the new uses of old!
2271 for( uint i = 1; i < prev->req(); i++ ) {
2272 worklist.push(phi); // Onto worklist once for each 'old' input
2273 }
2274 }
2275 } else {
2276 // Get new RegionNode merging old and new loop exits
2277 prev = old_new[prev->_idx];
2278 assert( prev, "just made this in step 7" );
2279 if( idx == 0) { // Updating control edge?
2280 phi = prev; // Just use existing control
2281 } else { // Else need a new Phi
2282 // Make a new Phi merging data values properly
2283 phi = PhiNode::make( prev, old );
2284 phi->set_req( 1, nnn );
2285 }
2286 }
2287 // If inserting a new Phi, check for prior hits
2288 if( idx != 0 ) {
2289 Node *hit = _igvn.hash_find_insert(phi);
2290 if( hit == nullptr ) {
2291 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2292 } else { // or
2293 // Remove the new phi from the graph and use the hit
2294 _igvn.remove_dead_node(phi);
2295 phi = hit; // Use existing phi
2296 }
2297 set_ctrl(phi, prev);
2298 }
2299 // Make 'use' use the Phi instead of the old loop body exit value
2300 assert(use->in(idx) == old, "old is still input of use");
2301 // We notify all uses of old, including use, and the indirect uses,
2302 // that may now be optimized because we have replaced old with phi.
2303 _igvn.add_users_to_worklist(old);
2304 if (idx == 0 &&
2305 use->depends_only_on_test()) {
2306 Node* pinned_clone = use->pin_array_access_node();
2307 if (pinned_clone != nullptr) {
2308 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2309 // into the region is left, an array load could become dependent on a condition that's not a range check for
2310 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2311 // floating above its range check.
2312 pinned_clone->set_req(0, phi);
2313 register_new_node_with_ctrl_of(pinned_clone, use);
2314 _igvn.replace_node(use, pinned_clone);
2315 continue;
2316 }
2317 }
2318 _igvn.replace_input_of(use, idx, phi);
2319 if( use->_idx >= new_counter ) { // If updating new phis
2320 // Not needed for correctness, but prevents a weak assert
2321 // in AddPNode from tripping (when we end up with different
2322 // base & derived Phis that will become the same after
2323 // IGVN does CSE).
2324 Node *hit = _igvn.hash_find_insert(use);
2325 if( hit ) // Go ahead and re-hash for hits.
2326 _igvn.replace_node( use, hit );
2327 }
2328 }
2329 }
2330 }
2331
2332 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2333 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2334 bool check_old_new) {
2335 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2336 Node* u = n->fast_out(j);
2337 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2338 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2339 Node* c = phase->get_ctrl(u);
2340 IdealLoopTree* u_loop = phase->get_loop(c);
2341 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2342 if (!loop->is_member(u_loop)) {
2343 if (outer_loop->is_member(u_loop)) {
2344 wq.push(u);
2345 } else {
2346 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2347 // the outer loop too
2348 Node* u_c = u->in(0);
2349 if (u_c != nullptr) {
2350 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2351 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2352 wq.push(u);
2353 }
2354 }
2355 }
2356 }
2357 }
2358 }
2359 }
2360
2361 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2362 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2363 Node_List& extra_data_nodes) {
2364 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2365 CountedLoopNode* cl = head->as_CountedLoop();
2366 Node* l = cl->outer_loop();
2367 Node* tail = cl->outer_loop_tail();
2368 IfNode* le = cl->outer_loop_end();
2369 Node* sfpt = cl->outer_safepoint();
2370 CountedLoopEndNode* cle = cl->loopexit();
2371 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2372 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2373 Node* cle_out = cle->proj_out(false);
2374
2375 Node* new_sfpt = nullptr;
2376 Node* new_cle_out = cle_out->clone();
2377 old_new.map(cle_out->_idx, new_cle_out);
2378 if (mode == CloneIncludesStripMined) {
2379 // clone outer loop body
2380 Node* new_l = l->clone();
2381 Node* new_tail = tail->clone();
2382 IfNode* new_le = le->clone()->as_If();
2383 new_sfpt = sfpt->clone();
2384
2385 set_loop(new_l, outer_loop->_parent);
2386 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2387 set_loop(new_cle_out, outer_loop->_parent);
2388 set_idom(new_cle_out, new_cle, dd);
2389 set_loop(new_sfpt, outer_loop->_parent);
2390 set_idom(new_sfpt, new_cle_out, dd);
2391 set_loop(new_le, outer_loop->_parent);
2392 set_idom(new_le, new_sfpt, dd);
2393 set_loop(new_tail, outer_loop->_parent);
2394 set_idom(new_tail, new_le, dd);
2395 set_idom(new_cl, new_l, dd);
2396
2397 old_new.map(l->_idx, new_l);
2398 old_new.map(tail->_idx, new_tail);
2399 old_new.map(le->_idx, new_le);
2400 old_new.map(sfpt->_idx, new_sfpt);
2401
2402 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2403 new_l->set_req(0, new_l);
2404 new_tail->set_req(0, new_le);
2405 new_le->set_req(0, new_sfpt);
2406 new_sfpt->set_req(0, new_cle_out);
2407 new_cle_out->set_req(0, new_cle);
2408 new_cl->set_req(LoopNode::EntryControl, new_l);
2409
2410 _igvn.register_new_node_with_optimizer(new_l);
2411 _igvn.register_new_node_with_optimizer(new_tail);
2412 _igvn.register_new_node_with_optimizer(new_le);
2413 } else {
2414 Node *newhead = old_new[loop->_head->_idx];
2415 newhead->as_Loop()->clear_strip_mined();
2416 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2417 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2418 }
2419 // Look at data node that were assigned a control in the outer
2420 // loop: they are kept in the outer loop by the safepoint so start
2421 // from the safepoint node's inputs.
2422 IdealLoopTree* outer_loop = get_loop(l);
2423 Node_Stack stack(2);
2424 stack.push(sfpt, 1);
2425 uint new_counter = C->unique();
2426 while (stack.size() > 0) {
2427 Node* n = stack.node();
2428 uint i = stack.index();
2429 while (i < n->req() &&
2430 (n->in(i) == nullptr ||
2431 !has_ctrl(n->in(i)) ||
2432 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2433 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2434 i++;
2435 }
2436 if (i < n->req()) {
2437 stack.set_index(i+1);
2438 stack.push(n->in(i), 0);
2439 } else {
2440 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2441 Node* m = n == sfpt ? new_sfpt : n->clone();
2442 if (m != nullptr) {
2443 for (uint i = 0; i < n->req(); i++) {
2444 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2445 m->set_req(i, old_new[m->in(i)->_idx]);
2446 }
2447 }
2448 } else {
2449 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2450 }
2451 if (n != sfpt) {
2452 extra_data_nodes.push(n);
2453 _igvn.register_new_node_with_optimizer(m);
2454 assert(get_ctrl(n) == cle_out, "what other control?");
2455 set_ctrl(m, new_cle_out);
2456 old_new.map(n->_idx, m);
2457 }
2458 stack.pop();
2459 }
2460 }
2461 if (mode == CloneIncludesStripMined) {
2462 _igvn.register_new_node_with_optimizer(new_sfpt);
2463 _igvn.register_new_node_with_optimizer(new_cle_out);
2464 }
2465 // Some other transformation may have pessimistically assigned some
2466 // data nodes to the outer loop. Set their control so they are out
2467 // of the outer loop.
2468 ResourceMark rm;
2469 Unique_Node_List wq;
2470 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2471 Node* old = extra_data_nodes.at(i);
2472 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2473 }
2474
2475 for (uint i = 0; i < loop->_body.size(); i++) {
2476 Node* old = loop->_body.at(i);
2477 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2478 }
2479
2480 Node* inner_out = sfpt->in(0);
2481 if (inner_out->outcnt() > 1) {
2482 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2483 }
2484
2485 Node* new_ctrl = cl->outer_loop_exit();
2486 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2487 for (uint i = 0; i < wq.size(); i++) {
2488 Node* n = wq.at(i);
2489 set_ctrl(n, new_ctrl);
2490 if (n->in(0) != nullptr) {
2491 _igvn.replace_input_of(n, 0, new_ctrl);
2492 }
2493 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2494 }
2495 } else {
2496 Node *newhead = old_new[loop->_head->_idx];
2497 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2498 }
2499 }
2500
2501 //------------------------------clone_loop-------------------------------------
2502 //
2503 // C L O N E A L O O P B O D Y
2504 //
2505 // This is the basic building block of the loop optimizations. It clones an
2506 // entire loop body. It makes an old_new loop body mapping; with this mapping
2507 // you can find the new-loop equivalent to an old-loop node. All new-loop
2508 // nodes are exactly equal to their old-loop counterparts, all edges are the
2509 // same. All exits from the old-loop now have a RegionNode that merges the
2510 // equivalent new-loop path. This is true even for the normal "loop-exit"
2511 // condition. All uses of loop-invariant old-loop values now come from (one
2512 // or more) Phis that merge their new-loop equivalents.
2513 //
2514 // This operation leaves the graph in an illegal state: there are two valid
2515 // control edges coming from the loop pre-header to both loop bodies. I'll
2516 // definitely have to hack the graph after running this transform.
2517 //
2518 // From this building block I will further edit edges to perform loop peeling
2519 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2520 //
2521 // Parameter side_by_size_idom:
2522 // When side_by_size_idom is null, the dominator tree is constructed for
2523 // the clone loop to dominate the original. Used in construction of
2524 // pre-main-post loop sequence.
2525 // When nonnull, the clone and original are side-by-side, both are
2526 // dominated by the side_by_side_idom node. Used in construction of
2527 // unswitched loops.
2528 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2529 CloneLoopMode mode, Node* side_by_side_idom) {
2530
2531 LoopNode* head = loop->_head->as_Loop();
2532 head->verify_strip_mined(1);
2533
2534 if (C->do_vector_loop() && PrintOpto) {
2535 const char* mname = C->method()->name()->as_quoted_ascii();
2536 if (mname != nullptr) {
2537 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2538 }
2539 }
2540
2541 CloneMap& cm = C->clone_map();
2542 if (C->do_vector_loop()) {
2543 cm.set_clone_idx(cm.max_gen()+1);
2544 #ifndef PRODUCT
2545 if (PrintOpto) {
2546 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2547 loop->dump_head();
2548 }
2549 #endif
2550 }
2551
2552 // Step 1: Clone the loop body. Make the old->new mapping.
2553 clone_loop_body(loop->_body, old_new, &cm);
2554
2555 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2556
2557 // Step 2: Fix the edges in the new body. If the old input is outside the
2558 // loop use it. If the old input is INside the loop, use the corresponding
2559 // new node instead.
2560 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2561
2562 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2563 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2564
2565 // Step 3: Now fix control uses. Loop varying control uses have already
2566 // been fixed up (as part of all input edges in Step 2). Loop invariant
2567 // control uses must be either an IfFalse or an IfTrue. Make a merge
2568 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2569 // refer to this.
2570 Node_List worklist;
2571 uint new_counter = C->unique();
2572 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2573
2574 // Step 4: If loop-invariant use is not control, it must be dominated by a
2575 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2576 // there if needed. Make a Phi there merging old and new used values.
2577 Node_List *split_if_set = nullptr;
2578 Node_List *split_bool_set = nullptr;
2579 Node_List *split_cex_set = nullptr;
2580 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2581
2582 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2583 Node* old = extra_data_nodes.at(i);
2584 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2585 split_bool_set, split_cex_set, worklist, new_counter,
2586 mode);
2587 }
2588
2589 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2590 // the loop uses a condition set in the loop. The original IF probably
2591 // takes control from one or more OLD Regions (which in turn get from NEW
2592 // Regions). In any case, there will be a set of Phis for each merge point
2593 // from the IF up to where the original BOOL def exists the loop.
2594 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2595
2596 }
2597
2598 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2599 if (split_if_set) {
2600 while (split_if_set->size()) {
2601 Node *iff = split_if_set->pop();
2602 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2603 if (iff->in(input)->is_Phi()) {
2604 Node *b = clone_iff(iff->in(input)->as_Phi());
2605 _igvn.replace_input_of(iff, input, b);
2606 }
2607 }
2608 }
2609 if (split_bool_set) {
2610 while (split_bool_set->size()) {
2611 Node *b = split_bool_set->pop();
2612 Node *phi = b->in(1);
2613 assert(phi->is_Phi(), "");
2614 CmpNode *cmp = clone_bool((PhiNode*) phi);
2615 _igvn.replace_input_of(b, 1, cmp);
2616 }
2617 }
2618 if (split_cex_set) {
2619 while (split_cex_set->size()) {
2620 Node *b = split_cex_set->pop();
2621 assert(b->in(0)->is_Region(), "");
2622 assert(b->in(1)->is_Phi(), "");
2623 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2624 split_up(b, b->in(0), nullptr);
2625 }
2626 }
2627 }
2628
2629 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2630 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2631 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2632 for(uint i = 0; i < body.size(); i++ ) {
2633 Node* old = body.at(i);
2634 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2635 split_bool_set, split_cex_set, worklist, new_counter,
2636 mode);
2637 }
2638 }
2639
2640 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2641 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2642 LoopNode* head = loop->_head->as_Loop();
2643 for(uint i = 0; i < body.size(); i++ ) {
2644 Node* old = body.at(i);
2645 if( !old->is_CFG() ) continue;
2646
2647 // Copy uses to a worklist, so I can munge the def-use info
2648 // with impunity.
2649 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2650 worklist.push(old->fast_out(j));
2651 }
2652
2653 while (worklist.size()) { // Visit all uses
2654 Node *use = worklist.pop();
2655 if (!has_node(use)) continue; // Ignore dead nodes
2656 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2657 if (!loop->is_member(use_loop) && use->is_CFG()) {
2658 // Both OLD and USE are CFG nodes here.
2659 assert(use->is_Proj(), "" );
2660 Node* nnn = old_new[old->_idx];
2661
2662 Node* newuse = nullptr;
2663 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2664 CountedLoopNode* cl = head->as_CountedLoop();
2665 CountedLoopEndNode* cle = cl->loopexit();
2666 Node* cle_out = cle->proj_out_or_null(false);
2667 if (use == cle_out) {
2668 IfNode* le = cl->outer_loop_end();
2669 use = le->proj_out(false);
2670 use_loop = get_loop(use);
2671 if (mode == CloneIncludesStripMined) {
2672 nnn = old_new[le->_idx];
2673 } else {
2674 newuse = old_new[cle_out->_idx];
2675 }
2676 }
2677 }
2678 if (newuse == nullptr) {
2679 newuse = use->clone();
2680 }
2681
2682 // Clone the loop exit control projection
2683 if (C->do_vector_loop() && cm != nullptr) {
2684 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2685 }
2686 newuse->set_req(0,nnn);
2687 _igvn.register_new_node_with_optimizer(newuse);
2688 set_loop(newuse, use_loop);
2689 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2690
2691 // We need a Region to merge the exit from the peeled body and the
2692 // exit from the old loop body.
2693 RegionNode *r = new RegionNode(3);
2694 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2695 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2696
2697 // The original user of 'use' uses 'r' instead.
2698 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2699 Node* useuse = use->last_out(l);
2700 _igvn.rehash_node_delayed(useuse);
2701 uint uses_found = 0;
2702 if (useuse->in(0) == use) {
2703 useuse->set_req(0, r);
2704 uses_found++;
2705 if (useuse->is_CFG()) {
2706 // This is not a dom_depth > dd_r because when new
2707 // control flow is constructed by a loop opt, a node and
2708 // its dominator can end up at the same dom_depth
2709 assert(dom_depth(useuse) >= dd_r, "");
2710 set_idom(useuse, r, dom_depth(useuse));
2711 }
2712 }
2713 for (uint k = 1; k < useuse->req(); k++) {
2714 if( useuse->in(k) == use ) {
2715 useuse->set_req(k, r);
2716 uses_found++;
2717 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2718 // This is not a dom_depth > dd_r because when new
2719 // control flow is constructed by a loop opt, a node
2720 // and its dominator can end up at the same dom_depth
2721 assert(dom_depth(useuse) >= dd_r , "");
2722 set_idom(useuse, r, dom_depth(useuse));
2723 }
2724 }
2725 }
2726 l -= uses_found; // we deleted 1 or more copies of this edge
2727 }
2728
2729 assert(use->is_Proj(), "loop exit should be projection");
2730 // lazy_replace() below moves all nodes that are:
2731 // - control dependent on the loop exit or
2732 // - have control set to the loop exit
2733 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2734 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2735 // becomes dead and is "replaced" by the region.
2736 Node* use_clone = use->clone();
2737 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2738 // Now finish up 'r'
2739 r->set_req(1, newuse);
2740 r->set_req(2, use_clone);
2741 _igvn.register_new_node_with_optimizer(r);
2742 set_loop(r, use_loop);
2743 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2744 lazy_replace(use, r);
2745 // Map the (cloned) old use to the new merge point
2746 old_new.map(use_clone->_idx, r);
2747 } // End of if a loop-exit test
2748 }
2749 }
2750 }
2751
2752 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2753 IdealLoopTree* parent, bool partial) {
2754 for(uint i = 0; i < body.size(); i++ ) {
2755 Node *old = body.at(i);
2756 Node *nnn = old_new[old->_idx];
2757 // Fix CFG/Loop controlling the new node
2758 if (has_ctrl(old)) {
2759 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2760 } else {
2761 set_loop(nnn, parent);
2762 if (old->outcnt() > 0) {
2763 Node* dom = idom(old);
2764 if (old_new[dom->_idx] != nullptr) {
2765 dom = old_new[dom->_idx];
2766 set_idom(nnn, dom, dd );
2767 }
2768 }
2769 }
2770 // Correct edges to the new node
2771 for (uint j = 0; j < nnn->req(); j++) {
2772 Node *n = nnn->in(j);
2773 if (n != nullptr) {
2774 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2775 if (loop->is_member(old_in_loop)) {
2776 if (old_new[n->_idx] != nullptr) {
2777 nnn->set_req(j, old_new[n->_idx]);
2778 } else {
2779 assert(!body.contains(n), "");
2780 assert(partial, "node not cloned");
2781 }
2782 }
2783 }
2784 }
2785 _igvn.hash_find_insert(nnn);
2786 }
2787 }
2788
2789 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2790 for (uint i = 0; i < body.size(); i++) {
2791 Node* old = body.at(i);
2792 Node* nnn = old->clone();
2793 old_new.map(old->_idx, nnn);
2794 if (C->do_vector_loop() && cm != nullptr) {
2795 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2796 }
2797 _igvn.register_new_node_with_optimizer(nnn);
2798 }
2799 }
2800
2801
2802 //---------------------- stride_of_possible_iv -------------------------------------
2803 // Looks for an iff/bool/comp with one operand of the compare
2804 // being a cycle involving an add and a phi,
2805 // with an optional truncation (left-shift followed by a right-shift)
2806 // of the add. Returns zero if not an iv.
2807 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2808 Node* trunc1 = nullptr;
2809 Node* trunc2 = nullptr;
2810 const TypeInteger* ttype = nullptr;
2811 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2812 return 0;
2813 }
2814 BoolNode* bl = iff->in(1)->as_Bool();
2815 Node* cmp = bl->in(1);
2816 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2817 return 0;
2818 }
2819 // Must have an invariant operand
2820 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2821 return 0;
2822 }
2823 Node* add2 = nullptr;
2824 Node* cmp1 = cmp->in(1);
2825 if (cmp1->is_Phi()) {
2826 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2827 Node* phi = cmp1;
2828 for (uint i = 1; i < phi->req(); i++) {
2829 Node* in = phi->in(i);
2830 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2831 &trunc1, &trunc2, &ttype, T_INT);
2832 if (add && add->in(1) == phi) {
2833 add2 = add->in(2);
2834 break;
2835 }
2836 }
2837 } else {
2838 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2839 Node* addtrunc = cmp1;
2840 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2841 &trunc1, &trunc2, &ttype, T_INT);
2842 if (add && add->in(1)->is_Phi()) {
2843 Node* phi = add->in(1);
2844 for (uint i = 1; i < phi->req(); i++) {
2845 if (phi->in(i) == addtrunc) {
2846 add2 = add->in(2);
2847 break;
2848 }
2849 }
2850 }
2851 }
2852 if (add2 != nullptr) {
2853 const TypeInt* add2t = _igvn.type(add2)->is_int();
2854 if (add2t->is_con()) {
2855 return add2t->get_con();
2856 }
2857 }
2858 return 0;
2859 }
2860
2861
2862 //---------------------- stay_in_loop -------------------------------------
2863 // Return the (unique) control output node that's in the loop (if it exists.)
2864 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2865 Node* unique = nullptr;
2866 if (!n) return nullptr;
2867 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2868 Node* use = n->fast_out(i);
2869 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2870 if (unique != nullptr) {
2871 return nullptr;
2872 }
2873 unique = use;
2874 }
2875 }
2876 return unique;
2877 }
2878
2879 //------------------------------ register_node -------------------------------------
2880 // Utility to register node "n" with PhaseIdealLoop
2881 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2882 _igvn.register_new_node_with_optimizer(n);
2883 loop->_body.push(n);
2884 if (n->is_CFG()) {
2885 set_loop(n, loop);
2886 set_idom(n, pred, ddepth);
2887 } else {
2888 set_ctrl(n, pred);
2889 }
2890 }
2891
2892 //------------------------------ proj_clone -------------------------------------
2893 // Utility to create an if-projection
2894 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2895 ProjNode* c = p->clone()->as_Proj();
2896 c->set_req(0, iff);
2897 return c;
2898 }
2899
2900 //------------------------------ short_circuit_if -------------------------------------
2901 // Force the iff control output to be the live_proj
2902 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2903 guarantee(live_proj != nullptr, "null projection");
2904 int proj_con = live_proj->_con;
2905 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2906 Node *con = _igvn.intcon(proj_con);
2907 set_ctrl(con, C->root());
2908 if (iff) {
2909 iff->set_req(1, con);
2910 }
2911 return con;
2912 }
2913
2914 //------------------------------ insert_if_before_proj -------------------------------------
2915 // Insert a new if before an if projection (* - new node)
2916 //
2917 // before
2918 // if(test)
2919 // / \
2920 // v v
2921 // other-proj proj (arg)
2922 //
2923 // after
2924 // if(test)
2925 // / \
2926 // / v
2927 // | * proj-clone
2928 // v |
2929 // other-proj v
2930 // * new_if(relop(cmp[IU](left,right)))
2931 // / \
2932 // v v
2933 // * new-proj proj
2934 // (returned)
2935 //
2936 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2937 IfNode* iff = proj->in(0)->as_If();
2938 IdealLoopTree *loop = get_loop(proj);
2939 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2940 uint ddepth = dom_depth(proj);
2941
2942 _igvn.rehash_node_delayed(iff);
2943 _igvn.rehash_node_delayed(proj);
2944
2945 proj->set_req(0, nullptr); // temporary disconnect
2946 ProjNode* proj2 = proj_clone(proj, iff);
2947 register_node(proj2, loop, iff, ddepth);
2948
2949 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2950 register_node(cmp, loop, proj2, ddepth);
2951
2952 BoolNode* bol = new BoolNode(cmp, relop);
2953 register_node(bol, loop, proj2, ddepth);
2954
2955 int opcode = iff->Opcode();
2956 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2957 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2958 register_node(new_if, loop, proj2, ddepth);
2959
2960 proj->set_req(0, new_if); // reattach
2961 set_idom(proj, new_if, ddepth);
2962
2963 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2964 guarantee(new_exit != nullptr, "null exit node");
2965 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2966
2967 return new_exit;
2968 }
2969
2970 //------------------------------ insert_region_before_proj -------------------------------------
2971 // Insert a region before an if projection (* - new node)
2972 //
2973 // before
2974 // if(test)
2975 // / |
2976 // v |
2977 // proj v
2978 // other-proj
2979 //
2980 // after
2981 // if(test)
2982 // / |
2983 // v |
2984 // * proj-clone v
2985 // | other-proj
2986 // v
2987 // * new-region
2988 // |
2989 // v
2990 // * dum_if
2991 // / \
2992 // v \
2993 // * dum-proj v
2994 // proj
2995 //
2996 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2997 IfNode* iff = proj->in(0)->as_If();
2998 IdealLoopTree *loop = get_loop(proj);
2999 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3000 uint ddepth = dom_depth(proj);
3001
3002 _igvn.rehash_node_delayed(iff);
3003 _igvn.rehash_node_delayed(proj);
3004
3005 proj->set_req(0, nullptr); // temporary disconnect
3006 ProjNode* proj2 = proj_clone(proj, iff);
3007 register_node(proj2, loop, iff, ddepth);
3008
3009 RegionNode* reg = new RegionNode(2);
3010 reg->set_req(1, proj2);
3011 register_node(reg, loop, iff, ddepth);
3012
3013 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3014 register_node(dum_if, loop, reg, ddepth);
3015
3016 proj->set_req(0, dum_if); // reattach
3017 set_idom(proj, dum_if, ddepth);
3018
3019 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3020 register_node(dum_proj, loop, dum_if, ddepth);
3021
3022 return reg;
3023 }
3024
3025 // Idea
3026 // ----
3027 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3028 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3029 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3030 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3031 // test alone instead:
3032 //
3033 // Before Partial Peeling:
3034 // Loop:
3035 // <peeled section>
3036 // Split off signed loop exit test
3037 // <-- CUT HERE -->
3038 // Unchanged unsigned loop exit test
3039 // <rest of unpeeled section>
3040 // goto Loop
3041 //
3042 // After Partial Peeling:
3043 // <cloned peeled section>
3044 // Cloned split off signed loop exit test
3045 // Loop:
3046 // Unchanged unsigned loop exit test
3047 // <rest of unpeeled section>
3048 // <peeled section>
3049 // Split off signed loop exit test
3050 // goto Loop
3051 //
3052 // Details
3053 // -------
3054 // Before:
3055 // if (i <u limit) Unsigned loop exit condition
3056 // / |
3057 // v v
3058 // exit-proj stay-in-loop-proj
3059 //
3060 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3061 // before the CmpU on the stay-in-loop path and keep both tests:
3062 //
3063 // if (i <u limit) Signed loop exit test
3064 // / |
3065 // / if (i <u limit) Unsigned loop exit test
3066 // / / |
3067 // v v v
3068 // exit-region stay-in-loop-proj
3069 //
3070 // Implementation
3071 // --------------
3072 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3073 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3074 // exit tests is preserved, and their loop nesting is correct.
3075 //
3076 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3077 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3078 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3079 //
3080 // if (stay-in-loop-const) Killed original unsigned loop exit test
3081 // / |
3082 // / v
3083 // / if (i < limit) Split off signed loop exit test
3084 // / / |
3085 // / / v
3086 // / / if (i <u limit) Cloned unsigned loop exit test
3087 // / / / |
3088 // v v v |
3089 // exit-region |
3090 // | |
3091 // dummy-if |
3092 // / | |
3093 // dead | |
3094 // v v
3095 // exit-proj stay-in-loop-proj
3096 //
3097 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3098 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3099 //
3100 // Requirements
3101 // ------------
3102 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3103 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3104 // Otherwise, we need to bail out (see comments in the code below).
3105 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3106 const bool Signed = true;
3107 const bool Unsigned = false;
3108
3109 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3110 if (bol->_test._test != BoolTest::lt) {
3111 return nullptr;
3112 }
3113 CmpNode* cmpu = bol->in(1)->as_Cmp();
3114 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3115
3116 int stride = stride_of_possible_iv(if_cmpu);
3117 if (stride == 0) {
3118 return nullptr;
3119 }
3120
3121 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3122 guarantee(lp_proj != nullptr, "null loop node");
3123
3124 ProjNode* lp_continue = lp_proj->as_Proj();
3125 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3126 if (!lp_exit->is_IfFalse()) {
3127 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3128 // We therefore can't add a single exit condition.
3129 return nullptr;
3130 }
3131 // The unsigned loop exit condition is
3132 // !(i <u limit)
3133 // = i >=u limit
3134 //
3135 // First, we note that for any x for which
3136 // 0 <= x <= INT_MAX
3137 // we can convert x to an unsigned int and still get the same guarantee:
3138 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3139 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3140 //
3141 // With that in mind, if
3142 // limit >= 0 (COND)
3143 // then the unsigned loop exit condition
3144 // i >=u limit (ULE)
3145 // is equivalent to
3146 // i < 0 || i >= limit (SLE-full)
3147 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3148 // (uint) i >=u MAX_INT >= limit >= 0
3149 // or otherwise
3150 // i >= limit >= 0
3151 // holds due to (LEMMA).
3152 //
3153 // For completeness, a counterexample with limit < 0:
3154 // Assume i = -3 and limit = -2:
3155 // i < 0
3156 // -2 < 0
3157 // is true and thus also "i < 0 || i >= limit". But
3158 // i >=u limit
3159 // -3 >=u -2
3160 // is false.
3161 Node* limit = cmpu->in(2);
3162 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3163 if (type_limit->_lo < 0) {
3164 return nullptr;
3165 }
3166
3167 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3168 // stride < 0:
3169 // i < 0 (SLE = SLE-negative)
3170 // stride > 0:
3171 // i >= limit (SLE = SLE-positive)
3172 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3173 //
3174 // Loop:
3175 // <peeled section>
3176 // i >= limit (SLE-positive)
3177 // <-- CUT HERE -->
3178 // i >=u limit (ULE)
3179 // <rest of unpeeled section>
3180 // goto Loop
3181 //
3182 // We exit the loop if:
3183 // (SLE) is true OR (ULE) is true
3184 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3185 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3186 // (SLE) IMPLIES (ULE)
3187 // This indeed holds when (COND) is given:
3188 // - stride > 0:
3189 // i >= limit // (SLE = SLE-positive)
3190 // i >= limit >= 0 // (COND)
3191 // i >=u limit >= 0 // (LEMMA)
3192 // which is the unsigned loop exit condition (ULE).
3193 // - stride < 0:
3194 // i < 0 // (SLE = SLE-negative)
3195 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3196 // MAX_INT >= limit >= 0 // (COND)
3197 // MAX_INT >=u limit >= 0 // (LEMMA)
3198 // and thus from (NEG) and (LEMMA):
3199 // i >=u limit
3200 // which is the unsigned loop exit condition (ULE).
3201 //
3202 //
3203 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3204 // <cloned peeled section>
3205 // i >= limit (SLE-positive)
3206 // Loop:
3207 // i >=u limit (ULE)
3208 // <rest of unpeeled section>
3209 // <peeled section>
3210 // i >= limit (SLE-positive)
3211 // goto Loop
3212 Node* rhs_cmpi;
3213 if (stride > 0) {
3214 rhs_cmpi = limit; // For i >= limit
3215 } else {
3216 rhs_cmpi = _igvn.makecon(TypeInt::ZERO); // For i < 0
3217 set_ctrl(rhs_cmpi, C->root());
3218 }
3219 // Create a new region on the exit path
3220 RegionNode* reg = insert_region_before_proj(lp_exit);
3221 guarantee(reg != nullptr, "null region node");
3222
3223 // Clone the if-cmpu-true-false using a signed compare
3224 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3225 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3226 reg->add_req(cmpi_exit);
3227
3228 // Clone the if-cmpu-true-false
3229 BoolTest::mask rel_u = bol->_test._test;
3230 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3231 reg->add_req(cmpu_exit);
3232
3233 // Force original if to stay in loop.
3234 short_circuit_if(if_cmpu, lp_continue);
3235
3236 return cmpi_exit->in(0)->as_If();
3237 }
3238
3239 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3240 // Remove a previously inserted signed compare loop exit.
3241 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3242 Node* lp_proj = stay_in_loop(if_cmp, loop);
3243 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3244 stay_in_loop(lp_proj, loop)->is_If() &&
3245 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3246 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3247 set_ctrl(con, C->root());
3248 if_cmp->set_req(1, con);
3249 }
3250
3251 //------------------------------ scheduled_nodelist -------------------------------------
3252 // Create a post order schedule of nodes that are in the
3253 // "member" set. The list is returned in "sched".
3254 // The first node in "sched" is the loop head, followed by
3255 // nodes which have no inputs in the "member" set, and then
3256 // followed by the nodes that have an immediate input dependence
3257 // on a node in "sched".
3258 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3259
3260 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3261 VectorSet visited;
3262 Node_Stack nstack(loop->_body.size());
3263
3264 Node* n = loop->_head; // top of stack is cached in "n"
3265 uint idx = 0;
3266 visited.set(n->_idx);
3267
3268 // Initially push all with no inputs from within member set
3269 for(uint i = 0; i < loop->_body.size(); i++ ) {
3270 Node *elt = loop->_body.at(i);
3271 if (member.test(elt->_idx)) {
3272 bool found = false;
3273 for (uint j = 0; j < elt->req(); j++) {
3274 Node* def = elt->in(j);
3275 if (def && member.test(def->_idx) && def != elt) {
3276 found = true;
3277 break;
3278 }
3279 }
3280 if (!found && elt != loop->_head) {
3281 nstack.push(n, idx);
3282 n = elt;
3283 assert(!visited.test(n->_idx), "not seen yet");
3284 visited.set(n->_idx);
3285 }
3286 }
3287 }
3288
3289 // traverse out's that are in the member set
3290 while (true) {
3291 if (idx < n->outcnt()) {
3292 Node* use = n->raw_out(idx);
3293 idx++;
3294 if (!visited.test_set(use->_idx)) {
3295 if (member.test(use->_idx)) {
3296 nstack.push(n, idx);
3297 n = use;
3298 idx = 0;
3299 }
3300 }
3301 } else {
3302 // All outputs processed
3303 sched.push(n);
3304 if (nstack.is_empty()) break;
3305 n = nstack.node();
3306 idx = nstack.index();
3307 nstack.pop();
3308 }
3309 }
3310 }
3311
3312
3313 //------------------------------ has_use_in_set -------------------------------------
3314 // Has a use in the vector set
3315 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3316 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3317 Node* use = n->fast_out(j);
3318 if (vset.test(use->_idx)) {
3319 return true;
3320 }
3321 }
3322 return false;
3323 }
3324
3325
3326 //------------------------------ has_use_internal_to_set -------------------------------------
3327 // Has use internal to the vector set (ie. not in a phi at the loop head)
3328 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3329 Node* head = loop->_head;
3330 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3331 Node* use = n->fast_out(j);
3332 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3333 return true;
3334 }
3335 }
3336 return false;
3337 }
3338
3339
3340 //------------------------------ clone_for_use_outside_loop -------------------------------------
3341 // clone "n" for uses that are outside of loop
3342 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3343 int cloned = 0;
3344 assert(worklist.size() == 0, "should be empty");
3345 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3346 Node* use = n->fast_out(j);
3347 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3348 worklist.push(use);
3349 }
3350 }
3351
3352 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3353 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3354 return -1;
3355 }
3356
3357 while( worklist.size() ) {
3358 Node *use = worklist.pop();
3359 if (!has_node(use) || use->in(0) == C->top()) continue;
3360 uint j;
3361 for (j = 0; j < use->req(); j++) {
3362 if (use->in(j) == n) break;
3363 }
3364 assert(j < use->req(), "must be there");
3365
3366 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3367 Node* n_clone = n->clone();
3368 _igvn.replace_input_of(use, j, n_clone);
3369 cloned++;
3370 Node* use_c;
3371 if (!use->is_Phi()) {
3372 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3373 } else {
3374 // Use in a phi is considered a use in the associated predecessor block
3375 use_c = use->in(0)->in(j);
3376 }
3377 set_ctrl(n_clone, use_c);
3378 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3379 get_loop(use_c)->_body.push(n_clone);
3380 _igvn.register_new_node_with_optimizer(n_clone);
3381 #ifndef PRODUCT
3382 if (TracePartialPeeling) {
3383 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3384 }
3385 #endif
3386 }
3387 return cloned;
3388 }
3389
3390
3391 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3392 // clone "n" for special uses that are in the not_peeled region.
3393 // If these def-uses occur in separate blocks, the code generator
3394 // marks the method as not compilable. For example, if a "BoolNode"
3395 // is in a different basic block than the "IfNode" that uses it, then
3396 // the compilation is aborted in the code generator.
3397 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3398 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3399 if (n->is_Phi() || n->is_Load()) {
3400 return;
3401 }
3402 assert(worklist.size() == 0, "should be empty");
3403 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3404 Node* use = n->fast_out(j);
3405 if ( not_peel.test(use->_idx) &&
3406 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3407 use->in(1) == n) {
3408 worklist.push(use);
3409 }
3410 }
3411 if (worklist.size() > 0) {
3412 // clone "n" and insert it between inputs of "n" and the use
3413 Node* n_clone = n->clone();
3414 loop->_body.push(n_clone);
3415 _igvn.register_new_node_with_optimizer(n_clone);
3416 set_ctrl(n_clone, get_ctrl(n));
3417 sink_list.push(n_clone);
3418 not_peel.set(n_clone->_idx);
3419 #ifndef PRODUCT
3420 if (TracePartialPeeling) {
3421 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3422 }
3423 #endif
3424 while( worklist.size() ) {
3425 Node *use = worklist.pop();
3426 _igvn.rehash_node_delayed(use);
3427 for (uint j = 1; j < use->req(); j++) {
3428 if (use->in(j) == n) {
3429 use->set_req(j, n_clone);
3430 }
3431 }
3432 }
3433 }
3434 }
3435
3436
3437 //------------------------------ insert_phi_for_loop -------------------------------------
3438 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3439 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3440 Node *phi = PhiNode::make(lp, back_edge_val);
3441 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3442 // Use existing phi if it already exists
3443 Node *hit = _igvn.hash_find_insert(phi);
3444 if( hit == nullptr ) {
3445 _igvn.register_new_node_with_optimizer(phi);
3446 set_ctrl(phi, lp);
3447 } else {
3448 // Remove the new phi from the graph and use the hit
3449 _igvn.remove_dead_node(phi);
3450 phi = hit;
3451 }
3452 _igvn.replace_input_of(use, idx, phi);
3453 }
3454
3455 #ifdef ASSERT
3456 //------------------------------ is_valid_loop_partition -------------------------------------
3457 // Validate the loop partition sets: peel and not_peel
3458 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3459 VectorSet& not_peel ) {
3460 uint i;
3461 // Check that peel_list entries are in the peel set
3462 for (i = 0; i < peel_list.size(); i++) {
3463 if (!peel.test(peel_list.at(i)->_idx)) {
3464 return false;
3465 }
3466 }
3467 // Check at loop members are in one of peel set or not_peel set
3468 for (i = 0; i < loop->_body.size(); i++ ) {
3469 Node *def = loop->_body.at(i);
3470 uint di = def->_idx;
3471 // Check that peel set elements are in peel_list
3472 if (peel.test(di)) {
3473 if (not_peel.test(di)) {
3474 return false;
3475 }
3476 // Must be in peel_list also
3477 bool found = false;
3478 for (uint j = 0; j < peel_list.size(); j++) {
3479 if (peel_list.at(j)->_idx == di) {
3480 found = true;
3481 break;
3482 }
3483 }
3484 if (!found) {
3485 return false;
3486 }
3487 } else if (not_peel.test(di)) {
3488 if (peel.test(di)) {
3489 return false;
3490 }
3491 } else {
3492 return false;
3493 }
3494 }
3495 return true;
3496 }
3497
3498 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3499 // Ensure a use outside of loop is of the right form
3500 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3501 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3502 return (use->is_Phi() &&
3503 use_c->is_Region() && use_c->req() == 3 &&
3504 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3505 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3506 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3507 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3508 }
3509
3510 //------------------------------ is_valid_clone_loop_form -------------------------------------
3511 // Ensure that all uses outside of loop are of the right form
3512 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3513 uint orig_exit_idx, uint clone_exit_idx) {
3514 uint len = peel_list.size();
3515 for (uint i = 0; i < len; i++) {
3516 Node *def = peel_list.at(i);
3517
3518 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3519 Node *use = def->fast_out(j);
3520 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3521 if (!loop->is_member(get_loop(use_c))) {
3522 // use is not in the loop, check for correct structure
3523 if (use->in(0) == def) {
3524 // Okay
3525 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3526 return false;
3527 }
3528 }
3529 }
3530 }
3531 return true;
3532 }
3533 #endif
3534
3535 //------------------------------ partial_peel -------------------------------------
3536 // Partially peel (aka loop rotation) the top portion of a loop (called
3537 // the peel section below) by cloning it and placing one copy just before
3538 // the new loop head and the other copy at the bottom of the new loop.
3539 //
3540 // before after where it came from
3541 //
3542 // stmt1 stmt1
3543 // loop: stmt2 clone
3544 // stmt2 if condA goto exitA clone
3545 // if condA goto exitA new_loop: new
3546 // stmt3 stmt3 clone
3547 // if !condB goto loop if condB goto exitB clone
3548 // exitB: stmt2 orig
3549 // stmt4 if !condA goto new_loop orig
3550 // exitA: goto exitA
3551 // exitB:
3552 // stmt4
3553 // exitA:
3554 //
3555 // Step 1: find the cut point: an exit test on probable
3556 // induction variable.
3557 // Step 2: schedule (with cloning) operations in the peel
3558 // section that can be executed after the cut into
3559 // the section that is not peeled. This may need
3560 // to clone operations into exit blocks. For
3561 // instance, a reference to A[i] in the not-peel
3562 // section and a reference to B[i] in an exit block
3563 // may cause a left-shift of i by 2 to be placed
3564 // in the peel block. This step will clone the left
3565 // shift into the exit block and sink the left shift
3566 // from the peel to the not-peel section.
3567 // Step 3: clone the loop, retarget the control, and insert
3568 // phis for values that are live across the new loop
3569 // head. This is very dependent on the graph structure
3570 // from clone_loop. It creates region nodes for
3571 // exit control and associated phi nodes for values
3572 // flow out of the loop through that exit. The region
3573 // node is dominated by the clone's control projection.
3574 // So the clone's peel section is placed before the
3575 // new loop head, and the clone's not-peel section is
3576 // forms the top part of the new loop. The original
3577 // peel section forms the tail of the new loop.
3578 // Step 4: update the dominator tree and recompute the
3579 // dominator depth.
3580 //
3581 // orig
3582 //
3583 // stmt1
3584 // |
3585 // v
3586 // predicates
3587 // |
3588 // v
3589 // loop<----+
3590 // | |
3591 // stmt2 |
3592 // | |
3593 // v |
3594 // ifA |
3595 // / | |
3596 // v v |
3597 // false true ^ <-- last_peel
3598 // / | |
3599 // / ===|==cut |
3600 // / stmt3 | <-- first_not_peel
3601 // / | |
3602 // | v |
3603 // v ifB |
3604 // exitA: / \ |
3605 // / \ |
3606 // v v |
3607 // false true |
3608 // / \ |
3609 // / ----+
3610 // |
3611 // v
3612 // exitB:
3613 // stmt4
3614 //
3615 //
3616 // after clone loop
3617 //
3618 // stmt1
3619 // |
3620 // v
3621 // predicates
3622 // / \
3623 // clone / \ orig
3624 // / \
3625 // / \
3626 // v v
3627 // +---->loop loop<----+
3628 // | | | |
3629 // | stmt2 stmt2 |
3630 // | | | |
3631 // | v v |
3632 // | ifA ifA |
3633 // | | \ / | |
3634 // | v v v v |
3635 // ^ true false false true ^ <-- last_peel
3636 // | | ^ \ / | |
3637 // | cut==|== \ \ / ===|==cut |
3638 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3639 // | | dom | | | |
3640 // | v \ 1v v2 v |
3641 // | ifB regionA ifB |
3642 // | / \ | / \ |
3643 // | / \ v / \ |
3644 // | v v exitA: v v |
3645 // | true false false true |
3646 // | / ^ \ / \ |
3647 // +---- \ \ / ----+
3648 // dom \ /
3649 // \ 1v v2
3650 // regionB
3651 // |
3652 // v
3653 // exitB:
3654 // stmt4
3655 //
3656 //
3657 // after partial peel
3658 //
3659 // stmt1
3660 // |
3661 // v
3662 // predicates
3663 // /
3664 // clone / orig
3665 // / TOP
3666 // / \
3667 // v v
3668 // TOP->loop loop----+
3669 // | | |
3670 // stmt2 stmt2 |
3671 // | | |
3672 // v v |
3673 // ifA ifA |
3674 // | \ / | |
3675 // v v v v |
3676 // true false false true | <-- last_peel
3677 // | ^ \ / +------|---+
3678 // +->newloop \ \ / === ==cut | |
3679 // | stmt3 \ \ / TOP | |
3680 // | | dom | | stmt3 | | <-- first_not_peel
3681 // | v \ 1v v2 v | |
3682 // | ifB regionA ifB ^ v
3683 // | / \ | / \ | |
3684 // | / \ v / \ | |
3685 // | v v exitA: v v | |
3686 // | true false false true | |
3687 // | / ^ \ / \ | |
3688 // | | \ \ / v | |
3689 // | | dom \ / TOP | |
3690 // | | \ 1v v2 | |
3691 // ^ v regionB | |
3692 // | | | | |
3693 // | | v ^ v
3694 // | | exitB: | |
3695 // | | stmt4 | |
3696 // | +------------>-----------------+ |
3697 // | |
3698 // +-----------------<---------------------+
3699 //
3700 //
3701 // final graph
3702 //
3703 // stmt1
3704 // |
3705 // v
3706 // predicates
3707 // |
3708 // v
3709 // stmt2 clone
3710 // |
3711 // v
3712 // ........> ifA clone
3713 // : / |
3714 // dom / |
3715 // : v v
3716 // : false true
3717 // : | |
3718 // : | v
3719 // : | newloop<-----+
3720 // : | | |
3721 // : | stmt3 clone |
3722 // : | | |
3723 // : | v |
3724 // : | ifB |
3725 // : | / \ |
3726 // : | v v |
3727 // : | false true |
3728 // : | | | |
3729 // : | v stmt2 |
3730 // : | exitB: | |
3731 // : | stmt4 v |
3732 // : | ifA orig |
3733 // : | / \ |
3734 // : | / \ |
3735 // : | v v |
3736 // : | false true |
3737 // : | / \ |
3738 // : v v -----+
3739 // RegionA
3740 // |
3741 // v
3742 // exitA
3743 //
3744 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3745
3746 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3747 if (!loop->_head->is_Loop()) {
3748 return false;
3749 }
3750 LoopNode *head = loop->_head->as_Loop();
3751
3752 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3753 return false;
3754 }
3755
3756 // Check for complex exit control
3757 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3758 Node *n = loop->_body.at(ii);
3759 int opc = n->Opcode();
3760 if (n->is_Call() ||
3761 opc == Op_Catch ||
3762 opc == Op_CatchProj ||
3763 opc == Op_Jump ||
3764 opc == Op_JumpProj) {
3765 #ifndef PRODUCT
3766 if (TracePartialPeeling) {
3767 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3768 }
3769 #endif
3770 return false;
3771 }
3772 }
3773
3774 int dd = dom_depth(head);
3775
3776 // Step 1: find cut point
3777
3778 // Walk up dominators to loop head looking for first loop exit
3779 // which is executed on every path thru loop.
3780 IfNode *peel_if = nullptr;
3781 IfNode *peel_if_cmpu = nullptr;
3782
3783 Node *iff = loop->tail();
3784 while (iff != head) {
3785 if (iff->is_If()) {
3786 Node *ctrl = get_ctrl(iff->in(1));
3787 if (ctrl->is_top()) return false; // Dead test on live IF.
3788 // If loop-varying exit-test, check for induction variable
3789 if (loop->is_member(get_loop(ctrl)) &&
3790 loop->is_loop_exit(iff) &&
3791 is_possible_iv_test(iff)) {
3792 Node* cmp = iff->in(1)->in(1);
3793 if (cmp->Opcode() == Op_CmpI) {
3794 peel_if = iff->as_If();
3795 } else {
3796 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3797 peel_if_cmpu = iff->as_If();
3798 }
3799 }
3800 }
3801 iff = idom(iff);
3802 }
3803
3804 // Prefer signed compare over unsigned compare.
3805 IfNode* new_peel_if = nullptr;
3806 if (peel_if == nullptr) {
3807 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3808 return false; // No peel point found
3809 }
3810 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3811 if (new_peel_if == nullptr) {
3812 return false; // No peel point found
3813 }
3814 peel_if = new_peel_if;
3815 }
3816 Node* last_peel = stay_in_loop(peel_if, loop);
3817 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3818 if (first_not_peeled == nullptr || first_not_peeled == head) {
3819 return false;
3820 }
3821
3822 #ifndef PRODUCT
3823 if (TraceLoopOpts) {
3824 tty->print("PartialPeel ");
3825 loop->dump_head();
3826 }
3827
3828 if (TracePartialPeeling) {
3829 tty->print_cr("before partial peel one iteration");
3830 Node_List wl;
3831 Node* t = head->in(2);
3832 while (true) {
3833 wl.push(t);
3834 if (t == head) break;
3835 t = idom(t);
3836 }
3837 while (wl.size() > 0) {
3838 Node* tt = wl.pop();
3839 tt->dump();
3840 if (tt == last_peel) tty->print_cr("-- cut --");
3841 }
3842 }
3843 #endif
3844
3845 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3846
3847 VectorSet peel;
3848 VectorSet not_peel;
3849 Node_List peel_list;
3850 Node_List worklist;
3851 Node_List sink_list;
3852
3853 uint estimate = loop->est_loop_clone_sz(1);
3854 if (exceeding_node_budget(estimate)) {
3855 return false;
3856 }
3857
3858 // Set of cfg nodes to peel are those that are executable from
3859 // the head through last_peel.
3860 assert(worklist.size() == 0, "should be empty");
3861 worklist.push(head);
3862 peel.set(head->_idx);
3863 while (worklist.size() > 0) {
3864 Node *n = worklist.pop();
3865 if (n != last_peel) {
3866 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3867 Node* use = n->fast_out(j);
3868 if (use->is_CFG() &&
3869 loop->is_member(get_loop(use)) &&
3870 !peel.test_set(use->_idx)) {
3871 worklist.push(use);
3872 }
3873 }
3874 }
3875 }
3876
3877 // Set of non-cfg nodes to peel are those that are control
3878 // dependent on the cfg nodes.
3879 for (uint i = 0; i < loop->_body.size(); i++) {
3880 Node *n = loop->_body.at(i);
3881 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3882 if (peel.test(n_c->_idx)) {
3883 peel.set(n->_idx);
3884 } else {
3885 not_peel.set(n->_idx);
3886 }
3887 }
3888
3889 // Step 2: move operations from the peeled section down into the
3890 // not-peeled section
3891
3892 // Get a post order schedule of nodes in the peel region
3893 // Result in right-most operand.
3894 scheduled_nodelist(loop, peel, peel_list);
3895
3896 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3897
3898 // For future check for too many new phis
3899 uint old_phi_cnt = 0;
3900 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3901 Node* use = head->fast_out(j);
3902 if (use->is_Phi()) old_phi_cnt++;
3903 }
3904
3905 #ifndef PRODUCT
3906 if (TracePartialPeeling) {
3907 tty->print_cr("\npeeled list");
3908 }
3909 #endif
3910
3911 // Evacuate nodes in peel region into the not_peeled region if possible
3912 bool too_many_clones = false;
3913 uint new_phi_cnt = 0;
3914 uint cloned_for_outside_use = 0;
3915 for (uint i = 0; i < peel_list.size();) {
3916 Node* n = peel_list.at(i);
3917 #ifndef PRODUCT
3918 if (TracePartialPeeling) n->dump();
3919 #endif
3920 bool incr = true;
3921 if (!n->is_CFG()) {
3922 if (has_use_in_set(n, not_peel)) {
3923 // If not used internal to the peeled region,
3924 // move "n" from peeled to not_peeled region.
3925 if (!has_use_internal_to_set(n, peel, loop)) {
3926 // if not pinned and not a load (which maybe anti-dependent on a store)
3927 // and not a CMove (Matcher expects only bool->cmove).
3928 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3929 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3930 if (C->failing()) return false;
3931 if (new_clones == -1) {
3932 too_many_clones = true;
3933 break;
3934 }
3935 cloned_for_outside_use += new_clones;
3936 sink_list.push(n);
3937 peel.remove(n->_idx);
3938 not_peel.set(n->_idx);
3939 peel_list.remove(i);
3940 incr = false;
3941 #ifndef PRODUCT
3942 if (TracePartialPeeling) {
3943 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3944 n->_idx, get_ctrl(n)->_idx);
3945 }
3946 #endif
3947 }
3948 } else {
3949 // Otherwise check for special def-use cases that span
3950 // the peel/not_peel boundary such as bool->if
3951 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3952 new_phi_cnt++;
3953 }
3954 }
3955 }
3956 if (incr) i++;
3957 }
3958
3959 estimate += cloned_for_outside_use + new_phi_cnt;
3960 bool exceed_node_budget = !may_require_nodes(estimate);
3961 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3962
3963 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3964 #ifndef PRODUCT
3965 if (TracePartialPeeling && exceed_phi_limit) {
3966 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3967 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3968 }
3969 #endif
3970 if (new_peel_if != nullptr) {
3971 remove_cmpi_loop_exit(new_peel_if, loop);
3972 }
3973 // Inhibit more partial peeling on this loop
3974 assert(!head->is_partial_peel_loop(), "not partial peeled");
3975 head->mark_partial_peel_failed();
3976 if (cloned_for_outside_use > 0) {
3977 // Terminate this round of loop opts because
3978 // the graph outside this loop was changed.
3979 C->set_major_progress();
3980 return true;
3981 }
3982 return false;
3983 }
3984
3985 // Step 3: clone loop, retarget control, and insert new phis
3986
3987 // Create new loop head for new phis and to hang
3988 // the nodes being moved (sinked) from the peel region.
3989 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3990 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3991 _igvn.register_new_node_with_optimizer(new_head);
3992 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3993 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3994 set_loop(new_head, loop);
3995 loop->_body.push(new_head);
3996 not_peel.set(new_head->_idx);
3997 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3998 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3999
4000 while (sink_list.size() > 0) {
4001 Node* n = sink_list.pop();
4002 set_ctrl(n, new_head);
4003 }
4004
4005 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4006
4007 clone_loop(loop, old_new, dd, IgnoreStripMined);
4008
4009 const uint clone_exit_idx = 1;
4010 const uint orig_exit_idx = 2;
4011 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4012
4013 Node* head_clone = old_new[head->_idx];
4014 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4015 Node* orig_tail_clone = head_clone->in(2);
4016
4017 // Add phi if "def" node is in peel set and "use" is not
4018
4019 for (uint i = 0; i < peel_list.size(); i++) {
4020 Node *def = peel_list.at(i);
4021 if (!def->is_CFG()) {
4022 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4023 Node *use = def->fast_out(j);
4024 if (has_node(use) && use->in(0) != C->top() &&
4025 (!peel.test(use->_idx) ||
4026 (use->is_Phi() && use->in(0) == head)) ) {
4027 worklist.push(use);
4028 }
4029 }
4030 while( worklist.size() ) {
4031 Node *use = worklist.pop();
4032 for (uint j = 1; j < use->req(); j++) {
4033 Node* n = use->in(j);
4034 if (n == def) {
4035
4036 // "def" is in peel set, "use" is not in peel set
4037 // or "use" is in the entry boundary (a phi) of the peel set
4038
4039 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4040
4041 if ( loop->is_member(get_loop( use_c )) ) {
4042 // use is in loop
4043 if (old_new[use->_idx] != nullptr) { // null for dead code
4044 Node* use_clone = old_new[use->_idx];
4045 _igvn.replace_input_of(use, j, C->top());
4046 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4047 }
4048 } else {
4049 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4050 // use is not in the loop, check if the live range includes the cut
4051 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4052 if (not_peel.test(lp_if->_idx)) {
4053 assert(j == orig_exit_idx, "use from original loop");
4054 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4055 }
4056 }
4057 }
4058 }
4059 }
4060 }
4061 }
4062
4063 // Step 3b: retarget control
4064
4065 // Redirect control to the new loop head if a cloned node in
4066 // the not_peeled region has control that points into the peeled region.
4067 // This necessary because the cloned peeled region will be outside
4068 // the loop.
4069 // from to
4070 // cloned-peeled <---+
4071 // new_head_clone: | <--+
4072 // cloned-not_peeled in(0) in(0)
4073 // orig-peeled
4074
4075 for (uint i = 0; i < loop->_body.size(); i++) {
4076 Node *n = loop->_body.at(i);
4077 if (!n->is_CFG() && n->in(0) != nullptr &&
4078 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4079 Node* n_clone = old_new[n->_idx];
4080 if (n_clone->depends_only_on_test()) {
4081 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4082 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4083 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4084 // floating above its range check.
4085 Node* pinned_clone = n_clone->pin_array_access_node();
4086 if (pinned_clone != nullptr) {
4087 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4088 old_new.map(n->_idx, pinned_clone);
4089 _igvn.replace_node(n_clone, pinned_clone);
4090 n_clone = pinned_clone;
4091 }
4092 }
4093 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4094 }
4095 }
4096
4097 // Backedge of the surviving new_head (the clone) is original last_peel
4098 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4099
4100 // Cut first node in original not_peel set
4101 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4102 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4103 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4104
4105 // Copy head_clone back-branch info to original head
4106 // and remove original head's loop entry and
4107 // clone head's back-branch
4108 _igvn.rehash_node_delayed(head); // Multiple edge updates
4109 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4110 head->set_req(LoopNode::LoopBackControl, C->top());
4111 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4112
4113 // Similarly modify the phis
4114 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4115 Node* use = head->fast_out(k);
4116 if (use->is_Phi() && use->outcnt() > 0) {
4117 Node* use_clone = old_new[use->_idx];
4118 _igvn.rehash_node_delayed(use); // Multiple edge updates
4119 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4120 use->set_req(LoopNode::LoopBackControl, C->top());
4121 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4122 }
4123 }
4124
4125 // Step 4: update dominator tree and dominator depth
4126
4127 set_idom(head, orig_tail_clone, dd);
4128 recompute_dom_depth();
4129
4130 // Inhibit more partial peeling on this loop
4131 new_head_clone->set_partial_peel_loop();
4132 C->set_major_progress();
4133 loop->record_for_igvn();
4134
4135 #ifndef PRODUCT
4136 if (TracePartialPeeling) {
4137 tty->print_cr("\nafter partial peel one iteration");
4138 Node_List wl;
4139 Node* t = last_peel;
4140 while (true) {
4141 wl.push(t);
4142 if (t == head_clone) break;
4143 t = idom(t);
4144 }
4145 while (wl.size() > 0) {
4146 Node* tt = wl.pop();
4147 if (tt == head) tty->print_cr("orig head");
4148 else if (tt == new_head_clone) tty->print_cr("new head");
4149 else if (tt == head_clone) tty->print_cr("clone head");
4150 tt->dump();
4151 }
4152 }
4153 #endif
4154
4155 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4156
4157 return true;
4158 }
4159
4160 // Transform:
4161 //
4162 // loop<-----------------+
4163 // | |
4164 // stmt1 stmt2 .. stmtn |
4165 // | | | |
4166 // \ | / |
4167 // v v v |
4168 // region |
4169 // | |
4170 // shared_stmt |
4171 // | |
4172 // v |
4173 // if |
4174 // / \ |
4175 // | -----------+
4176 // v
4177 //
4178 // into:
4179 //
4180 // loop<-------------------+
4181 // | |
4182 // v |
4183 // +->loop |
4184 // | | |
4185 // | stmt1 stmt2 .. stmtn |
4186 // | | | | |
4187 // | | \ / |
4188 // | | v v |
4189 // | | region1 |
4190 // | | | |
4191 // | shared_stmt shared_stmt |
4192 // | | | |
4193 // | v v |
4194 // | if if |
4195 // | /\ / \ |
4196 // +-- | | -------+
4197 // \ /
4198 // v v
4199 // region2
4200 //
4201 // (region2 is shown to merge mirrored projections of the loop exit
4202 // ifs to make the diagram clearer but they really merge the same
4203 // projection)
4204 //
4205 // Conditions for this transformation to trigger:
4206 // - the path through stmt1 is frequent enough
4207 // - the inner loop will be turned into a counted loop after transformation
4208 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4209 if (!DuplicateBackedge) {
4210 return false;
4211 }
4212 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4213 if (!loop->_head->is_Loop()) {
4214 return false;
4215 }
4216
4217 uint estimate = loop->est_loop_clone_sz(1);
4218 if (exceeding_node_budget(estimate)) {
4219 return false;
4220 }
4221
4222 LoopNode *head = loop->_head->as_Loop();
4223
4224 Node* region = nullptr;
4225 IfNode* exit_test = nullptr;
4226 uint inner;
4227 float f;
4228 if (StressDuplicateBackedge) {
4229 if (head->is_strip_mined()) {
4230 return false;
4231 }
4232 Node* c = head->in(LoopNode::LoopBackControl);
4233
4234 while (c != head) {
4235 if (c->is_Region()) {
4236 region = c;
4237 }
4238 c = idom(c);
4239 }
4240
4241 if (region == nullptr) {
4242 return false;
4243 }
4244
4245 inner = 1;
4246 } else {
4247 // Is the shape of the loop that of a counted loop...
4248 Node* back_control = loop_exit_control(head, loop);
4249 if (back_control == nullptr) {
4250 return false;
4251 }
4252
4253 BoolTest::mask bt = BoolTest::illegal;
4254 float cl_prob = 0;
4255 Node* incr = nullptr;
4256 Node* limit = nullptr;
4257 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4258 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4259 return false;
4260 }
4261
4262 // With an extra phi for the candidate iv?
4263 // Or the region node is the loop head
4264 if (!incr->is_Phi() || incr->in(0) == head) {
4265 return false;
4266 }
4267
4268 PathFrequency pf(head, this);
4269 region = incr->in(0);
4270
4271 // Go over all paths for the extra phi's region and see if that
4272 // path is frequent enough and would match the expected iv shape
4273 // if the extra phi is removed
4274 inner = 0;
4275 for (uint i = 1; i < incr->req(); ++i) {
4276 Node* in = incr->in(i);
4277 Node* trunc1 = nullptr;
4278 Node* trunc2 = nullptr;
4279 const TypeInteger* iv_trunc_t = nullptr;
4280 Node* orig_in = in;
4281 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4282 continue;
4283 }
4284 assert(in->Opcode() == Op_AddI, "wrong increment code");
4285 Node* xphi = nullptr;
4286 Node* stride = loop_iv_stride(in, loop, xphi);
4287
4288 if (stride == nullptr) {
4289 continue;
4290 }
4291
4292 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4293 if (phi == nullptr ||
4294 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4295 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4296 return false;
4297 }
4298
4299 f = pf.to(region->in(i));
4300 if (f > 0.5) {
4301 inner = i;
4302 break;
4303 }
4304 }
4305
4306 if (inner == 0) {
4307 return false;
4308 }
4309
4310 exit_test = back_control->in(0)->as_If();
4311 }
4312
4313 if (idom(region)->is_Catch()) {
4314 return false;
4315 }
4316
4317 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4318 Unique_Node_List wq;
4319 wq.push(head->in(LoopNode::LoopBackControl));
4320 for (uint i = 0; i < wq.size(); i++) {
4321 Node* c = wq.at(i);
4322 assert(get_loop(c) == loop, "not in the right loop?");
4323 if (c->is_Region()) {
4324 if (c != region) {
4325 for (uint j = 1; j < c->req(); ++j) {
4326 wq.push(c->in(j));
4327 }
4328 }
4329 } else {
4330 wq.push(c->in(0));
4331 }
4332 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4333 }
4334
4335 Node* region_dom = idom(region);
4336
4337 // Can't do the transformation if this would cause a membar pair to
4338 // be split
4339 for (uint i = 0; i < wq.size(); i++) {
4340 Node* c = wq.at(i);
4341 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4342 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4343 if (!wq.member(c->as_MemBar()->leading_membar())) {
4344 return false;
4345 }
4346 }
4347 }
4348
4349 // Collect data nodes that need to be clones as well
4350 int dd = dom_depth(head);
4351
4352 for (uint i = 0; i < loop->_body.size(); ++i) {
4353 Node* n = loop->_body.at(i);
4354 if (has_ctrl(n)) {
4355 Node* c = get_ctrl(n);
4356 if (wq.member(c)) {
4357 wq.push(n);
4358 }
4359 } else {
4360 set_idom(n, idom(n), dd);
4361 }
4362 }
4363
4364 // clone shared_stmt
4365 clone_loop_body(wq, old_new, nullptr);
4366
4367 Node* region_clone = old_new[region->_idx];
4368 region_clone->set_req(inner, C->top());
4369 set_idom(region, region->in(inner), dd);
4370
4371 // Prepare the outer loop
4372 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4373 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4374 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4375 set_idom(head, outer_head, dd);
4376
4377 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4378
4379 // Make one of the shared_stmt copies only reachable from stmt1, the
4380 // other only from stmt2..stmtn.
4381 Node* dom = nullptr;
4382 for (uint i = 1; i < region->req(); ++i) {
4383 if (i != inner) {
4384 _igvn.replace_input_of(region, i, C->top());
4385 }
4386 Node* in = region_clone->in(i);
4387 if (in->is_top()) {
4388 continue;
4389 }
4390 if (dom == nullptr) {
4391 dom = in;
4392 } else {
4393 dom = dom_lca(dom, in);
4394 }
4395 }
4396
4397 set_idom(region_clone, dom, dd);
4398
4399 // Set up the outer loop
4400 for (uint i = 0; i < head->outcnt(); i++) {
4401 Node* u = head->raw_out(i);
4402 if (u->is_Phi()) {
4403 Node* outer_phi = u->clone();
4404 outer_phi->set_req(0, outer_head);
4405 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4406 if (backedge == nullptr) {
4407 backedge = u->in(LoopNode::LoopBackControl);
4408 }
4409 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4410 register_new_node(outer_phi, outer_head);
4411 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4412 }
4413 }
4414
4415 // create control and data nodes for out of loop uses (including region2)
4416 Node_List worklist;
4417 uint new_counter = C->unique();
4418 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4419
4420 Node_List *split_if_set = nullptr;
4421 Node_List *split_bool_set = nullptr;
4422 Node_List *split_cex_set = nullptr;
4423 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4424 split_if_set, split_bool_set, split_cex_set);
4425
4426 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4427
4428 if (exit_test != nullptr) {
4429 float cnt = exit_test->_fcnt;
4430 if (cnt != COUNT_UNKNOWN) {
4431 exit_test->_fcnt = cnt * f;
4432 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4433 }
4434 }
4435
4436 C->set_major_progress();
4437
4438 return true;
4439 }
4440
4441 // AutoVectorize the loop: replace scalar ops with vector ops.
4442 PhaseIdealLoop::AutoVectorizeStatus
4443 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4444 // Counted loop only
4445 if (!lpt->is_counted()) {
4446 return AutoVectorizeStatus::Impossible;
4447 }
4448
4449 // Main-loop only
4450 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4451 if (!cl->is_main_loop()) {
4452 return AutoVectorizeStatus::Impossible;
4453 }
4454
4455 VLoop vloop(lpt, false);
4456 if (!vloop.check_preconditions()) {
4457 return AutoVectorizeStatus::TriedAndFailed;
4458 }
4459
4460 // Ensure the shared data is cleared before each use
4461 vshared.clear();
4462
4463 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4464 if (!vloop_analyzer.success()) {
4465 return AutoVectorizeStatus::TriedAndFailed;
4466 }
4467
4468 SuperWord sw(vloop_analyzer);
4469 if (!sw.transform_loop()) {
4470 return AutoVectorizeStatus::TriedAndFailed;
4471 }
4472
4473 return AutoVectorizeStatus::Success;
4474 }
4475
4476 // Returns true if the Reduction node is unordered.
4477 static bool is_unordered_reduction(Node* n) {
4478 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4479 }
4480
4481 // Having ReductionNodes in the loop is expensive. They need to recursively
4482 // fold together the vector values, for every vectorized loop iteration. If
4483 // we encounter the following pattern, we can vector accumulate the values
4484 // inside the loop, and only have a single UnorderedReduction after the loop.
4485 //
4486 // Note: UnorderedReduction represents a ReductionNode which does not require
4487 // calculating in strict order.
4488 //
4489 // CountedLoop init
4490 // | |
4491 // +------+ | +-----------------------+
4492 // | | | |
4493 // PhiNode (s) |
4494 // | |
4495 // | Vector |
4496 // | | |
4497 // UnorderedReduction (first_ur) |
4498 // | |
4499 // ... Vector |
4500 // | | |
4501 // UnorderedReduction (last_ur) |
4502 // | |
4503 // +---------------------+
4504 //
4505 // We patch the graph to look like this:
4506 //
4507 // CountedLoop identity_vector
4508 // | |
4509 // +-------+ | +---------------+
4510 // | | | |
4511 // PhiNode (v) |
4512 // | |
4513 // | Vector |
4514 // | | |
4515 // VectorAccumulator |
4516 // | |
4517 // ... Vector |
4518 // | | |
4519 // init VectorAccumulator |
4520 // | | | |
4521 // UnorderedReduction +-----------+
4522 //
4523 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4524 // use vector_accumulators, which do the same reductions, but only element
4525 // wise. This is a single operation per vector_accumulator, rather than many
4526 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4527 // after the loop, and also reduce the init value into it.
4528 //
4529 // We can not do this with all reductions. Some reductions do not allow the
4530 // reordering of operations (for example float addition/multiplication require
4531 // strict order).
4532 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4533 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4534
4535 // Find all Phi nodes with an unordered Reduction on backedge.
4536 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4537 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4538 Node* phi = cl->fast_out(j);
4539 // We have a phi with a single use, and an unordered Reduction on the backedge.
4540 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4541 continue;
4542 }
4543
4544 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4545 assert(!last_ur->requires_strict_order(), "must be");
4546
4547 // Determine types
4548 const TypeVect* vec_t = last_ur->vect_type();
4549 uint vector_length = vec_t->length();
4550 BasicType bt = vec_t->element_basic_type();
4551 const Type* bt_t = Type::get_const_basic_type(bt);
4552
4553 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4554 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4555 const int vopc = VectorNode::opcode(sopc, bt);
4556 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4557 DEBUG_ONLY( last_ur->dump(); )
4558 assert(false, "do not have normal vector op for this reduction");
4559 continue; // not implemented -> fails
4560 }
4561
4562 // Traverse up the chain of unordered Reductions, checking that it loops back to
4563 // the phi. Check that all unordered Reductions only have a single use, except for
4564 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4565 // are outside the loop.
4566 ReductionNode* current = last_ur;
4567 ReductionNode* first_ur = nullptr;
4568 while (true) {
4569 assert(!current->requires_strict_order(), "sanity");
4570
4571 // Expect no ctrl and a vector_input from within the loop.
4572 Node* ctrl = current->in(0);
4573 Node* vector_input = current->in(2);
4574 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4575 DEBUG_ONLY( current->dump(1); )
4576 assert(false, "reduction has ctrl or bad vector_input");
4577 break; // Chain traversal fails.
4578 }
4579
4580 assert(current->vect_type() != nullptr, "must have vector type");
4581 if (current->vect_type() != last_ur->vect_type()) {
4582 // Reductions do not have the same vector type (length and element type).
4583 break; // Chain traversal fails.
4584 }
4585
4586 // Expect single use of an unordered Reduction, except for last_ur.
4587 if (current == last_ur) {
4588 // Expect all uses to be outside the loop, except phi.
4589 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4590 Node* use = current->fast_out(k);
4591 if (use != phi && ctrl_or_self(use) == cl) {
4592 DEBUG_ONLY( current->dump(-1); )
4593 assert(false, "reduction has use inside loop");
4594 // Should not be allowed by SuperWord::mark_reductions
4595 return; // bail out of optimization
4596 }
4597 }
4598 } else {
4599 if (current->outcnt() != 1) {
4600 break; // Chain traversal fails.
4601 }
4602 }
4603
4604 // Expect another unordered Reduction or phi as the scalar input.
4605 Node* scalar_input = current->in(1);
4606 if (is_unordered_reduction(scalar_input) &&
4607 scalar_input->Opcode() == current->Opcode()) {
4608 // Move up the unordered Reduction chain.
4609 current = scalar_input->as_Reduction();
4610 assert(!current->requires_strict_order(), "must be");
4611 } else if (scalar_input == phi) {
4612 // Chain terminates at phi.
4613 first_ur = current;
4614 current = nullptr;
4615 break; // Success.
4616 } else {
4617 // scalar_input is neither phi nor a matching reduction
4618 // Can for example be scalar reduction when we have
4619 // partial vectorization.
4620 break; // Chain traversal fails.
4621 }
4622 }
4623 if (current != nullptr) {
4624 // Chain traversal was not successful.
4625 continue;
4626 }
4627 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4628
4629 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4630 set_ctrl(identity_scalar, C->root());
4631 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t);
4632 register_new_node(identity_vector, C->root());
4633 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4634 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4635
4636 // Turn the scalar phi into a vector phi.
4637 _igvn.rehash_node_delayed(phi);
4638 Node* init = phi->in(1); // Remember init before replacing it.
4639 phi->set_req_X(1, identity_vector, &_igvn);
4640 phi->as_Type()->set_type(vec_t);
4641 _igvn.set_type(phi, vec_t);
4642
4643 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4644 current = first_ur;
4645 while (true) {
4646 // Create vector_accumulator to replace current.
4647 Node* last_vector_accumulator = current->in(1);
4648 Node* vector_input = current->in(2);
4649 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4650 register_new_node(vector_accumulator, cl);
4651 _igvn.replace_node(current, vector_accumulator);
4652 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4653 if (current == last_ur) {
4654 break;
4655 }
4656 current = vector_accumulator->unique_out()->as_Reduction();
4657 assert(!current->requires_strict_order(), "must be");
4658 }
4659
4660 // Create post-loop reduction.
4661 Node* last_accumulator = phi->in(2);
4662 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4663
4664 // Take over uses of last_accumulator that are not in the loop.
4665 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4666 Node* use = last_accumulator->out(i);
4667 if (use != phi && use != post_loop_reduction) {
4668 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4669 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4670 --i;
4671 }
4672 }
4673 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4674 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4675
4676 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4677 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4678 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4679 }
4680 }
4681
4682 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4683 for (uint i = 0; i < _data_nodes.size(); i++) {
4684 clone(_data_nodes[i], new_ctrl);
4685 }
4686 }
4687
4688 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4689 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4690 Node* clone = node->clone();
4691 _phase->igvn().register_new_node_with_optimizer(clone);
4692 _orig_to_new.put(node, clone);
4693 _phase->set_ctrl(clone, new_ctrl);
4694 if (node->is_CastII()) {
4695 clone->set_req(0, new_ctrl);
4696 }
4697 }
4698
4699 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4700 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4701 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4702 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4703 for (uint i = 1; i < node->req(); i++) {
4704 Node** cloned_input = _orig_to_new.get(node->in(i));
4705 if (cloned_input != nullptr) {
4706 // Input was also cloned -> rewire clone to the cloned input.
4707 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4708 }
4709 }
4710 });
4711 }
4712
4713 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4714 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4715 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4716 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4717 Node* new_ctrl) {
4718 for (uint i = 0; i < _data_nodes.size(); i++) {
4719 Node* data_node = _data_nodes[i];
4720 if (data_node->is_Opaque1()) {
4721 transform_opaque_node(transform_strategy, data_node);
4722 } else {
4723 clone(data_node, new_ctrl);
4724 }
4725 }
4726 }
4727
4728 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4729 Node* transformed_node;
4730 if (node->is_OpaqueLoopInit()) {
4731 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4732 } else {
4733 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4734 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4735 }
4736 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4737 _orig_to_new.put(node, transformed_node);
4738 }