1 /*
2 * Copyright (c) 1999, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/divnode.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/matcher.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/movenode.hpp"
39 #include "opto/opaquenode.hpp"
40 #include "opto/rootnode.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/subtypenode.hpp"
43 #include "opto/superword.hpp"
44 #include "opto/vectornode.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
52 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
53 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
54 // so disable this for now
55 return nullptr;
56 }
57
58 // Splitting range check CastIIs through a loop induction Phi can
59 // cause new Phis to be created that are left unrelated to the loop
60 // induction Phi and prevent optimizations (vectorization)
61 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
62 n->in(1) == region->as_CountedLoop()->phi()) {
63 return nullptr;
64 }
65
66 if (cannot_split_division(n, region)) {
67 return nullptr;
68 }
69
70 int wins = 0;
71 assert(!n->is_CFG(), "");
72 assert(region->is_Region(), "");
73
74 const Type* type = n->bottom_type();
75 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
76 Node* phi;
77 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
78 int iid = t_oop->instance_id();
79 int index = C->get_alias_index(t_oop);
80 int offset = t_oop->offset();
81 phi = new PhiNode(region, type, nullptr, iid, index, offset);
82 } else {
83 phi = PhiNode::make_blank(region, n);
84 }
85 uint old_unique = C->unique();
86 for (uint i = 1; i < region->req(); i++) {
87 Node* x;
88 Node* the_clone = nullptr;
89 if (region->in(i) == C->top()) {
90 x = C->top(); // Dead path? Use a dead data op
91 } else {
92 x = n->clone(); // Else clone up the data op
93 the_clone = x; // Remember for possible deletion.
94 // Alter data node to use pre-phi inputs
95 if (n->in(0) == region)
96 x->set_req( 0, region->in(i) );
97 for (uint j = 1; j < n->req(); j++) {
98 Node* in = n->in(j);
99 if (in->is_Phi() && in->in(0) == region)
100 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
101 }
102 }
103 // Check for a 'win' on some paths
104 const Type* t = x->Value(&_igvn);
105
106 bool singleton = t->singleton();
107
108 // A TOP singleton indicates that there are no possible values incoming
109 // along a particular edge. In most cases, this is OK, and the Phi will
110 // be eliminated later in an Ideal call. However, we can't allow this to
111 // happen if the singleton occurs on loop entry, as the elimination of
112 // the PhiNode may cause the resulting node to migrate back to a previous
113 // loop iteration.
114 if (singleton && t == Type::TOP) {
115 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
116 // irreducible loop may not be indicated by an affirmative is_Loop());
117 // therefore, the only top we can split thru a phi is on a backedge of
118 // a loop.
119 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
120 }
121
122 if (singleton) {
123 wins++;
124 x = makecon(t);
125 } else {
126 // We now call Identity to try to simplify the cloned node.
127 // Note that some Identity methods call phase->type(this).
128 // Make sure that the type array is big enough for
129 // our new node, even though we may throw the node away.
130 // (Note: This tweaking with igvn only works because x is a new node.)
131 _igvn.set_type(x, t);
132 // If x is a TypeNode, capture any more-precise type permanently into Node
133 // otherwise it will be not updated during igvn->transform since
134 // igvn->type(x) is set to x->Value() already.
135 x->raise_bottom_type(t);
136 Node* y = x->Identity(&_igvn);
137 if (y != x) {
138 wins++;
139 x = y;
140 } else {
141 y = _igvn.hash_find(x);
142 if (y == nullptr) {
143 y = similar_subtype_check(x, region->in(i));
144 }
145 if (y) {
146 wins++;
147 x = y;
148 } else {
149 // Else x is a new node we are keeping
150 // We do not need register_new_node_with_optimizer
151 // because set_type has already been called.
152 _igvn._worklist.push(x);
153 }
154 }
155 }
156
157 phi->set_req( i, x );
158
159 if (the_clone == nullptr) {
160 continue;
161 }
162
163 if (the_clone != x) {
164 _igvn.remove_dead_node(the_clone);
165 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
166 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
167 // it is not a win if 'x' moved from an outer to an inner loop
168 // this edge case can only happen for Load nodes
169 wins = 0;
170 break;
171 }
172 }
173 // Too few wins?
174 if (wins <= policy) {
175 _igvn.remove_dead_node(phi);
176 return nullptr;
177 }
178
179 // Record Phi
180 register_new_node( phi, region );
181
182 for (uint i2 = 1; i2 < phi->req(); i2++) {
183 Node *x = phi->in(i2);
184 // If we commoned up the cloned 'x' with another existing Node,
185 // the existing Node picks up a new use. We need to make the
186 // existing Node occur higher up so it dominates its uses.
187 Node *old_ctrl;
188 IdealLoopTree *old_loop;
189
190 if (x->is_Con()) {
191 assert(get_ctrl(x) == C->root(), "constant control is not root");
192 continue;
193 }
194 // The occasional new node
195 if (x->_idx >= old_unique) { // Found a new, unplaced node?
196 old_ctrl = nullptr;
197 old_loop = nullptr; // Not in any prior loop
198 } else {
199 old_ctrl = get_ctrl(x);
200 old_loop = get_loop(old_ctrl); // Get prior loop
201 }
202 // New late point must dominate new use
203 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
204 if (new_ctrl == old_ctrl) // Nothing is changed
205 continue;
206
207 IdealLoopTree *new_loop = get_loop(new_ctrl);
208
209 // Don't move x into a loop if its uses are
210 // outside of loop. Otherwise x will be cloned
211 // for each use outside of this loop.
212 IdealLoopTree *use_loop = get_loop(region);
213 if (!new_loop->is_member(use_loop) &&
214 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
215 // Take early control, later control will be recalculated
216 // during next iteration of loop optimizations.
217 new_ctrl = get_early_ctrl(x);
218 new_loop = get_loop(new_ctrl);
219 }
220 // Set new location
221 set_ctrl(x, new_ctrl);
222 // If changing loop bodies, see if we need to collect into new body
223 if (old_loop != new_loop) {
224 if (old_loop && !old_loop->_child)
225 old_loop->_body.yank(x);
226 if (!new_loop->_child)
227 new_loop->_body.push(x); // Collect body info
228 }
229 }
230
231 return phi;
232 }
233
234 // Test whether node 'x' can move into an inner loop relative to node 'n'.
235 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
236 // BUT it can also return true and 'x' is in the outer loop
237 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
238 IdealLoopTree* n_loop_tree = get_loop(n_loop);
239 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
240 // x_loop_tree should be outer or same loop as n_loop_tree
241 return !x_loop_tree->is_member(n_loop_tree);
242 }
243
244 // Subtype checks that carry profile data don't common so look for a replacement by following edges
245 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
246 if (x->is_SubTypeCheck()) {
247 Node* in1 = x->in(1);
248 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
249 Node* u = in1->fast_out(i);
250 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
251 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
252 Node* bol = u->fast_out(j);
253 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
254 Node* iff = bol->fast_out(k);
255 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
256 // unrelated profile
257 if (iff->is_If() && is_dominator(iff, r_in)) {
258 return u;
259 }
260 }
261 }
262 }
263 }
264 }
265 return nullptr;
266 }
267
268 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
269 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
270 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
271 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
272 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
273 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
274 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
275 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
276 const Type* zero;
277 switch (n->Opcode()) {
278 case Op_DivI:
279 case Op_ModI:
280 case Op_UDivI:
281 case Op_UModI:
282 zero = TypeInt::ZERO;
283 break;
284 case Op_DivL:
285 case Op_ModL:
286 case Op_UDivL:
287 case Op_UModL:
288 zero = TypeLong::ZERO;
289 break;
290 default:
291 return false;
292 }
293
294 if (n->in(0) != nullptr) {
295 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
296 return true;
297 }
298
299 Node* divisor = n->in(2);
300 return is_divisor_loop_phi(divisor, region) &&
301 loop_phi_backedge_type_contains_zero(divisor, zero);
302 }
303
304 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
305 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
306 }
307
308 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
309 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
310 }
311
312 //------------------------------dominated_by------------------------------------
313 // Replace the dominated test with an obvious true or false. Place it on the
314 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
315 // live path up to the dominating control.
316 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
317 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
318
319 // prevdom is the dominating projection of the dominating test.
320 assert(iff->Opcode() == Op_If ||
321 iff->Opcode() == Op_CountedLoopEnd ||
322 iff->Opcode() == Op_LongCountedLoopEnd ||
323 iff->Opcode() == Op_RangeCheck ||
324 iff->Opcode() == Op_ParsePredicate,
325 "Check this code when new subtype is added");
326
327 int pop = prevdom->Opcode();
328 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
329 if (flip) {
330 if (pop == Op_IfTrue)
331 pop = Op_IfFalse;
332 else
333 pop = Op_IfTrue;
334 }
335 // 'con' is set to true or false to kill the dominated test.
336 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
337 // Hack the dominated test
338 _igvn.replace_input_of(iff, 1, con);
339
340 // If I don't have a reachable TRUE and FALSE path following the IfNode then
341 // I can assume this path reaches an infinite loop. In this case it's not
342 // important to optimize the data Nodes - either the whole compilation will
343 // be tossed or this path (and all data Nodes) will go dead.
344 if (iff->outcnt() != 2) {
345 return;
346 }
347
348 // Make control-dependent data Nodes on the live path (path that will remain
349 // once the dominated IF is removed) become control-dependent on the
350 // dominating projection.
351 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
352
353 if (dp == nullptr) {
354 return;
355 }
356
357 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
358 }
359
360 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
361 IdealLoopTree* old_loop = get_loop(source);
362
363 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
364 Node* out = source->fast_out(i); // Control-dependent node
365 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
366 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
367 assert(out->in(0) == source, "must be control dependent on source");
368 _igvn.replace_input_of(out, 0, dominator);
369 if (pin_array_access_nodes) {
370 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
371 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
372 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
373 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
374 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
375 // dominating check.
376 Node* clone = out->pin_array_access_node();
377 if (clone != nullptr) {
378 clone = _igvn.register_new_node_with_optimizer(clone, out);
379 _igvn.replace_node(out, clone);
380 out = clone;
381 }
382 }
383 set_early_ctrl(out, false);
384 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
385 if (old_loop != new_loop) {
386 if (!old_loop->_child) {
387 old_loop->_body.yank(out);
388 }
389 if (!new_loop->_child) {
390 new_loop->_body.push(out);
391 }
392 }
393 --i;
394 --imax;
395 }
396 }
397 }
398
399 //------------------------------has_local_phi_input----------------------------
400 // Return TRUE if 'n' has Phi inputs from its local block and no other
401 // block-local inputs (all non-local-phi inputs come from earlier blocks)
402 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
403 Node *n_ctrl = get_ctrl(n);
404 // See if some inputs come from a Phi in this block, or from before
405 // this block.
406 uint i;
407 for( i = 1; i < n->req(); i++ ) {
408 Node *phi = n->in(i);
409 if( phi->is_Phi() && phi->in(0) == n_ctrl )
410 break;
411 }
412 if( i >= n->req() )
413 return nullptr; // No Phi inputs; nowhere to clone thru
414
415 // Check for inputs created between 'n' and the Phi input. These
416 // must split as well; they have already been given the chance
417 // (courtesy of a post-order visit) and since they did not we must
418 // recover the 'cost' of splitting them by being very profitable
419 // when splitting 'n'. Since this is unlikely we simply give up.
420 for( i = 1; i < n->req(); i++ ) {
421 Node *m = n->in(i);
422 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
423 // We allow the special case of AddP's with no local inputs.
424 // This allows us to split-up address expressions.
425 if (m->is_AddP() &&
426 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
427 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
428 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
429 // Move the AddP up to the dominating point. That's fine because control of m's inputs
430 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
431 Node* c = find_non_split_ctrl(idom(n_ctrl));
432 if (c->is_OuterStripMinedLoop()) {
433 c->as_Loop()->verify_strip_mined(1);
434 c = c->in(LoopNode::EntryControl);
435 }
436 set_ctrl_and_loop(m, c);
437 continue;
438 }
439 return nullptr;
440 }
441 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
442 }
443
444 return n_ctrl;
445 }
446
447 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
448 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
449 assert(bt == T_INT || bt == T_LONG, "only for integers");
450 int n_op = n->Opcode();
451
452 if (n_op == Op_LShift(bt)) {
453 // Scale is loop invariant
454 Node* scale = n->in(2);
455 Node* scale_ctrl = get_ctrl(scale);
456 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
457 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
458 return nullptr;
459 }
460 const TypeInt* scale_t = scale->bottom_type()->isa_int();
461 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
462 return nullptr; // Dont bother with byte/short masking
463 }
464 // Add must vary with loop (else shift would be loop-invariant)
465 Node* add = n->in(1);
466 Node* add_ctrl = get_ctrl(add);
467 IdealLoopTree* add_loop = get_loop(add_ctrl);
468 if (n_loop != add_loop) {
469 return nullptr; // happens w/ evil ZKM loops
470 }
471
472 // Convert I-V into I+ (0-V); same for V-I
473 if (add->Opcode() == Op_Sub(bt) &&
474 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
475 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
476 Node* zero = integercon(0, bt);
477 Node* neg = SubNode::make(zero, add->in(2), bt);
478 register_new_node_with_ctrl_of(neg, add->in(2));
479 add = AddNode::make(add->in(1), neg, bt);
480 register_new_node(add, add_ctrl);
481 }
482 if (add->Opcode() != Op_Add(bt)) return nullptr;
483 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
484 // See if one add input is loop invariant
485 Node* add_var = add->in(1);
486 Node* add_var_ctrl = get_ctrl(add_var);
487 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
488 Node* add_invar = add->in(2);
489 Node* add_invar_ctrl = get_ctrl(add_invar);
490 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
491 if (add_invar_loop == n_loop) {
492 // Swap to find the invariant part
493 add_invar = add_var;
494 add_invar_ctrl = add_var_ctrl;
495 add_invar_loop = add_var_loop;
496 add_var = add->in(2);
497 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
498 return nullptr;
499 }
500 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
501 return nullptr; // No invariant part of the add?
502 }
503
504 // Yes! Reshape address expression!
505 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
506 Node* inv_scale_ctrl =
507 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
508 add_invar_ctrl : scale_ctrl;
509 register_new_node(inv_scale, inv_scale_ctrl);
510 Node* var_scale = LShiftNode::make(add_var, scale, bt);
511 register_new_node(var_scale, n_ctrl);
512 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
513 register_new_node(var_add, n_ctrl);
514 _igvn.replace_node(n, var_add);
515 return var_add;
516 }
517 return nullptr;
518 }
519
520 //------------------------------remix_address_expressions----------------------
521 // Rework addressing expressions to get the most loop-invariant stuff
522 // moved out. We'd like to do all associative operators, but it's especially
523 // important (common) to do address expressions.
524 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
525 if (!has_ctrl(n)) return nullptr;
526 Node* n_ctrl = get_ctrl(n);
527 IdealLoopTree* n_loop = get_loop(n_ctrl);
528
529 // See if 'n' mixes loop-varying and loop-invariant inputs and
530 // itself is loop-varying.
531
532 // Only interested in binary ops (and AddP)
533 if (n->req() < 3 || n->req() > 4) return nullptr;
534
535 Node* n1_ctrl = get_ctrl(n->in( 1));
536 Node* n2_ctrl = get_ctrl(n->in( 2));
537 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
538 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
539 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
540 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
541
542 // Does one of my inputs spin in a tighter loop than self?
543 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
544 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
545 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
546 return nullptr; // Leave well enough alone
547 }
548
549 // Is at least one of my inputs loop-invariant?
550 if (n1_loop == n_loop &&
551 n2_loop == n_loop &&
552 n3_loop == n_loop) {
553 return nullptr; // No loop-invariant inputs
554 }
555
556 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
557 if (res != nullptr) {
558 return res;
559 }
560 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
561 if (res != nullptr) {
562 return res;
563 }
564
565 int n_op = n->Opcode();
566 // Replace (I+V) with (V+I)
567 if (n_op == Op_AddI ||
568 n_op == Op_AddL ||
569 n_op == Op_AddF ||
570 n_op == Op_AddD ||
571 n_op == Op_MulI ||
572 n_op == Op_MulL ||
573 n_op == Op_MulF ||
574 n_op == Op_MulD) {
575 if (n2_loop == n_loop) {
576 assert(n1_loop != n_loop, "");
577 n->swap_edges(1, 2);
578 }
579 }
580
581 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
582 // but not if I2 is a constant. Skip for irreducible loops.
583 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
584 if (n2_loop == n_loop && n3_loop != n_loop) {
585 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
586 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
587 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
588 IdealLoopTree* n22loop = get_loop(n22_ctrl);
589 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
590 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
591 n23_loop == n_loop) {
592 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
593 // Stuff new AddP in the loop preheader
594 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
595 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
596 register_new_node(add2, n_ctrl);
597 _igvn.replace_node(n, add2);
598 return add2;
599 }
600 }
601 }
602
603 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
604 if (n2_loop != n_loop && n3_loop == n_loop) {
605 if (n->in(3)->Opcode() == Op_AddX) {
606 Node* V = n->in(3)->in(1);
607 Node* I = n->in(3)->in(2);
608 if (is_member(n_loop,get_ctrl(V))) {
609 } else {
610 Node *tmp = V; V = I; I = tmp;
611 }
612 if (!is_member(n_loop,get_ctrl(I))) {
613 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
614 // Stuff new AddP in the loop preheader
615 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
616 Node* add2 = new AddPNode(n->in(1), add1, V);
617 register_new_node(add2, n_ctrl);
618 _igvn.replace_node(n, add2);
619 return add2;
620 }
621 }
622 }
623 }
624
625 return nullptr;
626 }
627
628 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
629 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
630 assert(n->Opcode() == Op_AddI, "sanity");
631 Node * nn = nullptr;
632 Node * in1 = n->in(1);
633 Node * in2 = n->in(2);
634 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
635 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
636 if (loop_n->is_counted() &&
637 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
638 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
639 Matcher::match_rule_supported(Op_MulAddS2I)) {
640 Node* mul_in1 = in1->in(1);
641 Node* mul_in2 = in1->in(2);
642 Node* mul_in3 = in2->in(1);
643 Node* mul_in4 = in2->in(2);
644 if (mul_in1->Opcode() == Op_LoadS &&
645 mul_in2->Opcode() == Op_LoadS &&
646 mul_in3->Opcode() == Op_LoadS &&
647 mul_in4->Opcode() == Op_LoadS) {
648 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
649 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
650 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
651 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
652 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
653 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
654 // All nodes should be in the same counted loop.
655 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
656 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
657 Node* adr1 = mul_in1->in(MemNode::Address);
658 Node* adr2 = mul_in2->in(MemNode::Address);
659 Node* adr3 = mul_in3->in(MemNode::Address);
660 Node* adr4 = mul_in4->in(MemNode::Address);
661 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
662 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
663 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
664 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
665 register_new_node_with_ctrl_of(nn, n);
666 _igvn.replace_node(n, nn);
667 return nn;
668 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
669 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
670 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
671 register_new_node_with_ctrl_of(nn, n);
672 _igvn.replace_node(n, nn);
673 return nn;
674 }
675 }
676 }
677 }
678 }
679 }
680 return nn;
681 }
682
683 //------------------------------conditional_move-------------------------------
684 // Attempt to replace a Phi with a conditional move. We have some pretty
685 // strict profitability requirements. All Phis at the merge point must
686 // be converted, so we can remove the control flow. We need to limit the
687 // number of c-moves to a small handful. All code that was in the side-arms
688 // of the CFG diamond is now speculatively executed. This code has to be
689 // "cheap enough". We are pretty much limited to CFG diamonds that merge
690 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
691 Node *PhaseIdealLoop::conditional_move( Node *region ) {
692
693 assert(region->is_Region(), "sanity check");
694 if (region->req() != 3) return nullptr;
695
696 // Check for CFG diamond
697 Node *lp = region->in(1);
698 Node *rp = region->in(2);
699 if (!lp || !rp) return nullptr;
700 Node *lp_c = lp->in(0);
701 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
702 IfNode *iff = lp_c->as_If();
703
704 // Check for ops pinned in an arm of the diamond.
705 // Can't remove the control flow in this case
706 if (lp->outcnt() > 1) return nullptr;
707 if (rp->outcnt() > 1) return nullptr;
708
709 IdealLoopTree* r_loop = get_loop(region);
710 assert(r_loop == get_loop(iff), "sanity");
711 // Always convert to CMOVE if all results are used only outside this loop.
712 bool used_inside_loop = (r_loop == _ltree_root);
713
714 // Check profitability
715 int cost = 0;
716 int phis = 0;
717 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
718 Node *out = region->fast_out(i);
719 if (!out->is_Phi()) continue; // Ignore other control edges, etc
720 phis++;
721 PhiNode* phi = out->as_Phi();
722 BasicType bt = phi->type()->basic_type();
723 switch (bt) {
724 case T_DOUBLE:
725 case T_FLOAT:
726 if (C->use_cmove()) {
727 continue; //TODO: maybe we want to add some cost
728 }
729 cost += Matcher::float_cmove_cost(); // Could be very expensive
730 break;
731 case T_LONG: {
732 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
733 }
734 case T_INT: // These all CMOV fine
735 case T_ADDRESS: { // (RawPtr)
736 cost++;
737 break;
738 }
739 case T_NARROWOOP: // Fall through
740 case T_OBJECT: { // Base oops are OK, but not derived oops
741 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
742 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
743 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
744 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
745 // have a Phi for the base here that we convert to a CMOVE all is well
746 // and good. But if the base is dead, we'll not make a CMOVE. Later
747 // the allocator will have to produce a base by creating a CMOVE of the
748 // relevant bases. This puts the allocator in the business of
749 // manufacturing expensive instructions, generally a bad plan.
750 // Just Say No to Conditionally-Moved Derived Pointers.
751 if (tp && tp->offset() != 0)
752 return nullptr;
753 cost++;
754 break;
755 }
756 default:
757 return nullptr; // In particular, can't do memory or I/O
758 }
759 // Add in cost any speculative ops
760 for (uint j = 1; j < region->req(); j++) {
761 Node *proj = region->in(j);
762 Node *inp = phi->in(j);
763 if (get_ctrl(inp) == proj) { // Found local op
764 cost++;
765 // Check for a chain of dependent ops; these will all become
766 // speculative in a CMOV.
767 for (uint k = 1; k < inp->req(); k++)
768 if (get_ctrl(inp->in(k)) == proj)
769 cost += ConditionalMoveLimit; // Too much speculative goo
770 }
771 }
772 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
773 // This will likely Split-If, a higher-payoff operation.
774 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
775 Node* use = phi->fast_out(k);
776 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
777 cost += ConditionalMoveLimit;
778 // Is there a use inside the loop?
779 // Note: check only basic types since CMoveP is pinned.
780 if (!used_inside_loop && is_java_primitive(bt)) {
781 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
782 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
783 used_inside_loop = true;
784 }
785 }
786 }
787 }//for
788 Node* bol = iff->in(1);
789 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
790 if (bol->is_OpaqueTemplateAssertionPredicate()) {
791 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
792 return nullptr;
793 }
794 if (bol->is_OpaqueMultiversioning()) {
795 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
796 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
797 // and will make the multiversion_if constant fold in the next IGVN round.
798 return nullptr;
799 }
800 if (!bol->is_Bool()) {
801 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
802 return nullptr;
803 }
804 int cmp_op = bol->in(1)->Opcode();
805 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
806 return nullptr;
807 }
808 // It is expensive to generate flags from a float compare.
809 // Avoid duplicated float compare.
810 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
811
812 float infrequent_prob = PROB_UNLIKELY_MAG(3);
813 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
814 if (used_inside_loop) {
815 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
816
817 // BlockLayoutByFrequency optimization moves infrequent branch
818 // from hot path. No point in CMOV'ing in such case (110 is used
819 // instead of 100 to take into account not exactness of float value).
820 if (BlockLayoutByFrequency) {
821 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
822 }
823 }
824 // Check for highly predictable branch. No point in CMOV'ing if
825 // we are going to predict accurately all the time.
826 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
827 //keep going
828 } else if (iff->_prob < infrequent_prob ||
829 iff->_prob > (1.0f - infrequent_prob))
830 return nullptr;
831
832 // --------------
833 // Now replace all Phis with CMOV's
834 Node *cmov_ctrl = iff->in(0);
835 uint flip = (lp->Opcode() == Op_IfTrue);
836 Node_List wq;
837 while (1) {
838 PhiNode* phi = nullptr;
839 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
840 Node *out = region->fast_out(i);
841 if (out->is_Phi()) {
842 phi = out->as_Phi();
843 break;
844 }
845 }
846 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
847 break;
848 }
849 // Move speculative ops
850 wq.push(phi);
851 while (wq.size() > 0) {
852 Node *n = wq.pop();
853 for (uint j = 1; j < n->req(); j++) {
854 Node* m = n->in(j);
855 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
856 set_ctrl(m, cmov_ctrl);
857 wq.push(m);
858 }
859 }
860 }
861 Node* cmov = CMoveNode::make(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 (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 if (C->failing()) {
1554 return;
1555 }
1556
1557 try_move_store_after_loop(n);
1558 }
1559
1560 // Transform:
1561 //
1562 // if (some_condition) {
1563 // // body 1
1564 // } else {
1565 // // body 2
1566 // }
1567 // if (some_condition) {
1568 // // body 3
1569 // } else {
1570 // // body 4
1571 // }
1572 //
1573 // into:
1574 //
1575 //
1576 // if (some_condition) {
1577 // // body 1
1578 // // body 3
1579 // } else {
1580 // // body 2
1581 // // body 4
1582 // }
1583 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1584 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1585 Node *n_ctrl = n->in(0);
1586 IfNode* dom_if = idom(n_ctrl)->as_If();
1587 if (n->in(1) != dom_if->in(1)) {
1588 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1589 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1590 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1591 _igvn.replace_input_of(n, 1, dom_if->in(1));
1592 }
1593 ProjNode* dom_proj_true = dom_if->proj_out(1);
1594 ProjNode* dom_proj_false = dom_if->proj_out(0);
1595
1596 // Now split the IF
1597 RegionNode* new_false_region;
1598 RegionNode* new_true_region;
1599 do_split_if(n, &new_false_region, &new_true_region);
1600 assert(new_false_region->req() == new_true_region->req(), "");
1601 #ifdef ASSERT
1602 for (uint i = 1; i < new_false_region->req(); ++i) {
1603 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1604 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");
1605 }
1606 #endif
1607 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1608
1609 // We now have:
1610 // if (some_condition) {
1611 // // body 1
1612 // if (some_condition) {
1613 // body3: // new_true_region
1614 // // body3
1615 // } else {
1616 // goto body4;
1617 // }
1618 // } else {
1619 // // body 2
1620 // if (some_condition) {
1621 // goto body3;
1622 // } else {
1623 // body4: // new_false_region
1624 // // body4;
1625 // }
1626 // }
1627 //
1628
1629 // clone pinned nodes thru the resulting regions
1630 push_pinned_nodes_thru_region(dom_if, new_true_region);
1631 push_pinned_nodes_thru_region(dom_if, new_false_region);
1632
1633 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1634 // 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
1635 // unrelated control dependency.
1636 for (uint i = 1; i < new_false_region->req(); i++) {
1637 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1638 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1639 } else {
1640 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1641 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1642 }
1643 }
1644 return true;
1645 }
1646 return false;
1647 }
1648
1649 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1650 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1651 Node* u = region->out(i);
1652 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1653 continue;
1654 }
1655 assert(u->in(0) == region, "not a control dependent node?");
1656 uint j = 1;
1657 for (; j < u->req(); ++j) {
1658 Node* in = u->in(j);
1659 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1660 break;
1661 }
1662 }
1663 if (j == u->req()) {
1664 Node *phi = PhiNode::make_blank(region, u);
1665 for (uint k = 1; k < region->req(); ++k) {
1666 Node* clone = u->clone();
1667 clone->set_req(0, region->in(k));
1668 register_new_node(clone, region->in(k));
1669 phi->init_req(k, clone);
1670 }
1671 register_new_node(phi, region);
1672 _igvn.replace_node(u, phi);
1673 --i;
1674 }
1675 }
1676 }
1677
1678 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1679 if (!dom->is_CountedLoopEnd()) {
1680 return true;
1681 }
1682 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1683 CountedLoopNode* cl = le->loopnode();
1684 if (cl == nullptr) {
1685 return true;
1686 }
1687 if (!cl->is_main_loop()) {
1688 return true;
1689 }
1690 if (cl->is_canonical_loop_entry() == nullptr) {
1691 return true;
1692 }
1693 // Further unrolling is possible so loop exit condition might change
1694 return false;
1695 }
1696
1697 // See if a shared loop-varying computation has no loop-varying uses.
1698 // Happens if something is only used for JVM state in uncommon trap exits,
1699 // like various versions of induction variable+offset. Clone the
1700 // computation per usage to allow it to sink out of the loop.
1701 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1702 if (has_ctrl(n) &&
1703 !n->is_Phi() &&
1704 !n->is_Bool() &&
1705 !n->is_Proj() &&
1706 !n->is_MergeMem() &&
1707 !n->is_CMove() &&
1708 !n->is_OpaqueNotNull() &&
1709 !n->is_OpaqueInitializedAssertionPredicate() &&
1710 !n->is_OpaqueTemplateAssertionPredicate() &&
1711 !n->is_Type()) {
1712 Node *n_ctrl = get_ctrl(n);
1713 IdealLoopTree *n_loop = get_loop(n_ctrl);
1714
1715 if (n->in(0) != nullptr) {
1716 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1717 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1718 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1719 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1720 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1721 Node* maybe_pinned_n = n;
1722 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1723 if (n->depends_only_on_test()) {
1724 Node* pinned_clone = n->pin_array_access_node();
1725 if (pinned_clone != nullptr) {
1726 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1727 // range check for that access. If that condition is replaced by an identical dominating one, then an
1728 // unpinned load would risk floating above its range check.
1729 register_new_node(pinned_clone, n_ctrl);
1730 maybe_pinned_n = pinned_clone;
1731 _igvn.replace_node(n, pinned_clone);
1732 }
1733 }
1734 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1735 }
1736 }
1737 if (n_loop != _ltree_root && n->outcnt() > 1) {
1738 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1739 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1740 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1741 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1742 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1743 if (n->is_Store() || n->is_LoadStore()) {
1744 assert(false, "no node with a side effect");
1745 C->record_failure("no node with a side effect");
1746 return;
1747 }
1748 Node* outer_loop_clone = nullptr;
1749 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1750 Node* u = n->last_out(j); // Clone private computation per use
1751 _igvn.rehash_node_delayed(u);
1752 Node* x = nullptr;
1753 if (n->depends_only_on_test()) {
1754 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1755 // range check for that access. If that condition is replaced by an identical dominating one, then an
1756 // unpinned load would risk floating above its range check.
1757 x = n->pin_array_access_node();
1758 }
1759 if (x == nullptr) {
1760 x = n->clone();
1761 }
1762 Node* x_ctrl = nullptr;
1763 if (u->is_Phi()) {
1764 // Replace all uses of normal nodes. Replace Phi uses
1765 // individually, so the separate Nodes can sink down
1766 // different paths.
1767 uint k = 1;
1768 while (u->in(k) != n) k++;
1769 u->set_req(k, x);
1770 // x goes next to Phi input path
1771 x_ctrl = u->in(0)->in(k);
1772 // Find control for 'x' next to use but not inside inner loops.
1773 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1774 --j;
1775 } else { // Normal use
1776 if (has_ctrl(u)) {
1777 x_ctrl = get_ctrl(u);
1778 } else {
1779 x_ctrl = u->in(0);
1780 }
1781 // Find control for 'x' next to use but not inside inner loops.
1782 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1783 // Replace all uses
1784 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1785 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1786 // anymore now that we're going to pin n as well
1787 _igvn.replace_node(u, x);
1788 --j;
1789 } else {
1790 int nb = u->replace_edge(n, x, &_igvn);
1791 j -= nb;
1792 }
1793 }
1794
1795 if (n->is_Load()) {
1796 // For loads, add a control edge to a CFG node outside of the loop
1797 // to force them to not combine and return back inside the loop
1798 // during GVN optimization (4641526).
1799 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1800
1801 IdealLoopTree* x_loop = get_loop(x_ctrl);
1802 Node* x_head = x_loop->_head;
1803 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1804 // Do not add duplicate LoadNodes to the outer strip mined loop
1805 if (outer_loop_clone != nullptr) {
1806 _igvn.replace_node(x, outer_loop_clone);
1807 continue;
1808 }
1809 outer_loop_clone = x;
1810 }
1811 x->set_req(0, x_ctrl);
1812 } else if (n->in(0) != nullptr){
1813 x->set_req(0, x_ctrl);
1814 }
1815 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1816 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1817 register_new_node(x, x_ctrl);
1818
1819 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1820 // All AddP nodes must keep the same base after sinking so:
1821 // 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,
1822 // their bases remain the same.
1823 // (see 2- below)
1824 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1825 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1826 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1827 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1828 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1829 assert(!x->is_Load(), "load should be pinned");
1830 // Use a cast node to pin clone out of loop
1831 Node* cast = nullptr;
1832 for (uint k = 0; k < x->req(); k++) {
1833 Node* in = x->in(k);
1834 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1835 const Type* in_t = _igvn.type(in);
1836 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1837 ConstraintCastNode::UnconditionalDependency, nullptr);
1838 }
1839 if (cast != nullptr) {
1840 Node* prev = _igvn.hash_find_insert(cast);
1841 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1842 cast->destruct(&_igvn);
1843 cast = prev;
1844 } else {
1845 register_new_node(cast, x_ctrl);
1846 }
1847 x->replace_edge(in, cast);
1848 // Chain of AddP nodes:
1849 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1850 if (x->is_AddP() && k == AddPNode::Base) {
1851 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1852 }
1853 break;
1854 }
1855 }
1856 assert(cast != nullptr, "must have added a cast to pin the node");
1857 }
1858 }
1859 _igvn.remove_dead_node(n);
1860 }
1861 _dom_lca_tags_round = 0;
1862 }
1863 }
1864 }
1865
1866 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1867 ResourceMark rm;
1868 Node_List wq;
1869 wq.push(x);
1870 while (wq.size() != 0) {
1871 Node* n = wq.pop();
1872 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1873 Node* u = n->fast_out(i);
1874 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1875 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1876 wq.push(u);
1877 }
1878 }
1879 }
1880 }
1881
1882 // Compute the early control of a node by following its inputs until we reach
1883 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1884 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1885 Node* early_ctrl = nullptr;
1886 ResourceMark rm;
1887 Unique_Node_List wq;
1888 wq.push(n);
1889 for (uint i = 0; i < wq.size(); i++) {
1890 Node* m = wq.at(i);
1891 Node* c = nullptr;
1892 if (m->is_CFG()) {
1893 c = m;
1894 } else if (m->pinned()) {
1895 c = m->in(0);
1896 } else {
1897 for (uint j = 0; j < m->req(); j++) {
1898 Node* in = m->in(j);
1899 if (in != nullptr) {
1900 wq.push(in);
1901 }
1902 }
1903 }
1904 if (c != nullptr) {
1905 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1906 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1907 early_ctrl = c;
1908 }
1909 }
1910 }
1911 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1912 return early_ctrl;
1913 }
1914
1915 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1916 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1917 Node* u = n->fast_out(i);
1918 if (u->is_Opaque1()) {
1919 return false; // Found loop limit, bugfix for 4677003
1920 }
1921 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1922 // get_late_ctrl_with_anti_dep() use their own tag
1923 _dom_lca_tags_round++;
1924 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1925
1926 if (u->is_Phi()) {
1927 for (uint j = 1; j < u->req(); ++j) {
1928 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1929 return false;
1930 }
1931 }
1932 } else {
1933 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1934 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1935 return false;
1936 }
1937 }
1938 }
1939 return true;
1940 }
1941
1942 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1943 if (n->is_Load()) {
1944 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1945 }
1946 IdealLoopTree *u_loop = get_loop(ctrl);
1947 if (u_loop == n_loop) {
1948 return false; // Found loop-varying use
1949 }
1950 if (n_loop->is_member(u_loop)) {
1951 return false; // Found use in inner loop
1952 }
1953 // 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
1954 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1955 // test of the pre loop above the point in the graph where it's pinned.
1956 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1957 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1958 if (is_dominator(pre_loop->loopexit(), ctrl)) {
1959 return false;
1960 }
1961 }
1962 return true;
1963 }
1964
1965 //------------------------------split_if_with_blocks---------------------------
1966 // Check for aggressive application of 'split-if' optimization,
1967 // using basic block level info.
1968 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1969 Node* root = C->root();
1970 visited.set(root->_idx); // first, mark root as visited
1971 // Do pre-visit work for root
1972 Node* n = split_if_with_blocks_pre(root);
1973 uint cnt = n->outcnt();
1974 uint i = 0;
1975
1976 while (true) {
1977 // Visit all children
1978 if (i < cnt) {
1979 Node* use = n->raw_out(i);
1980 ++i;
1981 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1982 // Now do pre-visit work for this use
1983 use = split_if_with_blocks_pre(use);
1984 nstack.push(n, i); // Save parent and next use's index.
1985 n = use; // Process all children of current use.
1986 cnt = use->outcnt();
1987 i = 0;
1988 }
1989 }
1990 else {
1991 // All of n's children have been processed, complete post-processing.
1992 if (cnt != 0 && !n->is_Con()) {
1993 assert(has_node(n), "no dead nodes");
1994 split_if_with_blocks_post(n);
1995 if (C->failing()) {
1996 return;
1997 }
1998 }
1999 if (must_throttle_split_if()) {
2000 nstack.clear();
2001 }
2002 if (nstack.is_empty()) {
2003 // Finished all nodes on stack.
2004 break;
2005 }
2006 // Get saved parent node and next use's index. Visit the rest of uses.
2007 n = nstack.node();
2008 cnt = n->outcnt();
2009 i = nstack.index();
2010 nstack.pop();
2011 }
2012 }
2013 }
2014
2015
2016 //=============================================================================
2017 //
2018 // C L O N E A L O O P B O D Y
2019 //
2020
2021 //------------------------------clone_iff--------------------------------------
2022 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2023 // "Nearly" because all Nodes have been cloned from the original in the loop,
2024 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2025 // through the Phi recursively, and return a Bool.
2026 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2027
2028 // Convert this Phi into a Phi merging Bools
2029 uint i;
2030 for (i = 1; i < phi->req(); i++) {
2031 Node* b = phi->in(i);
2032 if (b->is_Phi()) {
2033 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2034 } else {
2035 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2036 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2037 }
2038 }
2039 Node* n = phi->in(1);
2040 Node* sample_opaque = nullptr;
2041 Node *sample_bool = nullptr;
2042 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2043 sample_opaque = n;
2044 sample_bool = n->in(1);
2045 assert(sample_bool->is_Bool(), "wrong type");
2046 } else {
2047 sample_bool = n;
2048 }
2049 Node *sample_cmp = sample_bool->in(1);
2050
2051 // Make Phis to merge the Cmp's inputs.
2052 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2053 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2054 for (i = 1; i < phi->req(); i++) {
2055 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2056 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2057 phi1->set_req(i, n1);
2058 phi2->set_req(i, n2);
2059 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2060 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2061 }
2062 // See if these Phis have been made before.
2063 // Register with optimizer
2064 Node *hit1 = _igvn.hash_find_insert(phi1);
2065 if (hit1) { // Hit, toss just made Phi
2066 _igvn.remove_dead_node(phi1); // Remove new phi
2067 assert(hit1->is_Phi(), "" );
2068 phi1 = (PhiNode*)hit1; // Use existing phi
2069 } else { // Miss
2070 _igvn.register_new_node_with_optimizer(phi1);
2071 }
2072 Node *hit2 = _igvn.hash_find_insert(phi2);
2073 if (hit2) { // Hit, toss just made Phi
2074 _igvn.remove_dead_node(phi2); // Remove new phi
2075 assert(hit2->is_Phi(), "" );
2076 phi2 = (PhiNode*)hit2; // Use existing phi
2077 } else { // Miss
2078 _igvn.register_new_node_with_optimizer(phi2);
2079 }
2080 // Register Phis with loop/block info
2081 set_ctrl(phi1, phi->in(0));
2082 set_ctrl(phi2, phi->in(0));
2083 // Make a new Cmp
2084 Node *cmp = sample_cmp->clone();
2085 cmp->set_req(1, phi1);
2086 cmp->set_req(2, phi2);
2087 _igvn.register_new_node_with_optimizer(cmp);
2088 set_ctrl(cmp, phi->in(0));
2089
2090 // Make a new Bool
2091 Node *b = sample_bool->clone();
2092 b->set_req(1,cmp);
2093 _igvn.register_new_node_with_optimizer(b);
2094 set_ctrl(b, phi->in(0));
2095
2096 if (sample_opaque != nullptr) {
2097 Node* opaque = sample_opaque->clone();
2098 opaque->set_req(1, b);
2099 _igvn.register_new_node_with_optimizer(opaque);
2100 set_ctrl(opaque, phi->in(0));
2101 return opaque;
2102 }
2103
2104 assert(b->is_Bool(), "");
2105 return b;
2106 }
2107
2108 //------------------------------clone_bool-------------------------------------
2109 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2110 // "Nearly" because all Nodes have been cloned from the original in the loop,
2111 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2112 // through the Phi recursively, and return a Bool.
2113 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2114 uint i;
2115 // Convert this Phi into a Phi merging Bools
2116 for( i = 1; i < phi->req(); i++ ) {
2117 Node *b = phi->in(i);
2118 if( b->is_Phi() ) {
2119 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2120 } else {
2121 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2122 }
2123 }
2124
2125 Node *sample_cmp = phi->in(1);
2126
2127 // Make Phis to merge the Cmp's inputs.
2128 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2129 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2130 for( uint j = 1; j < phi->req(); j++ ) {
2131 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2132 Node *n1, *n2;
2133 if( cmp_top->is_Cmp() ) {
2134 n1 = cmp_top->in(1);
2135 n2 = cmp_top->in(2);
2136 } else {
2137 n1 = n2 = cmp_top;
2138 }
2139 phi1->set_req( j, n1 );
2140 phi2->set_req( j, n2 );
2141 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2142 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2143 }
2144
2145 // See if these Phis have been made before.
2146 // Register with optimizer
2147 Node *hit1 = _igvn.hash_find_insert(phi1);
2148 if( hit1 ) { // Hit, toss just made Phi
2149 _igvn.remove_dead_node(phi1); // Remove new phi
2150 assert( hit1->is_Phi(), "" );
2151 phi1 = (PhiNode*)hit1; // Use existing phi
2152 } else { // Miss
2153 _igvn.register_new_node_with_optimizer(phi1);
2154 }
2155 Node *hit2 = _igvn.hash_find_insert(phi2);
2156 if( hit2 ) { // Hit, toss just made Phi
2157 _igvn.remove_dead_node(phi2); // Remove new phi
2158 assert( hit2->is_Phi(), "" );
2159 phi2 = (PhiNode*)hit2; // Use existing phi
2160 } else { // Miss
2161 _igvn.register_new_node_with_optimizer(phi2);
2162 }
2163 // Register Phis with loop/block info
2164 set_ctrl(phi1, phi->in(0));
2165 set_ctrl(phi2, phi->in(0));
2166 // Make a new Cmp
2167 Node *cmp = sample_cmp->clone();
2168 cmp->set_req( 1, phi1 );
2169 cmp->set_req( 2, phi2 );
2170 _igvn.register_new_node_with_optimizer(cmp);
2171 set_ctrl(cmp, phi->in(0));
2172
2173 assert( cmp->is_Cmp(), "" );
2174 return (CmpNode*)cmp;
2175 }
2176
2177 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2178 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2179 Node_List*& split_if_set, Node_List*& split_bool_set,
2180 Node_List*& split_cex_set, Node_List& worklist,
2181 uint new_counter, CloneLoopMode mode) {
2182 Node* nnn = old_new[old->_idx];
2183 // Copy uses to a worklist, so I can munge the def-use info
2184 // with impunity.
2185 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2186 worklist.push(old->fast_out(j));
2187
2188 while( worklist.size() ) {
2189 Node *use = worklist.pop();
2190 if (!has_node(use)) continue; // Ignore dead nodes
2191 if (use->in(0) == C->top()) continue;
2192 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2193 // Check for data-use outside of loop - at least one of OLD or USE
2194 // must not be a CFG node.
2195 #ifdef ASSERT
2196 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2197 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2198 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2199 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2200 }
2201 #endif
2202 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2203
2204 // If the Data use is an IF, that means we have an IF outside the
2205 // loop that is switching on a condition that is set inside the
2206 // loop. Happens if people set a loop-exit flag; then test the flag
2207 // in the loop to break the loop, then test is again outside the
2208 // loop to determine which way the loop exited.
2209 //
2210 // For several uses we need to make sure that there is no phi between,
2211 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2212 // to avoid such a phi in between.
2213 // For example, it is unexpected that there is a Phi between an
2214 // AllocateArray node and its ValidLengthTest input that could cause
2215 // split if to break.
2216 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2217 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2218 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2219 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2220 // Since this code is highly unlikely, we lazily build the worklist
2221 // of such Nodes to go split.
2222 if (!split_if_set) {
2223 split_if_set = new Node_List();
2224 }
2225 split_if_set->push(use);
2226 }
2227 if (use->is_Bool()) {
2228 if (!split_bool_set) {
2229 split_bool_set = new Node_List();
2230 }
2231 split_bool_set->push(use);
2232 }
2233 if (use->Opcode() == Op_CreateEx) {
2234 if (!split_cex_set) {
2235 split_cex_set = new Node_List();
2236 }
2237 split_cex_set->push(use);
2238 }
2239
2240
2241 // Get "block" use is in
2242 uint idx = 0;
2243 while( use->in(idx) != old ) idx++;
2244 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2245 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2246 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2247 ? prev->in(2)
2248 : idom(prev);
2249 if( use->is_Phi() ) // Phi use is in prior block
2250 cfg = prev->in(idx); // NOT in block of Phi itself
2251 if (cfg->is_top()) { // Use is dead?
2252 _igvn.replace_input_of(use, idx, C->top());
2253 continue;
2254 }
2255
2256 // If use is referenced through control edge... (idx == 0)
2257 if (mode == IgnoreStripMined && idx == 0) {
2258 LoopNode *head = loop->_head->as_Loop();
2259 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2260 // That node is outside the inner loop, leave it outside the
2261 // outer loop as well to not confuse verification code.
2262 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2263 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2264 continue;
2265 }
2266 }
2267
2268 while(!outer_loop->is_member(get_loop(cfg))) {
2269 prev = cfg;
2270 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2271 }
2272 // If the use occurs after merging several exits from the loop, then
2273 // old value must have dominated all those exits. Since the same old
2274 // value was used on all those exits we did not need a Phi at this
2275 // merge point. NOW we do need a Phi here. Each loop exit value
2276 // is now merged with the peeled body exit; each exit gets its own
2277 // private Phi and those Phis need to be merged here.
2278 Node *phi;
2279 if( prev->is_Region() ) {
2280 if( idx == 0 ) { // Updating control edge?
2281 phi = prev; // Just use existing control
2282 } else { // Else need a new Phi
2283 phi = PhiNode::make( prev, old );
2284 // Now recursively fix up the new uses of old!
2285 for( uint i = 1; i < prev->req(); i++ ) {
2286 worklist.push(phi); // Onto worklist once for each 'old' input
2287 }
2288 }
2289 } else {
2290 // Get new RegionNode merging old and new loop exits
2291 prev = old_new[prev->_idx];
2292 assert( prev, "just made this in step 7" );
2293 if( idx == 0) { // Updating control edge?
2294 phi = prev; // Just use existing control
2295 } else { // Else need a new Phi
2296 // Make a new Phi merging data values properly
2297 phi = PhiNode::make( prev, old );
2298 phi->set_req( 1, nnn );
2299 }
2300 }
2301 // If inserting a new Phi, check for prior hits
2302 if( idx != 0 ) {
2303 Node *hit = _igvn.hash_find_insert(phi);
2304 if( hit == nullptr ) {
2305 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2306 } else { // or
2307 // Remove the new phi from the graph and use the hit
2308 _igvn.remove_dead_node(phi);
2309 phi = hit; // Use existing phi
2310 }
2311 set_ctrl(phi, prev);
2312 }
2313 // Make 'use' use the Phi instead of the old loop body exit value
2314 assert(use->in(idx) == old, "old is still input of use");
2315 // We notify all uses of old, including use, and the indirect uses,
2316 // that may now be optimized because we have replaced old with phi.
2317 _igvn.add_users_to_worklist(old);
2318 if (idx == 0 &&
2319 use->depends_only_on_test()) {
2320 Node* pinned_clone = use->pin_array_access_node();
2321 if (pinned_clone != nullptr) {
2322 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2323 // into the region is left, an array load could become dependent on a condition that's not a range check for
2324 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2325 // floating above its range check.
2326 pinned_clone->set_req(0, phi);
2327 register_new_node_with_ctrl_of(pinned_clone, use);
2328 _igvn.replace_node(use, pinned_clone);
2329 continue;
2330 }
2331 }
2332 _igvn.replace_input_of(use, idx, phi);
2333 if( use->_idx >= new_counter ) { // If updating new phis
2334 // Not needed for correctness, but prevents a weak assert
2335 // in AddPNode from tripping (when we end up with different
2336 // base & derived Phis that will become the same after
2337 // IGVN does CSE).
2338 Node *hit = _igvn.hash_find_insert(use);
2339 if( hit ) // Go ahead and re-hash for hits.
2340 _igvn.replace_node( use, hit );
2341 }
2342 }
2343 }
2344 }
2345
2346 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2347 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2348 bool check_old_new) {
2349 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2350 Node* u = n->fast_out(j);
2351 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2352 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2353 Node* c = phase->get_ctrl(u);
2354 IdealLoopTree* u_loop = phase->get_loop(c);
2355 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2356 if (!loop->is_member(u_loop)) {
2357 if (outer_loop->is_member(u_loop)) {
2358 wq.push(u);
2359 } else {
2360 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2361 // the outer loop too
2362 Node* u_c = u->in(0);
2363 if (u_c != nullptr) {
2364 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2365 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2366 wq.push(u);
2367 }
2368 }
2369 }
2370 }
2371 }
2372 }
2373 }
2374
2375 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2376 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2377 Node_List& extra_data_nodes) {
2378 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2379 CountedLoopNode* cl = head->as_CountedLoop();
2380 Node* l = cl->outer_loop();
2381 Node* tail = cl->outer_loop_tail();
2382 IfNode* le = cl->outer_loop_end();
2383 Node* sfpt = cl->outer_safepoint();
2384 CountedLoopEndNode* cle = cl->loopexit();
2385 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2386 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2387 Node* cle_out = cle->proj_out(false);
2388
2389 Node* new_sfpt = nullptr;
2390 Node* new_cle_out = cle_out->clone();
2391 old_new.map(cle_out->_idx, new_cle_out);
2392 if (mode == CloneIncludesStripMined) {
2393 // clone outer loop body
2394 Node* new_l = l->clone();
2395 Node* new_tail = tail->clone();
2396 IfNode* new_le = le->clone()->as_If();
2397 new_sfpt = sfpt->clone();
2398
2399 set_loop(new_l, outer_loop->_parent);
2400 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2401 set_loop(new_cle_out, outer_loop->_parent);
2402 set_idom(new_cle_out, new_cle, dd);
2403 set_loop(new_sfpt, outer_loop->_parent);
2404 set_idom(new_sfpt, new_cle_out, dd);
2405 set_loop(new_le, outer_loop->_parent);
2406 set_idom(new_le, new_sfpt, dd);
2407 set_loop(new_tail, outer_loop->_parent);
2408 set_idom(new_tail, new_le, dd);
2409 set_idom(new_cl, new_l, dd);
2410
2411 old_new.map(l->_idx, new_l);
2412 old_new.map(tail->_idx, new_tail);
2413 old_new.map(le->_idx, new_le);
2414 old_new.map(sfpt->_idx, new_sfpt);
2415
2416 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2417 new_l->set_req(0, new_l);
2418 new_tail->set_req(0, new_le);
2419 new_le->set_req(0, new_sfpt);
2420 new_sfpt->set_req(0, new_cle_out);
2421 new_cle_out->set_req(0, new_cle);
2422 new_cl->set_req(LoopNode::EntryControl, new_l);
2423
2424 _igvn.register_new_node_with_optimizer(new_l);
2425 _igvn.register_new_node_with_optimizer(new_tail);
2426 _igvn.register_new_node_with_optimizer(new_le);
2427 } else {
2428 Node *newhead = old_new[loop->_head->_idx];
2429 newhead->as_Loop()->clear_strip_mined();
2430 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2431 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2432 }
2433 // Look at data node that were assigned a control in the outer
2434 // loop: they are kept in the outer loop by the safepoint so start
2435 // from the safepoint node's inputs.
2436 IdealLoopTree* outer_loop = get_loop(l);
2437 Node_Stack stack(2);
2438 stack.push(sfpt, 1);
2439 uint new_counter = C->unique();
2440 while (stack.size() > 0) {
2441 Node* n = stack.node();
2442 uint i = stack.index();
2443 while (i < n->req() &&
2444 (n->in(i) == nullptr ||
2445 !has_ctrl(n->in(i)) ||
2446 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2447 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2448 i++;
2449 }
2450 if (i < n->req()) {
2451 stack.set_index(i+1);
2452 stack.push(n->in(i), 0);
2453 } else {
2454 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2455 Node* m = n == sfpt ? new_sfpt : n->clone();
2456 if (m != nullptr) {
2457 for (uint i = 0; i < n->req(); i++) {
2458 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2459 m->set_req(i, old_new[m->in(i)->_idx]);
2460 }
2461 }
2462 } else {
2463 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2464 }
2465 if (n != sfpt) {
2466 extra_data_nodes.push(n);
2467 _igvn.register_new_node_with_optimizer(m);
2468 assert(get_ctrl(n) == cle_out, "what other control?");
2469 set_ctrl(m, new_cle_out);
2470 old_new.map(n->_idx, m);
2471 }
2472 stack.pop();
2473 }
2474 }
2475 if (mode == CloneIncludesStripMined) {
2476 _igvn.register_new_node_with_optimizer(new_sfpt);
2477 _igvn.register_new_node_with_optimizer(new_cle_out);
2478 }
2479 // Some other transformation may have pessimistically assigned some
2480 // data nodes to the outer loop. Set their control so they are out
2481 // of the outer loop.
2482 ResourceMark rm;
2483 Unique_Node_List wq;
2484 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2485 Node* old = extra_data_nodes.at(i);
2486 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2487 }
2488
2489 for (uint i = 0; i < loop->_body.size(); i++) {
2490 Node* old = loop->_body.at(i);
2491 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2492 }
2493
2494 Node* inner_out = sfpt->in(0);
2495 if (inner_out->outcnt() > 1) {
2496 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2497 }
2498
2499 Node* new_ctrl = cl->outer_loop_exit();
2500 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2501 for (uint i = 0; i < wq.size(); i++) {
2502 Node* n = wq.at(i);
2503 set_ctrl(n, new_ctrl);
2504 if (n->in(0) != nullptr) {
2505 _igvn.replace_input_of(n, 0, new_ctrl);
2506 }
2507 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2508 }
2509 } else {
2510 Node *newhead = old_new[loop->_head->_idx];
2511 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2512 }
2513 }
2514
2515 //------------------------------clone_loop-------------------------------------
2516 //
2517 // C L O N E A L O O P B O D Y
2518 //
2519 // This is the basic building block of the loop optimizations. It clones an
2520 // entire loop body. It makes an old_new loop body mapping; with this mapping
2521 // you can find the new-loop equivalent to an old-loop node. All new-loop
2522 // nodes are exactly equal to their old-loop counterparts, all edges are the
2523 // same. All exits from the old-loop now have a RegionNode that merges the
2524 // equivalent new-loop path. This is true even for the normal "loop-exit"
2525 // condition. All uses of loop-invariant old-loop values now come from (one
2526 // or more) Phis that merge their new-loop equivalents.
2527 //
2528 // This operation leaves the graph in an illegal state: there are two valid
2529 // control edges coming from the loop pre-header to both loop bodies. I'll
2530 // definitely have to hack the graph after running this transform.
2531 //
2532 // From this building block I will further edit edges to perform loop peeling
2533 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2534 //
2535 // Parameter side_by_size_idom:
2536 // When side_by_size_idom is null, the dominator tree is constructed for
2537 // the clone loop to dominate the original. Used in construction of
2538 // pre-main-post loop sequence.
2539 // When nonnull, the clone and original are side-by-side, both are
2540 // dominated by the side_by_side_idom node. Used in construction of
2541 // unswitched loops.
2542 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2543 CloneLoopMode mode, Node* side_by_side_idom) {
2544
2545 LoopNode* head = loop->_head->as_Loop();
2546 head->verify_strip_mined(1);
2547
2548 if (C->do_vector_loop() && PrintOpto) {
2549 const char* mname = C->method()->name()->as_quoted_ascii();
2550 if (mname != nullptr) {
2551 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2552 }
2553 }
2554
2555 CloneMap& cm = C->clone_map();
2556 if (C->do_vector_loop()) {
2557 cm.set_clone_idx(cm.max_gen()+1);
2558 #ifndef PRODUCT
2559 if (PrintOpto) {
2560 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2561 loop->dump_head();
2562 }
2563 #endif
2564 }
2565
2566 // Step 1: Clone the loop body. Make the old->new mapping.
2567 clone_loop_body(loop->_body, old_new, &cm);
2568
2569 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2570
2571 // Step 2: Fix the edges in the new body. If the old input is outside the
2572 // loop use it. If the old input is INside the loop, use the corresponding
2573 // new node instead.
2574 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2575
2576 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2577 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2578
2579 // Step 3: Now fix control uses. Loop varying control uses have already
2580 // been fixed up (as part of all input edges in Step 2). Loop invariant
2581 // control uses must be either an IfFalse or an IfTrue. Make a merge
2582 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2583 // refer to this.
2584 Node_List worklist;
2585 uint new_counter = C->unique();
2586 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2587
2588 // Step 4: If loop-invariant use is not control, it must be dominated by a
2589 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2590 // there if needed. Make a Phi there merging old and new used values.
2591 Node_List *split_if_set = nullptr;
2592 Node_List *split_bool_set = nullptr;
2593 Node_List *split_cex_set = nullptr;
2594 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2595
2596 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2597 Node* old = extra_data_nodes.at(i);
2598 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2599 split_bool_set, split_cex_set, worklist, new_counter,
2600 mode);
2601 }
2602
2603 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2604 // the loop uses a condition set in the loop. The original IF probably
2605 // takes control from one or more OLD Regions (which in turn get from NEW
2606 // Regions). In any case, there will be a set of Phis for each merge point
2607 // from the IF up to where the original BOOL def exists the loop.
2608 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2609
2610 }
2611
2612 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2613 if (split_if_set) {
2614 while (split_if_set->size()) {
2615 Node *iff = split_if_set->pop();
2616 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2617 if (iff->in(input)->is_Phi()) {
2618 Node *b = clone_iff(iff->in(input)->as_Phi());
2619 _igvn.replace_input_of(iff, input, b);
2620 }
2621 }
2622 }
2623 if (split_bool_set) {
2624 while (split_bool_set->size()) {
2625 Node *b = split_bool_set->pop();
2626 Node *phi = b->in(1);
2627 assert(phi->is_Phi(), "");
2628 CmpNode *cmp = clone_bool((PhiNode*) phi);
2629 _igvn.replace_input_of(b, 1, cmp);
2630 }
2631 }
2632 if (split_cex_set) {
2633 while (split_cex_set->size()) {
2634 Node *b = split_cex_set->pop();
2635 assert(b->in(0)->is_Region(), "");
2636 assert(b->in(1)->is_Phi(), "");
2637 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2638 split_up(b, b->in(0), nullptr);
2639 }
2640 }
2641 }
2642
2643 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2644 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2645 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2646 for(uint i = 0; i < body.size(); i++ ) {
2647 Node* old = body.at(i);
2648 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2649 split_bool_set, split_cex_set, worklist, new_counter,
2650 mode);
2651 }
2652 }
2653
2654 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2655 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2656 LoopNode* head = loop->_head->as_Loop();
2657 for(uint i = 0; i < body.size(); i++ ) {
2658 Node* old = body.at(i);
2659 if( !old->is_CFG() ) continue;
2660
2661 // Copy uses to a worklist, so I can munge the def-use info
2662 // with impunity.
2663 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2664 worklist.push(old->fast_out(j));
2665 }
2666
2667 while (worklist.size()) { // Visit all uses
2668 Node *use = worklist.pop();
2669 if (!has_node(use)) continue; // Ignore dead nodes
2670 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2671 if (!loop->is_member(use_loop) && use->is_CFG()) {
2672 // Both OLD and USE are CFG nodes here.
2673 assert(use->is_Proj(), "" );
2674 Node* nnn = old_new[old->_idx];
2675
2676 Node* newuse = nullptr;
2677 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2678 CountedLoopNode* cl = head->as_CountedLoop();
2679 CountedLoopEndNode* cle = cl->loopexit();
2680 Node* cle_out = cle->proj_out_or_null(false);
2681 if (use == cle_out) {
2682 IfNode* le = cl->outer_loop_end();
2683 use = le->proj_out(false);
2684 use_loop = get_loop(use);
2685 if (mode == CloneIncludesStripMined) {
2686 nnn = old_new[le->_idx];
2687 } else {
2688 newuse = old_new[cle_out->_idx];
2689 }
2690 }
2691 }
2692 if (newuse == nullptr) {
2693 newuse = use->clone();
2694 }
2695
2696 // Clone the loop exit control projection
2697 if (C->do_vector_loop() && cm != nullptr) {
2698 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2699 }
2700 newuse->set_req(0,nnn);
2701 _igvn.register_new_node_with_optimizer(newuse);
2702 set_loop(newuse, use_loop);
2703 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2704
2705 // We need a Region to merge the exit from the peeled body and the
2706 // exit from the old loop body.
2707 RegionNode *r = new RegionNode(3);
2708 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2709 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2710
2711 // The original user of 'use' uses 'r' instead.
2712 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2713 Node* useuse = use->last_out(l);
2714 _igvn.rehash_node_delayed(useuse);
2715 uint uses_found = 0;
2716 if (useuse->in(0) == use) {
2717 useuse->set_req(0, r);
2718 uses_found++;
2719 if (useuse->is_CFG()) {
2720 // This is not a dom_depth > dd_r because when new
2721 // control flow is constructed by a loop opt, a node and
2722 // its dominator can end up at the same dom_depth
2723 assert(dom_depth(useuse) >= dd_r, "");
2724 set_idom(useuse, r, dom_depth(useuse));
2725 }
2726 }
2727 for (uint k = 1; k < useuse->req(); k++) {
2728 if( useuse->in(k) == use ) {
2729 useuse->set_req(k, r);
2730 uses_found++;
2731 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2732 // This is not a dom_depth > dd_r because when new
2733 // control flow is constructed by a loop opt, a node
2734 // and its dominator can end up at the same dom_depth
2735 assert(dom_depth(useuse) >= dd_r , "");
2736 set_idom(useuse, r, dom_depth(useuse));
2737 }
2738 }
2739 }
2740 l -= uses_found; // we deleted 1 or more copies of this edge
2741 }
2742
2743 assert(use->is_Proj(), "loop exit should be projection");
2744 // lazy_replace() below moves all nodes that are:
2745 // - control dependent on the loop exit or
2746 // - have control set to the loop exit
2747 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2748 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2749 // becomes dead and is "replaced" by the region.
2750 Node* use_clone = use->clone();
2751 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2752 // Now finish up 'r'
2753 r->set_req(1, newuse);
2754 r->set_req(2, use_clone);
2755 _igvn.register_new_node_with_optimizer(r);
2756 set_loop(r, use_loop);
2757 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2758 lazy_replace(use, r);
2759 // Map the (cloned) old use to the new merge point
2760 old_new.map(use_clone->_idx, r);
2761 } // End of if a loop-exit test
2762 }
2763 }
2764 }
2765
2766 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2767 IdealLoopTree* parent, bool partial) {
2768 for(uint i = 0; i < body.size(); i++ ) {
2769 Node *old = body.at(i);
2770 Node *nnn = old_new[old->_idx];
2771 // Fix CFG/Loop controlling the new node
2772 if (has_ctrl(old)) {
2773 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2774 } else {
2775 set_loop(nnn, parent);
2776 if (old->outcnt() > 0) {
2777 Node* dom = idom(old);
2778 if (old_new[dom->_idx] != nullptr) {
2779 dom = old_new[dom->_idx];
2780 set_idom(nnn, dom, dd );
2781 }
2782 }
2783 }
2784 // Correct edges to the new node
2785 for (uint j = 0; j < nnn->req(); j++) {
2786 Node *n = nnn->in(j);
2787 if (n != nullptr) {
2788 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2789 if (loop->is_member(old_in_loop)) {
2790 if (old_new[n->_idx] != nullptr) {
2791 nnn->set_req(j, old_new[n->_idx]);
2792 } else {
2793 assert(!body.contains(n), "");
2794 assert(partial, "node not cloned");
2795 }
2796 }
2797 }
2798 }
2799 _igvn.hash_find_insert(nnn);
2800 }
2801 }
2802
2803 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2804 for (uint i = 0; i < body.size(); i++) {
2805 Node* old = body.at(i);
2806 Node* nnn = old->clone();
2807 old_new.map(old->_idx, nnn);
2808 if (C->do_vector_loop() && cm != nullptr) {
2809 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2810 }
2811 _igvn.register_new_node_with_optimizer(nnn);
2812 }
2813 }
2814
2815
2816 //---------------------- stride_of_possible_iv -------------------------------------
2817 // Looks for an iff/bool/comp with one operand of the compare
2818 // being a cycle involving an add and a phi,
2819 // with an optional truncation (left-shift followed by a right-shift)
2820 // of the add. Returns zero if not an iv.
2821 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2822 Node* trunc1 = nullptr;
2823 Node* trunc2 = nullptr;
2824 const TypeInteger* ttype = nullptr;
2825 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2826 return 0;
2827 }
2828 BoolNode* bl = iff->in(1)->as_Bool();
2829 Node* cmp = bl->in(1);
2830 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2831 return 0;
2832 }
2833 // Must have an invariant operand
2834 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2835 return 0;
2836 }
2837 Node* add2 = nullptr;
2838 Node* cmp1 = cmp->in(1);
2839 if (cmp1->is_Phi()) {
2840 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2841 Node* phi = cmp1;
2842 for (uint i = 1; i < phi->req(); i++) {
2843 Node* in = phi->in(i);
2844 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2845 &trunc1, &trunc2, &ttype, T_INT);
2846 if (add && add->in(1) == phi) {
2847 add2 = add->in(2);
2848 break;
2849 }
2850 }
2851 } else {
2852 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2853 Node* addtrunc = cmp1;
2854 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2855 &trunc1, &trunc2, &ttype, T_INT);
2856 if (add && add->in(1)->is_Phi()) {
2857 Node* phi = add->in(1);
2858 for (uint i = 1; i < phi->req(); i++) {
2859 if (phi->in(i) == addtrunc) {
2860 add2 = add->in(2);
2861 break;
2862 }
2863 }
2864 }
2865 }
2866 if (add2 != nullptr) {
2867 const TypeInt* add2t = _igvn.type(add2)->is_int();
2868 if (add2t->is_con()) {
2869 return add2t->get_con();
2870 }
2871 }
2872 return 0;
2873 }
2874
2875
2876 //---------------------- stay_in_loop -------------------------------------
2877 // Return the (unique) control output node that's in the loop (if it exists.)
2878 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2879 Node* unique = nullptr;
2880 if (!n) return nullptr;
2881 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2882 Node* use = n->fast_out(i);
2883 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2884 if (unique != nullptr) {
2885 return nullptr;
2886 }
2887 unique = use;
2888 }
2889 }
2890 return unique;
2891 }
2892
2893 //------------------------------ register_node -------------------------------------
2894 // Utility to register node "n" with PhaseIdealLoop
2895 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2896 _igvn.register_new_node_with_optimizer(n);
2897 loop->_body.push(n);
2898 if (n->is_CFG()) {
2899 set_loop(n, loop);
2900 set_idom(n, pred, ddepth);
2901 } else {
2902 set_ctrl(n, pred);
2903 }
2904 }
2905
2906 //------------------------------ proj_clone -------------------------------------
2907 // Utility to create an if-projection
2908 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2909 ProjNode* c = p->clone()->as_Proj();
2910 c->set_req(0, iff);
2911 return c;
2912 }
2913
2914 //------------------------------ short_circuit_if -------------------------------------
2915 // Force the iff control output to be the live_proj
2916 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2917 guarantee(live_proj != nullptr, "null projection");
2918 int proj_con = live_proj->_con;
2919 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2920 Node* con = intcon(proj_con);
2921 if (iff) {
2922 iff->set_req(1, con);
2923 }
2924 return con;
2925 }
2926
2927 //------------------------------ insert_if_before_proj -------------------------------------
2928 // Insert a new if before an if projection (* - new node)
2929 //
2930 // before
2931 // if(test)
2932 // / \
2933 // v v
2934 // other-proj proj (arg)
2935 //
2936 // after
2937 // if(test)
2938 // / \
2939 // / v
2940 // | * proj-clone
2941 // v |
2942 // other-proj v
2943 // * new_if(relop(cmp[IU](left,right)))
2944 // / \
2945 // v v
2946 // * new-proj proj
2947 // (returned)
2948 //
2949 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2950 IfNode* iff = proj->in(0)->as_If();
2951 IdealLoopTree *loop = get_loop(proj);
2952 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2953 uint ddepth = dom_depth(proj);
2954
2955 _igvn.rehash_node_delayed(iff);
2956 _igvn.rehash_node_delayed(proj);
2957
2958 proj->set_req(0, nullptr); // temporary disconnect
2959 ProjNode* proj2 = proj_clone(proj, iff);
2960 register_node(proj2, loop, iff, ddepth);
2961
2962 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2963 register_node(cmp, loop, proj2, ddepth);
2964
2965 BoolNode* bol = new BoolNode(cmp, relop);
2966 register_node(bol, loop, proj2, ddepth);
2967
2968 int opcode = iff->Opcode();
2969 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2970 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2971 register_node(new_if, loop, proj2, ddepth);
2972
2973 proj->set_req(0, new_if); // reattach
2974 set_idom(proj, new_if, ddepth);
2975
2976 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2977 guarantee(new_exit != nullptr, "null exit node");
2978 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2979
2980 return new_exit;
2981 }
2982
2983 //------------------------------ insert_region_before_proj -------------------------------------
2984 // Insert a region before an if projection (* - new node)
2985 //
2986 // before
2987 // if(test)
2988 // / |
2989 // v |
2990 // proj v
2991 // other-proj
2992 //
2993 // after
2994 // if(test)
2995 // / |
2996 // v |
2997 // * proj-clone v
2998 // | other-proj
2999 // v
3000 // * new-region
3001 // |
3002 // v
3003 // * dum_if
3004 // / \
3005 // v \
3006 // * dum-proj v
3007 // proj
3008 //
3009 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3010 IfNode* iff = proj->in(0)->as_If();
3011 IdealLoopTree *loop = get_loop(proj);
3012 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3013 uint ddepth = dom_depth(proj);
3014
3015 _igvn.rehash_node_delayed(iff);
3016 _igvn.rehash_node_delayed(proj);
3017
3018 proj->set_req(0, nullptr); // temporary disconnect
3019 ProjNode* proj2 = proj_clone(proj, iff);
3020 register_node(proj2, loop, iff, ddepth);
3021
3022 RegionNode* reg = new RegionNode(2);
3023 reg->set_req(1, proj2);
3024 register_node(reg, loop, iff, ddepth);
3025
3026 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3027 register_node(dum_if, loop, reg, ddepth);
3028
3029 proj->set_req(0, dum_if); // reattach
3030 set_idom(proj, dum_if, ddepth);
3031
3032 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3033 register_node(dum_proj, loop, dum_if, ddepth);
3034
3035 return reg;
3036 }
3037
3038 // Idea
3039 // ----
3040 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3041 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3042 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3043 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3044 // test alone instead:
3045 //
3046 // Before Partial Peeling:
3047 // Loop:
3048 // <peeled section>
3049 // Split off signed loop exit test
3050 // <-- CUT HERE -->
3051 // Unchanged unsigned loop exit test
3052 // <rest of unpeeled section>
3053 // goto Loop
3054 //
3055 // After Partial Peeling:
3056 // <cloned peeled section>
3057 // Cloned split off signed loop exit test
3058 // Loop:
3059 // Unchanged unsigned loop exit test
3060 // <rest of unpeeled section>
3061 // <peeled section>
3062 // Split off signed loop exit test
3063 // goto Loop
3064 //
3065 // Details
3066 // -------
3067 // Before:
3068 // if (i <u limit) Unsigned loop exit condition
3069 // / |
3070 // v v
3071 // exit-proj stay-in-loop-proj
3072 //
3073 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3074 // before the CmpU on the stay-in-loop path and keep both tests:
3075 //
3076 // if (i <u limit) Signed loop exit test
3077 // / |
3078 // / if (i <u limit) Unsigned loop exit test
3079 // / / |
3080 // v v v
3081 // exit-region stay-in-loop-proj
3082 //
3083 // Implementation
3084 // --------------
3085 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3086 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3087 // exit tests is preserved, and their loop nesting is correct.
3088 //
3089 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3090 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3091 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3092 //
3093 // if (stay-in-loop-const) Killed original unsigned loop exit test
3094 // / |
3095 // / v
3096 // / if (i < limit) Split off signed loop exit test
3097 // / / |
3098 // / / v
3099 // / / if (i <u limit) Cloned unsigned loop exit test
3100 // / / / |
3101 // v v v |
3102 // exit-region |
3103 // | |
3104 // dummy-if |
3105 // / | |
3106 // dead | |
3107 // v v
3108 // exit-proj stay-in-loop-proj
3109 //
3110 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3111 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3112 //
3113 // Requirements
3114 // ------------
3115 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3116 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3117 // Otherwise, we need to bail out (see comments in the code below).
3118 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3119 const bool Signed = true;
3120 const bool Unsigned = false;
3121
3122 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3123 if (bol->_test._test != BoolTest::lt) {
3124 return nullptr;
3125 }
3126 CmpNode* cmpu = bol->in(1)->as_Cmp();
3127 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3128
3129 int stride = stride_of_possible_iv(if_cmpu);
3130 if (stride == 0) {
3131 return nullptr;
3132 }
3133
3134 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3135 guarantee(lp_proj != nullptr, "null loop node");
3136
3137 ProjNode* lp_continue = lp_proj->as_Proj();
3138 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3139 if (!lp_exit->is_IfFalse()) {
3140 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3141 // We therefore can't add a single exit condition.
3142 return nullptr;
3143 }
3144 // The unsigned loop exit condition is
3145 // !(i <u limit)
3146 // = i >=u limit
3147 //
3148 // First, we note that for any x for which
3149 // 0 <= x <= INT_MAX
3150 // we can convert x to an unsigned int and still get the same guarantee:
3151 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3152 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3153 //
3154 // With that in mind, if
3155 // limit >= 0 (COND)
3156 // then the unsigned loop exit condition
3157 // i >=u limit (ULE)
3158 // is equivalent to
3159 // i < 0 || i >= limit (SLE-full)
3160 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3161 // (uint) i >=u MAX_INT >= limit >= 0
3162 // or otherwise
3163 // i >= limit >= 0
3164 // holds due to (LEMMA).
3165 //
3166 // For completeness, a counterexample with limit < 0:
3167 // Assume i = -3 and limit = -2:
3168 // i < 0
3169 // -2 < 0
3170 // is true and thus also "i < 0 || i >= limit". But
3171 // i >=u limit
3172 // -3 >=u -2
3173 // is false.
3174 Node* limit = cmpu->in(2);
3175 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3176 if (type_limit->_lo < 0) {
3177 return nullptr;
3178 }
3179
3180 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3181 // stride < 0:
3182 // i < 0 (SLE = SLE-negative)
3183 // stride > 0:
3184 // i >= limit (SLE = SLE-positive)
3185 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3186 //
3187 // Loop:
3188 // <peeled section>
3189 // i >= limit (SLE-positive)
3190 // <-- CUT HERE -->
3191 // i >=u limit (ULE)
3192 // <rest of unpeeled section>
3193 // goto Loop
3194 //
3195 // We exit the loop if:
3196 // (SLE) is true OR (ULE) is true
3197 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3198 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3199 // (SLE) IMPLIES (ULE)
3200 // This indeed holds when (COND) is given:
3201 // - stride > 0:
3202 // i >= limit // (SLE = SLE-positive)
3203 // i >= limit >= 0 // (COND)
3204 // i >=u limit >= 0 // (LEMMA)
3205 // which is the unsigned loop exit condition (ULE).
3206 // - stride < 0:
3207 // i < 0 // (SLE = SLE-negative)
3208 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3209 // MAX_INT >= limit >= 0 // (COND)
3210 // MAX_INT >=u limit >= 0 // (LEMMA)
3211 // and thus from (NEG) and (LEMMA):
3212 // i >=u limit
3213 // which is the unsigned loop exit condition (ULE).
3214 //
3215 //
3216 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3217 // <cloned peeled section>
3218 // i >= limit (SLE-positive)
3219 // Loop:
3220 // i >=u limit (ULE)
3221 // <rest of unpeeled section>
3222 // <peeled section>
3223 // i >= limit (SLE-positive)
3224 // goto Loop
3225 Node* rhs_cmpi;
3226 if (stride > 0) {
3227 rhs_cmpi = limit; // For i >= limit
3228 } else {
3229 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3230 }
3231 // Create a new region on the exit path
3232 RegionNode* reg = insert_region_before_proj(lp_exit);
3233 guarantee(reg != nullptr, "null region node");
3234
3235 // Clone the if-cmpu-true-false using a signed compare
3236 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3237 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3238 reg->add_req(cmpi_exit);
3239
3240 // Clone the if-cmpu-true-false
3241 BoolTest::mask rel_u = bol->_test._test;
3242 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3243 reg->add_req(cmpu_exit);
3244
3245 // Force original if to stay in loop.
3246 short_circuit_if(if_cmpu, lp_continue);
3247
3248 return cmpi_exit->in(0)->as_If();
3249 }
3250
3251 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3252 // Remove a previously inserted signed compare loop exit.
3253 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3254 Node* lp_proj = stay_in_loop(if_cmp, loop);
3255 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3256 stay_in_loop(lp_proj, loop)->is_If() &&
3257 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3258 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3259 if_cmp->set_req(1, con);
3260 }
3261
3262 //------------------------------ scheduled_nodelist -------------------------------------
3263 // Create a post order schedule of nodes that are in the
3264 // "member" set. The list is returned in "sched".
3265 // The first node in "sched" is the loop head, followed by
3266 // nodes which have no inputs in the "member" set, and then
3267 // followed by the nodes that have an immediate input dependence
3268 // on a node in "sched".
3269 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3270
3271 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3272 VectorSet visited;
3273 Node_Stack nstack(loop->_body.size());
3274
3275 Node* n = loop->_head; // top of stack is cached in "n"
3276 uint idx = 0;
3277 visited.set(n->_idx);
3278
3279 // Initially push all with no inputs from within member set
3280 for(uint i = 0; i < loop->_body.size(); i++ ) {
3281 Node *elt = loop->_body.at(i);
3282 if (member.test(elt->_idx)) {
3283 bool found = false;
3284 for (uint j = 0; j < elt->req(); j++) {
3285 Node* def = elt->in(j);
3286 if (def && member.test(def->_idx) && def != elt) {
3287 found = true;
3288 break;
3289 }
3290 }
3291 if (!found && elt != loop->_head) {
3292 nstack.push(n, idx);
3293 n = elt;
3294 assert(!visited.test(n->_idx), "not seen yet");
3295 visited.set(n->_idx);
3296 }
3297 }
3298 }
3299
3300 // traverse out's that are in the member set
3301 while (true) {
3302 if (idx < n->outcnt()) {
3303 Node* use = n->raw_out(idx);
3304 idx++;
3305 if (!visited.test_set(use->_idx)) {
3306 if (member.test(use->_idx)) {
3307 nstack.push(n, idx);
3308 n = use;
3309 idx = 0;
3310 }
3311 }
3312 } else {
3313 // All outputs processed
3314 sched.push(n);
3315 if (nstack.is_empty()) break;
3316 n = nstack.node();
3317 idx = nstack.index();
3318 nstack.pop();
3319 }
3320 }
3321 }
3322
3323
3324 //------------------------------ has_use_in_set -------------------------------------
3325 // Has a use in the vector set
3326 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3327 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3328 Node* use = n->fast_out(j);
3329 if (vset.test(use->_idx)) {
3330 return true;
3331 }
3332 }
3333 return false;
3334 }
3335
3336
3337 //------------------------------ has_use_internal_to_set -------------------------------------
3338 // Has use internal to the vector set (ie. not in a phi at the loop head)
3339 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3340 Node* head = loop->_head;
3341 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3342 Node* use = n->fast_out(j);
3343 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3344 return true;
3345 }
3346 }
3347 return false;
3348 }
3349
3350
3351 //------------------------------ clone_for_use_outside_loop -------------------------------------
3352 // clone "n" for uses that are outside of loop
3353 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3354 int cloned = 0;
3355 assert(worklist.size() == 0, "should be empty");
3356 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3357 Node* use = n->fast_out(j);
3358 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3359 worklist.push(use);
3360 }
3361 }
3362
3363 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3364 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3365 return -1;
3366 }
3367
3368 while( worklist.size() ) {
3369 Node *use = worklist.pop();
3370 if (!has_node(use) || use->in(0) == C->top()) continue;
3371 uint j;
3372 for (j = 0; j < use->req(); j++) {
3373 if (use->in(j) == n) break;
3374 }
3375 assert(j < use->req(), "must be there");
3376
3377 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3378 Node* n_clone = n->clone();
3379 _igvn.replace_input_of(use, j, n_clone);
3380 cloned++;
3381 Node* use_c;
3382 if (!use->is_Phi()) {
3383 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3384 } else {
3385 // Use in a phi is considered a use in the associated predecessor block
3386 use_c = use->in(0)->in(j);
3387 }
3388 set_ctrl(n_clone, use_c);
3389 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3390 get_loop(use_c)->_body.push(n_clone);
3391 _igvn.register_new_node_with_optimizer(n_clone);
3392 #ifndef PRODUCT
3393 if (TracePartialPeeling) {
3394 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3395 }
3396 #endif
3397 }
3398 return cloned;
3399 }
3400
3401
3402 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3403 // clone "n" for special uses that are in the not_peeled region.
3404 // If these def-uses occur in separate blocks, the code generator
3405 // marks the method as not compilable. For example, if a "BoolNode"
3406 // is in a different basic block than the "IfNode" that uses it, then
3407 // the compilation is aborted in the code generator.
3408 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3409 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3410 if (n->is_Phi() || n->is_Load()) {
3411 return;
3412 }
3413 assert(worklist.size() == 0, "should be empty");
3414 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3415 Node* use = n->fast_out(j);
3416 if ( not_peel.test(use->_idx) &&
3417 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3418 use->in(1) == n) {
3419 worklist.push(use);
3420 }
3421 }
3422 if (worklist.size() > 0) {
3423 // clone "n" and insert it between inputs of "n" and the use
3424 Node* n_clone = n->clone();
3425 loop->_body.push(n_clone);
3426 _igvn.register_new_node_with_optimizer(n_clone);
3427 set_ctrl(n_clone, get_ctrl(n));
3428 sink_list.push(n_clone);
3429 not_peel.set(n_clone->_idx);
3430 #ifndef PRODUCT
3431 if (TracePartialPeeling) {
3432 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3433 }
3434 #endif
3435 while( worklist.size() ) {
3436 Node *use = worklist.pop();
3437 _igvn.rehash_node_delayed(use);
3438 for (uint j = 1; j < use->req(); j++) {
3439 if (use->in(j) == n) {
3440 use->set_req(j, n_clone);
3441 }
3442 }
3443 }
3444 }
3445 }
3446
3447
3448 //------------------------------ insert_phi_for_loop -------------------------------------
3449 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3450 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3451 Node *phi = PhiNode::make(lp, back_edge_val);
3452 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3453 // Use existing phi if it already exists
3454 Node *hit = _igvn.hash_find_insert(phi);
3455 if( hit == nullptr ) {
3456 _igvn.register_new_node_with_optimizer(phi);
3457 set_ctrl(phi, lp);
3458 } else {
3459 // Remove the new phi from the graph and use the hit
3460 _igvn.remove_dead_node(phi);
3461 phi = hit;
3462 }
3463 _igvn.replace_input_of(use, idx, phi);
3464 }
3465
3466 #ifdef ASSERT
3467 //------------------------------ is_valid_loop_partition -------------------------------------
3468 // Validate the loop partition sets: peel and not_peel
3469 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3470 VectorSet& not_peel ) {
3471 uint i;
3472 // Check that peel_list entries are in the peel set
3473 for (i = 0; i < peel_list.size(); i++) {
3474 if (!peel.test(peel_list.at(i)->_idx)) {
3475 return false;
3476 }
3477 }
3478 // Check at loop members are in one of peel set or not_peel set
3479 for (i = 0; i < loop->_body.size(); i++ ) {
3480 Node *def = loop->_body.at(i);
3481 uint di = def->_idx;
3482 // Check that peel set elements are in peel_list
3483 if (peel.test(di)) {
3484 if (not_peel.test(di)) {
3485 return false;
3486 }
3487 // Must be in peel_list also
3488 bool found = false;
3489 for (uint j = 0; j < peel_list.size(); j++) {
3490 if (peel_list.at(j)->_idx == di) {
3491 found = true;
3492 break;
3493 }
3494 }
3495 if (!found) {
3496 return false;
3497 }
3498 } else if (not_peel.test(di)) {
3499 if (peel.test(di)) {
3500 return false;
3501 }
3502 } else {
3503 return false;
3504 }
3505 }
3506 return true;
3507 }
3508
3509 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3510 // Ensure a use outside of loop is of the right form
3511 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3512 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3513 return (use->is_Phi() &&
3514 use_c->is_Region() && use_c->req() == 3 &&
3515 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3516 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3517 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3518 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3519 }
3520
3521 //------------------------------ is_valid_clone_loop_form -------------------------------------
3522 // Ensure that all uses outside of loop are of the right form
3523 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3524 uint orig_exit_idx, uint clone_exit_idx) {
3525 uint len = peel_list.size();
3526 for (uint i = 0; i < len; i++) {
3527 Node *def = peel_list.at(i);
3528
3529 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3530 Node *use = def->fast_out(j);
3531 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3532 if (!loop->is_member(get_loop(use_c))) {
3533 // use is not in the loop, check for correct structure
3534 if (use->in(0) == def) {
3535 // Okay
3536 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3537 return false;
3538 }
3539 }
3540 }
3541 }
3542 return true;
3543 }
3544 #endif
3545
3546 //------------------------------ partial_peel -------------------------------------
3547 // Partially peel (aka loop rotation) the top portion of a loop (called
3548 // the peel section below) by cloning it and placing one copy just before
3549 // the new loop head and the other copy at the bottom of the new loop.
3550 //
3551 // before after where it came from
3552 //
3553 // stmt1 stmt1
3554 // loop: stmt2 clone
3555 // stmt2 if condA goto exitA clone
3556 // if condA goto exitA new_loop: new
3557 // stmt3 stmt3 clone
3558 // if !condB goto loop if condB goto exitB clone
3559 // exitB: stmt2 orig
3560 // stmt4 if !condA goto new_loop orig
3561 // exitA: goto exitA
3562 // exitB:
3563 // stmt4
3564 // exitA:
3565 //
3566 // Step 1: find the cut point: an exit test on probable
3567 // induction variable.
3568 // Step 2: schedule (with cloning) operations in the peel
3569 // section that can be executed after the cut into
3570 // the section that is not peeled. This may need
3571 // to clone operations into exit blocks. For
3572 // instance, a reference to A[i] in the not-peel
3573 // section and a reference to B[i] in an exit block
3574 // may cause a left-shift of i by 2 to be placed
3575 // in the peel block. This step will clone the left
3576 // shift into the exit block and sink the left shift
3577 // from the peel to the not-peel section.
3578 // Step 3: clone the loop, retarget the control, and insert
3579 // phis for values that are live across the new loop
3580 // head. This is very dependent on the graph structure
3581 // from clone_loop. It creates region nodes for
3582 // exit control and associated phi nodes for values
3583 // flow out of the loop through that exit. The region
3584 // node is dominated by the clone's control projection.
3585 // So the clone's peel section is placed before the
3586 // new loop head, and the clone's not-peel section is
3587 // forms the top part of the new loop. The original
3588 // peel section forms the tail of the new loop.
3589 // Step 4: update the dominator tree and recompute the
3590 // dominator depth.
3591 //
3592 // orig
3593 //
3594 // stmt1
3595 // |
3596 // v
3597 // predicates
3598 // |
3599 // v
3600 // loop<----+
3601 // | |
3602 // stmt2 |
3603 // | |
3604 // v |
3605 // ifA |
3606 // / | |
3607 // v v |
3608 // false true ^ <-- last_peel
3609 // / | |
3610 // / ===|==cut |
3611 // / stmt3 | <-- first_not_peel
3612 // / | |
3613 // | v |
3614 // v ifB |
3615 // exitA: / \ |
3616 // / \ |
3617 // v v |
3618 // false true |
3619 // / \ |
3620 // / ----+
3621 // |
3622 // v
3623 // exitB:
3624 // stmt4
3625 //
3626 //
3627 // after clone loop
3628 //
3629 // stmt1
3630 // |
3631 // v
3632 // predicates
3633 // / \
3634 // clone / \ orig
3635 // / \
3636 // / \
3637 // v v
3638 // +---->loop loop<----+
3639 // | | | |
3640 // | stmt2 stmt2 |
3641 // | | | |
3642 // | v v |
3643 // | ifA ifA |
3644 // | | \ / | |
3645 // | v v v v |
3646 // ^ true false false true ^ <-- last_peel
3647 // | | ^ \ / | |
3648 // | cut==|== \ \ / ===|==cut |
3649 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3650 // | | dom | | | |
3651 // | v \ 1v v2 v |
3652 // | ifB regionA ifB |
3653 // | / \ | / \ |
3654 // | / \ v / \ |
3655 // | v v exitA: v v |
3656 // | true false false true |
3657 // | / ^ \ / \ |
3658 // +---- \ \ / ----+
3659 // dom \ /
3660 // \ 1v v2
3661 // regionB
3662 // |
3663 // v
3664 // exitB:
3665 // stmt4
3666 //
3667 //
3668 // after partial peel
3669 //
3670 // stmt1
3671 // |
3672 // v
3673 // predicates
3674 // /
3675 // clone / orig
3676 // / TOP
3677 // / \
3678 // v v
3679 // TOP->loop loop----+
3680 // | | |
3681 // stmt2 stmt2 |
3682 // | | |
3683 // v v |
3684 // ifA ifA |
3685 // | \ / | |
3686 // v v v v |
3687 // true false false true | <-- last_peel
3688 // | ^ \ / +------|---+
3689 // +->newloop \ \ / === ==cut | |
3690 // | stmt3 \ \ / TOP | |
3691 // | | dom | | stmt3 | | <-- first_not_peel
3692 // | v \ 1v v2 v | |
3693 // | ifB regionA ifB ^ v
3694 // | / \ | / \ | |
3695 // | / \ v / \ | |
3696 // | v v exitA: v v | |
3697 // | true false false true | |
3698 // | / ^ \ / \ | |
3699 // | | \ \ / v | |
3700 // | | dom \ / TOP | |
3701 // | | \ 1v v2 | |
3702 // ^ v regionB | |
3703 // | | | | |
3704 // | | v ^ v
3705 // | | exitB: | |
3706 // | | stmt4 | |
3707 // | +------------>-----------------+ |
3708 // | |
3709 // +-----------------<---------------------+
3710 //
3711 //
3712 // final graph
3713 //
3714 // stmt1
3715 // |
3716 // v
3717 // predicates
3718 // |
3719 // v
3720 // stmt2 clone
3721 // |
3722 // v
3723 // ........> ifA clone
3724 // : / |
3725 // dom / |
3726 // : v v
3727 // : false true
3728 // : | |
3729 // : | v
3730 // : | newloop<-----+
3731 // : | | |
3732 // : | stmt3 clone |
3733 // : | | |
3734 // : | v |
3735 // : | ifB |
3736 // : | / \ |
3737 // : | v v |
3738 // : | false true |
3739 // : | | | |
3740 // : | v stmt2 |
3741 // : | exitB: | |
3742 // : | stmt4 v |
3743 // : | ifA orig |
3744 // : | / \ |
3745 // : | / \ |
3746 // : | v v |
3747 // : | false true |
3748 // : | / \ |
3749 // : v v -----+
3750 // RegionA
3751 // |
3752 // v
3753 // exitA
3754 //
3755 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3756
3757 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3758 if (!loop->_head->is_Loop()) {
3759 return false;
3760 }
3761 LoopNode *head = loop->_head->as_Loop();
3762
3763 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3764 return false;
3765 }
3766
3767 // Check for complex exit control
3768 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3769 Node *n = loop->_body.at(ii);
3770 int opc = n->Opcode();
3771 if (n->is_Call() ||
3772 opc == Op_Catch ||
3773 opc == Op_CatchProj ||
3774 opc == Op_Jump ||
3775 opc == Op_JumpProj) {
3776 #ifndef PRODUCT
3777 if (TracePartialPeeling) {
3778 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3779 }
3780 #endif
3781 return false;
3782 }
3783 }
3784
3785 int dd = dom_depth(head);
3786
3787 // Step 1: find cut point
3788
3789 // Walk up dominators to loop head looking for first loop exit
3790 // which is executed on every path thru loop.
3791 IfNode *peel_if = nullptr;
3792 IfNode *peel_if_cmpu = nullptr;
3793
3794 Node *iff = loop->tail();
3795 while (iff != head) {
3796 if (iff->is_If()) {
3797 Node *ctrl = get_ctrl(iff->in(1));
3798 if (ctrl->is_top()) return false; // Dead test on live IF.
3799 // If loop-varying exit-test, check for induction variable
3800 if (loop->is_member(get_loop(ctrl)) &&
3801 loop->is_loop_exit(iff) &&
3802 is_possible_iv_test(iff)) {
3803 Node* cmp = iff->in(1)->in(1);
3804 if (cmp->Opcode() == Op_CmpI) {
3805 peel_if = iff->as_If();
3806 } else {
3807 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3808 peel_if_cmpu = iff->as_If();
3809 }
3810 }
3811 }
3812 iff = idom(iff);
3813 }
3814
3815 // Prefer signed compare over unsigned compare.
3816 IfNode* new_peel_if = nullptr;
3817 if (peel_if == nullptr) {
3818 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3819 return false; // No peel point found
3820 }
3821 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3822 if (new_peel_if == nullptr) {
3823 return false; // No peel point found
3824 }
3825 peel_if = new_peel_if;
3826 }
3827 Node* last_peel = stay_in_loop(peel_if, loop);
3828 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3829 if (first_not_peeled == nullptr || first_not_peeled == head) {
3830 return false;
3831 }
3832
3833 #ifndef PRODUCT
3834 if (TraceLoopOpts) {
3835 tty->print("PartialPeel ");
3836 loop->dump_head();
3837 }
3838
3839 if (TracePartialPeeling) {
3840 tty->print_cr("before partial peel one iteration");
3841 Node_List wl;
3842 Node* t = head->in(2);
3843 while (true) {
3844 wl.push(t);
3845 if (t == head) break;
3846 t = idom(t);
3847 }
3848 while (wl.size() > 0) {
3849 Node* tt = wl.pop();
3850 tt->dump();
3851 if (tt == last_peel) tty->print_cr("-- cut --");
3852 }
3853 }
3854 #endif
3855
3856 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3857
3858 VectorSet peel;
3859 VectorSet not_peel;
3860 Node_List peel_list;
3861 Node_List worklist;
3862 Node_List sink_list;
3863
3864 uint estimate = loop->est_loop_clone_sz(1);
3865 if (exceeding_node_budget(estimate)) {
3866 return false;
3867 }
3868
3869 // Set of cfg nodes to peel are those that are executable from
3870 // the head through last_peel.
3871 assert(worklist.size() == 0, "should be empty");
3872 worklist.push(head);
3873 peel.set(head->_idx);
3874 while (worklist.size() > 0) {
3875 Node *n = worklist.pop();
3876 if (n != last_peel) {
3877 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3878 Node* use = n->fast_out(j);
3879 if (use->is_CFG() &&
3880 loop->is_member(get_loop(use)) &&
3881 !peel.test_set(use->_idx)) {
3882 worklist.push(use);
3883 }
3884 }
3885 }
3886 }
3887
3888 // Set of non-cfg nodes to peel are those that are control
3889 // dependent on the cfg nodes.
3890 for (uint i = 0; i < loop->_body.size(); i++) {
3891 Node *n = loop->_body.at(i);
3892 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3893 if (peel.test(n_c->_idx)) {
3894 peel.set(n->_idx);
3895 } else {
3896 not_peel.set(n->_idx);
3897 }
3898 }
3899
3900 // Step 2: move operations from the peeled section down into the
3901 // not-peeled section
3902
3903 // Get a post order schedule of nodes in the peel region
3904 // Result in right-most operand.
3905 scheduled_nodelist(loop, peel, peel_list);
3906
3907 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3908
3909 // For future check for too many new phis
3910 uint old_phi_cnt = 0;
3911 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3912 Node* use = head->fast_out(j);
3913 if (use->is_Phi()) old_phi_cnt++;
3914 }
3915
3916 #ifndef PRODUCT
3917 if (TracePartialPeeling) {
3918 tty->print_cr("\npeeled list");
3919 }
3920 #endif
3921
3922 // Evacuate nodes in peel region into the not_peeled region if possible
3923 bool too_many_clones = false;
3924 uint new_phi_cnt = 0;
3925 uint cloned_for_outside_use = 0;
3926 for (uint i = 0; i < peel_list.size();) {
3927 Node* n = peel_list.at(i);
3928 #ifndef PRODUCT
3929 if (TracePartialPeeling) n->dump();
3930 #endif
3931 bool incr = true;
3932 if (!n->is_CFG()) {
3933 if (has_use_in_set(n, not_peel)) {
3934 // If not used internal to the peeled region,
3935 // move "n" from peeled to not_peeled region.
3936 if (!has_use_internal_to_set(n, peel, loop)) {
3937 // if not pinned and not a load (which maybe anti-dependent on a store)
3938 // and not a CMove (Matcher expects only bool->cmove).
3939 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3940 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3941 if (C->failing()) return false;
3942 if (new_clones == -1) {
3943 too_many_clones = true;
3944 break;
3945 }
3946 cloned_for_outside_use += new_clones;
3947 sink_list.push(n);
3948 peel.remove(n->_idx);
3949 not_peel.set(n->_idx);
3950 peel_list.remove(i);
3951 incr = false;
3952 #ifndef PRODUCT
3953 if (TracePartialPeeling) {
3954 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3955 n->_idx, get_ctrl(n)->_idx);
3956 }
3957 #endif
3958 }
3959 } else {
3960 // Otherwise check for special def-use cases that span
3961 // the peel/not_peel boundary such as bool->if
3962 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3963 new_phi_cnt++;
3964 }
3965 }
3966 }
3967 if (incr) i++;
3968 }
3969
3970 estimate += cloned_for_outside_use + new_phi_cnt;
3971 bool exceed_node_budget = !may_require_nodes(estimate);
3972 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3973
3974 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3975 #ifndef PRODUCT
3976 if (TracePartialPeeling && exceed_phi_limit) {
3977 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3978 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3979 }
3980 #endif
3981 if (new_peel_if != nullptr) {
3982 remove_cmpi_loop_exit(new_peel_if, loop);
3983 }
3984 // Inhibit more partial peeling on this loop
3985 assert(!head->is_partial_peel_loop(), "not partial peeled");
3986 head->mark_partial_peel_failed();
3987 if (cloned_for_outside_use > 0) {
3988 // Terminate this round of loop opts because
3989 // the graph outside this loop was changed.
3990 C->set_major_progress();
3991 return true;
3992 }
3993 return false;
3994 }
3995
3996 // Step 3: clone loop, retarget control, and insert new phis
3997
3998 // Create new loop head for new phis and to hang
3999 // the nodes being moved (sinked) from the peel region.
4000 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4001 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4002 _igvn.register_new_node_with_optimizer(new_head);
4003 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4004 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4005 set_loop(new_head, loop);
4006 loop->_body.push(new_head);
4007 not_peel.set(new_head->_idx);
4008 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4009 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4010
4011 while (sink_list.size() > 0) {
4012 Node* n = sink_list.pop();
4013 set_ctrl(n, new_head);
4014 }
4015
4016 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4017
4018 clone_loop(loop, old_new, dd, IgnoreStripMined);
4019
4020 const uint clone_exit_idx = 1;
4021 const uint orig_exit_idx = 2;
4022 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4023
4024 Node* head_clone = old_new[head->_idx];
4025 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4026 Node* orig_tail_clone = head_clone->in(2);
4027
4028 // Add phi if "def" node is in peel set and "use" is not
4029
4030 for (uint i = 0; i < peel_list.size(); i++) {
4031 Node *def = peel_list.at(i);
4032 if (!def->is_CFG()) {
4033 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4034 Node *use = def->fast_out(j);
4035 if (has_node(use) && use->in(0) != C->top() &&
4036 (!peel.test(use->_idx) ||
4037 (use->is_Phi() && use->in(0) == head)) ) {
4038 worklist.push(use);
4039 }
4040 }
4041 while( worklist.size() ) {
4042 Node *use = worklist.pop();
4043 for (uint j = 1; j < use->req(); j++) {
4044 Node* n = use->in(j);
4045 if (n == def) {
4046
4047 // "def" is in peel set, "use" is not in peel set
4048 // or "use" is in the entry boundary (a phi) of the peel set
4049
4050 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4051
4052 if ( loop->is_member(get_loop( use_c )) ) {
4053 // use is in loop
4054 if (old_new[use->_idx] != nullptr) { // null for dead code
4055 Node* use_clone = old_new[use->_idx];
4056 _igvn.replace_input_of(use, j, C->top());
4057 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4058 }
4059 } else {
4060 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4061 // use is not in the loop, check if the live range includes the cut
4062 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4063 if (not_peel.test(lp_if->_idx)) {
4064 assert(j == orig_exit_idx, "use from original loop");
4065 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4066 }
4067 }
4068 }
4069 }
4070 }
4071 }
4072 }
4073
4074 // Step 3b: retarget control
4075
4076 // Redirect control to the new loop head if a cloned node in
4077 // the not_peeled region has control that points into the peeled region.
4078 // This necessary because the cloned peeled region will be outside
4079 // the loop.
4080 // from to
4081 // cloned-peeled <---+
4082 // new_head_clone: | <--+
4083 // cloned-not_peeled in(0) in(0)
4084 // orig-peeled
4085
4086 for (uint i = 0; i < loop->_body.size(); i++) {
4087 Node *n = loop->_body.at(i);
4088 if (!n->is_CFG() && n->in(0) != nullptr &&
4089 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4090 Node* n_clone = old_new[n->_idx];
4091 if (n_clone->depends_only_on_test()) {
4092 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4093 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4094 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4095 // floating above its range check.
4096 Node* pinned_clone = n_clone->pin_array_access_node();
4097 if (pinned_clone != nullptr) {
4098 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4099 old_new.map(n->_idx, pinned_clone);
4100 _igvn.replace_node(n_clone, pinned_clone);
4101 n_clone = pinned_clone;
4102 }
4103 }
4104 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4105 }
4106 }
4107
4108 // Backedge of the surviving new_head (the clone) is original last_peel
4109 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4110
4111 // Cut first node in original not_peel set
4112 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4113 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4114 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4115
4116 // Copy head_clone back-branch info to original head
4117 // and remove original head's loop entry and
4118 // clone head's back-branch
4119 _igvn.rehash_node_delayed(head); // Multiple edge updates
4120 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4121 head->set_req(LoopNode::LoopBackControl, C->top());
4122 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4123
4124 // Similarly modify the phis
4125 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4126 Node* use = head->fast_out(k);
4127 if (use->is_Phi() && use->outcnt() > 0) {
4128 Node* use_clone = old_new[use->_idx];
4129 _igvn.rehash_node_delayed(use); // Multiple edge updates
4130 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4131 use->set_req(LoopNode::LoopBackControl, C->top());
4132 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4133 }
4134 }
4135
4136 // Step 4: update dominator tree and dominator depth
4137
4138 set_idom(head, orig_tail_clone, dd);
4139 recompute_dom_depth();
4140
4141 // Inhibit more partial peeling on this loop
4142 new_head_clone->set_partial_peel_loop();
4143 C->set_major_progress();
4144 loop->record_for_igvn();
4145
4146 #ifndef PRODUCT
4147 if (TracePartialPeeling) {
4148 tty->print_cr("\nafter partial peel one iteration");
4149 Node_List wl;
4150 Node* t = last_peel;
4151 while (true) {
4152 wl.push(t);
4153 if (t == head_clone) break;
4154 t = idom(t);
4155 }
4156 while (wl.size() > 0) {
4157 Node* tt = wl.pop();
4158 if (tt == head) tty->print_cr("orig head");
4159 else if (tt == new_head_clone) tty->print_cr("new head");
4160 else if (tt == head_clone) tty->print_cr("clone head");
4161 tt->dump();
4162 }
4163 }
4164 #endif
4165
4166 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4167
4168 return true;
4169 }
4170
4171 // Transform:
4172 //
4173 // loop<-----------------+
4174 // | |
4175 // stmt1 stmt2 .. stmtn |
4176 // | | | |
4177 // \ | / |
4178 // v v v |
4179 // region |
4180 // | |
4181 // shared_stmt |
4182 // | |
4183 // v |
4184 // if |
4185 // / \ |
4186 // | -----------+
4187 // v
4188 //
4189 // into:
4190 //
4191 // loop<-------------------+
4192 // | |
4193 // v |
4194 // +->loop |
4195 // | | |
4196 // | stmt1 stmt2 .. stmtn |
4197 // | | | | |
4198 // | | \ / |
4199 // | | v v |
4200 // | | region1 |
4201 // | | | |
4202 // | shared_stmt shared_stmt |
4203 // | | | |
4204 // | v v |
4205 // | if if |
4206 // | /\ / \ |
4207 // +-- | | -------+
4208 // \ /
4209 // v v
4210 // region2
4211 //
4212 // (region2 is shown to merge mirrored projections of the loop exit
4213 // ifs to make the diagram clearer but they really merge the same
4214 // projection)
4215 //
4216 // Conditions for this transformation to trigger:
4217 // - the path through stmt1 is frequent enough
4218 // - the inner loop will be turned into a counted loop after transformation
4219 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4220 if (!DuplicateBackedge) {
4221 return false;
4222 }
4223 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4224 if (!loop->_head->is_Loop()) {
4225 return false;
4226 }
4227
4228 uint estimate = loop->est_loop_clone_sz(1);
4229 if (exceeding_node_budget(estimate)) {
4230 return false;
4231 }
4232
4233 LoopNode *head = loop->_head->as_Loop();
4234
4235 Node* region = nullptr;
4236 IfNode* exit_test = nullptr;
4237 uint inner;
4238 float f;
4239 if (StressDuplicateBackedge) {
4240 if (head->is_strip_mined()) {
4241 return false;
4242 }
4243 Node* c = head->in(LoopNode::LoopBackControl);
4244
4245 while (c != head) {
4246 if (c->is_Region()) {
4247 region = c;
4248 }
4249 c = idom(c);
4250 }
4251
4252 if (region == nullptr) {
4253 return false;
4254 }
4255
4256 inner = 1;
4257 } else {
4258 // Is the shape of the loop that of a counted loop...
4259 Node* back_control = loop_exit_control(head, loop);
4260 if (back_control == nullptr) {
4261 return false;
4262 }
4263
4264 BoolTest::mask bt = BoolTest::illegal;
4265 float cl_prob = 0;
4266 Node* incr = nullptr;
4267 Node* limit = nullptr;
4268 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4269 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4270 return false;
4271 }
4272
4273 // With an extra phi for the candidate iv?
4274 // Or the region node is the loop head
4275 if (!incr->is_Phi() || incr->in(0) == head) {
4276 return false;
4277 }
4278
4279 PathFrequency pf(head, this);
4280 region = incr->in(0);
4281
4282 // Go over all paths for the extra phi's region and see if that
4283 // path is frequent enough and would match the expected iv shape
4284 // if the extra phi is removed
4285 inner = 0;
4286 for (uint i = 1; i < incr->req(); ++i) {
4287 Node* in = incr->in(i);
4288 Node* trunc1 = nullptr;
4289 Node* trunc2 = nullptr;
4290 const TypeInteger* iv_trunc_t = nullptr;
4291 Node* orig_in = in;
4292 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4293 continue;
4294 }
4295 assert(in->Opcode() == Op_AddI, "wrong increment code");
4296 Node* xphi = nullptr;
4297 Node* stride = loop_iv_stride(in, loop, xphi);
4298
4299 if (stride == nullptr) {
4300 continue;
4301 }
4302
4303 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4304 if (phi == nullptr ||
4305 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4306 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4307 return false;
4308 }
4309
4310 f = pf.to(region->in(i));
4311 if (f > 0.5) {
4312 inner = i;
4313 break;
4314 }
4315 }
4316
4317 if (inner == 0) {
4318 return false;
4319 }
4320
4321 exit_test = back_control->in(0)->as_If();
4322 }
4323
4324 if (idom(region)->is_Catch()) {
4325 return false;
4326 }
4327
4328 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4329 Unique_Node_List wq;
4330 wq.push(head->in(LoopNode::LoopBackControl));
4331 for (uint i = 0; i < wq.size(); i++) {
4332 Node* c = wq.at(i);
4333 assert(get_loop(c) == loop, "not in the right loop?");
4334 if (c->is_Region()) {
4335 if (c != region) {
4336 for (uint j = 1; j < c->req(); ++j) {
4337 wq.push(c->in(j));
4338 }
4339 }
4340 } else {
4341 wq.push(c->in(0));
4342 }
4343 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4344 }
4345
4346 Node* region_dom = idom(region);
4347
4348 // Can't do the transformation if this would cause a membar pair to
4349 // be split
4350 for (uint i = 0; i < wq.size(); i++) {
4351 Node* c = wq.at(i);
4352 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4353 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4354 if (!wq.member(c->as_MemBar()->leading_membar())) {
4355 return false;
4356 }
4357 }
4358 }
4359
4360 // Collect data nodes that need to be clones as well
4361 int dd = dom_depth(head);
4362
4363 for (uint i = 0; i < loop->_body.size(); ++i) {
4364 Node* n = loop->_body.at(i);
4365 if (has_ctrl(n)) {
4366 Node* c = get_ctrl(n);
4367 if (wq.member(c)) {
4368 wq.push(n);
4369 }
4370 } else {
4371 set_idom(n, idom(n), dd);
4372 }
4373 }
4374
4375 // clone shared_stmt
4376 clone_loop_body(wq, old_new, nullptr);
4377
4378 Node* region_clone = old_new[region->_idx];
4379 region_clone->set_req(inner, C->top());
4380 set_idom(region, region->in(inner), dd);
4381
4382 // Prepare the outer loop
4383 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4384 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4385 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4386 set_idom(head, outer_head, dd);
4387
4388 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4389
4390 // Make one of the shared_stmt copies only reachable from stmt1, the
4391 // other only from stmt2..stmtn.
4392 Node* dom = nullptr;
4393 for (uint i = 1; i < region->req(); ++i) {
4394 if (i != inner) {
4395 _igvn.replace_input_of(region, i, C->top());
4396 }
4397 Node* in = region_clone->in(i);
4398 if (in->is_top()) {
4399 continue;
4400 }
4401 if (dom == nullptr) {
4402 dom = in;
4403 } else {
4404 dom = dom_lca(dom, in);
4405 }
4406 }
4407
4408 set_idom(region_clone, dom, dd);
4409
4410 // Set up the outer loop
4411 for (uint i = 0; i < head->outcnt(); i++) {
4412 Node* u = head->raw_out(i);
4413 if (u->is_Phi()) {
4414 Node* outer_phi = u->clone();
4415 outer_phi->set_req(0, outer_head);
4416 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4417 if (backedge == nullptr) {
4418 backedge = u->in(LoopNode::LoopBackControl);
4419 }
4420 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4421 register_new_node(outer_phi, outer_head);
4422 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4423 }
4424 }
4425
4426 // create control and data nodes for out of loop uses (including region2)
4427 Node_List worklist;
4428 uint new_counter = C->unique();
4429 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4430
4431 Node_List *split_if_set = nullptr;
4432 Node_List *split_bool_set = nullptr;
4433 Node_List *split_cex_set = nullptr;
4434 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4435 split_if_set, split_bool_set, split_cex_set);
4436
4437 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4438
4439 if (exit_test != nullptr) {
4440 float cnt = exit_test->_fcnt;
4441 if (cnt != COUNT_UNKNOWN) {
4442 exit_test->_fcnt = cnt * f;
4443 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4444 }
4445 }
4446
4447 C->set_major_progress();
4448
4449 return true;
4450 }
4451
4452 // AutoVectorize the loop: replace scalar ops with vector ops.
4453 PhaseIdealLoop::AutoVectorizeStatus
4454 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4455 // Counted loop only
4456 if (!lpt->is_counted()) {
4457 return AutoVectorizeStatus::Impossible;
4458 }
4459
4460 // Main-loop only
4461 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4462 if (!cl->is_main_loop()) {
4463 return AutoVectorizeStatus::Impossible;
4464 }
4465
4466 VLoop vloop(lpt, false);
4467 if (!vloop.check_preconditions()) {
4468 return AutoVectorizeStatus::TriedAndFailed;
4469 }
4470
4471 // Ensure the shared data is cleared before each use
4472 vshared.clear();
4473
4474 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4475 if (!vloop_analyzer.success()) {
4476 return AutoVectorizeStatus::TriedAndFailed;
4477 }
4478
4479 SuperWord sw(vloop_analyzer);
4480 if (!sw.transform_loop()) {
4481 return AutoVectorizeStatus::TriedAndFailed;
4482 }
4483
4484 return AutoVectorizeStatus::Success;
4485 }
4486
4487 // Just before insert_pre_post_loops, we can multiversion the loop:
4488 //
4489 // multiversion_if
4490 // | |
4491 // fast_loop slow_loop
4492 //
4493 // In the fast_loop we can make speculative assumptions, and put the
4494 // conditions into the multiversion_if. If the conditions hold at runtime,
4495 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4496 // instead which does not make any of the speculative assumptions.
4497 //
4498 // Note: we only multiversion the loop if the loop does not have any
4499 // auto vectorization check Predicate. If we have that predicate,
4500 // then we can simply add the speculative assumption checks to
4501 // that Predicate. This means we do not need to duplicate the
4502 // loop - we have a smaller graph and save compile time. Should
4503 // the conditions ever fail, then we deopt / trap at the Predicate
4504 // and recompile without that Predicate. At that point we will
4505 // multiversion the loop, so that we can still have speculative
4506 // runtime checks.
4507 //
4508 // We perform the multiversioning when the loop is still in its single
4509 // iteration form, even before we insert pre and post loops. This makes
4510 // the cloning much simpler. However, this means that both the fast
4511 // and the slow loop have to be optimized independently (adding pre
4512 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4513 // we may end up not needing any speculative assumptions in the fast_loop
4514 // and then rejecting the slow_loop by constant folding the multiversion_if.
4515 //
4516 // Therefore, we "delay" the optimization of the slow_loop until we add
4517 // at least one speculative assumption for the fast_loop. If we never
4518 // add such a speculative runtime check, the OpaqueMultiversioningNode
4519 // of the multiversion_if constant folds to true after loop opts, and the
4520 // multiversion_if folds away the "delayed" slow_loop. If we add any
4521 // speculative assumption, then we notify the OpaqueMultiversioningNode
4522 // with "notify_slow_loop_that_it_can_resume_optimizations".
4523 //
4524 // Note: new runtime checks can be added to the multiversion_if with
4525 // PhaseIdealLoop::create_new_if_for_multiversion
4526 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4527 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4528 LoopNode* outer_loop = cl->skip_strip_mined();
4529 Node* entry = outer_loop->in(LoopNode::EntryControl);
4530
4531 // Check we have multiversioning enabled, and are not already multiversioned.
4532 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4533
4534 // Check that we do not have a parse-predicate where we can add the runtime checks
4535 // during auto-vectorization.
4536 const Predicates predicates(entry);
4537 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4538 if (predicate_block->has_parse_predicate()) { return; }
4539
4540 // Check node budget.
4541 uint estimate = lpt->est_loop_clone_sz(2);
4542 if (!may_require_nodes(estimate)) { return; }
4543
4544 do_multiversioning(lpt, old_new);
4545 }
4546
4547 // Returns true if the Reduction node is unordered.
4548 static bool is_unordered_reduction(Node* n) {
4549 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4550 }
4551
4552 // Having ReductionNodes in the loop is expensive. They need to recursively
4553 // fold together the vector values, for every vectorized loop iteration. If
4554 // we encounter the following pattern, we can vector accumulate the values
4555 // inside the loop, and only have a single UnorderedReduction after the loop.
4556 //
4557 // Note: UnorderedReduction represents a ReductionNode which does not require
4558 // calculating in strict order.
4559 //
4560 // CountedLoop init
4561 // | |
4562 // +------+ | +-----------------------+
4563 // | | | |
4564 // PhiNode (s) |
4565 // | |
4566 // | Vector |
4567 // | | |
4568 // UnorderedReduction (first_ur) |
4569 // | |
4570 // ... Vector |
4571 // | | |
4572 // UnorderedReduction (last_ur) |
4573 // | |
4574 // +---------------------+
4575 //
4576 // We patch the graph to look like this:
4577 //
4578 // CountedLoop identity_vector
4579 // | |
4580 // +-------+ | +---------------+
4581 // | | | |
4582 // PhiNode (v) |
4583 // | |
4584 // | Vector |
4585 // | | |
4586 // VectorAccumulator |
4587 // | |
4588 // ... Vector |
4589 // | | |
4590 // init VectorAccumulator |
4591 // | | | |
4592 // UnorderedReduction +-----------+
4593 //
4594 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4595 // use vector_accumulators, which do the same reductions, but only element
4596 // wise. This is a single operation per vector_accumulator, rather than many
4597 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4598 // after the loop, and also reduce the init value into it.
4599 //
4600 // We can not do this with all reductions. Some reductions do not allow the
4601 // reordering of operations (for example float addition/multiplication require
4602 // strict order).
4603 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4604 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4605
4606 // Find all Phi nodes with an unordered Reduction on backedge.
4607 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4608 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4609 Node* phi = cl->fast_out(j);
4610 // We have a phi with a single use, and an unordered Reduction on the backedge.
4611 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4612 continue;
4613 }
4614
4615 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4616 assert(!last_ur->requires_strict_order(), "must be");
4617
4618 // Determine types
4619 const TypeVect* vec_t = last_ur->vect_type();
4620 uint vector_length = vec_t->length();
4621 BasicType bt = vec_t->element_basic_type();
4622
4623 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4624 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4625 const int vopc = VectorNode::opcode(sopc, bt);
4626 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4627 DEBUG_ONLY( last_ur->dump(); )
4628 assert(false, "do not have normal vector op for this reduction");
4629 continue; // not implemented -> fails
4630 }
4631
4632 // Traverse up the chain of unordered Reductions, checking that it loops back to
4633 // the phi. Check that all unordered Reductions only have a single use, except for
4634 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4635 // are outside the loop.
4636 ReductionNode* current = last_ur;
4637 ReductionNode* first_ur = nullptr;
4638 while (true) {
4639 assert(!current->requires_strict_order(), "sanity");
4640
4641 // Expect no ctrl and a vector_input from within the loop.
4642 Node* ctrl = current->in(0);
4643 Node* vector_input = current->in(2);
4644 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4645 DEBUG_ONLY( current->dump(1); )
4646 assert(false, "reduction has ctrl or bad vector_input");
4647 break; // Chain traversal fails.
4648 }
4649
4650 assert(current->vect_type() != nullptr, "must have vector type");
4651 if (current->vect_type() != last_ur->vect_type()) {
4652 // Reductions do not have the same vector type (length and element type).
4653 break; // Chain traversal fails.
4654 }
4655
4656 // Expect single use of an unordered Reduction, except for last_ur.
4657 if (current == last_ur) {
4658 // Expect all uses to be outside the loop, except phi.
4659 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4660 Node* use = current->fast_out(k);
4661 if (use != phi && ctrl_or_self(use) == cl) {
4662 DEBUG_ONLY( current->dump(-1); )
4663 assert(false, "reduction has use inside loop");
4664 // Should not be allowed by SuperWord::mark_reductions
4665 return; // bail out of optimization
4666 }
4667 }
4668 } else {
4669 if (current->outcnt() != 1) {
4670 break; // Chain traversal fails.
4671 }
4672 }
4673
4674 // Expect another unordered Reduction or phi as the scalar input.
4675 Node* scalar_input = current->in(1);
4676 if (is_unordered_reduction(scalar_input) &&
4677 scalar_input->Opcode() == current->Opcode()) {
4678 // Move up the unordered Reduction chain.
4679 current = scalar_input->as_Reduction();
4680 assert(!current->requires_strict_order(), "must be");
4681 } else if (scalar_input == phi) {
4682 // Chain terminates at phi.
4683 first_ur = current;
4684 current = nullptr;
4685 break; // Success.
4686 } else {
4687 // scalar_input is neither phi nor a matching reduction
4688 // Can for example be scalar reduction when we have
4689 // partial vectorization.
4690 break; // Chain traversal fails.
4691 }
4692 }
4693 if (current != nullptr) {
4694 // Chain traversal was not successful.
4695 continue;
4696 }
4697 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4698
4699 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4700 set_root_as_ctrl(identity_scalar);
4701 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4702 register_new_node(identity_vector, C->root());
4703 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4704 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4705
4706 // Turn the scalar phi into a vector phi.
4707 _igvn.rehash_node_delayed(phi);
4708 Node* init = phi->in(1); // Remember init before replacing it.
4709 phi->set_req_X(1, identity_vector, &_igvn);
4710 phi->as_Type()->set_type(vec_t);
4711 _igvn.set_type(phi, vec_t);
4712
4713 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4714 current = first_ur;
4715 while (true) {
4716 // Create vector_accumulator to replace current.
4717 Node* last_vector_accumulator = current->in(1);
4718 Node* vector_input = current->in(2);
4719 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4720 register_new_node(vector_accumulator, cl);
4721 _igvn.replace_node(current, vector_accumulator);
4722 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4723 if (current == last_ur) {
4724 break;
4725 }
4726 current = vector_accumulator->unique_out()->as_Reduction();
4727 assert(!current->requires_strict_order(), "must be");
4728 }
4729
4730 // Create post-loop reduction.
4731 Node* last_accumulator = phi->in(2);
4732 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4733
4734 // Take over uses of last_accumulator that are not in the loop.
4735 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4736 Node* use = last_accumulator->out(i);
4737 if (use != phi && use != post_loop_reduction) {
4738 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4739 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4740 --i;
4741 }
4742 }
4743 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4744 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4745
4746 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4747 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4748 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4749 }
4750 }
4751
4752 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4753 for (uint i = 0; i < _data_nodes.size(); i++) {
4754 clone(_data_nodes[i], new_ctrl);
4755 }
4756 }
4757
4758 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4759 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4760 Node* clone = node->clone();
4761 _phase->igvn().register_new_node_with_optimizer(clone);
4762 _orig_to_new.put(node, clone);
4763 _phase->set_ctrl(clone, new_ctrl);
4764 if (node->is_CastII()) {
4765 clone->set_req(0, new_ctrl);
4766 }
4767 }
4768
4769 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4770 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4771 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4772 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4773 for (uint i = 1; i < node->req(); i++) {
4774 Node** cloned_input = _orig_to_new.get(node->in(i));
4775 if (cloned_input != nullptr) {
4776 // Input was also cloned -> rewire clone to the cloned input.
4777 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4778 }
4779 }
4780 });
4781 }
4782
4783 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4784 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4785 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4786 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4787 Node* new_ctrl) {
4788 for (uint i = 0; i < _data_nodes.size(); i++) {
4789 Node* data_node = _data_nodes[i];
4790 if (data_node->is_Opaque1()) {
4791 transform_opaque_node(transform_strategy, data_node);
4792 } else {
4793 clone(data_node, new_ctrl);
4794 }
4795 }
4796 }
4797
4798 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4799 Node* transformed_node;
4800 if (node->is_OpaqueLoopInit()) {
4801 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4802 } else {
4803 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4804 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4805 }
4806 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4807 _orig_to_new.put(node, transformed_node);
4808 }