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/vectornode.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
52 // ConvI2L may have type information on it which is unsafe to push up
53 // so disable this for now
54 return nullptr;
55 }
56
57 // Splitting range check CastIIs through a loop induction Phi can
58 // cause new Phis to be created that are left unrelated to the loop
59 // induction Phi and prevent optimizations (vectorization)
60 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
61 n->in(1) == region->as_CountedLoop()->phi()) {
62 return nullptr;
63 }
64
65 if (cannot_split_division(n, region)) {
66 return nullptr;
67 }
68
69 int wins = 0;
70 assert(!n->is_CFG(), "");
71 assert(region->is_Region(), "");
72
73 const Type* type = n->bottom_type();
74 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
75 Node* phi;
76 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
77 int iid = t_oop->instance_id();
78 int index = C->get_alias_index(t_oop);
79 int offset = t_oop->offset();
80 phi = new PhiNode(region, type, nullptr, iid, index, offset);
81 } else {
82 phi = PhiNode::make_blank(region, n);
83 }
84 uint old_unique = C->unique();
85 for (uint i = 1; i < region->req(); i++) {
86 Node* x;
87 Node* the_clone = nullptr;
88 if (region->in(i) == C->top()) {
89 x = C->top(); // Dead path? Use a dead data op
90 } else {
91 x = n->clone(); // Else clone up the data op
92 the_clone = x; // Remember for possible deletion.
93 // Alter data node to use pre-phi inputs
94 if (n->in(0) == region)
95 x->set_req( 0, region->in(i) );
96 for (uint j = 1; j < n->req(); j++) {
97 Node* in = n->in(j);
98 if (in->is_Phi() && in->in(0) == region)
99 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
100 }
101 }
102 // Check for a 'win' on some paths
103 const Type* t = x->Value(&_igvn);
104
105 bool singleton = t->singleton();
106
107 // A TOP singleton indicates that there are no possible values incoming
108 // along a particular edge. In most cases, this is OK, and the Phi will
109 // be eliminated later in an Ideal call. However, we can't allow this to
110 // happen if the singleton occurs on loop entry, as the elimination of
111 // the PhiNode may cause the resulting node to migrate back to a previous
112 // loop iteration.
113 if (singleton && t == Type::TOP) {
114 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
115 // irreducible loop may not be indicated by an affirmative is_Loop());
116 // therefore, the only top we can split thru a phi is on a backedge of
117 // a loop.
118 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
119 }
120
121 if (singleton) {
122 wins++;
123 x = ((PhaseGVN&)_igvn).makecon(t);
124 } else {
125 // We now call Identity to try to simplify the cloned node.
126 // Note that some Identity methods call phase->type(this).
127 // Make sure that the type array is big enough for
128 // our new node, even though we may throw the node away.
129 // (Note: This tweaking with igvn only works because x is a new node.)
130 _igvn.set_type(x, t);
131 // If x is a TypeNode, capture any more-precise type permanently into Node
132 // otherwise it will be not updated during igvn->transform since
133 // igvn->type(x) is set to x->Value() already.
134 x->raise_bottom_type(t);
135 Node* y = x->Identity(&_igvn);
136 if (y != x) {
137 wins++;
138 x = y;
139 } else {
140 y = _igvn.hash_find(x);
141 if (y == nullptr) {
142 y = similar_subtype_check(x, region->in(i));
143 }
144 if (y) {
145 wins++;
146 x = y;
147 } else {
148 // Else x is a new node we are keeping
149 // We do not need register_new_node_with_optimizer
150 // because set_type has already been called.
151 _igvn._worklist.push(x);
152 }
153 }
154 }
155
156 phi->set_req( i, x );
157
158 if (the_clone == nullptr) {
159 continue;
160 }
161
162 if (the_clone != x) {
163 _igvn.remove_dead_node(the_clone);
164 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
165 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
166 // it is not a win if 'x' moved from an outer to an inner loop
167 // this edge case can only happen for Load nodes
168 wins = 0;
169 break;
170 }
171 }
172 // Too few wins?
173 if (wins <= policy) {
174 _igvn.remove_dead_node(phi);
175 return nullptr;
176 }
177
178 // Record Phi
179 register_new_node( phi, region );
180
181 for (uint i2 = 1; i2 < phi->req(); i2++) {
182 Node *x = phi->in(i2);
183 // If we commoned up the cloned 'x' with another existing Node,
184 // the existing Node picks up a new use. We need to make the
185 // existing Node occur higher up so it dominates its uses.
186 Node *old_ctrl;
187 IdealLoopTree *old_loop;
188
189 if (x->is_Con()) {
190 // Constant's control is always root.
191 set_ctrl(x, C->root());
192 continue;
193 }
194 // The occasional new node
195 if (x->_idx >= old_unique) { // Found a new, unplaced node?
196 old_ctrl = nullptr;
197 old_loop = nullptr; // Not in any prior loop
198 } else {
199 old_ctrl = get_ctrl(x);
200 old_loop = get_loop(old_ctrl); // Get prior loop
201 }
202 // New late point must dominate new use
203 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
204 if (new_ctrl == old_ctrl) // Nothing is changed
205 continue;
206
207 IdealLoopTree *new_loop = get_loop(new_ctrl);
208
209 // Don't move x into a loop if its uses are
210 // outside of loop. Otherwise x will be cloned
211 // for each use outside of this loop.
212 IdealLoopTree *use_loop = get_loop(region);
213 if (!new_loop->is_member(use_loop) &&
214 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
215 // Take early control, later control will be recalculated
216 // during next iteration of loop optimizations.
217 new_ctrl = get_early_ctrl(x);
218 new_loop = get_loop(new_ctrl);
219 }
220 // Set new location
221 set_ctrl(x, new_ctrl);
222 // If changing loop bodies, see if we need to collect into new body
223 if (old_loop != new_loop) {
224 if (old_loop && !old_loop->_child)
225 old_loop->_body.yank(x);
226 if (!new_loop->_child)
227 new_loop->_body.push(x); // Collect body info
228 }
229 }
230
231 return phi;
232 }
233
234 // Test whether node 'x' can move into an inner loop relative to node 'n'.
235 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
236 // BUT it can also return true and 'x' is in the outer loop
237 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
238 IdealLoopTree* n_loop_tree = get_loop(n_loop);
239 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
240 // x_loop_tree should be outer or same loop as n_loop_tree
241 return !x_loop_tree->is_member(n_loop_tree);
242 }
243
244 // Subtype checks that carry profile data don't common so look for a replacement by following edges
245 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
246 if (x->is_SubTypeCheck()) {
247 Node* in1 = x->in(1);
248 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
249 Node* u = in1->fast_out(i);
250 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
251 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
252 Node* bol = u->fast_out(j);
253 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
254 Node* iff = bol->fast_out(k);
255 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
256 // unrelated profile
257 if (iff->is_If() && is_dominator(iff, r_in)) {
258 return u;
259 }
260 }
261 }
262 }
263 }
264 }
265 return nullptr;
266 }
267
268 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
269 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
270 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
271 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
272 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
273 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
274 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
275 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
276 const Type* zero;
277 switch (n->Opcode()) {
278 case Op_DivI:
279 case Op_ModI:
280 zero = TypeInt::ZERO;
281 break;
282 case Op_DivL:
283 case Op_ModL:
284 zero = TypeLong::ZERO;
285 break;
286 default:
287 return false;
288 }
289
290 assert(n->in(0) == nullptr, "divisions with zero check should already have bailed out earlier in split-if");
291 Node* divisor = n->in(2);
292 return is_divisor_counted_loop_phi(divisor, region) &&
293 loop_phi_backedge_type_contains_zero(divisor, zero);
294 }
295
296 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) {
297 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop;
298 }
299
300 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
301 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
302 }
303
304 //------------------------------dominated_by------------------------------------
305 // Replace the dominated test with an obvious true or false. Place it on the
306 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
307 // live path up to the dominating control.
308 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
309 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
310
311 // prevdom is the dominating projection of the dominating test.
312 assert(iff->Opcode() == Op_If ||
313 iff->Opcode() == Op_CountedLoopEnd ||
314 iff->Opcode() == Op_LongCountedLoopEnd ||
315 iff->Opcode() == Op_RangeCheck ||
316 iff->Opcode() == Op_ParsePredicate,
317 "Check this code when new subtype is added");
318
319 int pop = prevdom->Opcode();
320 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
321 if (flip) {
322 if (pop == Op_IfTrue)
323 pop = Op_IfFalse;
324 else
325 pop = Op_IfTrue;
326 }
327 // 'con' is set to true or false to kill the dominated test.
328 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
329 set_ctrl(con, C->root()); // Constant gets a new use
330 // Hack the dominated test
331 _igvn.replace_input_of(iff, 1, con);
332
333 // If I dont have a reachable TRUE and FALSE path following the IfNode then
334 // I can assume this path reaches an infinite loop. In this case it's not
335 // important to optimize the data Nodes - either the whole compilation will
336 // be tossed or this path (and all data Nodes) will go dead.
337 if (iff->outcnt() != 2) return;
338
339 // Make control-dependent data Nodes on the live path (path that will remain
340 // once the dominated IF is removed) become control-dependent on the
341 // dominating projection.
342 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
343
344 // Loop predicates may have depending checks which should not
345 // be skipped. For example, range check predicate has two checks
346 // for lower and upper bounds.
347 if (dp == nullptr)
348 return;
349
350 ProjNode* dp_proj = dp->as_Proj();
351 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
352 if (exclude_loop_predicate &&
353 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr ||
354 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr ||
355 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != nullptr)) {
356 // If this is a range check (IfNode::is_range_check), do not
357 // reorder because Compile::allow_range_check_smearing might have
358 // changed the check.
359 return; // Let IGVN transformation change control dependence.
360 }
361
362 IdealLoopTree* old_loop = get_loop(dp);
363
364 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
365 Node* cd = dp->fast_out(i); // Control-dependent node
366 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
367 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
368 assert(cd->in(0) == dp, "");
369 _igvn.replace_input_of(cd, 0, prevdom);
370 set_early_ctrl(cd, false);
371 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
372 if (old_loop != new_loop) {
373 if (!old_loop->_child) {
374 old_loop->_body.yank(cd);
375 }
376 if (!new_loop->_child) {
377 new_loop->_body.push(cd);
378 }
379 }
380 --i;
381 --imax;
382 }
383 }
384 }
385
386 //------------------------------has_local_phi_input----------------------------
387 // Return TRUE if 'n' has Phi inputs from its local block and no other
388 // block-local inputs (all non-local-phi inputs come from earlier blocks)
389 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
390 Node *n_ctrl = get_ctrl(n);
391 // See if some inputs come from a Phi in this block, or from before
392 // this block.
393 uint i;
394 for( i = 1; i < n->req(); i++ ) {
395 Node *phi = n->in(i);
396 if( phi->is_Phi() && phi->in(0) == n_ctrl )
397 break;
398 }
399 if( i >= n->req() )
400 return nullptr; // No Phi inputs; nowhere to clone thru
401
402 // Check for inputs created between 'n' and the Phi input. These
403 // must split as well; they have already been given the chance
404 // (courtesy of a post-order visit) and since they did not we must
405 // recover the 'cost' of splitting them by being very profitable
406 // when splitting 'n'. Since this is unlikely we simply give up.
407 for( i = 1; i < n->req(); i++ ) {
408 Node *m = n->in(i);
409 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
410 // We allow the special case of AddP's with no local inputs.
411 // This allows us to split-up address expressions.
412 if (m->is_AddP() &&
413 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
414 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
415 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
416 // Move the AddP up to the dominating point. That's fine because control of m's inputs
417 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
418 Node* c = find_non_split_ctrl(idom(n_ctrl));
419 if (c->is_OuterStripMinedLoop()) {
420 c->as_Loop()->verify_strip_mined(1);
421 c = c->in(LoopNode::EntryControl);
422 }
423 set_ctrl_and_loop(m, c);
424 continue;
425 }
426 return nullptr;
427 }
428 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
429 }
430
431 return n_ctrl;
432 }
433
434 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
435 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
436 assert(bt == T_INT || bt == T_LONG, "only for integers");
437 int n_op = n->Opcode();
438
439 if (n_op == Op_LShift(bt)) {
440 // Scale is loop invariant
441 Node* scale = n->in(2);
442 Node* scale_ctrl = get_ctrl(scale);
443 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
444 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
445 return nullptr;
446 }
447 const TypeInt* scale_t = scale->bottom_type()->isa_int();
448 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
449 return nullptr; // Dont bother with byte/short masking
450 }
451 // Add must vary with loop (else shift would be loop-invariant)
452 Node* add = n->in(1);
453 Node* add_ctrl = get_ctrl(add);
454 IdealLoopTree* add_loop = get_loop(add_ctrl);
455 if (n_loop != add_loop) {
456 return nullptr; // happens w/ evil ZKM loops
457 }
458
459 // Convert I-V into I+ (0-V); same for V-I
460 if (add->Opcode() == Op_Sub(bt) &&
461 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
462 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
463 Node* zero = _igvn.integercon(0, bt);
464 set_ctrl(zero, C->root());
465 Node* neg = SubNode::make(zero, add->in(2), bt);
466 register_new_node(neg, get_ctrl(add->in(2)));
467 add = AddNode::make(add->in(1), neg, bt);
468 register_new_node(add, add_ctrl);
469 }
470 if (add->Opcode() != Op_Add(bt)) return nullptr;
471 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
472 // See if one add input is loop invariant
473 Node* add_var = add->in(1);
474 Node* add_var_ctrl = get_ctrl(add_var);
475 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
476 Node* add_invar = add->in(2);
477 Node* add_invar_ctrl = get_ctrl(add_invar);
478 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
479 if (add_invar_loop == n_loop) {
480 // Swap to find the invariant part
481 add_invar = add_var;
482 add_invar_ctrl = add_var_ctrl;
483 add_invar_loop = add_var_loop;
484 add_var = add->in(2);
485 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
486 return nullptr;
487 }
488 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
489 return nullptr; // No invariant part of the add?
490 }
491
492 // Yes! Reshape address expression!
493 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
494 Node* inv_scale_ctrl =
495 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
496 add_invar_ctrl : scale_ctrl;
497 register_new_node(inv_scale, inv_scale_ctrl);
498 Node* var_scale = LShiftNode::make(add_var, scale, bt);
499 register_new_node(var_scale, n_ctrl);
500 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
501 register_new_node(var_add, n_ctrl);
502 _igvn.replace_node(n, var_add);
503 return var_add;
504 }
505 return nullptr;
506 }
507
508 //------------------------------remix_address_expressions----------------------
509 // Rework addressing expressions to get the most loop-invariant stuff
510 // moved out. We'd like to do all associative operators, but it's especially
511 // important (common) to do address expressions.
512 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
513 if (!has_ctrl(n)) return nullptr;
514 Node* n_ctrl = get_ctrl(n);
515 IdealLoopTree* n_loop = get_loop(n_ctrl);
516
517 // See if 'n' mixes loop-varying and loop-invariant inputs and
518 // itself is loop-varying.
519
520 // Only interested in binary ops (and AddP)
521 if (n->req() < 3 || n->req() > 4) return nullptr;
522
523 Node* n1_ctrl = get_ctrl(n->in( 1));
524 Node* n2_ctrl = get_ctrl(n->in( 2));
525 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
526 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
527 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
528 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
529
530 // Does one of my inputs spin in a tighter loop than self?
531 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
532 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
533 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
534 return nullptr; // Leave well enough alone
535 }
536
537 // Is at least one of my inputs loop-invariant?
538 if (n1_loop == n_loop &&
539 n2_loop == n_loop &&
540 n3_loop == n_loop) {
541 return nullptr; // No loop-invariant inputs
542 }
543
544 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
545 if (res != nullptr) {
546 return res;
547 }
548 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
549 if (res != nullptr) {
550 return res;
551 }
552
553 int n_op = n->Opcode();
554 // Replace (I+V) with (V+I)
555 if (n_op == Op_AddI ||
556 n_op == Op_AddL ||
557 n_op == Op_AddF ||
558 n_op == Op_AddD ||
559 n_op == Op_MulI ||
560 n_op == Op_MulL ||
561 n_op == Op_MulF ||
562 n_op == Op_MulD) {
563 if (n2_loop == n_loop) {
564 assert(n1_loop != n_loop, "");
565 n->swap_edges(1, 2);
566 }
567 }
568
569 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
570 // but not if I2 is a constant.
571 if (n_op == Op_AddP) {
572 if (n2_loop == n_loop && n3_loop != n_loop) {
573 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
574 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
575 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
576 IdealLoopTree* n22loop = get_loop(n22_ctrl);
577 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
578 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
579 n23_loop == n_loop) {
580 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
581 // Stuff new AddP in the loop preheader
582 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
583 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
584 register_new_node(add2, n_ctrl);
585 _igvn.replace_node(n, add2);
586 return add2;
587 }
588 }
589 }
590
591 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
592 if (n2_loop != n_loop && n3_loop == n_loop) {
593 if (n->in(3)->Opcode() == Op_AddX) {
594 Node* V = n->in(3)->in(1);
595 Node* I = n->in(3)->in(2);
596 if (is_member(n_loop,get_ctrl(V))) {
597 } else {
598 Node *tmp = V; V = I; I = tmp;
599 }
600 if (!is_member(n_loop,get_ctrl(I))) {
601 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
602 // Stuff new AddP in the loop preheader
603 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
604 Node* add2 = new AddPNode(n->in(1), add1, V);
605 register_new_node(add2, n_ctrl);
606 _igvn.replace_node(n, add2);
607 return add2;
608 }
609 }
610 }
611 }
612
613 return nullptr;
614 }
615
616 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
617 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
618 assert(n->Opcode() == Op_AddI, "sanity");
619 Node * nn = nullptr;
620 Node * in1 = n->in(1);
621 Node * in2 = n->in(2);
622 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
623 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
624 if (loop_n->is_counted() &&
625 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
626 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
627 Matcher::match_rule_supported(Op_MulAddS2I)) {
628 Node* mul_in1 = in1->in(1);
629 Node* mul_in2 = in1->in(2);
630 Node* mul_in3 = in2->in(1);
631 Node* mul_in4 = in2->in(2);
632 if (mul_in1->Opcode() == Op_LoadS &&
633 mul_in2->Opcode() == Op_LoadS &&
634 mul_in3->Opcode() == Op_LoadS &&
635 mul_in4->Opcode() == Op_LoadS) {
636 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
637 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
638 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
639 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
640 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
641 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
642 // All nodes should be in the same counted loop.
643 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
644 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
645 Node* adr1 = mul_in1->in(MemNode::Address);
646 Node* adr2 = mul_in2->in(MemNode::Address);
647 Node* adr3 = mul_in3->in(MemNode::Address);
648 Node* adr4 = mul_in4->in(MemNode::Address);
649 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
650 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
651 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
652 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
653 register_new_node(nn, get_ctrl(n));
654 _igvn.replace_node(n, nn);
655 return nn;
656 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
657 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
658 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
659 register_new_node(nn, get_ctrl(n));
660 _igvn.replace_node(n, nn);
661 return nn;
662 }
663 }
664 }
665 }
666 }
667 }
668 return nn;
669 }
670
671 //------------------------------conditional_move-------------------------------
672 // Attempt to replace a Phi with a conditional move. We have some pretty
673 // strict profitability requirements. All Phis at the merge point must
674 // be converted, so we can remove the control flow. We need to limit the
675 // number of c-moves to a small handful. All code that was in the side-arms
676 // of the CFG diamond is now speculatively executed. This code has to be
677 // "cheap enough". We are pretty much limited to CFG diamonds that merge
678 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
679 Node *PhaseIdealLoop::conditional_move( Node *region ) {
680
681 assert(region->is_Region(), "sanity check");
682 if (region->req() != 3) return nullptr;
683
684 // Check for CFG diamond
685 Node *lp = region->in(1);
686 Node *rp = region->in(2);
687 if (!lp || !rp) return nullptr;
688 Node *lp_c = lp->in(0);
689 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
690 IfNode *iff = lp_c->as_If();
691
692 // Check for ops pinned in an arm of the diamond.
693 // Can't remove the control flow in this case
694 if (lp->outcnt() > 1) return nullptr;
695 if (rp->outcnt() > 1) return nullptr;
696
697 IdealLoopTree* r_loop = get_loop(region);
698 assert(r_loop == get_loop(iff), "sanity");
699 // Always convert to CMOVE if all results are used only outside this loop.
700 bool used_inside_loop = (r_loop == _ltree_root);
701
702 // Check profitability
703 int cost = 0;
704 int phis = 0;
705 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
706 Node *out = region->fast_out(i);
707 if (!out->is_Phi()) continue; // Ignore other control edges, etc
708 phis++;
709 PhiNode* phi = out->as_Phi();
710 BasicType bt = phi->type()->basic_type();
711 switch (bt) {
712 case T_DOUBLE:
713 case T_FLOAT:
714 if (C->use_cmove()) {
715 continue; //TODO: maybe we want to add some cost
716 }
717 cost += Matcher::float_cmove_cost(); // Could be very expensive
718 break;
719 case T_LONG: {
720 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
721 }
722 case T_INT: // These all CMOV fine
723 case T_ADDRESS: { // (RawPtr)
724 cost++;
725 break;
726 }
727 case T_NARROWOOP: // Fall through
728 case T_OBJECT: { // Base oops are OK, but not derived oops
729 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
730 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
731 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
732 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
733 // have a Phi for the base here that we convert to a CMOVE all is well
734 // and good. But if the base is dead, we'll not make a CMOVE. Later
735 // the allocator will have to produce a base by creating a CMOVE of the
736 // relevant bases. This puts the allocator in the business of
737 // manufacturing expensive instructions, generally a bad plan.
738 // Just Say No to Conditionally-Moved Derived Pointers.
739 if (tp && tp->offset() != 0)
740 return nullptr;
741 cost++;
742 break;
743 }
744 default:
745 return nullptr; // In particular, can't do memory or I/O
746 }
747 // Add in cost any speculative ops
748 for (uint j = 1; j < region->req(); j++) {
749 Node *proj = region->in(j);
750 Node *inp = phi->in(j);
751 if (get_ctrl(inp) == proj) { // Found local op
752 cost++;
753 // Check for a chain of dependent ops; these will all become
754 // speculative in a CMOV.
755 for (uint k = 1; k < inp->req(); k++)
756 if (get_ctrl(inp->in(k)) == proj)
757 cost += ConditionalMoveLimit; // Too much speculative goo
758 }
759 }
760 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
761 // This will likely Split-If, a higher-payoff operation.
762 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
763 Node* use = phi->fast_out(k);
764 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
765 cost += ConditionalMoveLimit;
766 // Is there a use inside the loop?
767 // Note: check only basic types since CMoveP is pinned.
768 if (!used_inside_loop && is_java_primitive(bt)) {
769 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
770 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
771 used_inside_loop = true;
772 }
773 }
774 }
775 }//for
776 Node* bol = iff->in(1);
777 if (bol->Opcode() == Op_Opaque4) {
778 return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
779 }
780 assert(bol->Opcode() == Op_Bool, "Unexpected node");
781 int cmp_op = bol->in(1)->Opcode();
782 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
783 return nullptr;
784 }
785 // It is expensive to generate flags from a float compare.
786 // Avoid duplicated float compare.
787 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
788
789 float infrequent_prob = PROB_UNLIKELY_MAG(3);
790 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
791 if (used_inside_loop) {
792 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
793
794 // BlockLayoutByFrequency optimization moves infrequent branch
795 // from hot path. No point in CMOV'ing in such case (110 is used
796 // instead of 100 to take into account not exactness of float value).
797 if (BlockLayoutByFrequency) {
798 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
799 }
800 }
801 // Check for highly predictable branch. No point in CMOV'ing if
802 // we are going to predict accurately all the time.
803 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
804 //keep going
805 } else if (iff->_prob < infrequent_prob ||
806 iff->_prob > (1.0f - infrequent_prob))
807 return nullptr;
808
809 // --------------
810 // Now replace all Phis with CMOV's
811 Node *cmov_ctrl = iff->in(0);
812 uint flip = (lp->Opcode() == Op_IfTrue);
813 Node_List wq;
814 while (1) {
815 PhiNode* phi = nullptr;
816 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
817 Node *out = region->fast_out(i);
818 if (out->is_Phi()) {
819 phi = out->as_Phi();
820 break;
821 }
822 }
823 if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
824 break;
825 }
826 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
827 // Move speculative ops
828 wq.push(phi);
829 while (wq.size() > 0) {
830 Node *n = wq.pop();
831 for (uint j = 1; j < n->req(); j++) {
832 Node* m = n->in(j);
833 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
834 #ifndef PRODUCT
835 if (PrintOpto && VerifyLoopOptimizations) {
836 tty->print(" speculate: ");
837 m->dump();
838 }
839 #endif
840 set_ctrl(m, cmov_ctrl);
841 wq.push(m);
842 }
843 }
844 }
845 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
846 register_new_node( cmov, cmov_ctrl );
847 _igvn.replace_node( phi, cmov );
848 #ifndef PRODUCT
849 if (TraceLoopOpts) {
850 tty->print("CMOV ");
851 r_loop->dump_head();
852 if (Verbose) {
853 bol->in(1)->dump(1);
854 cmov->dump(1);
855 }
856 }
857 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
858 #endif
859 }
860
861 // The useless CFG diamond will fold up later; see the optimization in
862 // RegionNode::Ideal.
863 _igvn._worklist.push(region);
864
865 return iff->in(1);
866 }
867
868 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
869 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
870 Node* u = m->fast_out(i);
871 if (u->is_CFG()) {
872 if (u->is_NeverBranch()) {
873 u = u->as_NeverBranch()->proj_out(0);
874 enqueue_cfg_uses(u, wq);
875 } else {
876 wq.push(u);
877 }
878 }
879 }
880 }
881
882 // Try moving a store out of a loop, right before the loop
883 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
884 // Store has to be first in the loop body
885 IdealLoopTree *n_loop = get_loop(n_ctrl);
886 if (n->is_Store() && n_loop != _ltree_root &&
887 n_loop->is_loop() && n_loop->_head->is_Loop() &&
888 n->in(0) != nullptr) {
889 Node* address = n->in(MemNode::Address);
890 Node* value = n->in(MemNode::ValueIn);
891 Node* mem = n->in(MemNode::Memory);
892 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
893 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
894
895 // - address and value must be loop invariant
896 // - memory must be a memory Phi for the loop
897 // - Store must be the only store on this memory slice in the
898 // loop: if there's another store following this one then value
899 // written at iteration i by the second store could be overwritten
900 // at iteration i+n by the first store: it's not safe to move the
901 // first store out of the loop
902 // - nothing must observe the memory Phi: it guarantees no read
903 // before the store, we are also guaranteed the store post
904 // dominates the loop head (ignoring a possible early
905 // exit). Otherwise there would be extra Phi involved between the
906 // loop's Phi and the store.
907 // - there must be no early exit from the loop before the Store
908 // (such an exit most of the time would be an extra use of the
909 // memory Phi but sometimes is a bottom memory Phi that takes the
910 // store as input).
911
912 if (!n_loop->is_member(address_loop) &&
913 !n_loop->is_member(value_loop) &&
914 mem->is_Phi() && mem->in(0) == n_loop->_head &&
915 mem->outcnt() == 1 &&
916 mem->in(LoopNode::LoopBackControl) == n) {
917
918 assert(n_loop->_tail != nullptr, "need a tail");
919 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
920
921 // Verify that there's no early exit of the loop before the store.
922 bool ctrl_ok = false;
923 {
924 // Follow control from loop head until n, we exit the loop or
925 // we reach the tail
926 ResourceMark rm;
927 Unique_Node_List wq;
928 wq.push(n_loop->_head);
929
930 for (uint next = 0; next < wq.size(); ++next) {
931 Node *m = wq.at(next);
932 if (m == n->in(0)) {
933 ctrl_ok = true;
934 continue;
935 }
936 assert(!has_ctrl(m), "should be CFG");
937 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
938 ctrl_ok = false;
939 break;
940 }
941 enqueue_cfg_uses(m, wq);
942 if (wq.size() > 10) {
943 ctrl_ok = false;
944 break;
945 }
946 }
947 }
948 if (ctrl_ok) {
949 // move the Store
950 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
951 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
952 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
953 // Disconnect the phi now. An empty phi can confuse other
954 // optimizations in this pass of loop opts.
955 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
956 n_loop->_body.yank(mem);
957
958 set_ctrl_and_loop(n, n->in(0));
959
960 return n;
961 }
962 }
963 }
964 return nullptr;
965 }
966
967 // Try moving a store out of a loop, right after the loop
968 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
969 if (n->is_Store() && n->in(0) != nullptr) {
970 Node *n_ctrl = get_ctrl(n);
971 IdealLoopTree *n_loop = get_loop(n_ctrl);
972 // Store must be in a loop
973 if (n_loop != _ltree_root && !n_loop->_irreducible) {
974 Node* address = n->in(MemNode::Address);
975 Node* value = n->in(MemNode::ValueIn);
976 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
977 // address must be loop invariant
978 if (!n_loop->is_member(address_loop)) {
979 // Store must be last on this memory slice in the loop and
980 // nothing in the loop must observe it
981 Node* phi = nullptr;
982 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
983 Node* u = n->fast_out(i);
984 if (has_ctrl(u)) { // control use?
985 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
986 if (!n_loop->is_member(u_loop)) {
987 continue;
988 }
989 if (u->is_Phi() && u->in(0) == n_loop->_head) {
990 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
991 // multiple phis on the same slice are possible
992 if (phi != nullptr) {
993 return;
994 }
995 phi = u;
996 continue;
997 }
998 }
999 return;
1000 }
1001 if (phi != nullptr) {
1002 // Nothing in the loop before the store (next iteration)
1003 // must observe the stored value
1004 bool mem_ok = true;
1005 {
1006 ResourceMark rm;
1007 Unique_Node_List wq;
1008 wq.push(phi);
1009 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1010 Node *m = wq.at(next);
1011 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1012 Node* u = m->fast_out(i);
1013 if (u->is_Store() || u->is_Phi()) {
1014 if (u != n) {
1015 wq.push(u);
1016 mem_ok = (wq.size() <= 10);
1017 }
1018 } else {
1019 mem_ok = false;
1020 break;
1021 }
1022 }
1023 }
1024 }
1025 if (mem_ok) {
1026 // Move the store out of the loop if the LCA of all
1027 // users (except for the phi) is outside the loop.
1028 Node* hook = new Node(1);
1029 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1030 _igvn.rehash_node_delayed(phi);
1031 int count = phi->replace_edge(n, hook, &_igvn);
1032 assert(count > 0, "inconsistent phi");
1033
1034 // Compute latest point this store can go
1035 Node* lca = get_late_ctrl(n, get_ctrl(n));
1036 if (lca->is_OuterStripMinedLoop()) {
1037 lca = lca->in(LoopNode::EntryControl);
1038 }
1039 if (n_loop->is_member(get_loop(lca))) {
1040 // LCA is in the loop - bail out
1041 _igvn.replace_node(hook, n);
1042 return;
1043 }
1044 #ifdef ASSERT
1045 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1046 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1047 n_loop->_head->as_Loop()->verify_strip_mined(1);
1048 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1049 IdealLoopTree* outer_loop = get_loop(outer);
1050 assert(n_loop->_parent == outer_loop, "broken loop tree");
1051 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1052 }
1053 #endif
1054 lca = place_outside_loop(lca, n_loop);
1055 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1056 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop");
1057
1058 // Move store out of the loop
1059 _igvn.replace_node(hook, n->in(MemNode::Memory));
1060 _igvn.replace_input_of(n, 0, lca);
1061 set_ctrl_and_loop(n, lca);
1062
1063 // Disconnect the phi now. An empty phi can confuse other
1064 // optimizations in this pass of loop opts..
1065 if (phi->in(LoopNode::LoopBackControl) == phi) {
1066 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1067 n_loop->_body.yank(phi);
1068 }
1069 }
1070 }
1071 }
1072 }
1073 }
1074 }
1075
1076 //------------------------------split_if_with_blocks_pre-----------------------
1077 // Do the real work in a non-recursive function. Data nodes want to be
1078 // cloned in the pre-order so they can feed each other nicely.
1079 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1080 // Cloning these guys is unlikely to win
1081 int n_op = n->Opcode();
1082 if (n_op == Op_MergeMem) {
1083 return n;
1084 }
1085 if (n->is_Proj()) {
1086 return n;
1087 }
1088 // Do not clone-up CmpFXXX variations, as these are always
1089 // followed by a CmpI
1090 if (n->is_Cmp()) {
1091 return n;
1092 }
1093 // Attempt to use a conditional move instead of a phi/branch
1094 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1095 Node *cmov = conditional_move( n );
1096 if (cmov) {
1097 return cmov;
1098 }
1099 }
1100 if (n->is_CFG() || n->is_LoadStore()) {
1101 return n;
1102 }
1103 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1104 if (!C->major_progress()) { // If chance of no more loop opts...
1105 _igvn._worklist.push(n); // maybe we'll remove them
1106 }
1107 return n;
1108 }
1109
1110 if (n->is_Con()) {
1111 return n; // No cloning for Con nodes
1112 }
1113
1114 Node *n_ctrl = get_ctrl(n);
1115 if (!n_ctrl) {
1116 return n; // Dead node
1117 }
1118
1119 Node* res = try_move_store_before_loop(n, n_ctrl);
1120 if (res != nullptr) {
1121 return n;
1122 }
1123
1124 // Attempt to remix address expressions for loop invariants
1125 Node *m = remix_address_expressions( n );
1126 if( m ) return m;
1127
1128 if (n_op == Op_AddI) {
1129 Node *nn = convert_add_to_muladd( n );
1130 if ( nn ) return nn;
1131 }
1132
1133 if (n->is_ConstraintCast()) {
1134 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1135 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1136 // Node control inputs don't necessarily agree with loop control info (due to
1137 // transformations happened in between), thus additional dominance check is needed
1138 // to keep loop info valid.
1139 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1140 _igvn.replace_node(n, dom_cast);
1141 return dom_cast;
1142 }
1143 }
1144
1145 // Determine if the Node has inputs from some local Phi.
1146 // Returns the block to clone thru.
1147 Node *n_blk = has_local_phi_input( n );
1148 if( !n_blk ) return n;
1149
1150 // Do not clone the trip counter through on a CountedLoop
1151 // (messes up the canonical shape).
1152 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1153 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1154 return n;
1155 }
1156 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1157 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1158 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1159 return n;
1160 }
1161
1162 // Check for having no control input; not pinned. Allow
1163 // dominating control.
1164 if (n->in(0)) {
1165 Node *dom = idom(n_blk);
1166 if (dom_lca(n->in(0), dom) != n->in(0)) {
1167 return n;
1168 }
1169 }
1170 // Policy: when is it profitable. You must get more wins than
1171 // policy before it is considered profitable. Policy is usually 0,
1172 // so 1 win is considered profitable. Big merges will require big
1173 // cloning, so get a larger policy.
1174 int policy = n_blk->req() >> 2;
1175
1176 // If the loop is a candidate for range check elimination,
1177 // delay splitting through it's phi until a later loop optimization
1178 if (n_blk->is_BaseCountedLoop()) {
1179 IdealLoopTree *lp = get_loop(n_blk);
1180 if (lp && lp->_rce_candidate) {
1181 return n;
1182 }
1183 }
1184
1185 if (must_throttle_split_if()) return n;
1186
1187 // Split 'n' through the merge point if it is profitable
1188 Node *phi = split_thru_phi( n, n_blk, policy );
1189 if (!phi) return n;
1190
1191 // Found a Phi to split thru!
1192 // Replace 'n' with the new phi
1193 _igvn.replace_node( n, phi );
1194 // Moved a load around the loop, 'en-registering' something.
1195 if (n_blk->is_Loop() && n->is_Load() &&
1196 !phi->in(LoopNode::LoopBackControl)->is_Load())
1197 C->set_major_progress();
1198
1199 return phi;
1200 }
1201
1202 static bool merge_point_too_heavy(Compile* C, Node* region) {
1203 // Bail out if the region and its phis have too many users.
1204 int weight = 0;
1205 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1206 weight += region->fast_out(i)->outcnt();
1207 }
1208 int nodes_left = C->max_node_limit() - C->live_nodes();
1209 if (weight * 8 > nodes_left) {
1210 if (PrintOpto) {
1211 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1212 }
1213 return true;
1214 } else {
1215 return false;
1216 }
1217 }
1218
1219 static bool merge_point_safe(Node* region) {
1220 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1221 // having a PhiNode input. This sidesteps the dangerous case where the split
1222 // ConvI2LNode may become TOP if the input Value() does not
1223 // overlap the ConvI2L range, leaving a node which may not dominate its
1224 // uses.
1225 // A better fix for this problem can be found in the BugTraq entry, but
1226 // expediency for Mantis demands this hack.
1227 #ifdef _LP64
1228 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1229 Node* n = region->fast_out(i);
1230 if (n->is_Phi()) {
1231 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1232 Node* m = n->fast_out(j);
1233 if (m->Opcode() == Op_ConvI2L)
1234 return false;
1235 if (m->is_CastII()) {
1236 return false;
1237 }
1238 }
1239 }
1240 }
1241 #endif
1242 return true;
1243 }
1244
1245
1246 //------------------------------place_outside_loop---------------------------------
1247 // Place some computation outside of this loop on the path to the use passed as argument
1248 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1249 Node* head = loop->_head;
1250 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1251 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1252 loop = loop->_parent;
1253 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1254 }
1255
1256 // Pick control right outside the loop
1257 for (;;) {
1258 Node* dom = idom(useblock);
1259 if (loop->is_member(get_loop(dom)) ||
1260 // NeverBranch nodes are not assigned to the loop when constructed
1261 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) {
1262 break;
1263 }
1264 useblock = dom;
1265 }
1266 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1267 return useblock;
1268 }
1269
1270
1271 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1272 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1273 return false;
1274 }
1275 if (!n->in(0)->is_Region()) {
1276 return false;
1277 }
1278
1279 Node* region = n->in(0);
1280 Node* dom = idom(region);
1281 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1282 return false;
1283 }
1284 IfNode* dom_if = dom->as_If();
1285 Node* proj_true = dom_if->proj_out(1);
1286 Node* proj_false = dom_if->proj_out(0);
1287
1288 for (uint i = 1; i < region->req(); i++) {
1289 if (is_dominator(proj_true, region->in(i))) {
1290 continue;
1291 }
1292 if (is_dominator(proj_false, region->in(i))) {
1293 continue;
1294 }
1295 return false;
1296 }
1297
1298 return true;
1299 }
1300
1301
1302 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1303 if (must_throttle_split_if()) {
1304 return false;
1305 }
1306
1307 // Do not do 'split-if' if irreducible loops are present.
1308 if (_has_irreducible_loops) {
1309 return false;
1310 }
1311
1312 if (merge_point_too_heavy(C, n_ctrl)) {
1313 return false;
1314 }
1315
1316 // Do not do 'split-if' if some paths are dead. First do dead code
1317 // elimination and then see if its still profitable.
1318 for (uint i = 1; i < n_ctrl->req(); i++) {
1319 if (n_ctrl->in(i) == C->top()) {
1320 return false;
1321 }
1322 }
1323
1324 // If trying to do a 'Split-If' at the loop head, it is only
1325 // profitable if the cmp folds up on BOTH paths. Otherwise we
1326 // risk peeling a loop forever.
1327
1328 // CNC - Disabled for now. Requires careful handling of loop
1329 // body selection for the cloned code. Also, make sure we check
1330 // for any input path not being in the same loop as n_ctrl. For
1331 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1332 // because the alternative loop entry points won't be converted
1333 // into LoopNodes.
1334 IdealLoopTree *n_loop = get_loop(n_ctrl);
1335 for (uint j = 1; j < n_ctrl->req(); j++) {
1336 if (get_loop(n_ctrl->in(j)) != n_loop) {
1337 return false;
1338 }
1339 }
1340
1341 // Check for safety of the merge point.
1342 if (!merge_point_safe(n_ctrl)) {
1343 return false;
1344 }
1345
1346 return true;
1347 }
1348
1349 // Detect if the node is the inner strip-mined loop
1350 // Return: null if it's not the case, or the exit of outer strip-mined loop
1351 static Node* is_inner_of_stripmined_loop(const Node* out) {
1352 Node* out_le = nullptr;
1353
1354 if (out->is_CountedLoopEnd()) {
1355 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1356
1357 if (loop != nullptr && loop->is_strip_mined()) {
1358 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1359 }
1360 }
1361
1362 return out_le;
1363 }
1364
1365 //------------------------------split_if_with_blocks_post----------------------
1366 // Do the real work in a non-recursive function. CFG hackery wants to be
1367 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1368 // info.
1369 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1370
1371 // Cloning Cmp through Phi's involves the split-if transform.
1372 // FastLock is not used by an If
1373 if (n->is_Cmp() && !n->is_FastLock()) {
1374 Node *n_ctrl = get_ctrl(n);
1375 // Determine if the Node has inputs from some local Phi.
1376 // Returns the block to clone thru.
1377 Node *n_blk = has_local_phi_input(n);
1378 if (n_blk != n_ctrl) {
1379 return;
1380 }
1381
1382 if (!can_split_if(n_ctrl)) {
1383 return;
1384 }
1385
1386 if (n->outcnt() != 1) {
1387 return; // Multiple bool's from 1 compare?
1388 }
1389 Node *bol = n->unique_out();
1390 assert(bol->is_Bool(), "expect a bool here");
1391 if (bol->outcnt() != 1) {
1392 return;// Multiple branches from 1 compare?
1393 }
1394 Node *iff = bol->unique_out();
1395
1396 // Check some safety conditions
1397 if (iff->is_If()) { // Classic split-if?
1398 if (iff->in(0) != n_ctrl) {
1399 return; // Compare must be in same blk as if
1400 }
1401 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1402 // Can't split CMove with different control.
1403 if (get_ctrl(iff) != n_ctrl) {
1404 return;
1405 }
1406 if (get_ctrl(iff->in(2)) == n_ctrl ||
1407 get_ctrl(iff->in(3)) == n_ctrl) {
1408 return; // Inputs not yet split-up
1409 }
1410 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1411 return; // Loop-invar test gates loop-varying CMOVE
1412 }
1413 } else {
1414 return; // some other kind of node, such as an Allocate
1415 }
1416
1417 // When is split-if profitable? Every 'win' on means some control flow
1418 // goes dead, so it's almost always a win.
1419 int policy = 0;
1420 // Split compare 'n' through the merge point if it is profitable
1421 Node *phi = split_thru_phi( n, n_ctrl, policy);
1422 if (!phi) {
1423 return;
1424 }
1425
1426 // Found a Phi to split thru!
1427 // Replace 'n' with the new phi
1428 _igvn.replace_node(n, phi);
1429
1430 // Now split the bool up thru the phi
1431 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1432 guarantee(bolphi != nullptr, "null boolean phi node");
1433
1434 _igvn.replace_node(bol, bolphi);
1435 assert(iff->in(1) == bolphi, "");
1436
1437 if (bolphi->Value(&_igvn)->singleton()) {
1438 return;
1439 }
1440
1441 // Conditional-move? Must split up now
1442 if (!iff->is_If()) {
1443 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1444 _igvn.replace_node(iff, cmovphi);
1445 return;
1446 }
1447
1448 // Now split the IF
1449 do_split_if(iff);
1450 return;
1451 }
1452
1453 // Two identical ifs back to back can be merged
1454 if (try_merge_identical_ifs(n)) {
1455 return;
1456 }
1457
1458 // Check for an IF ready to split; one that has its
1459 // condition codes input coming from a Phi at the block start.
1460 int n_op = n->Opcode();
1461
1462 // Check for an IF being dominated by another IF same test
1463 if (n_op == Op_If ||
1464 n_op == Op_RangeCheck) {
1465 Node *bol = n->in(1);
1466 uint max = bol->outcnt();
1467 // Check for same test used more than once?
1468 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1469 // Search up IDOMs to see if this IF is dominated.
1470 Node* cmp = bol->in(1);
1471 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1472
1473 // Now search up IDOMs till cutoff, looking for a dominating test
1474 Node *prevdom = n;
1475 Node *dom = idom(prevdom);
1476 while (dom != cutoff) {
1477 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1478 safe_for_if_replacement(dom)) {
1479 // It's invalid to move control dependent data nodes in the inner
1480 // strip-mined loop, because:
1481 // 1) break validation of LoopNode::verify_strip_mined()
1482 // 2) move code with side-effect in strip-mined loop
1483 // Move to the exit of outer strip-mined loop in that case.
1484 Node* out_le = is_inner_of_stripmined_loop(dom);
1485 if (out_le != nullptr) {
1486 prevdom = out_le;
1487 }
1488 // Replace the dominated test with an obvious true or false.
1489 // Place it on the IGVN worklist for later cleanup.
1490 C->set_major_progress();
1491 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
1492 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1493 return;
1494 }
1495 prevdom = dom;
1496 dom = idom(prevdom);
1497 }
1498 }
1499 }
1500
1501 try_sink_out_of_loop(n);
1502
1503 try_move_store_after_loop(n);
1504 }
1505
1506 // Transform:
1507 //
1508 // if (some_condition) {
1509 // // body 1
1510 // } else {
1511 // // body 2
1512 // }
1513 // if (some_condition) {
1514 // // body 3
1515 // } else {
1516 // // body 4
1517 // }
1518 //
1519 // into:
1520 //
1521 //
1522 // if (some_condition) {
1523 // // body 1
1524 // // body 3
1525 // } else {
1526 // // body 2
1527 // // body 4
1528 // }
1529 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1530 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1531 Node *n_ctrl = n->in(0);
1532 IfNode* dom_if = idom(n_ctrl)->as_If();
1533 if (n->in(1) != dom_if->in(1)) {
1534 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1535 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1536 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1537 _igvn.replace_input_of(n, 1, dom_if->in(1));
1538 }
1539 ProjNode* dom_proj_true = dom_if->proj_out(1);
1540 ProjNode* dom_proj_false = dom_if->proj_out(0);
1541
1542 // Now split the IF
1543 RegionNode* new_false_region;
1544 RegionNode* new_true_region;
1545 do_split_if(n, &new_false_region, &new_true_region);
1546 assert(new_false_region->req() == new_true_region->req(), "");
1547 #ifdef ASSERT
1548 for (uint i = 1; i < new_false_region->req(); ++i) {
1549 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1550 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");
1551 }
1552 #endif
1553 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1554
1555 // We now have:
1556 // if (some_condition) {
1557 // // body 1
1558 // if (some_condition) {
1559 // body3: // new_true_region
1560 // // body3
1561 // } else {
1562 // goto body4;
1563 // }
1564 // } else {
1565 // // body 2
1566 // if (some_condition) {
1567 // goto body3;
1568 // } else {
1569 // body4: // new_false_region
1570 // // body4;
1571 // }
1572 // }
1573 //
1574
1575 // clone pinned nodes thru the resulting regions
1576 push_pinned_nodes_thru_region(dom_if, new_true_region);
1577 push_pinned_nodes_thru_region(dom_if, new_false_region);
1578
1579 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1580 // 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
1581 // unrelated control dependency.
1582 for (uint i = 1; i < new_false_region->req(); i++) {
1583 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1584 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1585 } else {
1586 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1587 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1588 }
1589 }
1590 return true;
1591 }
1592 return false;
1593 }
1594
1595 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1596 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1597 Node* u = region->out(i);
1598 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1599 continue;
1600 }
1601 assert(u->in(0) == region, "not a control dependent node?");
1602 uint j = 1;
1603 for (; j < u->req(); ++j) {
1604 Node* in = u->in(j);
1605 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1606 break;
1607 }
1608 }
1609 if (j == u->req()) {
1610 Node *phi = PhiNode::make_blank(region, u);
1611 for (uint k = 1; k < region->req(); ++k) {
1612 Node* clone = u->clone();
1613 clone->set_req(0, region->in(k));
1614 register_new_node(clone, region->in(k));
1615 phi->init_req(k, clone);
1616 }
1617 register_new_node(phi, region);
1618 _igvn.replace_node(u, phi);
1619 --i;
1620 }
1621 }
1622 }
1623
1624 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1625 if (!dom->is_CountedLoopEnd()) {
1626 return true;
1627 }
1628 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1629 CountedLoopNode* cl = le->loopnode();
1630 if (cl == nullptr) {
1631 return true;
1632 }
1633 if (!cl->is_main_loop()) {
1634 return true;
1635 }
1636 if (cl->is_canonical_loop_entry() == nullptr) {
1637 return true;
1638 }
1639 // Further unrolling is possible so loop exit condition might change
1640 return false;
1641 }
1642
1643 // See if a shared loop-varying computation has no loop-varying uses.
1644 // Happens if something is only used for JVM state in uncommon trap exits,
1645 // like various versions of induction variable+offset. Clone the
1646 // computation per usage to allow it to sink out of the loop.
1647 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1648 if (has_ctrl(n) &&
1649 !n->is_Phi() &&
1650 !n->is_Bool() &&
1651 !n->is_Proj() &&
1652 !n->is_MergeMem() &&
1653 !n->is_CMove() &&
1654 n->Opcode() != Op_Opaque4 &&
1655 !n->is_Type()) {
1656 Node *n_ctrl = get_ctrl(n);
1657 IdealLoopTree *n_loop = get_loop(n_ctrl);
1658
1659 if (n->in(0) != nullptr) {
1660 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1661 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1662 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1663 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1664 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1665 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1666 }
1667 }
1668 if (n_loop != _ltree_root && n->outcnt() > 1) {
1669 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1670 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1671 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1672 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1673 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1674 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1675 Node* outer_loop_clone = nullptr;
1676 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1677 Node* u = n->last_out(j); // Clone private computation per use
1678 _igvn.rehash_node_delayed(u);
1679 Node* x = n->clone(); // Clone computation
1680 Node* x_ctrl = nullptr;
1681 if (u->is_Phi()) {
1682 // Replace all uses of normal nodes. Replace Phi uses
1683 // individually, so the separate Nodes can sink down
1684 // different paths.
1685 uint k = 1;
1686 while (u->in(k) != n) k++;
1687 u->set_req(k, x);
1688 // x goes next to Phi input path
1689 x_ctrl = u->in(0)->in(k);
1690 // Find control for 'x' next to use but not inside inner loops.
1691 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1692 --j;
1693 } else { // Normal use
1694 if (has_ctrl(u)) {
1695 x_ctrl = get_ctrl(u);
1696 } else {
1697 x_ctrl = u->in(0);
1698 }
1699 // Find control for 'x' next to use but not inside inner loops.
1700 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1701 // Replace all uses
1702 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1703 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1704 // anymore now that we're going to pin n as well
1705 _igvn.replace_node(u, x);
1706 --j;
1707 } else {
1708 int nb = u->replace_edge(n, x, &_igvn);
1709 j -= nb;
1710 }
1711 }
1712
1713 if (n->is_Load()) {
1714 // For loads, add a control edge to a CFG node outside of the loop
1715 // to force them to not combine and return back inside the loop
1716 // during GVN optimization (4641526).
1717 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1718
1719 IdealLoopTree* x_loop = get_loop(x_ctrl);
1720 Node* x_head = x_loop->_head;
1721 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1722 // Do not add duplicate LoadNodes to the outer strip mined loop
1723 if (outer_loop_clone != nullptr) {
1724 _igvn.replace_node(x, outer_loop_clone);
1725 continue;
1726 }
1727 outer_loop_clone = x;
1728 }
1729 x->set_req(0, x_ctrl);
1730 } else if (n->in(0) != nullptr){
1731 x->set_req(0, x_ctrl);
1732 }
1733 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1734 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1735 register_new_node(x, x_ctrl);
1736
1737 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1738 // All AddP nodes must keep the same base after sinking so:
1739 // 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,
1740 // their bases remain the same.
1741 // (see 2- below)
1742 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1743 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1744 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1745 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1746 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1747 assert(!x->is_Load(), "load should be pinned");
1748 // Use a cast node to pin clone out of loop
1749 Node* cast = nullptr;
1750 for (uint k = 0; k < x->req(); k++) {
1751 Node* in = x->in(k);
1752 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1753 const Type* in_t = _igvn.type(in);
1754 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1755 ConstraintCastNode::UnconditionalDependency, nullptr);
1756 }
1757 if (cast != nullptr) {
1758 Node* prev = _igvn.hash_find_insert(cast);
1759 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1760 cast->destruct(&_igvn);
1761 cast = prev;
1762 } else {
1763 register_new_node(cast, x_ctrl);
1764 }
1765 x->replace_edge(in, cast);
1766 // Chain of AddP nodes:
1767 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1768 if (x->is_AddP() && k == AddPNode::Base) {
1769 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1770 }
1771 break;
1772 }
1773 }
1774 assert(cast != nullptr, "must have added a cast to pin the node");
1775 }
1776 }
1777 _igvn.remove_dead_node(n);
1778 }
1779 _dom_lca_tags_round = 0;
1780 }
1781 }
1782 }
1783
1784 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1785 ResourceMark rm;
1786 Node_List wq;
1787 wq.push(x);
1788 while (wq.size() != 0) {
1789 Node* n = wq.pop();
1790 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1791 Node* u = n->fast_out(i);
1792 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1793 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1794 wq.push(u);
1795 }
1796 }
1797 }
1798 }
1799
1800 // Compute the early control of a node by following its inputs until we reach
1801 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1802 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1803 Node* early_ctrl = nullptr;
1804 ResourceMark rm;
1805 Unique_Node_List wq;
1806 wq.push(n);
1807 for (uint i = 0; i < wq.size(); i++) {
1808 Node* m = wq.at(i);
1809 Node* c = nullptr;
1810 if (m->is_CFG()) {
1811 c = m;
1812 } else if (m->pinned()) {
1813 c = m->in(0);
1814 } else {
1815 for (uint j = 0; j < m->req(); j++) {
1816 Node* in = m->in(j);
1817 if (in != nullptr) {
1818 wq.push(in);
1819 }
1820 }
1821 }
1822 if (c != nullptr) {
1823 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1824 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1825 early_ctrl = c;
1826 }
1827 }
1828 }
1829 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1830 return early_ctrl;
1831 }
1832
1833 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1834 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1835 Node* u = n->fast_out(i);
1836 if (u->is_Opaque1()) {
1837 return false; // Found loop limit, bugfix for 4677003
1838 }
1839 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1840 // get_late_ctrl_with_anti_dep() use their own tag
1841 _dom_lca_tags_round++;
1842 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1843
1844 if (u->is_Phi()) {
1845 for (uint j = 1; j < u->req(); ++j) {
1846 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1847 return false;
1848 }
1849 }
1850 } else {
1851 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1852 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1853 return false;
1854 }
1855 }
1856 }
1857 return true;
1858 }
1859
1860 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1861 if (n->is_Load()) {
1862 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1863 }
1864 IdealLoopTree *u_loop = get_loop(ctrl);
1865 if (u_loop == n_loop) {
1866 return false; // Found loop-varying use
1867 }
1868 if (n_loop->is_member(u_loop)) {
1869 return false; // Found use in inner loop
1870 }
1871 // 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
1872 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1873 // test of the pre loop above the point in the graph where it's pinned.
1874 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop() &&
1875 u_loop->_head->is_CountedLoop() && u_loop->_head->as_CountedLoop()->is_main_loop() &&
1876 n_loop->_next == get_loop(u_loop->_head->as_CountedLoop()->skip_strip_mined())) {
1877 return false;
1878 }
1879 return true;
1880 }
1881
1882 //------------------------------split_if_with_blocks---------------------------
1883 // Check for aggressive application of 'split-if' optimization,
1884 // using basic block level info.
1885 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1886 Node* root = C->root();
1887 visited.set(root->_idx); // first, mark root as visited
1888 // Do pre-visit work for root
1889 Node* n = split_if_with_blocks_pre(root);
1890 uint cnt = n->outcnt();
1891 uint i = 0;
1892
1893 while (true) {
1894 // Visit all children
1895 if (i < cnt) {
1896 Node* use = n->raw_out(i);
1897 ++i;
1898 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1899 // Now do pre-visit work for this use
1900 use = split_if_with_blocks_pre(use);
1901 nstack.push(n, i); // Save parent and next use's index.
1902 n = use; // Process all children of current use.
1903 cnt = use->outcnt();
1904 i = 0;
1905 }
1906 }
1907 else {
1908 // All of n's children have been processed, complete post-processing.
1909 if (cnt != 0 && !n->is_Con()) {
1910 assert(has_node(n), "no dead nodes");
1911 split_if_with_blocks_post(n);
1912 }
1913 if (must_throttle_split_if()) {
1914 nstack.clear();
1915 }
1916 if (nstack.is_empty()) {
1917 // Finished all nodes on stack.
1918 break;
1919 }
1920 // Get saved parent node and next use's index. Visit the rest of uses.
1921 n = nstack.node();
1922 cnt = n->outcnt();
1923 i = nstack.index();
1924 nstack.pop();
1925 }
1926 }
1927 }
1928
1929
1930 //=============================================================================
1931 //
1932 // C L O N E A L O O P B O D Y
1933 //
1934
1935 //------------------------------clone_iff--------------------------------------
1936 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1937 // "Nearly" because all Nodes have been cloned from the original in the loop,
1938 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1939 // through the Phi recursively, and return a Bool.
1940 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
1941
1942 // Convert this Phi into a Phi merging Bools
1943 uint i;
1944 for (i = 1; i < phi->req(); i++) {
1945 Node *b = phi->in(i);
1946 if (b->is_Phi()) {
1947 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
1948 } else {
1949 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1950 }
1951 }
1952
1953 Node* n = phi->in(1);
1954 Node* sample_opaque = nullptr;
1955 Node *sample_bool = nullptr;
1956 if (n->Opcode() == Op_Opaque4) {
1957 sample_opaque = n;
1958 sample_bool = n->in(1);
1959 assert(sample_bool->is_Bool(), "wrong type");
1960 } else {
1961 sample_bool = n;
1962 }
1963 Node *sample_cmp = sample_bool->in(1);
1964
1965 // Make Phis to merge the Cmp's inputs.
1966 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1967 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1968 for (i = 1; i < phi->req(); i++) {
1969 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1970 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1971 phi1->set_req(i, n1);
1972 phi2->set_req(i, n2);
1973 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1974 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1975 }
1976 // See if these Phis have been made before.
1977 // Register with optimizer
1978 Node *hit1 = _igvn.hash_find_insert(phi1);
1979 if (hit1) { // Hit, toss just made Phi
1980 _igvn.remove_dead_node(phi1); // Remove new phi
1981 assert(hit1->is_Phi(), "" );
1982 phi1 = (PhiNode*)hit1; // Use existing phi
1983 } else { // Miss
1984 _igvn.register_new_node_with_optimizer(phi1);
1985 }
1986 Node *hit2 = _igvn.hash_find_insert(phi2);
1987 if (hit2) { // Hit, toss just made Phi
1988 _igvn.remove_dead_node(phi2); // Remove new phi
1989 assert(hit2->is_Phi(), "" );
1990 phi2 = (PhiNode*)hit2; // Use existing phi
1991 } else { // Miss
1992 _igvn.register_new_node_with_optimizer(phi2);
1993 }
1994 // Register Phis with loop/block info
1995 set_ctrl(phi1, phi->in(0));
1996 set_ctrl(phi2, phi->in(0));
1997 // Make a new Cmp
1998 Node *cmp = sample_cmp->clone();
1999 cmp->set_req(1, phi1);
2000 cmp->set_req(2, phi2);
2001 _igvn.register_new_node_with_optimizer(cmp);
2002 set_ctrl(cmp, phi->in(0));
2003
2004 // Make a new Bool
2005 Node *b = sample_bool->clone();
2006 b->set_req(1,cmp);
2007 _igvn.register_new_node_with_optimizer(b);
2008 set_ctrl(b, phi->in(0));
2009
2010 if (sample_opaque != nullptr) {
2011 Node* opaque = sample_opaque->clone();
2012 opaque->set_req(1, b);
2013 _igvn.register_new_node_with_optimizer(opaque);
2014 set_ctrl(opaque, phi->in(0));
2015 return opaque;
2016 }
2017
2018 assert(b->is_Bool(), "");
2019 return b;
2020 }
2021
2022 //------------------------------clone_bool-------------------------------------
2023 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2024 // "Nearly" because all Nodes have been cloned from the original in the loop,
2025 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2026 // through the Phi recursively, and return a Bool.
2027 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2028 uint i;
2029 // Convert this Phi into a Phi merging Bools
2030 for( i = 1; i < phi->req(); i++ ) {
2031 Node *b = phi->in(i);
2032 if( b->is_Phi() ) {
2033 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2034 } else {
2035 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2036 }
2037 }
2038
2039 Node *sample_cmp = phi->in(1);
2040
2041 // Make Phis to merge the Cmp's inputs.
2042 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2043 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2044 for( uint j = 1; j < phi->req(); j++ ) {
2045 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2046 Node *n1, *n2;
2047 if( cmp_top->is_Cmp() ) {
2048 n1 = cmp_top->in(1);
2049 n2 = cmp_top->in(2);
2050 } else {
2051 n1 = n2 = cmp_top;
2052 }
2053 phi1->set_req( j, n1 );
2054 phi2->set_req( j, n2 );
2055 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2056 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2057 }
2058
2059 // See if these Phis have been made before.
2060 // Register with optimizer
2061 Node *hit1 = _igvn.hash_find_insert(phi1);
2062 if( hit1 ) { // Hit, toss just made Phi
2063 _igvn.remove_dead_node(phi1); // Remove new phi
2064 assert( hit1->is_Phi(), "" );
2065 phi1 = (PhiNode*)hit1; // Use existing phi
2066 } else { // Miss
2067 _igvn.register_new_node_with_optimizer(phi1);
2068 }
2069 Node *hit2 = _igvn.hash_find_insert(phi2);
2070 if( hit2 ) { // Hit, toss just made Phi
2071 _igvn.remove_dead_node(phi2); // Remove new phi
2072 assert( hit2->is_Phi(), "" );
2073 phi2 = (PhiNode*)hit2; // Use existing phi
2074 } else { // Miss
2075 _igvn.register_new_node_with_optimizer(phi2);
2076 }
2077 // Register Phis with loop/block info
2078 set_ctrl(phi1, phi->in(0));
2079 set_ctrl(phi2, phi->in(0));
2080 // Make a new Cmp
2081 Node *cmp = sample_cmp->clone();
2082 cmp->set_req( 1, phi1 );
2083 cmp->set_req( 2, phi2 );
2084 _igvn.register_new_node_with_optimizer(cmp);
2085 set_ctrl(cmp, phi->in(0));
2086
2087 assert( cmp->is_Cmp(), "" );
2088 return (CmpNode*)cmp;
2089 }
2090
2091 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2092 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2093 Node_List*& split_if_set, Node_List*& split_bool_set,
2094 Node_List*& split_cex_set, Node_List& worklist,
2095 uint new_counter, CloneLoopMode mode) {
2096 Node* nnn = old_new[old->_idx];
2097 // Copy uses to a worklist, so I can munge the def-use info
2098 // with impunity.
2099 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2100 worklist.push(old->fast_out(j));
2101
2102 while( worklist.size() ) {
2103 Node *use = worklist.pop();
2104 if (!has_node(use)) continue; // Ignore dead nodes
2105 if (use->in(0) == C->top()) continue;
2106 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2107 // Check for data-use outside of loop - at least one of OLD or USE
2108 // must not be a CFG node.
2109 #ifdef ASSERT
2110 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2111 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2112 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2113 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2114 }
2115 #endif
2116 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2117
2118 // If the Data use is an IF, that means we have an IF outside of the
2119 // loop that is switching on a condition that is set inside of the
2120 // loop. Happens if people set a loop-exit flag; then test the flag
2121 // in the loop to break the loop, then test is again outside of the
2122 // loop to determine which way the loop exited.
2123 // Loop predicate If node connects to Bool node through Opaque1 node.
2124 //
2125 // If the use is an AllocateArray through its ValidLengthTest input,
2126 // make sure the Bool/Cmp input is cloned down to avoid a Phi between
2127 // the AllocateArray node and its ValidLengthTest input that could cause
2128 // split if to break.
2129 if (use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque4 ||
2130 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2131 // Since this code is highly unlikely, we lazily build the worklist
2132 // of such Nodes to go split.
2133 if (!split_if_set) {
2134 split_if_set = new Node_List();
2135 }
2136 split_if_set->push(use);
2137 }
2138 if (use->is_Bool()) {
2139 if (!split_bool_set) {
2140 split_bool_set = new Node_List();
2141 }
2142 split_bool_set->push(use);
2143 }
2144 if (use->Opcode() == Op_CreateEx) {
2145 if (!split_cex_set) {
2146 split_cex_set = new Node_List();
2147 }
2148 split_cex_set->push(use);
2149 }
2150
2151
2152 // Get "block" use is in
2153 uint idx = 0;
2154 while( use->in(idx) != old ) idx++;
2155 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2156 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2157 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2158 ? prev->in(2)
2159 : idom(prev);
2160 if( use->is_Phi() ) // Phi use is in prior block
2161 cfg = prev->in(idx); // NOT in block of Phi itself
2162 if (cfg->is_top()) { // Use is dead?
2163 _igvn.replace_input_of(use, idx, C->top());
2164 continue;
2165 }
2166
2167 // If use is referenced through control edge... (idx == 0)
2168 if (mode == IgnoreStripMined && idx == 0) {
2169 LoopNode *head = loop->_head->as_Loop();
2170 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2171 // That node is outside the inner loop, leave it outside the
2172 // outer loop as well to not confuse verification code.
2173 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2174 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2175 continue;
2176 }
2177 }
2178
2179 while(!outer_loop->is_member(get_loop(cfg))) {
2180 prev = cfg;
2181 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2182 }
2183 // If the use occurs after merging several exits from the loop, then
2184 // old value must have dominated all those exits. Since the same old
2185 // value was used on all those exits we did not need a Phi at this
2186 // merge point. NOW we do need a Phi here. Each loop exit value
2187 // is now merged with the peeled body exit; each exit gets its own
2188 // private Phi and those Phis need to be merged here.
2189 Node *phi;
2190 if( prev->is_Region() ) {
2191 if( idx == 0 ) { // Updating control edge?
2192 phi = prev; // Just use existing control
2193 } else { // Else need a new Phi
2194 phi = PhiNode::make( prev, old );
2195 // Now recursively fix up the new uses of old!
2196 for( uint i = 1; i < prev->req(); i++ ) {
2197 worklist.push(phi); // Onto worklist once for each 'old' input
2198 }
2199 }
2200 } else {
2201 // Get new RegionNode merging old and new loop exits
2202 prev = old_new[prev->_idx];
2203 assert( prev, "just made this in step 7" );
2204 if( idx == 0) { // Updating control edge?
2205 phi = prev; // Just use existing control
2206 } else { // Else need a new Phi
2207 // Make a new Phi merging data values properly
2208 phi = PhiNode::make( prev, old );
2209 phi->set_req( 1, nnn );
2210 }
2211 }
2212 // If inserting a new Phi, check for prior hits
2213 if( idx != 0 ) {
2214 Node *hit = _igvn.hash_find_insert(phi);
2215 if( hit == nullptr ) {
2216 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2217 } else { // or
2218 // Remove the new phi from the graph and use the hit
2219 _igvn.remove_dead_node(phi);
2220 phi = hit; // Use existing phi
2221 }
2222 set_ctrl(phi, prev);
2223 }
2224 // Make 'use' use the Phi instead of the old loop body exit value
2225 assert(use->in(idx) == old, "old is still input of use");
2226 // We notify all uses of old, including use, and the indirect uses,
2227 // that may now be optimized because we have replaced old with phi.
2228 _igvn.add_users_to_worklist(old);
2229 _igvn.replace_input_of(use, idx, phi);
2230 if( use->_idx >= new_counter ) { // If updating new phis
2231 // Not needed for correctness, but prevents a weak assert
2232 // in AddPNode from tripping (when we end up with different
2233 // base & derived Phis that will become the same after
2234 // IGVN does CSE).
2235 Node *hit = _igvn.hash_find_insert(use);
2236 if( hit ) // Go ahead and re-hash for hits.
2237 _igvn.replace_node( use, hit );
2238 }
2239 }
2240 }
2241 }
2242
2243 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2244 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2245 bool check_old_new) {
2246 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2247 Node* u = n->fast_out(j);
2248 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2249 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2250 Node* c = phase->get_ctrl(u);
2251 IdealLoopTree* u_loop = phase->get_loop(c);
2252 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2253 if (!loop->is_member(u_loop)) {
2254 if (outer_loop->is_member(u_loop)) {
2255 wq.push(u);
2256 } else {
2257 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2258 // the outer loop too
2259 Node* u_c = u->in(0);
2260 if (u_c != nullptr) {
2261 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2262 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2263 wq.push(u);
2264 }
2265 }
2266 }
2267 }
2268 }
2269 }
2270 }
2271
2272 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2273 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2274 Node_List& extra_data_nodes) {
2275 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2276 CountedLoopNode* cl = head->as_CountedLoop();
2277 Node* l = cl->outer_loop();
2278 Node* tail = cl->outer_loop_tail();
2279 IfNode* le = cl->outer_loop_end();
2280 Node* sfpt = cl->outer_safepoint();
2281 CountedLoopEndNode* cle = cl->loopexit();
2282 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2283 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2284 Node* cle_out = cle->proj_out(false);
2285
2286 Node* new_sfpt = nullptr;
2287 Node* new_cle_out = cle_out->clone();
2288 old_new.map(cle_out->_idx, new_cle_out);
2289 if (mode == CloneIncludesStripMined) {
2290 // clone outer loop body
2291 Node* new_l = l->clone();
2292 Node* new_tail = tail->clone();
2293 IfNode* new_le = le->clone()->as_If();
2294 new_sfpt = sfpt->clone();
2295
2296 set_loop(new_l, outer_loop->_parent);
2297 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2298 set_loop(new_cle_out, outer_loop->_parent);
2299 set_idom(new_cle_out, new_cle, dd);
2300 set_loop(new_sfpt, outer_loop->_parent);
2301 set_idom(new_sfpt, new_cle_out, dd);
2302 set_loop(new_le, outer_loop->_parent);
2303 set_idom(new_le, new_sfpt, dd);
2304 set_loop(new_tail, outer_loop->_parent);
2305 set_idom(new_tail, new_le, dd);
2306 set_idom(new_cl, new_l, dd);
2307
2308 old_new.map(l->_idx, new_l);
2309 old_new.map(tail->_idx, new_tail);
2310 old_new.map(le->_idx, new_le);
2311 old_new.map(sfpt->_idx, new_sfpt);
2312
2313 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2314 new_l->set_req(0, new_l);
2315 new_tail->set_req(0, new_le);
2316 new_le->set_req(0, new_sfpt);
2317 new_sfpt->set_req(0, new_cle_out);
2318 new_cle_out->set_req(0, new_cle);
2319 new_cl->set_req(LoopNode::EntryControl, new_l);
2320
2321 _igvn.register_new_node_with_optimizer(new_l);
2322 _igvn.register_new_node_with_optimizer(new_tail);
2323 _igvn.register_new_node_with_optimizer(new_le);
2324 } else {
2325 Node *newhead = old_new[loop->_head->_idx];
2326 newhead->as_Loop()->clear_strip_mined();
2327 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2328 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2329 }
2330 // Look at data node that were assigned a control in the outer
2331 // loop: they are kept in the outer loop by the safepoint so start
2332 // from the safepoint node's inputs.
2333 IdealLoopTree* outer_loop = get_loop(l);
2334 Node_Stack stack(2);
2335 stack.push(sfpt, 1);
2336 uint new_counter = C->unique();
2337 while (stack.size() > 0) {
2338 Node* n = stack.node();
2339 uint i = stack.index();
2340 while (i < n->req() &&
2341 (n->in(i) == nullptr ||
2342 !has_ctrl(n->in(i)) ||
2343 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2344 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2345 i++;
2346 }
2347 if (i < n->req()) {
2348 stack.set_index(i+1);
2349 stack.push(n->in(i), 0);
2350 } else {
2351 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2352 Node* m = n == sfpt ? new_sfpt : n->clone();
2353 if (m != nullptr) {
2354 for (uint i = 0; i < n->req(); i++) {
2355 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2356 m->set_req(i, old_new[m->in(i)->_idx]);
2357 }
2358 }
2359 } else {
2360 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2361 }
2362 if (n != sfpt) {
2363 extra_data_nodes.push(n);
2364 _igvn.register_new_node_with_optimizer(m);
2365 assert(get_ctrl(n) == cle_out, "what other control?");
2366 set_ctrl(m, new_cle_out);
2367 old_new.map(n->_idx, m);
2368 }
2369 stack.pop();
2370 }
2371 }
2372 if (mode == CloneIncludesStripMined) {
2373 _igvn.register_new_node_with_optimizer(new_sfpt);
2374 _igvn.register_new_node_with_optimizer(new_cle_out);
2375 }
2376 // Some other transformation may have pessimistically assigned some
2377 // data nodes to the outer loop. Set their control so they are out
2378 // of the outer loop.
2379 ResourceMark rm;
2380 Unique_Node_List wq;
2381 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2382 Node* old = extra_data_nodes.at(i);
2383 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2384 }
2385
2386 for (uint i = 0; i < loop->_body.size(); i++) {
2387 Node* old = loop->_body.at(i);
2388 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2389 }
2390
2391 Node* inner_out = sfpt->in(0);
2392 if (inner_out->outcnt() > 1) {
2393 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2394 }
2395
2396 Node* new_ctrl = cl->outer_loop_exit();
2397 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2398 for (uint i = 0; i < wq.size(); i++) {
2399 Node* n = wq.at(i);
2400 set_ctrl(n, new_ctrl);
2401 if (n->in(0) != nullptr) {
2402 _igvn.replace_input_of(n, 0, new_ctrl);
2403 }
2404 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2405 }
2406 } else {
2407 Node *newhead = old_new[loop->_head->_idx];
2408 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2409 }
2410 }
2411
2412 //------------------------------clone_loop-------------------------------------
2413 //
2414 // C L O N E A L O O P B O D Y
2415 //
2416 // This is the basic building block of the loop optimizations. It clones an
2417 // entire loop body. It makes an old_new loop body mapping; with this mapping
2418 // you can find the new-loop equivalent to an old-loop node. All new-loop
2419 // nodes are exactly equal to their old-loop counterparts, all edges are the
2420 // same. All exits from the old-loop now have a RegionNode that merges the
2421 // equivalent new-loop path. This is true even for the normal "loop-exit"
2422 // condition. All uses of loop-invariant old-loop values now come from (one
2423 // or more) Phis that merge their new-loop equivalents.
2424 //
2425 // This operation leaves the graph in an illegal state: there are two valid
2426 // control edges coming from the loop pre-header to both loop bodies. I'll
2427 // definitely have to hack the graph after running this transform.
2428 //
2429 // From this building block I will further edit edges to perform loop peeling
2430 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2431 //
2432 // Parameter side_by_size_idom:
2433 // When side_by_size_idom is null, the dominator tree is constructed for
2434 // the clone loop to dominate the original. Used in construction of
2435 // pre-main-post loop sequence.
2436 // When nonnull, the clone and original are side-by-side, both are
2437 // dominated by the side_by_side_idom node. Used in construction of
2438 // unswitched loops.
2439 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2440 CloneLoopMode mode, Node* side_by_side_idom) {
2441
2442 LoopNode* head = loop->_head->as_Loop();
2443 head->verify_strip_mined(1);
2444
2445 if (C->do_vector_loop() && PrintOpto) {
2446 const char* mname = C->method()->name()->as_quoted_ascii();
2447 if (mname != nullptr) {
2448 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2449 }
2450 }
2451
2452 CloneMap& cm = C->clone_map();
2453 if (C->do_vector_loop()) {
2454 cm.set_clone_idx(cm.max_gen()+1);
2455 #ifndef PRODUCT
2456 if (PrintOpto) {
2457 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2458 loop->dump_head();
2459 }
2460 #endif
2461 }
2462
2463 // Step 1: Clone the loop body. Make the old->new mapping.
2464 clone_loop_body(loop->_body, old_new, &cm);
2465
2466 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2467
2468 // Step 2: Fix the edges in the new body. If the old input is outside the
2469 // loop use it. If the old input is INside the loop, use the corresponding
2470 // new node instead.
2471 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2472
2473 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2474 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2475
2476 // Step 3: Now fix control uses. Loop varying control uses have already
2477 // been fixed up (as part of all input edges in Step 2). Loop invariant
2478 // control uses must be either an IfFalse or an IfTrue. Make a merge
2479 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2480 // refer to this.
2481 Node_List worklist;
2482 uint new_counter = C->unique();
2483 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2484
2485 // Step 4: If loop-invariant use is not control, it must be dominated by a
2486 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2487 // there if needed. Make a Phi there merging old and new used values.
2488 Node_List *split_if_set = nullptr;
2489 Node_List *split_bool_set = nullptr;
2490 Node_List *split_cex_set = nullptr;
2491 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2492
2493 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2494 Node* old = extra_data_nodes.at(i);
2495 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2496 split_bool_set, split_cex_set, worklist, new_counter,
2497 mode);
2498 }
2499
2500 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2501 // the loop uses a condition set in the loop. The original IF probably
2502 // takes control from one or more OLD Regions (which in turn get from NEW
2503 // Regions). In any case, there will be a set of Phis for each merge point
2504 // from the IF up to where the original BOOL def exists the loop.
2505 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2506
2507 }
2508
2509 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2510 if (split_if_set) {
2511 while (split_if_set->size()) {
2512 Node *iff = split_if_set->pop();
2513 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2514 if (iff->in(input)->is_Phi()) {
2515 Node *b = clone_iff(iff->in(input)->as_Phi());
2516 _igvn.replace_input_of(iff, input, b);
2517 }
2518 }
2519 }
2520 if (split_bool_set) {
2521 while (split_bool_set->size()) {
2522 Node *b = split_bool_set->pop();
2523 Node *phi = b->in(1);
2524 assert(phi->is_Phi(), "");
2525 CmpNode *cmp = clone_bool((PhiNode*) phi);
2526 _igvn.replace_input_of(b, 1, cmp);
2527 }
2528 }
2529 if (split_cex_set) {
2530 while (split_cex_set->size()) {
2531 Node *b = split_cex_set->pop();
2532 assert(b->in(0)->is_Region(), "");
2533 assert(b->in(1)->is_Phi(), "");
2534 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2535 split_up(b, b->in(0), nullptr);
2536 }
2537 }
2538 }
2539
2540 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2541 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2542 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2543 for(uint i = 0; i < body.size(); i++ ) {
2544 Node* old = body.at(i);
2545 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2546 split_bool_set, split_cex_set, worklist, new_counter,
2547 mode);
2548 }
2549 }
2550
2551 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2552 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2553 LoopNode* head = loop->_head->as_Loop();
2554 for(uint i = 0; i < body.size(); i++ ) {
2555 Node* old = body.at(i);
2556 if( !old->is_CFG() ) continue;
2557
2558 // Copy uses to a worklist, so I can munge the def-use info
2559 // with impunity.
2560 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2561 worklist.push(old->fast_out(j));
2562 }
2563
2564 while (worklist.size()) { // Visit all uses
2565 Node *use = worklist.pop();
2566 if (!has_node(use)) continue; // Ignore dead nodes
2567 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2568 if (!loop->is_member(use_loop) && use->is_CFG()) {
2569 // Both OLD and USE are CFG nodes here.
2570 assert(use->is_Proj(), "" );
2571 Node* nnn = old_new[old->_idx];
2572
2573 Node* newuse = nullptr;
2574 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2575 CountedLoopNode* cl = head->as_CountedLoop();
2576 CountedLoopEndNode* cle = cl->loopexit();
2577 Node* cle_out = cle->proj_out_or_null(false);
2578 if (use == cle_out) {
2579 IfNode* le = cl->outer_loop_end();
2580 use = le->proj_out(false);
2581 use_loop = get_loop(use);
2582 if (mode == CloneIncludesStripMined) {
2583 nnn = old_new[le->_idx];
2584 } else {
2585 newuse = old_new[cle_out->_idx];
2586 }
2587 }
2588 }
2589 if (newuse == nullptr) {
2590 newuse = use->clone();
2591 }
2592
2593 // Clone the loop exit control projection
2594 if (C->do_vector_loop() && cm != nullptr) {
2595 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2596 }
2597 newuse->set_req(0,nnn);
2598 _igvn.register_new_node_with_optimizer(newuse);
2599 set_loop(newuse, use_loop);
2600 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2601
2602 // We need a Region to merge the exit from the peeled body and the
2603 // exit from the old loop body.
2604 RegionNode *r = new RegionNode(3);
2605 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2606 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2607
2608 // The original user of 'use' uses 'r' instead.
2609 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2610 Node* useuse = use->last_out(l);
2611 _igvn.rehash_node_delayed(useuse);
2612 uint uses_found = 0;
2613 if (useuse->in(0) == use) {
2614 useuse->set_req(0, r);
2615 uses_found++;
2616 if (useuse->is_CFG()) {
2617 // This is not a dom_depth > dd_r because when new
2618 // control flow is constructed by a loop opt, a node and
2619 // its dominator can end up at the same dom_depth
2620 assert(dom_depth(useuse) >= dd_r, "");
2621 set_idom(useuse, r, dom_depth(useuse));
2622 }
2623 }
2624 for (uint k = 1; k < useuse->req(); k++) {
2625 if( useuse->in(k) == use ) {
2626 useuse->set_req(k, r);
2627 uses_found++;
2628 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2629 // This is not a dom_depth > dd_r because when new
2630 // control flow is constructed by a loop opt, a node
2631 // and its dominator can end up at the same dom_depth
2632 assert(dom_depth(useuse) >= dd_r , "");
2633 set_idom(useuse, r, dom_depth(useuse));
2634 }
2635 }
2636 }
2637 l -= uses_found; // we deleted 1 or more copies of this edge
2638 }
2639
2640 assert(use->is_Proj(), "loop exit should be projection");
2641 // lazy_replace() below moves all nodes that are:
2642 // - control dependent on the loop exit or
2643 // - have control set to the loop exit
2644 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2645 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2646 // becomes dead and is "replaced" by the region.
2647 Node* use_clone = use->clone();
2648 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2649 // Now finish up 'r'
2650 r->set_req(1, newuse);
2651 r->set_req(2, use_clone);
2652 _igvn.register_new_node_with_optimizer(r);
2653 set_loop(r, use_loop);
2654 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2655 lazy_replace(use, r);
2656 // Map the (cloned) old use to the new merge point
2657 old_new.map(use_clone->_idx, r);
2658 } // End of if a loop-exit test
2659 }
2660 }
2661 }
2662
2663 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2664 IdealLoopTree* parent, bool partial) {
2665 for(uint i = 0; i < body.size(); i++ ) {
2666 Node *old = body.at(i);
2667 Node *nnn = old_new[old->_idx];
2668 // Fix CFG/Loop controlling the new node
2669 if (has_ctrl(old)) {
2670 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2671 } else {
2672 set_loop(nnn, parent);
2673 if (old->outcnt() > 0) {
2674 Node* dom = idom(old);
2675 if (old_new[dom->_idx] != nullptr) {
2676 dom = old_new[dom->_idx];
2677 set_idom(nnn, dom, dd );
2678 }
2679 }
2680 }
2681 // Correct edges to the new node
2682 for (uint j = 0; j < nnn->req(); j++) {
2683 Node *n = nnn->in(j);
2684 if (n != nullptr) {
2685 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2686 if (loop->is_member(old_in_loop)) {
2687 if (old_new[n->_idx] != nullptr) {
2688 nnn->set_req(j, old_new[n->_idx]);
2689 } else {
2690 assert(!body.contains(n), "");
2691 assert(partial, "node not cloned");
2692 }
2693 }
2694 }
2695 }
2696 _igvn.hash_find_insert(nnn);
2697 }
2698 }
2699
2700 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2701 for (uint i = 0; i < body.size(); i++) {
2702 Node* old = body.at(i);
2703 Node* nnn = old->clone();
2704 old_new.map(old->_idx, nnn);
2705 if (C->do_vector_loop() && cm != nullptr) {
2706 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2707 }
2708 _igvn.register_new_node_with_optimizer(nnn);
2709 }
2710 }
2711
2712
2713 //---------------------- stride_of_possible_iv -------------------------------------
2714 // Looks for an iff/bool/comp with one operand of the compare
2715 // being a cycle involving an add and a phi,
2716 // with an optional truncation (left-shift followed by a right-shift)
2717 // of the add. Returns zero if not an iv.
2718 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2719 Node* trunc1 = nullptr;
2720 Node* trunc2 = nullptr;
2721 const TypeInteger* ttype = nullptr;
2722 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2723 return 0;
2724 }
2725 BoolNode* bl = iff->in(1)->as_Bool();
2726 Node* cmp = bl->in(1);
2727 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2728 return 0;
2729 }
2730 // Must have an invariant operand
2731 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2732 return 0;
2733 }
2734 Node* add2 = nullptr;
2735 Node* cmp1 = cmp->in(1);
2736 if (cmp1->is_Phi()) {
2737 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2738 Node* phi = cmp1;
2739 for (uint i = 1; i < phi->req(); i++) {
2740 Node* in = phi->in(i);
2741 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2742 &trunc1, &trunc2, &ttype, T_INT);
2743 if (add && add->in(1) == phi) {
2744 add2 = add->in(2);
2745 break;
2746 }
2747 }
2748 } else {
2749 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2750 Node* addtrunc = cmp1;
2751 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2752 &trunc1, &trunc2, &ttype, T_INT);
2753 if (add && add->in(1)->is_Phi()) {
2754 Node* phi = add->in(1);
2755 for (uint i = 1; i < phi->req(); i++) {
2756 if (phi->in(i) == addtrunc) {
2757 add2 = add->in(2);
2758 break;
2759 }
2760 }
2761 }
2762 }
2763 if (add2 != nullptr) {
2764 const TypeInt* add2t = _igvn.type(add2)->is_int();
2765 if (add2t->is_con()) {
2766 return add2t->get_con();
2767 }
2768 }
2769 return 0;
2770 }
2771
2772
2773 //---------------------- stay_in_loop -------------------------------------
2774 // Return the (unique) control output node that's in the loop (if it exists.)
2775 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2776 Node* unique = nullptr;
2777 if (!n) return nullptr;
2778 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2779 Node* use = n->fast_out(i);
2780 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2781 if (unique != nullptr) {
2782 return nullptr;
2783 }
2784 unique = use;
2785 }
2786 }
2787 return unique;
2788 }
2789
2790 //------------------------------ register_node -------------------------------------
2791 // Utility to register node "n" with PhaseIdealLoop
2792 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2793 _igvn.register_new_node_with_optimizer(n);
2794 loop->_body.push(n);
2795 if (n->is_CFG()) {
2796 set_loop(n, loop);
2797 set_idom(n, pred, ddepth);
2798 } else {
2799 set_ctrl(n, pred);
2800 }
2801 }
2802
2803 //------------------------------ proj_clone -------------------------------------
2804 // Utility to create an if-projection
2805 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2806 ProjNode* c = p->clone()->as_Proj();
2807 c->set_req(0, iff);
2808 return c;
2809 }
2810
2811 //------------------------------ short_circuit_if -------------------------------------
2812 // Force the iff control output to be the live_proj
2813 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2814 guarantee(live_proj != nullptr, "null projection");
2815 int proj_con = live_proj->_con;
2816 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2817 Node *con = _igvn.intcon(proj_con);
2818 set_ctrl(con, C->root());
2819 if (iff) {
2820 iff->set_req(1, con);
2821 }
2822 return con;
2823 }
2824
2825 //------------------------------ insert_if_before_proj -------------------------------------
2826 // Insert a new if before an if projection (* - new node)
2827 //
2828 // before
2829 // if(test)
2830 // / \
2831 // v v
2832 // other-proj proj (arg)
2833 //
2834 // after
2835 // if(test)
2836 // / \
2837 // / v
2838 // | * proj-clone
2839 // v |
2840 // other-proj v
2841 // * new_if(relop(cmp[IU](left,right)))
2842 // / \
2843 // v v
2844 // * new-proj proj
2845 // (returned)
2846 //
2847 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2848 IfNode* iff = proj->in(0)->as_If();
2849 IdealLoopTree *loop = get_loop(proj);
2850 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2851 uint ddepth = dom_depth(proj);
2852
2853 _igvn.rehash_node_delayed(iff);
2854 _igvn.rehash_node_delayed(proj);
2855
2856 proj->set_req(0, nullptr); // temporary disconnect
2857 ProjNode* proj2 = proj_clone(proj, iff);
2858 register_node(proj2, loop, iff, ddepth);
2859
2860 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2861 register_node(cmp, loop, proj2, ddepth);
2862
2863 BoolNode* bol = new BoolNode(cmp, relop);
2864 register_node(bol, loop, proj2, ddepth);
2865
2866 int opcode = iff->Opcode();
2867 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2868 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2869 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2870 register_node(new_if, loop, proj2, ddepth);
2871
2872 proj->set_req(0, new_if); // reattach
2873 set_idom(proj, new_if, ddepth);
2874
2875 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2876 guarantee(new_exit != nullptr, "null exit node");
2877 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2878
2879 return new_exit;
2880 }
2881
2882 //------------------------------ insert_region_before_proj -------------------------------------
2883 // Insert a region before an if projection (* - new node)
2884 //
2885 // before
2886 // if(test)
2887 // / |
2888 // v |
2889 // proj v
2890 // other-proj
2891 //
2892 // after
2893 // if(test)
2894 // / |
2895 // v |
2896 // * proj-clone v
2897 // | other-proj
2898 // v
2899 // * new-region
2900 // |
2901 // v
2902 // * dum_if
2903 // / \
2904 // v \
2905 // * dum-proj v
2906 // proj
2907 //
2908 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2909 IfNode* iff = proj->in(0)->as_If();
2910 IdealLoopTree *loop = get_loop(proj);
2911 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2912 uint ddepth = dom_depth(proj);
2913
2914 _igvn.rehash_node_delayed(iff);
2915 _igvn.rehash_node_delayed(proj);
2916
2917 proj->set_req(0, nullptr); // temporary disconnect
2918 ProjNode* proj2 = proj_clone(proj, iff);
2919 register_node(proj2, loop, iff, ddepth);
2920
2921 RegionNode* reg = new RegionNode(2);
2922 reg->set_req(1, proj2);
2923 register_node(reg, loop, iff, ddepth);
2924
2925 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
2926 register_node(dum_if, loop, reg, ddepth);
2927
2928 proj->set_req(0, dum_if); // reattach
2929 set_idom(proj, dum_if, ddepth);
2930
2931 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2932 register_node(dum_proj, loop, dum_if, ddepth);
2933
2934 return reg;
2935 }
2936
2937 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2938 // Clone a signed compare loop exit from an unsigned compare and
2939 // insert it before the unsigned cmp on the stay-in-loop path.
2940 // All new nodes inserted in the dominator tree between the original
2941 // if and it's projections. The original if test is replaced with
2942 // a constant to force the stay-in-loop path.
2943 //
2944 // This is done to make sure that the original if and it's projections
2945 // still dominate the same set of control nodes, that the ctrl() relation
2946 // from data nodes to them is preserved, and that their loop nesting is
2947 // preserved.
2948 //
2949 // before
2950 // if(i <u limit) unsigned compare loop exit
2951 // / |
2952 // v v
2953 // exit-proj stay-in-loop-proj
2954 //
2955 // after
2956 // if(stay-in-loop-const) original if
2957 // / |
2958 // / v
2959 // / if(i < limit) new signed test
2960 // / / |
2961 // / / v
2962 // / / if(i <u limit) new cloned unsigned test
2963 // / / / |
2964 // v v v |
2965 // region |
2966 // | |
2967 // dum-if |
2968 // / | |
2969 // ether | |
2970 // v v
2971 // exit-proj stay-in-loop-proj
2972 //
2973 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2974 const bool Signed = true;
2975 const bool Unsigned = false;
2976
2977 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2978 if (bol->_test._test != BoolTest::lt) return nullptr;
2979 CmpNode* cmpu = bol->in(1)->as_Cmp();
2980 if (cmpu->Opcode() != Op_CmpU) return nullptr;
2981 int stride = stride_of_possible_iv(if_cmpu);
2982 if (stride == 0) return nullptr;
2983
2984 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2985 guarantee(lp_proj != nullptr, "null loop node");
2986
2987 ProjNode* lp_continue = lp_proj->as_Proj();
2988 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2989 if (!lp_exit->is_IfFalse()) {
2990 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
2991 // We therefore can't add a single exit condition.
2992 return nullptr;
2993 }
2994 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
2995 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
2996 Node* limit = nullptr;
2997 if (stride > 0) {
2998 limit = cmpu->in(2);
2999 } else {
3000 limit = _igvn.makecon(TypeInt::ZERO);
3001 set_ctrl(limit, C->root());
3002 }
3003 // Create a new region on the exit path
3004 RegionNode* reg = insert_region_before_proj(lp_exit);
3005 guarantee(reg != nullptr, "null region node");
3006
3007 // Clone the if-cmpu-true-false using a signed compare
3008 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3009 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
3010 reg->add_req(cmpi_exit);
3011
3012 // Clone the if-cmpu-true-false
3013 BoolTest::mask rel_u = bol->_test._test;
3014 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3015 reg->add_req(cmpu_exit);
3016
3017 // Force original if to stay in loop.
3018 short_circuit_if(if_cmpu, lp_continue);
3019
3020 return cmpi_exit->in(0)->as_If();
3021 }
3022
3023 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3024 // Remove a previously inserted signed compare loop exit.
3025 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3026 Node* lp_proj = stay_in_loop(if_cmp, loop);
3027 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3028 stay_in_loop(lp_proj, loop)->is_If() &&
3029 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3030 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3031 set_ctrl(con, C->root());
3032 if_cmp->set_req(1, con);
3033 }
3034
3035 //------------------------------ scheduled_nodelist -------------------------------------
3036 // Create a post order schedule of nodes that are in the
3037 // "member" set. The list is returned in "sched".
3038 // The first node in "sched" is the loop head, followed by
3039 // nodes which have no inputs in the "member" set, and then
3040 // followed by the nodes that have an immediate input dependence
3041 // on a node in "sched".
3042 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3043
3044 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3045 VectorSet visited;
3046 Node_Stack nstack(loop->_body.size());
3047
3048 Node* n = loop->_head; // top of stack is cached in "n"
3049 uint idx = 0;
3050 visited.set(n->_idx);
3051
3052 // Initially push all with no inputs from within member set
3053 for(uint i = 0; i < loop->_body.size(); i++ ) {
3054 Node *elt = loop->_body.at(i);
3055 if (member.test(elt->_idx)) {
3056 bool found = false;
3057 for (uint j = 0; j < elt->req(); j++) {
3058 Node* def = elt->in(j);
3059 if (def && member.test(def->_idx) && def != elt) {
3060 found = true;
3061 break;
3062 }
3063 }
3064 if (!found && elt != loop->_head) {
3065 nstack.push(n, idx);
3066 n = elt;
3067 assert(!visited.test(n->_idx), "not seen yet");
3068 visited.set(n->_idx);
3069 }
3070 }
3071 }
3072
3073 // traverse out's that are in the member set
3074 while (true) {
3075 if (idx < n->outcnt()) {
3076 Node* use = n->raw_out(idx);
3077 idx++;
3078 if (!visited.test_set(use->_idx)) {
3079 if (member.test(use->_idx)) {
3080 nstack.push(n, idx);
3081 n = use;
3082 idx = 0;
3083 }
3084 }
3085 } else {
3086 // All outputs processed
3087 sched.push(n);
3088 if (nstack.is_empty()) break;
3089 n = nstack.node();
3090 idx = nstack.index();
3091 nstack.pop();
3092 }
3093 }
3094 }
3095
3096
3097 //------------------------------ has_use_in_set -------------------------------------
3098 // Has a use in the vector set
3099 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3100 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3101 Node* use = n->fast_out(j);
3102 if (vset.test(use->_idx)) {
3103 return true;
3104 }
3105 }
3106 return false;
3107 }
3108
3109
3110 //------------------------------ has_use_internal_to_set -------------------------------------
3111 // Has use internal to the vector set (ie. not in a phi at the loop head)
3112 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3113 Node* head = loop->_head;
3114 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3115 Node* use = n->fast_out(j);
3116 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3117 return true;
3118 }
3119 }
3120 return false;
3121 }
3122
3123
3124 //------------------------------ clone_for_use_outside_loop -------------------------------------
3125 // clone "n" for uses that are outside of loop
3126 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3127 int cloned = 0;
3128 assert(worklist.size() == 0, "should be empty");
3129 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3130 Node* use = n->fast_out(j);
3131 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3132 worklist.push(use);
3133 }
3134 }
3135
3136 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3137 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3138 return -1;
3139 }
3140
3141 while( worklist.size() ) {
3142 Node *use = worklist.pop();
3143 if (!has_node(use) || use->in(0) == C->top()) continue;
3144 uint j;
3145 for (j = 0; j < use->req(); j++) {
3146 if (use->in(j) == n) break;
3147 }
3148 assert(j < use->req(), "must be there");
3149
3150 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3151 Node* n_clone = n->clone();
3152 _igvn.replace_input_of(use, j, n_clone);
3153 cloned++;
3154 Node* use_c;
3155 if (!use->is_Phi()) {
3156 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3157 } else {
3158 // Use in a phi is considered a use in the associated predecessor block
3159 use_c = use->in(0)->in(j);
3160 }
3161 set_ctrl(n_clone, use_c);
3162 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3163 get_loop(use_c)->_body.push(n_clone);
3164 _igvn.register_new_node_with_optimizer(n_clone);
3165 #ifndef PRODUCT
3166 if (TracePartialPeeling) {
3167 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3168 }
3169 #endif
3170 }
3171 return cloned;
3172 }
3173
3174
3175 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3176 // clone "n" for special uses that are in the not_peeled region.
3177 // If these def-uses occur in separate blocks, the code generator
3178 // marks the method as not compilable. For example, if a "BoolNode"
3179 // is in a different basic block than the "IfNode" that uses it, then
3180 // the compilation is aborted in the code generator.
3181 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3182 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3183 if (n->is_Phi() || n->is_Load()) {
3184 return;
3185 }
3186 assert(worklist.size() == 0, "should be empty");
3187 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3188 Node* use = n->fast_out(j);
3189 if ( not_peel.test(use->_idx) &&
3190 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3191 use->in(1) == n) {
3192 worklist.push(use);
3193 }
3194 }
3195 if (worklist.size() > 0) {
3196 // clone "n" and insert it between inputs of "n" and the use
3197 Node* n_clone = n->clone();
3198 loop->_body.push(n_clone);
3199 _igvn.register_new_node_with_optimizer(n_clone);
3200 set_ctrl(n_clone, get_ctrl(n));
3201 sink_list.push(n_clone);
3202 not_peel.set(n_clone->_idx);
3203 #ifndef PRODUCT
3204 if (TracePartialPeeling) {
3205 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3206 }
3207 #endif
3208 while( worklist.size() ) {
3209 Node *use = worklist.pop();
3210 _igvn.rehash_node_delayed(use);
3211 for (uint j = 1; j < use->req(); j++) {
3212 if (use->in(j) == n) {
3213 use->set_req(j, n_clone);
3214 }
3215 }
3216 }
3217 }
3218 }
3219
3220
3221 //------------------------------ insert_phi_for_loop -------------------------------------
3222 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3223 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3224 Node *phi = PhiNode::make(lp, back_edge_val);
3225 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3226 // Use existing phi if it already exists
3227 Node *hit = _igvn.hash_find_insert(phi);
3228 if( hit == nullptr ) {
3229 _igvn.register_new_node_with_optimizer(phi);
3230 set_ctrl(phi, lp);
3231 } else {
3232 // Remove the new phi from the graph and use the hit
3233 _igvn.remove_dead_node(phi);
3234 phi = hit;
3235 }
3236 _igvn.replace_input_of(use, idx, phi);
3237 }
3238
3239 #ifdef ASSERT
3240 //------------------------------ is_valid_loop_partition -------------------------------------
3241 // Validate the loop partition sets: peel and not_peel
3242 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3243 VectorSet& not_peel ) {
3244 uint i;
3245 // Check that peel_list entries are in the peel set
3246 for (i = 0; i < peel_list.size(); i++) {
3247 if (!peel.test(peel_list.at(i)->_idx)) {
3248 return false;
3249 }
3250 }
3251 // Check at loop members are in one of peel set or not_peel set
3252 for (i = 0; i < loop->_body.size(); i++ ) {
3253 Node *def = loop->_body.at(i);
3254 uint di = def->_idx;
3255 // Check that peel set elements are in peel_list
3256 if (peel.test(di)) {
3257 if (not_peel.test(di)) {
3258 return false;
3259 }
3260 // Must be in peel_list also
3261 bool found = false;
3262 for (uint j = 0; j < peel_list.size(); j++) {
3263 if (peel_list.at(j)->_idx == di) {
3264 found = true;
3265 break;
3266 }
3267 }
3268 if (!found) {
3269 return false;
3270 }
3271 } else if (not_peel.test(di)) {
3272 if (peel.test(di)) {
3273 return false;
3274 }
3275 } else {
3276 return false;
3277 }
3278 }
3279 return true;
3280 }
3281
3282 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3283 // Ensure a use outside of loop is of the right form
3284 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3285 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3286 return (use->is_Phi() &&
3287 use_c->is_Region() && use_c->req() == 3 &&
3288 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3289 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3290 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3291 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3292 }
3293
3294 //------------------------------ is_valid_clone_loop_form -------------------------------------
3295 // Ensure that all uses outside of loop are of the right form
3296 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3297 uint orig_exit_idx, uint clone_exit_idx) {
3298 uint len = peel_list.size();
3299 for (uint i = 0; i < len; i++) {
3300 Node *def = peel_list.at(i);
3301
3302 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3303 Node *use = def->fast_out(j);
3304 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3305 if (!loop->is_member(get_loop(use_c))) {
3306 // use is not in the loop, check for correct structure
3307 if (use->in(0) == def) {
3308 // Okay
3309 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3310 return false;
3311 }
3312 }
3313 }
3314 }
3315 return true;
3316 }
3317 #endif
3318
3319 //------------------------------ partial_peel -------------------------------------
3320 // Partially peel (aka loop rotation) the top portion of a loop (called
3321 // the peel section below) by cloning it and placing one copy just before
3322 // the new loop head and the other copy at the bottom of the new loop.
3323 //
3324 // before after where it came from
3325 //
3326 // stmt1 stmt1
3327 // loop: stmt2 clone
3328 // stmt2 if condA goto exitA clone
3329 // if condA goto exitA new_loop: new
3330 // stmt3 stmt3 clone
3331 // if !condB goto loop if condB goto exitB clone
3332 // exitB: stmt2 orig
3333 // stmt4 if !condA goto new_loop orig
3334 // exitA: goto exitA
3335 // exitB:
3336 // stmt4
3337 // exitA:
3338 //
3339 // Step 1: find the cut point: an exit test on probable
3340 // induction variable.
3341 // Step 2: schedule (with cloning) operations in the peel
3342 // section that can be executed after the cut into
3343 // the section that is not peeled. This may need
3344 // to clone operations into exit blocks. For
3345 // instance, a reference to A[i] in the not-peel
3346 // section and a reference to B[i] in an exit block
3347 // may cause a left-shift of i by 2 to be placed
3348 // in the peel block. This step will clone the left
3349 // shift into the exit block and sink the left shift
3350 // from the peel to the not-peel section.
3351 // Step 3: clone the loop, retarget the control, and insert
3352 // phis for values that are live across the new loop
3353 // head. This is very dependent on the graph structure
3354 // from clone_loop. It creates region nodes for
3355 // exit control and associated phi nodes for values
3356 // flow out of the loop through that exit. The region
3357 // node is dominated by the clone's control projection.
3358 // So the clone's peel section is placed before the
3359 // new loop head, and the clone's not-peel section is
3360 // forms the top part of the new loop. The original
3361 // peel section forms the tail of the new loop.
3362 // Step 4: update the dominator tree and recompute the
3363 // dominator depth.
3364 //
3365 // orig
3366 //
3367 // stmt1
3368 // |
3369 // v
3370 // predicates
3371 // |
3372 // v
3373 // loop<----+
3374 // | |
3375 // stmt2 |
3376 // | |
3377 // v |
3378 // ifA |
3379 // / | |
3380 // v v |
3381 // false true ^ <-- last_peel
3382 // / | |
3383 // / ===|==cut |
3384 // / stmt3 | <-- first_not_peel
3385 // / | |
3386 // | v |
3387 // v ifB |
3388 // exitA: / \ |
3389 // / \ |
3390 // v v |
3391 // false true |
3392 // / \ |
3393 // / ----+
3394 // |
3395 // v
3396 // exitB:
3397 // stmt4
3398 //
3399 //
3400 // after clone loop
3401 //
3402 // stmt1
3403 // |
3404 // v
3405 // predicates
3406 // / \
3407 // clone / \ orig
3408 // / \
3409 // / \
3410 // v v
3411 // +---->loop loop<----+
3412 // | | | |
3413 // | stmt2 stmt2 |
3414 // | | | |
3415 // | v v |
3416 // | ifA ifA |
3417 // | | \ / | |
3418 // | v v v v |
3419 // ^ true false false true ^ <-- last_peel
3420 // | | ^ \ / | |
3421 // | cut==|== \ \ / ===|==cut |
3422 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3423 // | | dom | | | |
3424 // | v \ 1v v2 v |
3425 // | ifB regionA ifB |
3426 // | / \ | / \ |
3427 // | / \ v / \ |
3428 // | v v exitA: v v |
3429 // | true false false true |
3430 // | / ^ \ / \ |
3431 // +---- \ \ / ----+
3432 // dom \ /
3433 // \ 1v v2
3434 // regionB
3435 // |
3436 // v
3437 // exitB:
3438 // stmt4
3439 //
3440 //
3441 // after partial peel
3442 //
3443 // stmt1
3444 // |
3445 // v
3446 // predicates
3447 // /
3448 // clone / orig
3449 // / TOP
3450 // / \
3451 // v v
3452 // TOP->loop loop----+
3453 // | | |
3454 // stmt2 stmt2 |
3455 // | | |
3456 // v v |
3457 // ifA ifA |
3458 // | \ / | |
3459 // v v v v |
3460 // true false false true | <-- last_peel
3461 // | ^ \ / +------|---+
3462 // +->newloop \ \ / === ==cut | |
3463 // | stmt3 \ \ / TOP | |
3464 // | | dom | | stmt3 | | <-- first_not_peel
3465 // | v \ 1v v2 v | |
3466 // | ifB regionA ifB ^ v
3467 // | / \ | / \ | |
3468 // | / \ v / \ | |
3469 // | v v exitA: v v | |
3470 // | true false false true | |
3471 // | / ^ \ / \ | |
3472 // | | \ \ / v | |
3473 // | | dom \ / TOP | |
3474 // | | \ 1v v2 | |
3475 // ^ v regionB | |
3476 // | | | | |
3477 // | | v ^ v
3478 // | | exitB: | |
3479 // | | stmt4 | |
3480 // | +------------>-----------------+ |
3481 // | |
3482 // +-----------------<---------------------+
3483 //
3484 //
3485 // final graph
3486 //
3487 // stmt1
3488 // |
3489 // v
3490 // predicates
3491 // |
3492 // v
3493 // stmt2 clone
3494 // |
3495 // v
3496 // ........> ifA clone
3497 // : / |
3498 // dom / |
3499 // : v v
3500 // : false true
3501 // : | |
3502 // : | v
3503 // : | newloop<-----+
3504 // : | | |
3505 // : | stmt3 clone |
3506 // : | | |
3507 // : | v |
3508 // : | ifB |
3509 // : | / \ |
3510 // : | v v |
3511 // : | false true |
3512 // : | | | |
3513 // : | v stmt2 |
3514 // : | exitB: | |
3515 // : | stmt4 v |
3516 // : | ifA orig |
3517 // : | / \ |
3518 // : | / \ |
3519 // : | v v |
3520 // : | false true |
3521 // : | / \ |
3522 // : v v -----+
3523 // RegionA
3524 // |
3525 // v
3526 // exitA
3527 //
3528 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3529
3530 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3531 if (!loop->_head->is_Loop()) {
3532 return false;
3533 }
3534 LoopNode *head = loop->_head->as_Loop();
3535
3536 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3537 return false;
3538 }
3539
3540 // Check for complex exit control
3541 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3542 Node *n = loop->_body.at(ii);
3543 int opc = n->Opcode();
3544 if (n->is_Call() ||
3545 opc == Op_Catch ||
3546 opc == Op_CatchProj ||
3547 opc == Op_Jump ||
3548 opc == Op_JumpProj) {
3549 #ifndef PRODUCT
3550 if (TracePartialPeeling) {
3551 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3552 }
3553 #endif
3554 return false;
3555 }
3556 }
3557
3558 int dd = dom_depth(head);
3559
3560 // Step 1: find cut point
3561
3562 // Walk up dominators to loop head looking for first loop exit
3563 // which is executed on every path thru loop.
3564 IfNode *peel_if = nullptr;
3565 IfNode *peel_if_cmpu = nullptr;
3566
3567 Node *iff = loop->tail();
3568 while (iff != head) {
3569 if (iff->is_If()) {
3570 Node *ctrl = get_ctrl(iff->in(1));
3571 if (ctrl->is_top()) return false; // Dead test on live IF.
3572 // If loop-varying exit-test, check for induction variable
3573 if (loop->is_member(get_loop(ctrl)) &&
3574 loop->is_loop_exit(iff) &&
3575 is_possible_iv_test(iff)) {
3576 Node* cmp = iff->in(1)->in(1);
3577 if (cmp->Opcode() == Op_CmpI) {
3578 peel_if = iff->as_If();
3579 } else {
3580 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3581 peel_if_cmpu = iff->as_If();
3582 }
3583 }
3584 }
3585 iff = idom(iff);
3586 }
3587
3588 // Prefer signed compare over unsigned compare.
3589 IfNode* new_peel_if = nullptr;
3590 if (peel_if == nullptr) {
3591 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3592 return false; // No peel point found
3593 }
3594 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3595 if (new_peel_if == nullptr) {
3596 return false; // No peel point found
3597 }
3598 peel_if = new_peel_if;
3599 }
3600 Node* last_peel = stay_in_loop(peel_if, loop);
3601 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3602 if (first_not_peeled == nullptr || first_not_peeled == head) {
3603 return false;
3604 }
3605
3606 #ifndef PRODUCT
3607 if (TraceLoopOpts) {
3608 tty->print("PartialPeel ");
3609 loop->dump_head();
3610 }
3611
3612 if (TracePartialPeeling) {
3613 tty->print_cr("before partial peel one iteration");
3614 Node_List wl;
3615 Node* t = head->in(2);
3616 while (true) {
3617 wl.push(t);
3618 if (t == head) break;
3619 t = idom(t);
3620 }
3621 while (wl.size() > 0) {
3622 Node* tt = wl.pop();
3623 tt->dump();
3624 if (tt == last_peel) tty->print_cr("-- cut --");
3625 }
3626 }
3627 #endif
3628 VectorSet peel;
3629 VectorSet not_peel;
3630 Node_List peel_list;
3631 Node_List worklist;
3632 Node_List sink_list;
3633
3634 uint estimate = loop->est_loop_clone_sz(1);
3635 if (exceeding_node_budget(estimate)) {
3636 return false;
3637 }
3638
3639 // Set of cfg nodes to peel are those that are executable from
3640 // the head through last_peel.
3641 assert(worklist.size() == 0, "should be empty");
3642 worklist.push(head);
3643 peel.set(head->_idx);
3644 while (worklist.size() > 0) {
3645 Node *n = worklist.pop();
3646 if (n != last_peel) {
3647 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3648 Node* use = n->fast_out(j);
3649 if (use->is_CFG() &&
3650 loop->is_member(get_loop(use)) &&
3651 !peel.test_set(use->_idx)) {
3652 worklist.push(use);
3653 }
3654 }
3655 }
3656 }
3657
3658 // Set of non-cfg nodes to peel are those that are control
3659 // dependent on the cfg nodes.
3660 for (uint i = 0; i < loop->_body.size(); i++) {
3661 Node *n = loop->_body.at(i);
3662 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3663 if (peel.test(n_c->_idx)) {
3664 peel.set(n->_idx);
3665 } else {
3666 not_peel.set(n->_idx);
3667 }
3668 }
3669
3670 // Step 2: move operations from the peeled section down into the
3671 // not-peeled section
3672
3673 // Get a post order schedule of nodes in the peel region
3674 // Result in right-most operand.
3675 scheduled_nodelist(loop, peel, peel_list);
3676
3677 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3678
3679 // For future check for too many new phis
3680 uint old_phi_cnt = 0;
3681 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3682 Node* use = head->fast_out(j);
3683 if (use->is_Phi()) old_phi_cnt++;
3684 }
3685
3686 #ifndef PRODUCT
3687 if (TracePartialPeeling) {
3688 tty->print_cr("\npeeled list");
3689 }
3690 #endif
3691
3692 // Evacuate nodes in peel region into the not_peeled region if possible
3693 bool too_many_clones = false;
3694 uint new_phi_cnt = 0;
3695 uint cloned_for_outside_use = 0;
3696 for (uint i = 0; i < peel_list.size();) {
3697 Node* n = peel_list.at(i);
3698 #ifndef PRODUCT
3699 if (TracePartialPeeling) n->dump();
3700 #endif
3701 bool incr = true;
3702 if (!n->is_CFG()) {
3703 if (has_use_in_set(n, not_peel)) {
3704 // If not used internal to the peeled region,
3705 // move "n" from peeled to not_peeled region.
3706 if (!has_use_internal_to_set(n, peel, loop)) {
3707 // if not pinned and not a load (which maybe anti-dependent on a store)
3708 // and not a CMove (Matcher expects only bool->cmove).
3709 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3710 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3711 if (C->failing()) return false;
3712 if (new_clones == -1) {
3713 too_many_clones = true;
3714 break;
3715 }
3716 cloned_for_outside_use += new_clones;
3717 sink_list.push(n);
3718 peel.remove(n->_idx);
3719 not_peel.set(n->_idx);
3720 peel_list.remove(i);
3721 incr = false;
3722 #ifndef PRODUCT
3723 if (TracePartialPeeling) {
3724 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3725 n->_idx, get_ctrl(n)->_idx);
3726 }
3727 #endif
3728 }
3729 } else {
3730 // Otherwise check for special def-use cases that span
3731 // the peel/not_peel boundary such as bool->if
3732 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3733 new_phi_cnt++;
3734 }
3735 }
3736 }
3737 if (incr) i++;
3738 }
3739
3740 estimate += cloned_for_outside_use + new_phi_cnt;
3741 bool exceed_node_budget = !may_require_nodes(estimate);
3742 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3743
3744 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3745 #ifndef PRODUCT
3746 if (TracePartialPeeling && exceed_phi_limit) {
3747 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3748 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3749 }
3750 #endif
3751 if (new_peel_if != nullptr) {
3752 remove_cmpi_loop_exit(new_peel_if, loop);
3753 }
3754 // Inhibit more partial peeling on this loop
3755 assert(!head->is_partial_peel_loop(), "not partial peeled");
3756 head->mark_partial_peel_failed();
3757 if (cloned_for_outside_use > 0) {
3758 // Terminate this round of loop opts because
3759 // the graph outside this loop was changed.
3760 C->set_major_progress();
3761 return true;
3762 }
3763 return false;
3764 }
3765
3766 // Step 3: clone loop, retarget control, and insert new phis
3767
3768 // Create new loop head for new phis and to hang
3769 // the nodes being moved (sinked) from the peel region.
3770 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3771 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3772 _igvn.register_new_node_with_optimizer(new_head);
3773 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3774 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3775 set_loop(new_head, loop);
3776 loop->_body.push(new_head);
3777 not_peel.set(new_head->_idx);
3778 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3779 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3780
3781 while (sink_list.size() > 0) {
3782 Node* n = sink_list.pop();
3783 set_ctrl(n, new_head);
3784 }
3785
3786 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3787
3788 clone_loop(loop, old_new, dd, IgnoreStripMined);
3789
3790 const uint clone_exit_idx = 1;
3791 const uint orig_exit_idx = 2;
3792 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3793
3794 Node* head_clone = old_new[head->_idx];
3795 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3796 Node* orig_tail_clone = head_clone->in(2);
3797
3798 // Add phi if "def" node is in peel set and "use" is not
3799
3800 for (uint i = 0; i < peel_list.size(); i++) {
3801 Node *def = peel_list.at(i);
3802 if (!def->is_CFG()) {
3803 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3804 Node *use = def->fast_out(j);
3805 if (has_node(use) && use->in(0) != C->top() &&
3806 (!peel.test(use->_idx) ||
3807 (use->is_Phi() && use->in(0) == head)) ) {
3808 worklist.push(use);
3809 }
3810 }
3811 while( worklist.size() ) {
3812 Node *use = worklist.pop();
3813 for (uint j = 1; j < use->req(); j++) {
3814 Node* n = use->in(j);
3815 if (n == def) {
3816
3817 // "def" is in peel set, "use" is not in peel set
3818 // or "use" is in the entry boundary (a phi) of the peel set
3819
3820 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3821
3822 if ( loop->is_member(get_loop( use_c )) ) {
3823 // use is in loop
3824 if (old_new[use->_idx] != nullptr) { // null for dead code
3825 Node* use_clone = old_new[use->_idx];
3826 _igvn.replace_input_of(use, j, C->top());
3827 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3828 }
3829 } else {
3830 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3831 // use is not in the loop, check if the live range includes the cut
3832 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3833 if (not_peel.test(lp_if->_idx)) {
3834 assert(j == orig_exit_idx, "use from original loop");
3835 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3836 }
3837 }
3838 }
3839 }
3840 }
3841 }
3842 }
3843
3844 // Step 3b: retarget control
3845
3846 // Redirect control to the new loop head if a cloned node in
3847 // the not_peeled region has control that points into the peeled region.
3848 // This necessary because the cloned peeled region will be outside
3849 // the loop.
3850 // from to
3851 // cloned-peeled <---+
3852 // new_head_clone: | <--+
3853 // cloned-not_peeled in(0) in(0)
3854 // orig-peeled
3855
3856 for (uint i = 0; i < loop->_body.size(); i++) {
3857 Node *n = loop->_body.at(i);
3858 if (!n->is_CFG() && n->in(0) != nullptr &&
3859 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3860 Node* n_clone = old_new[n->_idx];
3861 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3862 }
3863 }
3864
3865 // Backedge of the surviving new_head (the clone) is original last_peel
3866 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3867
3868 // Cut first node in original not_peel set
3869 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3870 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3871 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3872
3873 // Copy head_clone back-branch info to original head
3874 // and remove original head's loop entry and
3875 // clone head's back-branch
3876 _igvn.rehash_node_delayed(head); // Multiple edge updates
3877 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3878 head->set_req(LoopNode::LoopBackControl, C->top());
3879 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3880
3881 // Similarly modify the phis
3882 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3883 Node* use = head->fast_out(k);
3884 if (use->is_Phi() && use->outcnt() > 0) {
3885 Node* use_clone = old_new[use->_idx];
3886 _igvn.rehash_node_delayed(use); // Multiple edge updates
3887 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3888 use->set_req(LoopNode::LoopBackControl, C->top());
3889 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3890 }
3891 }
3892
3893 // Step 4: update dominator tree and dominator depth
3894
3895 set_idom(head, orig_tail_clone, dd);
3896 recompute_dom_depth();
3897
3898 // Inhibit more partial peeling on this loop
3899 new_head_clone->set_partial_peel_loop();
3900 C->set_major_progress();
3901 loop->record_for_igvn();
3902
3903 #ifndef PRODUCT
3904 if (TracePartialPeeling) {
3905 tty->print_cr("\nafter partial peel one iteration");
3906 Node_List wl;
3907 Node* t = last_peel;
3908 while (true) {
3909 wl.push(t);
3910 if (t == head_clone) break;
3911 t = idom(t);
3912 }
3913 while (wl.size() > 0) {
3914 Node* tt = wl.pop();
3915 if (tt == head) tty->print_cr("orig head");
3916 else if (tt == new_head_clone) tty->print_cr("new head");
3917 else if (tt == head_clone) tty->print_cr("clone head");
3918 tt->dump();
3919 }
3920 }
3921 #endif
3922 return true;
3923 }
3924
3925 // Transform:
3926 //
3927 // loop<-----------------+
3928 // | |
3929 // stmt1 stmt2 .. stmtn |
3930 // | | | |
3931 // \ | / |
3932 // v v v |
3933 // region |
3934 // | |
3935 // shared_stmt |
3936 // | |
3937 // v |
3938 // if |
3939 // / \ |
3940 // | -----------+
3941 // v
3942 //
3943 // into:
3944 //
3945 // loop<-------------------+
3946 // | |
3947 // v |
3948 // +->loop |
3949 // | | |
3950 // | stmt1 stmt2 .. stmtn |
3951 // | | | | |
3952 // | | \ / |
3953 // | | v v |
3954 // | | region1 |
3955 // | | | |
3956 // | shared_stmt shared_stmt |
3957 // | | | |
3958 // | v v |
3959 // | if if |
3960 // | /\ / \ |
3961 // +-- | | -------+
3962 // \ /
3963 // v v
3964 // region2
3965 //
3966 // (region2 is shown to merge mirrored projections of the loop exit
3967 // ifs to make the diagram clearer but they really merge the same
3968 // projection)
3969 //
3970 // Conditions for this transformation to trigger:
3971 // - the path through stmt1 is frequent enough
3972 // - the inner loop will be turned into a counted loop after transformation
3973 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
3974 if (!DuplicateBackedge) {
3975 return false;
3976 }
3977 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
3978 if (!loop->_head->is_Loop()) {
3979 return false;
3980 }
3981
3982 uint estimate = loop->est_loop_clone_sz(1);
3983 if (exceeding_node_budget(estimate)) {
3984 return false;
3985 }
3986
3987 LoopNode *head = loop->_head->as_Loop();
3988
3989 Node* region = nullptr;
3990 IfNode* exit_test = nullptr;
3991 uint inner;
3992 float f;
3993 if (StressDuplicateBackedge) {
3994 if (head->is_strip_mined()) {
3995 return false;
3996 }
3997 Node* c = head->in(LoopNode::LoopBackControl);
3998
3999 while (c != head) {
4000 if (c->is_Region()) {
4001 region = c;
4002 }
4003 c = idom(c);
4004 }
4005
4006 if (region == nullptr) {
4007 return false;
4008 }
4009
4010 inner = 1;
4011 } else {
4012 // Is the shape of the loop that of a counted loop...
4013 Node* back_control = loop_exit_control(head, loop);
4014 if (back_control == nullptr) {
4015 return false;
4016 }
4017
4018 BoolTest::mask bt = BoolTest::illegal;
4019 float cl_prob = 0;
4020 Node* incr = nullptr;
4021 Node* limit = nullptr;
4022 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4023 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4024 return false;
4025 }
4026
4027 // With an extra phi for the candidate iv?
4028 // Or the region node is the loop head
4029 if (!incr->is_Phi() || incr->in(0) == head) {
4030 return false;
4031 }
4032
4033 PathFrequency pf(head, this);
4034 region = incr->in(0);
4035
4036 // Go over all paths for the extra phi's region and see if that
4037 // path is frequent enough and would match the expected iv shape
4038 // if the extra phi is removed
4039 inner = 0;
4040 for (uint i = 1; i < incr->req(); ++i) {
4041 Node* in = incr->in(i);
4042 Node* trunc1 = nullptr;
4043 Node* trunc2 = nullptr;
4044 const TypeInteger* iv_trunc_t = nullptr;
4045 Node* orig_in = in;
4046 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4047 continue;
4048 }
4049 assert(in->Opcode() == Op_AddI, "wrong increment code");
4050 Node* xphi = nullptr;
4051 Node* stride = loop_iv_stride(in, loop, xphi);
4052
4053 if (stride == nullptr) {
4054 continue;
4055 }
4056
4057 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4058 if (phi == nullptr ||
4059 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4060 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4061 return false;
4062 }
4063
4064 f = pf.to(region->in(i));
4065 if (f > 0.5) {
4066 inner = i;
4067 break;
4068 }
4069 }
4070
4071 if (inner == 0) {
4072 return false;
4073 }
4074
4075 exit_test = back_control->in(0)->as_If();
4076 }
4077
4078 if (idom(region)->is_Catch()) {
4079 return false;
4080 }
4081
4082 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4083 Unique_Node_List wq;
4084 wq.push(head->in(LoopNode::LoopBackControl));
4085 for (uint i = 0; i < wq.size(); i++) {
4086 Node* c = wq.at(i);
4087 assert(get_loop(c) == loop, "not in the right loop?");
4088 if (c->is_Region()) {
4089 if (c != region) {
4090 for (uint j = 1; j < c->req(); ++j) {
4091 wq.push(c->in(j));
4092 }
4093 }
4094 } else {
4095 wq.push(c->in(0));
4096 }
4097 assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
4098 }
4099
4100 Node* region_dom = idom(region);
4101
4102 // Can't do the transformation if this would cause a membar pair to
4103 // be split
4104 for (uint i = 0; i < wq.size(); i++) {
4105 Node* c = wq.at(i);
4106 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4107 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4108 if (!wq.member(c->as_MemBar()->leading_membar())) {
4109 return false;
4110 }
4111 }
4112 }
4113
4114 // Collect data nodes that need to be clones as well
4115 int dd = dom_depth(head);
4116
4117 for (uint i = 0; i < loop->_body.size(); ++i) {
4118 Node* n = loop->_body.at(i);
4119 if (has_ctrl(n)) {
4120 Node* c = get_ctrl(n);
4121 if (wq.member(c)) {
4122 wq.push(n);
4123 }
4124 } else {
4125 set_idom(n, idom(n), dd);
4126 }
4127 }
4128
4129 // clone shared_stmt
4130 clone_loop_body(wq, old_new, nullptr);
4131
4132 Node* region_clone = old_new[region->_idx];
4133 region_clone->set_req(inner, C->top());
4134 set_idom(region, region->in(inner), dd);
4135
4136 // Prepare the outer loop
4137 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4138 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4139 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4140 set_idom(head, outer_head, dd);
4141
4142 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4143
4144 // Make one of the shared_stmt copies only reachable from stmt1, the
4145 // other only from stmt2..stmtn.
4146 Node* dom = nullptr;
4147 for (uint i = 1; i < region->req(); ++i) {
4148 if (i != inner) {
4149 _igvn.replace_input_of(region, i, C->top());
4150 }
4151 Node* in = region_clone->in(i);
4152 if (in->is_top()) {
4153 continue;
4154 }
4155 if (dom == nullptr) {
4156 dom = in;
4157 } else {
4158 dom = dom_lca(dom, in);
4159 }
4160 }
4161
4162 set_idom(region_clone, dom, dd);
4163
4164 // Set up the outer loop
4165 for (uint i = 0; i < head->outcnt(); i++) {
4166 Node* u = head->raw_out(i);
4167 if (u->is_Phi()) {
4168 Node* outer_phi = u->clone();
4169 outer_phi->set_req(0, outer_head);
4170 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4171 if (backedge == nullptr) {
4172 backedge = u->in(LoopNode::LoopBackControl);
4173 }
4174 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4175 register_new_node(outer_phi, outer_head);
4176 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4177 }
4178 }
4179
4180 // create control and data nodes for out of loop uses (including region2)
4181 Node_List worklist;
4182 uint new_counter = C->unique();
4183 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4184
4185 Node_List *split_if_set = nullptr;
4186 Node_List *split_bool_set = nullptr;
4187 Node_List *split_cex_set = nullptr;
4188 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);
4189
4190 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4191
4192 if (exit_test != nullptr) {
4193 float cnt = exit_test->_fcnt;
4194 if (cnt != COUNT_UNKNOWN) {
4195 exit_test->_fcnt = cnt * f;
4196 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4197 }
4198 }
4199
4200 C->set_major_progress();
4201
4202 return true;
4203 }
4204
4205 // Having ReductionNodes in the loop is expensive. They need to recursively
4206 // fold together the vector values, for every vectorized loop iteration. If
4207 // we encounter the following pattern, we can vector accumulate the values
4208 // inside the loop, and only have a single UnorderedReduction after the loop.
4209 //
4210 // CountedLoop init
4211 // | |
4212 // +------+ | +-----------------------+
4213 // | | | |
4214 // PhiNode (s) |
4215 // | |
4216 // | Vector |
4217 // | | |
4218 // UnorderedReduction (first_ur) |
4219 // | |
4220 // ... Vector |
4221 // | | |
4222 // UnorderedReduction (last_ur) |
4223 // | |
4224 // +---------------------+
4225 //
4226 // We patch the graph to look like this:
4227 //
4228 // CountedLoop identity_vector
4229 // | |
4230 // +-------+ | +---------------+
4231 // | | | |
4232 // PhiNode (v) |
4233 // | |
4234 // | Vector |
4235 // | | |
4236 // VectorAccumulator |
4237 // | |
4238 // ... Vector |
4239 // | | |
4240 // init VectorAccumulator |
4241 // | | | |
4242 // UnorderedReduction +-----------+
4243 //
4244 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4245 // use vector_accumulators, which do the same reductions, but only element
4246 // wise. This is a single operation per vector_accumulator, rather than many
4247 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4248 // after the loop, and also reduce the init value into it.
4249 // We can not do this with all reductions. Some reductions do not allow the
4250 // reordering of operations (for example float addition).
4251 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4252 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4253
4254 // Find all Phi nodes with UnorderedReduction on backedge.
4255 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4256 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4257 Node* phi = cl->fast_out(j);
4258 // We have a phi with a single use, and a UnorderedReduction on the backedge.
4259 if (!phi->is_Phi() || phi->outcnt() != 1 || !phi->in(2)->is_UnorderedReduction()) {
4260 continue;
4261 }
4262
4263 UnorderedReductionNode* last_ur = phi->in(2)->as_UnorderedReduction();
4264
4265 // Determine types
4266 const TypeVect* vec_t = last_ur->vect_type();
4267 uint vector_length = vec_t->length();
4268 BasicType bt = vec_t->element_basic_type();
4269 const Type* bt_t = Type::get_const_basic_type(bt);
4270
4271 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4272 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4273 const int vopc = VectorNode::opcode(sopc, bt);
4274 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4275 DEBUG_ONLY( last_ur->dump(); )
4276 assert(false, "do not have normal vector op for this reduction");
4277 continue; // not implemented -> fails
4278 }
4279
4280 // Traverse up the chain of UnorderedReductions, checking that it loops back to
4281 // the phi. Check that all UnorderedReductions only have a single use, except for
4282 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4283 // are outside the loop.
4284 UnorderedReductionNode* current = last_ur;
4285 UnorderedReductionNode* first_ur = nullptr;
4286 while (true) {
4287 assert(current->is_UnorderedReduction(), "sanity");
4288
4289 // Expect no ctrl and a vector_input from within the loop.
4290 Node* ctrl = current->in(0);
4291 Node* vector_input = current->in(2);
4292 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4293 DEBUG_ONLY( current->dump(1); )
4294 assert(false, "reduction has ctrl or bad vector_input");
4295 break; // Chain traversal fails.
4296 }
4297
4298 assert(current->vect_type() != nullptr, "must have vector type");
4299 if (current->vect_type() != last_ur->vect_type()) {
4300 // Reductions do not have the same vector type (length and element type).
4301 break; // Chain traversal fails.
4302 }
4303
4304 // Expect single use of UnorderedReduction, except for last_ur.
4305 if (current == last_ur) {
4306 // Expect all uses to be outside the loop, except phi.
4307 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4308 Node* use = current->fast_out(k);
4309 if (use != phi && ctrl_or_self(use) == cl) {
4310 DEBUG_ONLY( current->dump(-1); )
4311 assert(false, "reduction has use inside loop");
4312 // Should not be allowed by SuperWord::mark_reductions
4313 return; // bail out of optimization
4314 }
4315 }
4316 } else {
4317 if (current->outcnt() != 1) {
4318 break; // Chain traversal fails.
4319 }
4320 }
4321
4322 // Expect another UnorderedReduction or phi as the scalar input.
4323 Node* scalar_input = current->in(1);
4324 if (scalar_input->is_UnorderedReduction() &&
4325 scalar_input->Opcode() == current->Opcode()) {
4326 // Move up the UnorderedReduction chain.
4327 current = scalar_input->as_UnorderedReduction();
4328 } else if (scalar_input == phi) {
4329 // Chain terminates at phi.
4330 first_ur = current;
4331 current = nullptr;
4332 break; // Success.
4333 } else {
4334 // scalar_input is neither phi nor a matching reduction
4335 // Can for example be scalar reduction when we have
4336 // partial vectorization.
4337 break; // Chain traversal fails.
4338 }
4339 }
4340 if (current != nullptr) {
4341 // Chain traversal was not successful.
4342 continue;
4343 }
4344 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4345
4346 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4347 set_ctrl(identity_scalar, C->root());
4348 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t);
4349 register_new_node(identity_vector, C->root());
4350 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4351 VectorNode::trace_new_vector(identity_vector, "UnorderedReduction");
4352
4353 // Turn the scalar phi into a vector phi.
4354 _igvn.rehash_node_delayed(phi);
4355 Node* init = phi->in(1); // Remember init before replacing it.
4356 phi->set_req_X(1, identity_vector, &_igvn);
4357 phi->as_Type()->set_type(vec_t);
4358 _igvn.set_type(phi, vec_t);
4359
4360 // Traverse down the chain of UnorderedReductions, and replace them with vector_accumulators.
4361 current = first_ur;
4362 while (true) {
4363 // Create vector_accumulator to replace current.
4364 Node* last_vector_accumulator = current->in(1);
4365 Node* vector_input = current->in(2);
4366 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4367 register_new_node(vector_accumulator, cl);
4368 _igvn.replace_node(current, vector_accumulator);
4369 VectorNode::trace_new_vector(vector_accumulator, "UnorderedReduction");
4370 if (current == last_ur) {
4371 break;
4372 }
4373 current = vector_accumulator->unique_out()->as_UnorderedReduction();
4374 }
4375
4376 // Create post-loop reduction.
4377 Node* last_accumulator = phi->in(2);
4378 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4379
4380 // Take over uses of last_accumulator that are not in the loop.
4381 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4382 Node* use = last_accumulator->out(i);
4383 if (use != phi && use != post_loop_reduction) {
4384 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4385 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4386 --i;
4387 }
4388 }
4389 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4390 VectorNode::trace_new_vector(post_loop_reduction, "UnorderedReduction");
4391
4392 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4393 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4394 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4395 }
4396 }