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