1 /*
2 * Copyright (c) 1999, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/inlinetypenode.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/matcher.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/opaquenode.hpp"
40 #include "opto/rootnode.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/subtypenode.hpp"
43 #include "opto/superword.hpp"
44 #include "opto/vectornode.hpp"
45 #include "utilities/checkedCast.hpp"
46 #include "utilities/macros.hpp"
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
52 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
53 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
54 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
55 // so disable this for now
56 return nullptr;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
63 n->in(1) == region->as_CountedLoop()->phi()) {
64 return nullptr;
65 }
66
67 // Inline types should not be split through Phis because they cannot be merged
68 // through Phi nodes but each value input needs to be merged individually.
69 if (n->is_InlineType()) {
70 return nullptr;
71 }
72
73 if (cannot_split_division(n, region)) {
74 return nullptr;
75 }
76
77 SplitThruPhiWins wins(region);
78 assert(!n->is_CFG(), "");
79 assert(region->is_Region(), "");
80
81 const Type* type = n->bottom_type();
82 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
83 Node* phi;
84 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
85 int iid = t_oop->instance_id();
86 int index = C->get_alias_index(t_oop);
87 int offset = t_oop->offset();
88 phi = new PhiNode(region, type, nullptr, iid, index, offset);
89 } else {
90 phi = PhiNode::make_blank(region, n);
91 }
92 uint old_unique = C->unique();
93 for (uint i = 1; i < region->req(); i++) {
94 Node* x;
95 Node* the_clone = nullptr;
96 if (region->in(i) == C->top()) {
97 x = C->top(); // Dead path? Use a dead data op
98 } else {
99 x = n->clone(); // Else clone up the data op
100 the_clone = x; // Remember for possible deletion.
101 // Alter data node to use pre-phi inputs
102 if (n->in(0) == region)
103 x->set_req( 0, region->in(i) );
104 for (uint j = 1; j < n->req(); j++) {
105 Node* in = n->in(j);
106 if (in->is_Phi() && in->in(0) == region)
107 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
108 }
109 }
110 // Check for a 'win' on some paths
111 const Type* t = x->Value(&_igvn);
112
113 bool singleton = t->singleton();
114
115 // A TOP singleton indicates that there are no possible values incoming
116 // along a particular edge. In most cases, this is OK, and the Phi will
117 // be eliminated later in an Ideal call. However, we can't allow this to
118 // happen if the singleton occurs on loop entry, as the elimination of
119 // the PhiNode may cause the resulting node to migrate back to a previous
120 // loop iteration.
121 if (singleton && t == Type::TOP) {
122 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
123 // irreducible loop may not be indicated by an affirmative is_Loop());
124 // therefore, the only top we can split thru a phi is on a backedge of
125 // a loop.
126 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
127 }
128
129 if (singleton) {
130 wins.add_win(i);
131 x = makecon(t);
132 } else {
133 // We now call Identity to try to simplify the cloned node.
134 // Note that some Identity methods call phase->type(this).
135 // Make sure that the type array is big enough for
136 // our new node, even though we may throw the node away.
137 // (Note: This tweaking with igvn only works because x is a new node.)
138 _igvn.set_type(x, t);
139 // If x is a TypeNode, capture any more-precise type permanently into Node
140 // otherwise it will be not updated during igvn->transform since
141 // igvn->type(x) is set to x->Value() already.
142 x->raise_bottom_type(t);
143 Node* y = x->Identity(&_igvn);
144 if (y != x) {
145 wins.add_win(i);
146 x = y;
147 } else {
148 y = _igvn.hash_find(x);
149 if (y == nullptr) {
150 y = similar_subtype_check(x, region->in(i));
151 }
152 if (y) {
153 wins.add_win(i);
154 x = y;
155 } else {
156 // Else x is a new node we are keeping
157 // We do not need register_new_node_with_optimizer
158 // because set_type has already been called.
159 _igvn._worklist.push(x);
160 }
161 }
162 }
163
164 phi->set_req( i, x );
165
166 if (the_clone == nullptr) {
167 continue;
168 }
169
170 if (the_clone != x) {
171 _igvn.remove_dead_node(the_clone, PhaseIterGVN::NodeOrigin::Speculative);
172 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
173 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
174 // it is not a win if 'x' moved from an outer to an inner loop
175 // this edge case can only happen for Load nodes
176 wins.reset();
177 break;
178 }
179 }
180 // Too few wins?
181 if (!wins.profitable(policy)) {
182 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Speculative);
183 return nullptr;
184 }
185
186 // Record Phi
187 register_new_node( phi, region );
188
189 for (uint i2 = 1; i2 < phi->req(); i2++) {
190 Node *x = phi->in(i2);
191 // If we commoned up the cloned 'x' with another existing Node,
192 // the existing Node picks up a new use. We need to make the
193 // existing Node occur higher up so it dominates its uses.
194 Node *old_ctrl;
195 IdealLoopTree *old_loop;
196
197 if (x->is_Con()) {
198 assert(get_ctrl(x) == C->root(), "constant control is not root");
199 continue;
200 }
201 // The occasional new node
202 if (x->_idx >= old_unique) { // Found a new, unplaced node?
203 old_ctrl = nullptr;
204 old_loop = nullptr; // Not in any prior loop
205 } else {
206 old_ctrl = get_ctrl(x);
207 old_loop = get_loop(old_ctrl); // Get prior loop
208 }
209 // New late point must dominate new use
210 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
211 if (new_ctrl == old_ctrl) // Nothing is changed
212 continue;
213
214 IdealLoopTree *new_loop = get_loop(new_ctrl);
215
216 // Don't move x into a loop if its uses are
217 // outside of loop. Otherwise x will be cloned
218 // for each use outside of this loop.
219 IdealLoopTree *use_loop = get_loop(region);
220 if (!new_loop->is_member(use_loop) &&
221 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
222 // Take early control, later control will be recalculated
223 // during next iteration of loop optimizations.
224 new_ctrl = get_early_ctrl(x);
225 new_loop = get_loop(new_ctrl);
226 }
227 // Set new location
228 set_ctrl(x, new_ctrl);
229 // If changing loop bodies, see if we need to collect into new body
230 if (old_loop != new_loop) {
231 if (old_loop && !old_loop->_child)
232 old_loop->_body.yank(x);
233 if (!new_loop->_child)
234 new_loop->_body.push(x); // Collect body info
235 }
236 }
237
238 split_thru_phi_yank_old_nodes(n, region);
239 _igvn.replace_node(n, phi);
240
241 #ifndef PRODUCT
242 if (TraceLoopOpts) {
243 tty->print_cr("Split %d %s through %d Phi in %d %s",
244 n->_idx, n->Name(), phi->_idx, region->_idx, region->Name());
245 }
246 #endif // !PRODUCT
247
248 return phi;
249 }
250
251 // If the region is a Loop, we are removing the old n,
252 // and need to yank it from the _body. If any phi we
253 // just split through now has no use any more, it also
254 // has to be removed.
255 void PhaseIdealLoop::split_thru_phi_yank_old_nodes(Node* n, Node* region) {
256 IdealLoopTree* region_loop = get_loop(region);
257 if (region->is_Loop() && region_loop->is_innermost()) {
258 region_loop->_body.yank(n);
259 for (uint j = 1; j < n->req(); j++) {
260 PhiNode* phi = n->in(j)->isa_Phi();
261 // Check that phi belongs to the region and only has n as a use.
262 if (phi != nullptr &&
263 phi->in(0) == region &&
264 phi->unique_multiple_edges_out_or_null() == n) {
265 assert(get_ctrl(phi) == region, "sanity");
266 assert(get_ctrl(n) == region, "sanity");
267 region_loop->_body.yank(phi);
268 }
269 }
270 }
271 }
272
273 // Test whether node 'x' can move into an inner loop relative to node 'n'.
274 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
275 // BUT it can also return true and 'x' is in the outer loop
276 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
277 IdealLoopTree* n_loop_tree = get_loop(n_loop);
278 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
279 // x_loop_tree should be outer or same loop as n_loop_tree
280 return !x_loop_tree->is_member(n_loop_tree);
281 }
282
283 // Subtype checks that carry profile data don't common so look for a replacement by following edges
284 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
285 if (x->is_SubTypeCheck()) {
286 Node* in1 = x->in(1);
287 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
288 Node* u = in1->fast_out(i);
289 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
290 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
291 Node* bol = u->fast_out(j);
292 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
293 Node* iff = bol->fast_out(k);
294 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
295 // unrelated profile
296 if (iff->is_If() && is_dominator(iff, r_in)) {
297 return u;
298 }
299 }
300 }
301 }
302 }
303 }
304 return nullptr;
305 }
306
307 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
308 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
309 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
310 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
311 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
312 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
313 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
314 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
315 const Type* zero;
316 switch (n->Opcode()) {
317 case Op_DivI:
318 case Op_ModI:
319 case Op_UDivI:
320 case Op_UModI:
321 zero = TypeInt::ZERO;
322 break;
323 case Op_DivL:
324 case Op_ModL:
325 case Op_UDivL:
326 case Op_UModL:
327 zero = TypeLong::ZERO;
328 break;
329 default:
330 return false;
331 }
332
333 if (n->in(0) != nullptr) {
334 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
335 return true;
336 }
337
338 Node* divisor = n->in(2);
339 return is_divisor_loop_phi(divisor, region) &&
340 loop_phi_backedge_type_contains_zero(divisor, zero);
341 }
342
343 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
344 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
345 }
346
347 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
348 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
349 }
350
351 //------------------------------dominated_by------------------------------------
352 // Replace the dominated test with an obvious true or false. Place it on the
353 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
354 // live path up to the dominating control.
355 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool prevdom_not_imply_this) {
356 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
357
358 // prevdom is the dominating projection of the dominating test.
359 assert(iff->Opcode() == Op_If ||
360 iff->Opcode() == Op_CountedLoopEnd ||
361 iff->Opcode() == Op_LongCountedLoopEnd ||
362 iff->Opcode() == Op_RangeCheck ||
363 iff->Opcode() == Op_ParsePredicate,
364 "Check this code when new subtype is added");
365
366 int pop = prevdom->Opcode();
367 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
368 if (flip) {
369 if (pop == Op_IfTrue)
370 pop = Op_IfFalse;
371 else
372 pop = Op_IfTrue;
373 }
374 // 'con' is set to true or false to kill the dominated test.
375 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
376 // Hack the dominated test
377 _igvn.replace_input_of(iff, 1, con);
378
379 // If I don't have a reachable TRUE and FALSE path following the IfNode then
380 // I can assume this path reaches an infinite loop. In this case it's not
381 // important to optimize the data Nodes - either the whole compilation will
382 // be tossed or this path (and all data Nodes) will go dead.
383 if (iff->outcnt() != 2) {
384 return;
385 }
386
387 // Make control-dependent data Nodes on the live path (path that will remain
388 // once the dominated IF is removed) become control-dependent on the
389 // dominating projection.
390 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
391
392 if (dp == nullptr) {
393 return;
394 }
395
396 rewire_safe_outputs_to_dominator(dp, prevdom, prevdom_not_imply_this);
397 }
398
399 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool dominator_not_imply_source) {
400 IdealLoopTree* old_loop = get_loop(source);
401
402 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
403 Node* out = source->fast_out(i); // Control-dependent node
404 if (out->depends_only_on_test()) {
405 assert(out->in(0) == source, "must be control dependent on source");
406 _igvn.replace_input_of(out, 0, dominator);
407 if (dominator_not_imply_source) {
408 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
409 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
410 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
411 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
412 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
413 // dominating check.
414 Node* clone = out->pin_node_under_control();
415 if (clone != nullptr) {
416 clone = _igvn.register_new_node_with_optimizer(clone, out);
417 _igvn.replace_node(out, clone);
418 out = clone;
419 }
420 }
421 set_early_ctrl(out, false);
422 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
423 if (old_loop != new_loop) {
424 if (!old_loop->_child) {
425 old_loop->_body.yank(out);
426 }
427 if (!new_loop->_child) {
428 new_loop->_body.push(out);
429 }
430 }
431 --i;
432 --imax;
433 }
434 }
435 }
436
437 //------------------------------has_local_phi_input----------------------------
438 // Return TRUE if 'n' has Phi inputs from its local block and no other
439 // block-local inputs (all non-local-phi inputs come from earlier blocks)
440 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
441 Node *n_ctrl = get_ctrl(n);
442 // See if some inputs come from a Phi in this block, or from before
443 // this block.
444 uint i;
445 for( i = 1; i < n->req(); i++ ) {
446 Node *phi = n->in(i);
447 if( phi->is_Phi() && phi->in(0) == n_ctrl )
448 break;
449 }
450 if( i >= n->req() )
451 return nullptr; // No Phi inputs; nowhere to clone thru
452
453 // Check for inputs created between 'n' and the Phi input. These
454 // must split as well; they have already been given the chance
455 // (courtesy of a post-order visit) and since they did not we must
456 // recover the 'cost' of splitting them by being very profitable
457 // when splitting 'n'. Since this is unlikely we simply give up.
458 for( i = 1; i < n->req(); i++ ) {
459 Node *m = n->in(i);
460 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
461 // We allow the special case of AddP's with no local inputs.
462 // This allows us to split-up address expressions.
463 if (m->is_AddP() &&
464 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
465 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
466 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
467 // Move the AddP up to the dominating point. That's fine because control of m's inputs
468 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
469 Node* c = find_non_split_ctrl(idom(n_ctrl));
470 if (c->is_OuterStripMinedLoop()) {
471 c->as_Loop()->verify_strip_mined(1);
472 c = c->in(LoopNode::EntryControl);
473 }
474 set_ctrl_and_loop(m, c);
475 continue;
476 }
477 return nullptr;
478 }
479 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
480 }
481
482 return n_ctrl;
483 }
484
485 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
486 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
487 assert(bt == T_INT || bt == T_LONG, "only for integers");
488 int n_op = n->Opcode();
489
490 if (n_op == Op_LShift(bt)) {
491 // Scale is loop invariant
492 Node* scale = n->in(2);
493 Node* scale_ctrl = get_ctrl(scale);
494 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
495 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
496 return nullptr;
497 }
498 const TypeInt* scale_t = scale->bottom_type()->isa_int();
499 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
500 return nullptr; // Dont bother with byte/short masking
501 }
502 // Add must vary with loop (else shift would be loop-invariant)
503 Node* add = n->in(1);
504 Node* add_ctrl = get_ctrl(add);
505 IdealLoopTree* add_loop = get_loop(add_ctrl);
506 if (n_loop != add_loop) {
507 return nullptr; // happens w/ evil ZKM loops
508 }
509
510 // Convert I-V into I+ (0-V); same for V-I
511 if (add->Opcode() == Op_Sub(bt) &&
512 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
513 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
514 Node* zero = integercon(0, bt);
515 Node* neg = SubNode::make(zero, add->in(2), bt);
516 register_new_node_with_ctrl_of(neg, add->in(2));
517 add = AddNode::make(add->in(1), neg, bt);
518 register_new_node(add, add_ctrl);
519 }
520 if (add->Opcode() != Op_Add(bt)) return nullptr;
521 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
522 // See if one add input is loop invariant
523 Node* add_var = add->in(1);
524 Node* add_var_ctrl = get_ctrl(add_var);
525 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
526 Node* add_invar = add->in(2);
527 Node* add_invar_ctrl = get_ctrl(add_invar);
528 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
529 if (add_invar_loop == n_loop) {
530 // Swap to find the invariant part
531 add_invar = add_var;
532 add_invar_ctrl = add_var_ctrl;
533 add_invar_loop = add_var_loop;
534 add_var = add->in(2);
535 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
536 return nullptr;
537 }
538 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
539 return nullptr; // No invariant part of the add?
540 }
541
542 // Yes! Reshape address expression!
543 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
544 Node* inv_scale_ctrl =
545 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
546 add_invar_ctrl : scale_ctrl;
547 register_new_node(inv_scale, inv_scale_ctrl);
548 Node* var_scale = LShiftNode::make(add_var, scale, bt);
549 register_new_node(var_scale, n_ctrl);
550 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
551 register_new_node(var_add, n_ctrl);
552 _igvn.replace_node(n, var_add);
553 return var_add;
554 }
555 return nullptr;
556 }
557
558 //------------------------------remix_address_expressions----------------------
559 // Rework addressing expressions to get the most loop-invariant stuff
560 // moved out. We'd like to do all associative operators, but it's especially
561 // important (common) to do address expressions.
562 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
563 if (!has_ctrl(n)) return nullptr;
564 Node* n_ctrl = get_ctrl(n);
565 IdealLoopTree* n_loop = get_loop(n_ctrl);
566
567 // See if 'n' mixes loop-varying and loop-invariant inputs and
568 // itself is loop-varying.
569
570 // Only interested in binary ops (and AddP)
571 if (n->req() < 3 || n->req() > 4) return nullptr;
572
573 Node* n1_ctrl = get_ctrl(n->in( 1));
574 Node* n2_ctrl = get_ctrl(n->in( 2));
575 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
576 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
577 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
578 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
579
580 // Does one of my inputs spin in a tighter loop than self?
581 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
582 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
583 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
584 return nullptr; // Leave well enough alone
585 }
586
587 // Is at least one of my inputs loop-invariant?
588 if (n1_loop == n_loop &&
589 n2_loop == n_loop &&
590 n3_loop == n_loop) {
591 return nullptr; // No loop-invariant inputs
592 }
593
594 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
595 if (res != nullptr) {
596 return res;
597 }
598 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
599 if (res != nullptr) {
600 return res;
601 }
602
603 int n_op = n->Opcode();
604 // Replace (I+V) with (V+I)
605 if (n_op == Op_AddI ||
606 n_op == Op_AddL ||
607 n_op == Op_AddF ||
608 n_op == Op_AddD ||
609 n_op == Op_MulI ||
610 n_op == Op_MulL ||
611 n_op == Op_MulF ||
612 n_op == Op_MulD) {
613 if (n2_loop == n_loop) {
614 assert(n1_loop != n_loop, "");
615 n->swap_edges(1, 2);
616 }
617 }
618
619 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
620 // but not if I2 is a constant. Skip for irreducible loops.
621 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
622 if (n2_loop == n_loop && n3_loop != n_loop) {
623 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
624 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
625 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
626 IdealLoopTree* n22loop = get_loop(n22_ctrl);
627 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
628 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
629 n23_loop == n_loop) {
630 Node* add1 = AddPNode::make_with_base(n->in(1), n->in(2)->in(2), n->in(3));
631 // Stuff new AddP in the loop preheader
632 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
633 Node* add2 = AddPNode::make_with_base(n->in(1), add1, n->in(2)->in(3));
634 register_new_node(add2, n_ctrl);
635 _igvn.replace_node(n, add2);
636 return add2;
637 }
638 }
639 }
640
641 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
642 if (n2_loop != n_loop && n3_loop == n_loop) {
643 if (n->in(3)->Opcode() == Op_AddX) {
644 Node* V = n->in(3)->in(1);
645 Node* I = n->in(3)->in(2);
646 if (ctrl_is_member(n_loop, V)) {
647 } else {
648 Node *tmp = V; V = I; I = tmp;
649 }
650 if (!ctrl_is_member(n_loop, I)) {
651 Node* add1 = AddPNode::make_with_base(n->in(1), n->in(2), I);
652 // Stuff new AddP in the loop preheader
653 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
654 Node* add2 = AddPNode::make_with_base(n->in(1), add1, V);
655 register_new_node(add2, n_ctrl);
656 _igvn.replace_node(n, add2);
657 return add2;
658 }
659 }
660 }
661 }
662
663 return nullptr;
664 }
665
666 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
667 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
668 assert(n->Opcode() == Op_AddI, "sanity");
669 Node * nn = nullptr;
670 Node * in1 = n->in(1);
671 Node * in2 = n->in(2);
672 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
673 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
674 if (loop_n->is_counted() &&
675 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
676 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
677 Matcher::match_rule_supported(Op_MulAddS2I)) {
678 Node* mul_in1 = in1->in(1);
679 Node* mul_in2 = in1->in(2);
680 Node* mul_in3 = in2->in(1);
681 Node* mul_in4 = in2->in(2);
682 if (mul_in1->Opcode() == Op_LoadS &&
683 mul_in2->Opcode() == Op_LoadS &&
684 mul_in3->Opcode() == Op_LoadS &&
685 mul_in4->Opcode() == Op_LoadS) {
686 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
687 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
688 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
689 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
690 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
691 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
692 // All nodes should be in the same counted loop.
693 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
694 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
695 Node* adr1 = mul_in1->in(MemNode::Address);
696 Node* adr2 = mul_in2->in(MemNode::Address);
697 Node* adr3 = mul_in3->in(MemNode::Address);
698 Node* adr4 = mul_in4->in(MemNode::Address);
699 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
700 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
701 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
702 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
703 register_new_node_with_ctrl_of(nn, n);
704 _igvn.replace_node(n, nn);
705 return nn;
706 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
707 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
708 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
709 register_new_node_with_ctrl_of(nn, n);
710 _igvn.replace_node(n, nn);
711 return nn;
712 }
713 }
714 }
715 }
716 }
717 }
718 return nn;
719 }
720
721 //------------------------------conditional_move-------------------------------
722 // Attempt to replace a Phi with a conditional move. We have some pretty
723 // strict profitability requirements. All Phis at the merge point must
724 // be converted, so we can remove the control flow. We need to limit the
725 // number of c-moves to a small handful. All code that was in the side-arms
726 // of the CFG diamond is now speculatively executed. This code has to be
727 // "cheap enough". We are pretty much limited to CFG diamonds that merge
728 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
729 Node *PhaseIdealLoop::conditional_move( Node *region ) {
730
731 assert(region->is_Region(), "sanity check");
732 if (region->req() != 3) return nullptr;
733
734 // Check for CFG diamond
735 Node *lp = region->in(1);
736 Node *rp = region->in(2);
737 if (!lp || !rp) return nullptr;
738 Node *lp_c = lp->in(0);
739 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
740 IfNode *iff = lp_c->as_If();
741
742 // Check for ops pinned in an arm of the diamond.
743 // Can't remove the control flow in this case
744 if (lp->outcnt() > 1) return nullptr;
745 if (rp->outcnt() > 1) return nullptr;
746
747 IdealLoopTree* r_loop = get_loop(region);
748 assert(r_loop == get_loop(iff), "sanity");
749 // Always convert to CMOVE if all results are used only outside this loop.
750 bool used_inside_loop = (r_loop == _ltree_root);
751
752 // Check profitability
753 int cost = 0;
754 int phis = 0;
755 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
756 Node *out = region->fast_out(i);
757 if (!out->is_Phi()) continue; // Ignore other control edges, etc
758 phis++;
759 PhiNode* phi = out->as_Phi();
760 BasicType bt = phi->type()->basic_type();
761 switch (bt) {
762 case T_DOUBLE:
763 case T_FLOAT:
764 if (C->use_cmove()) {
765 continue; //TODO: maybe we want to add some cost
766 }
767 cost += Matcher::float_cmove_cost(); // Could be very expensive
768 break;
769 case T_LONG: {
770 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
771 }
772 case T_INT: // These all CMOV fine
773 case T_ADDRESS: { // (RawPtr)
774 cost++;
775 break;
776 }
777 case T_NARROWOOP: // Fall through
778 case T_OBJECT: { // Base oops are OK, but not derived oops
779 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
780 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
781 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
782 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
783 // have a Phi for the base here that we convert to a CMOVE all is well
784 // and good. But if the base is dead, we'll not make a CMOVE. Later
785 // the allocator will have to produce a base by creating a CMOVE of the
786 // relevant bases. This puts the allocator in the business of
787 // manufacturing expensive instructions, generally a bad plan.
788 // Just Say No to Conditionally-Moved Derived Pointers.
789 if (tp && tp->offset() != 0)
790 return nullptr;
791 cost++;
792 break;
793 }
794 default:
795 return nullptr; // In particular, can't do memory or I/O
796 }
797 // Add in cost any speculative ops
798 for (uint j = 1; j < region->req(); j++) {
799 Node *proj = region->in(j);
800 Node *inp = phi->in(j);
801 if (get_ctrl(inp) == proj) { // Found local op
802 cost++;
803 // Check for a chain of dependent ops; these will all become
804 // speculative in a CMOV.
805 for (uint k = 1; k < inp->req(); k++)
806 if (get_ctrl(inp->in(k)) == proj)
807 cost += ConditionalMoveLimit; // Too much speculative goo
808 }
809 }
810 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
811 // This will likely Split-If, a higher-payoff operation.
812 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
813 Node* use = phi->fast_out(k);
814 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
815 cost += ConditionalMoveLimit;
816 // Is there a use inside the loop?
817 // Note: check only basic types since CMoveP is pinned.
818 if (!used_inside_loop && is_java_primitive(bt)) {
819 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
820 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
821 used_inside_loop = true;
822 }
823 }
824 }
825 }//for
826 Node* bol = iff->in(1);
827 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
828 if (bol->is_OpaqueTemplateAssertionPredicate()) {
829 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
830 return nullptr;
831 }
832 if (bol->is_OpaqueMultiversioning()) {
833 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
834 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
835 // and will make the multiversion_if constant fold in the next IGVN round.
836 return nullptr;
837 }
838 if (!bol->is_Bool()) {
839 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
840 return nullptr;
841 }
842 int cmp_op = bol->in(1)->Opcode();
843 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
844 return nullptr;
845 }
846 // It is expensive to generate flags from a float compare.
847 // Avoid duplicated float compare.
848 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
849
850 float infrequent_prob = PROB_UNLIKELY_MAG(3);
851 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
852 if (used_inside_loop) {
853 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
854
855 // BlockLayoutByFrequency optimization moves infrequent branch
856 // from hot path. No point in CMOV'ing in such case (110 is used
857 // instead of 100 to take into account not exactness of float value).
858 if (BlockLayoutByFrequency) {
859 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
860 }
861 }
862 // Check for highly predictable branch. No point in CMOV'ing if
863 // we are going to predict accurately all the time.
864 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
865 //keep going
866 } else if (iff->_prob < infrequent_prob ||
867 iff->_prob > (1.0f - infrequent_prob))
868 return nullptr;
869
870 // --------------
871 // Now replace all Phis with CMOV's
872 Node *cmov_ctrl = iff->in(0);
873 uint flip = (lp->Opcode() == Op_IfTrue);
874 Node_List wq;
875 while (1) {
876 PhiNode* phi = nullptr;
877 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
878 Node *out = region->fast_out(i);
879 if (out->is_Phi()) {
880 phi = out->as_Phi();
881 break;
882 }
883 }
884 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
885 break;
886 }
887 // Move speculative ops
888 wq.push(phi);
889 while (wq.size() > 0) {
890 Node *n = wq.pop();
891 for (uint j = 1; j < n->req(); j++) {
892 Node* m = n->in(j);
893 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
894 set_ctrl(m, cmov_ctrl);
895 wq.push(m);
896 }
897 }
898 }
899 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
900 register_new_node(cmov, cmov_ctrl);
901 _igvn.replace_node(phi, cmov);
902 #ifndef PRODUCT
903 if (TraceLoopOpts) {
904 tty->print("CMOV ");
905 r_loop->dump_head();
906 if (Verbose) {
907 bol->in(1)->dump(1);
908 cmov->dump(1);
909 }
910 }
911 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
912 #endif
913 }
914
915 // The useless CFG diamond will fold up later; see the optimization in
916 // RegionNode::Ideal.
917 _igvn._worklist.push(region);
918
919 return iff->in(1);
920 }
921
922 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
923 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
924 Node* u = m->fast_out(i);
925 if (u->is_CFG()) {
926 if (u->is_NeverBranch()) {
927 u = u->as_NeverBranch()->proj_out(0);
928 enqueue_cfg_uses(u, wq);
929 } else {
930 wq.push(u);
931 }
932 }
933 }
934 }
935
936 // Try moving a store out of a loop, right before the loop
937 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
938 // Store has to be first in the loop body
939 IdealLoopTree *n_loop = get_loop(n_ctrl);
940 if (n->is_Store() && n_loop != _ltree_root &&
941 n_loop->is_loop() && n_loop->_head->is_Loop() &&
942 n->in(0) != nullptr) {
943 Node* address = n->in(MemNode::Address);
944 Node* value = n->in(MemNode::ValueIn);
945 Node* mem = n->in(MemNode::Memory);
946
947 // - address and value must be loop invariant
948 // - memory must be a memory Phi for the loop
949 // - Store must be the only store on this memory slice in the
950 // loop: if there's another store following this one then value
951 // written at iteration i by the second store could be overwritten
952 // at iteration i+n by the first store: it's not safe to move the
953 // first store out of the loop
954 // - nothing must observe the memory Phi: it guarantees no read
955 // before the store, we are also guaranteed the store post
956 // dominates the loop head (ignoring a possible early
957 // exit). Otherwise there would be extra Phi involved between the
958 // loop's Phi and the store.
959 // - there must be no early exit from the loop before the Store
960 // (such an exit most of the time would be an extra use of the
961 // memory Phi but sometimes is a bottom memory Phi that takes the
962 // store as input).
963
964 if (!ctrl_is_member(n_loop, address) &&
965 !ctrl_is_member(n_loop, value) &&
966 mem->is_Phi() && mem->in(0) == n_loop->_head &&
967 mem->outcnt() == 1 &&
968 mem->in(LoopNode::LoopBackControl) == n) {
969
970 assert(n_loop->_tail != nullptr, "need a tail");
971 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
972
973 // Verify that there's no early exit of the loop before the store.
974 bool ctrl_ok = false;
975 {
976 // Follow control from loop head until n, we exit the loop or
977 // we reach the tail
978 ResourceMark rm;
979 Unique_Node_List wq;
980 wq.push(n_loop->_head);
981
982 for (uint next = 0; next < wq.size(); ++next) {
983 Node *m = wq.at(next);
984 if (m == n->in(0)) {
985 ctrl_ok = true;
986 continue;
987 }
988 assert(!has_ctrl(m), "should be CFG");
989 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
990 ctrl_ok = false;
991 break;
992 }
993 enqueue_cfg_uses(m, wq);
994 if (wq.size() > 10) {
995 ctrl_ok = false;
996 break;
997 }
998 }
999 }
1000 if (ctrl_ok) {
1001 // move the Store
1002 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
1003 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
1004 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
1005 // Disconnect the phi now. An empty phi can confuse other
1006 // optimizations in this pass of loop opts.
1007 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
1008 n_loop->_body.yank(mem);
1009
1010 set_ctrl_and_loop(n, n->in(0));
1011
1012 return n;
1013 }
1014 }
1015 }
1016 return nullptr;
1017 }
1018
1019 // Try moving a store out of a loop, right after the loop
1020 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
1021 if (n->is_Store() && n->in(0) != nullptr) {
1022 Node *n_ctrl = get_ctrl(n);
1023 IdealLoopTree *n_loop = get_loop(n_ctrl);
1024 // Store must be in a loop
1025 if (n_loop != _ltree_root && !n_loop->_irreducible) {
1026 Node* address = n->in(MemNode::Address);
1027 Node* value = n->in(MemNode::ValueIn);
1028 // address must be loop invariant
1029 if (!ctrl_is_member(n_loop, address)) {
1030 // Store must be last on this memory slice in the loop and
1031 // nothing in the loop must observe it
1032 Node* phi = nullptr;
1033 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1034 Node* u = n->fast_out(i);
1035 if (has_ctrl(u)) { // control use?
1036 if (!ctrl_is_member(n_loop, u)) {
1037 continue;
1038 }
1039 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1040 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1041 // multiple phis on the same slice are possible
1042 if (phi != nullptr) {
1043 return;
1044 }
1045 phi = u;
1046 continue;
1047 }
1048 }
1049 return;
1050 }
1051 if (phi != nullptr) {
1052 // Nothing in the loop before the store (next iteration)
1053 // must observe the stored value
1054 bool mem_ok = true;
1055 {
1056 ResourceMark rm;
1057 Unique_Node_List wq;
1058 wq.push(phi);
1059 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1060 Node *m = wq.at(next);
1061 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1062 Node* u = m->fast_out(i);
1063 if (u->is_Store() || u->is_Phi()) {
1064 if (u != n) {
1065 wq.push(u);
1066 mem_ok = (wq.size() <= 10);
1067 }
1068 } else {
1069 mem_ok = false;
1070 break;
1071 }
1072 }
1073 }
1074 }
1075 if (mem_ok) {
1076 // Move the store out of the loop if the LCA of all
1077 // users (except for the phi) is outside the loop.
1078 Node* hook = new Node(1);
1079 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1080 _igvn.rehash_node_delayed(phi);
1081 int count = phi->replace_edge(n, hook, &_igvn);
1082 assert(count > 0, "inconsistent phi");
1083
1084 // Compute latest point this store can go
1085 Node* lca = get_late_ctrl(n, get_ctrl(n));
1086 if (lca->is_OuterStripMinedLoop()) {
1087 lca = lca->in(LoopNode::EntryControl);
1088 }
1089 if (n_loop->is_member(get_loop(lca))) {
1090 // LCA is in the loop - bail out
1091 _igvn.replace_node(hook, n);
1092 return;
1093 }
1094 #ifdef ASSERT
1095 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1096 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1097 n_loop->_head->as_Loop()->verify_strip_mined(1);
1098 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1099 IdealLoopTree* outer_loop = get_loop(outer);
1100 assert(n_loop->_parent == outer_loop, "broken loop tree");
1101 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1102 }
1103 #endif
1104 lca = place_outside_loop(lca, n_loop);
1105 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1106 assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1107
1108 // Move store out of the loop
1109 _igvn.replace_node(hook, n->in(MemNode::Memory));
1110 _igvn.replace_input_of(n, 0, lca);
1111 set_ctrl_and_loop(n, lca);
1112
1113 // Disconnect the phi now. An empty phi can confuse other
1114 // optimizations in this pass of loop opts..
1115 if (phi->in(LoopNode::LoopBackControl) == phi) {
1116 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1117 n_loop->_body.yank(phi);
1118 }
1119 }
1120 }
1121 }
1122 }
1123 }
1124 }
1125
1126 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1127 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1128 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1129 // flat array checks out of loops, mainly to enable loop unswitching.
1130 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1131 // Skip checks for more than one array
1132 if (n->req() > 3) {
1133 return;
1134 }
1135 Node* mem = n->in(FlatArrayCheckNode::Memory);
1136 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1137 IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1138 IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1139
1140 // Check if array is loop invariant
1141 if (!check_loop->is_member(ary_loop)) {
1142 // Walk up memory graph from the check until we leave the loop
1143 VectorSet wq;
1144 wq.set(mem->_idx);
1145 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1146 if (mem->is_Phi()) {
1147 mem = mem->in(1);
1148 } else if (mem->is_MergeMem()) {
1149 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1150 } else if (mem->is_Proj()) {
1151 mem = mem->in(0);
1152 } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1153 mem = mem->in(TypeFunc::Memory);
1154 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1155 mem = mem->in(MemNode::Memory);
1156 } else {
1157 #ifdef ASSERT
1158 mem->dump();
1159 #endif
1160 ShouldNotReachHere();
1161 }
1162 if (wq.test_set(mem->_idx)) {
1163 return;
1164 }
1165 }
1166 // Replace memory input and re-compute ctrl to move the check out of the loop
1167 _igvn.replace_input_of(n, 1, mem);
1168 set_ctrl_and_loop(n, get_early_ctrl(n));
1169 Node* bol = n->unique_out();
1170 set_ctrl_and_loop(bol, get_early_ctrl(bol));
1171 }
1172 }
1173
1174 //------------------------------split_if_with_blocks_pre-----------------------
1175 // Do the real work in a non-recursive function. Data nodes want to be
1176 // cloned in the pre-order so they can feed each other nicely.
1177 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1178 // Cloning these guys is unlikely to win
1179 int n_op = n->Opcode();
1180 if (n_op == Op_MergeMem) {
1181 return n;
1182 }
1183 if (n->is_Proj()) {
1184 return n;
1185 }
1186
1187 if (n->isa_FlatArrayCheck()) {
1188 move_flat_array_check_out_of_loop(n);
1189 return n;
1190 }
1191
1192 // Do not clone-up CmpFXXX variations, as these are always
1193 // followed by a CmpI
1194 if (n->is_Cmp()) {
1195 return n;
1196 }
1197 // Attempt to use a conditional move instead of a phi/branch
1198 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1199 Node *cmov = conditional_move( n );
1200 if (cmov) {
1201 return cmov;
1202 }
1203 }
1204 if (n->is_CFG() || n->is_LoadStore()) {
1205 return n;
1206 }
1207 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1208 if (!C->major_progress()) { // If chance of no more loop opts...
1209 _igvn._worklist.push(n); // maybe we'll remove them
1210 }
1211 return n;
1212 }
1213
1214 if (n->is_Con()) {
1215 return n; // No cloning for Con nodes
1216 }
1217
1218 Node *n_ctrl = get_ctrl(n);
1219 if (!n_ctrl) {
1220 return n; // Dead node
1221 }
1222
1223 Node* res = try_move_store_before_loop(n, n_ctrl);
1224 if (res != nullptr) {
1225 return n;
1226 }
1227
1228 // Attempt to remix address expressions for loop invariants
1229 Node *m = remix_address_expressions( n );
1230 if( m ) return m;
1231
1232 if (n_op == Op_AddI) {
1233 Node *nn = convert_add_to_muladd( n );
1234 if ( nn ) return nn;
1235 }
1236
1237 if (n->is_ConstraintCast() && n->as_ConstraintCast()->dependency().narrows_type()) {
1238 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1239 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1240 // Node control inputs don't necessarily agree with loop control info (due to
1241 // transformations happened in between), thus additional dominance check is needed
1242 // to keep loop info valid.
1243 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1244 _igvn.replace_node(n, dom_cast);
1245 return dom_cast;
1246 }
1247 }
1248
1249 // Determine if the Node has inputs from some local Phi.
1250 // Returns the block to clone thru.
1251 Node *n_blk = has_local_phi_input( n );
1252 if( !n_blk ) return n;
1253
1254 // Do not clone the trip counter through on a CountedLoop
1255 // (messes up the canonical shape).
1256 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1257 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1258 return n;
1259 }
1260 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1261 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1262 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1263 return n;
1264 }
1265
1266 // Check for having no control input; not pinned. Allow
1267 // dominating control.
1268 if (n->in(0)) {
1269 Node *dom = idom(n_blk);
1270 if (dom_lca(n->in(0), dom) != n->in(0)) {
1271 return n;
1272 }
1273 }
1274 // Policy: when is it profitable. You must get more wins than
1275 // policy before it is considered profitable. Policy is usually 0,
1276 // so 1 win is considered profitable. Big merges will require big
1277 // cloning, so get a larger policy.
1278 int policy = n_blk->req() >> 2;
1279
1280 // If the loop is a candidate for range check elimination,
1281 // delay splitting through it's phi until a later loop optimization
1282 if (n_blk->is_BaseCountedLoop()) {
1283 IdealLoopTree *lp = get_loop(n_blk);
1284 if (lp && lp->_rce_candidate) {
1285 return n;
1286 }
1287 }
1288
1289 if (must_throttle_split_if()) return n;
1290
1291 // Split 'n' through the merge point if it is profitable, replacing it with a new phi.
1292 Node* phi = split_thru_phi(n, n_blk, policy);
1293 if (phi == nullptr) { return n; }
1294
1295 // Moved a load around the loop, 'en-registering' something.
1296 if (n_blk->is_Loop() && n->is_Load() &&
1297 !phi->in(LoopNode::LoopBackControl)->is_Load())
1298 C->set_major_progress();
1299
1300 return phi;
1301 }
1302
1303 static bool merge_point_too_heavy(Compile* C, Node* region) {
1304 // Bail out if the region and its phis have too many users.
1305 int weight = 0;
1306 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1307 weight += region->fast_out(i)->outcnt();
1308 }
1309 int nodes_left = C->max_node_limit() - C->live_nodes();
1310 if (weight * 8 > nodes_left) {
1311 if (PrintOpto) {
1312 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1313 }
1314 return true;
1315 } else {
1316 return false;
1317 }
1318 }
1319
1320 static bool merge_point_safe(Node* region) {
1321 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1322 // having a PhiNode input. This sidesteps the dangerous case where the split
1323 // ConvI2LNode may become TOP if the input Value() does not
1324 // overlap the ConvI2L range, leaving a node which may not dominate its
1325 // uses.
1326 // A better fix for this problem can be found in the BugTraq entry, but
1327 // expediency for Mantis demands this hack.
1328 #ifdef _LP64
1329 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1330 Node* n = region->fast_out(i);
1331 if (n->is_Phi()) {
1332 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1333 Node* m = n->fast_out(j);
1334 if (m->Opcode() == Op_ConvI2L)
1335 return false;
1336 if (m->is_CastII()) {
1337 return false;
1338 }
1339 }
1340 }
1341 }
1342 #endif
1343 return true;
1344 }
1345
1346
1347 //------------------------------place_outside_loop---------------------------------
1348 // Place some computation outside of this loop on the path to the use passed as argument
1349 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1350 Node* head = loop->_head;
1351 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1352 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1353 loop = loop->_parent;
1354 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1355 }
1356
1357 // Pick control right outside the loop
1358 for (;;) {
1359 Node* dom = idom(useblock);
1360 if (loop->is_member(get_loop(dom))) {
1361 break;
1362 }
1363 useblock = dom;
1364 }
1365 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1366 return useblock;
1367 }
1368
1369
1370 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1371 if (!n->is_If()) {
1372 return false;
1373 }
1374 if (n->outcnt() != n->as_If()->required_outcnt()) {
1375 assert(false, "malformed IfNode with %d outputs", n->outcnt());
1376 return false;
1377 }
1378 if (n->is_BaseCountedLoopEnd()) {
1379 return false;
1380 }
1381 if (!n->in(0)->is_Region()) {
1382 return false;
1383 }
1384
1385 Node* region = n->in(0);
1386 Node* dom = idom(region);
1387 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1388 return false;
1389 }
1390 IfNode* dom_if = dom->as_If();
1391 IfTrueNode* proj_true = dom_if->true_proj();
1392 IfFalseNode* proj_false = dom_if->false_proj();
1393
1394 for (uint i = 1; i < region->req(); i++) {
1395 if (is_dominator(proj_true, region->in(i))) {
1396 continue;
1397 }
1398 if (is_dominator(proj_false, region->in(i))) {
1399 continue;
1400 }
1401 return false;
1402 }
1403
1404 return true;
1405 }
1406
1407
1408 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1409 if (must_throttle_split_if()) {
1410 return false;
1411 }
1412
1413 // Do not do 'split-if' if irreducible loops are present.
1414 if (_has_irreducible_loops) {
1415 return false;
1416 }
1417
1418 if (merge_point_too_heavy(C, n_ctrl)) {
1419 return false;
1420 }
1421
1422 // Do not do 'split-if' if some paths are dead. First do dead code
1423 // elimination and then see if its still profitable.
1424 for (uint i = 1; i < n_ctrl->req(); i++) {
1425 if (n_ctrl->in(i) == C->top()) {
1426 return false;
1427 }
1428 }
1429
1430 // If trying to do a 'Split-If' at the loop head, it is only
1431 // profitable if the cmp folds up on BOTH paths. Otherwise we
1432 // risk peeling a loop forever.
1433
1434 // CNC - Disabled for now. Requires careful handling of loop
1435 // body selection for the cloned code. Also, make sure we check
1436 // for any input path not being in the same loop as n_ctrl. For
1437 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1438 // because the alternative loop entry points won't be converted
1439 // into LoopNodes.
1440 IdealLoopTree *n_loop = get_loop(n_ctrl);
1441 for (uint j = 1; j < n_ctrl->req(); j++) {
1442 if (get_loop(n_ctrl->in(j)) != n_loop) {
1443 return false;
1444 }
1445 }
1446
1447 // Check for safety of the merge point.
1448 if (!merge_point_safe(n_ctrl)) {
1449 return false;
1450 }
1451
1452 return true;
1453 }
1454
1455 // Detect if the node is the inner strip-mined loop
1456 // Return: null if it's not the case, or the exit of outer strip-mined loop
1457 static Node* is_inner_of_stripmined_loop(const Node* out) {
1458 Node* out_le = nullptr;
1459
1460 if (out->is_CountedLoopEnd()) {
1461 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1462
1463 if (loop != nullptr && loop->is_strip_mined()) {
1464 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1465 }
1466 }
1467
1468 return out_le;
1469 }
1470
1471 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) {
1472 // If the CmpP is a subtype check for a value that has just been
1473 // loaded from an array, the subtype check guarantees the value
1474 // can't be stored in a flat array and the load of the value
1475 // happens with a flat array check then: push the type check
1476 // through the phi of the flat array check. This needs special
1477 // logic because the subtype check's input is not a phi but a
1478 // LoadKlass that must first be cloned through the phi.
1479 if (n->Opcode() != Op_CmpP) {
1480 return false;
1481 }
1482
1483 Node* klassptr = n->in(1);
1484 Node* klasscon = n->in(2);
1485
1486 if (klassptr->is_DecodeNarrowPtr()) {
1487 klassptr = klassptr->in(1);
1488 }
1489
1490 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1491 return false;
1492 }
1493
1494 if (!klasscon->is_Con()) {
1495 return false;
1496 }
1497
1498 Node* addr = klassptr->in(MemNode::Address);
1499
1500 if (!addr->is_AddP()) {
1501 return false;
1502 }
1503
1504 intptr_t offset;
1505 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1506
1507 if (obj == nullptr) {
1508 return false;
1509 }
1510
1511 // TODO 8378077: The code below does not work anymore with off-heap accesses which set their bases to top with
1512 // JDK-8373343. Also: flat_array_element_type_check() was introduced with JDK-8228622 for a specific check to enable
1513 // split-if but JDK-8245729 changed how that check looks like. Is it still relevant? This should be revisited.
1514 if (addr->in(AddPNode::Base)->is_top()) {
1515 return false;
1516 }
1517
1518 if (obj->Opcode() == Op_CastPP) {
1519 obj = obj->in(1);
1520 }
1521
1522 if (!obj->is_Phi()) {
1523 return false;
1524 }
1525
1526 Node* region = obj->in(0);
1527
1528 Node* phi = PhiNode::make_blank(region, n->in(1));
1529 for (uint i = 1; i < region->req(); i++) {
1530 Node* in = obj->in(i);
1531 Node* ctrl = region->in(i);
1532 if (addr->in(AddPNode::Base) != obj) {
1533 Node* cast = addr->in(AddPNode::Base);
1534 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph");
1535 Node* cast_clone = cast->clone();
1536 cast_clone->set_req(0, ctrl);
1537 cast_clone->set_req(1, in);
1538 register_new_node(cast_clone, ctrl);
1539 const Type* tcast = cast_clone->Value(&_igvn);
1540 _igvn.set_type(cast_clone, tcast);
1541 cast_clone->as_Type()->set_type(tcast);
1542 in = cast_clone;
1543 }
1544 Node* addr_clone = addr->clone();
1545 addr_clone->set_req(AddPNode::Base, in);
1546 addr_clone->set_req(AddPNode::Address, in);
1547 register_new_node(addr_clone, ctrl);
1548 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1549 Node* klassptr_clone = klassptr->clone();
1550 klassptr_clone->set_req(2, addr_clone);
1551 register_new_node(klassptr_clone, ctrl);
1552 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1553 if (klassptr != n->in(1)) {
1554 Node* decode = n->in(1);
1555 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1556 Node* decode_clone = decode->clone();
1557 decode_clone->set_req(1, klassptr_clone);
1558 register_new_node(decode_clone, ctrl);
1559 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1560 klassptr_clone = decode_clone;
1561 }
1562 phi->set_req(i, klassptr_clone);
1563 }
1564 register_new_node(phi, region);
1565 Node* orig = n->in(1);
1566 _igvn.replace_input_of(n, 1, phi);
1567 split_if_with_blocks_post(n);
1568 if (n->outcnt() != 0) {
1569 _igvn.replace_input_of(n, 1, orig);
1570 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Graph);
1571 }
1572 return true;
1573 }
1574
1575 //------------------------------split_if_with_blocks_post----------------------
1576 // Do the real work in a non-recursive function. CFG hackery wants to be
1577 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1578 // info.
1579 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1580
1581 if (flat_array_element_type_check(n)) {
1582 return;
1583 }
1584
1585 // Cloning Cmp through Phi's involves the split-if transform.
1586 // FastLock is not used by an If
1587 if (n->is_Cmp() && !n->is_FastLock()) {
1588 Node *n_ctrl = get_ctrl(n);
1589 // Determine if the Node has inputs from some local Phi.
1590 // Returns the block to clone thru.
1591 Node *n_blk = has_local_phi_input(n);
1592 if (n_blk != n_ctrl) {
1593 return;
1594 }
1595
1596 if (!can_split_if(n_ctrl)) {
1597 return;
1598 }
1599
1600 if (n->outcnt() != 1) {
1601 return; // Multiple bool's from 1 compare?
1602 }
1603 Node *bol = n->unique_out();
1604 assert(bol->is_Bool(), "expect a bool here");
1605 if (bol->outcnt() != 1) {
1606 return;// Multiple branches from 1 compare?
1607 }
1608 Node *iff = bol->unique_out();
1609
1610 // Check some safety conditions
1611 if (iff->is_If()) { // Classic split-if?
1612 if (iff->outcnt() != iff->as_If()->required_outcnt()) {
1613 assert(false, "malformed IfNode with %d outputs", iff->outcnt());
1614 return;
1615 } else if (iff->in(0) != n_ctrl) {
1616 return; // Compare must be in same blk as if
1617 }
1618 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1619 // Can't split CMove with different control.
1620 if (get_ctrl(iff) != n_ctrl) {
1621 return;
1622 }
1623 if (get_ctrl(iff->in(2)) == n_ctrl ||
1624 get_ctrl(iff->in(3)) == n_ctrl) {
1625 return; // Inputs not yet split-up
1626 }
1627 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1628 return; // Loop-invar test gates loop-varying CMOVE
1629 }
1630 } else {
1631 return; // some other kind of node, such as an Allocate
1632 }
1633
1634 // When is split-if profitable? Every 'win' on means some control flow
1635 // goes dead, so it's almost always a win.
1636 int policy = 0;
1637 // Split compare 'n' through the merge point if it is profitable
1638 Node *phi = split_thru_phi( n, n_ctrl, policy);
1639 if (!phi) {
1640 return;
1641 }
1642
1643 // Now split the bool up thru the phi
1644 Node* bolphi = split_thru_phi(bol, n_ctrl, -1);
1645 guarantee(bolphi != nullptr, "null boolean phi node");
1646 assert(iff->in(1) == bolphi, "");
1647
1648 if (bolphi->Value(&_igvn)->singleton()) {
1649 return;
1650 }
1651
1652 // Conditional-move? Must split up now
1653 if (!iff->is_If()) {
1654 Node* cmovphi = split_thru_phi(iff, n_ctrl, -1);
1655 return;
1656 }
1657
1658 // Now split the IF
1659 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1660 #ifndef PRODUCT
1661 if (TraceLoopOpts || TraceSplitIf) {
1662 tty->print_cr("Split-If: %d %s", iff->_idx, iff->Name());
1663 }
1664 #endif
1665 do_split_if(iff);
1666 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1667 return;
1668 }
1669
1670 // Two identical ifs back to back can be merged
1671 if (try_merge_identical_ifs(n)) {
1672 return;
1673 }
1674
1675 // Check for an IF ready to split; one that has its
1676 // condition codes input coming from a Phi at the block start.
1677 int n_op = n->Opcode();
1678
1679 // Check for an IF being dominated by another IF same test
1680 if (n_op == Op_If ||
1681 n_op == Op_RangeCheck) {
1682 Node *bol = n->in(1);
1683 uint max = bol->outcnt();
1684 // Check for same test used more than once?
1685 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1686 // Search up IDOMs to see if this IF is dominated.
1687 Node* cmp = bol->in(1);
1688 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1689
1690 // Now search up IDOMs till cutoff, looking for a dominating test
1691 Node *prevdom = n;
1692 Node *dom = idom(prevdom);
1693 while (dom != cutoff) {
1694 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1695 safe_for_if_replacement(dom)) {
1696 // It's invalid to move control dependent data nodes in the inner
1697 // strip-mined loop, because:
1698 // 1) break validation of LoopNode::verify_strip_mined()
1699 // 2) move code with side-effect in strip-mined loop
1700 // Move to the exit of outer strip-mined loop in that case.
1701 Node* out_le = is_inner_of_stripmined_loop(dom);
1702 if (out_le != nullptr) {
1703 prevdom = out_le;
1704 }
1705 // Replace the dominated test with an obvious true or false.
1706 // Place it on the IGVN worklist for later cleanup.
1707 C->set_major_progress();
1708 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1709 // to prevent an array load from floating above its range check. There are three cases:
1710 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1711 // all its array accesses at that point.
1712 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1713 // accesses would start to float, since we don't pin at that point.
1714 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1715 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1716 prevdom->in(0)->Opcode() != Op_RangeCheck;
1717 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1718 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1719 return;
1720 }
1721 prevdom = dom;
1722 dom = idom(prevdom);
1723 }
1724 }
1725 }
1726
1727 try_sink_out_of_loop(n);
1728 if (C->failing()) {
1729 return;
1730 }
1731
1732 try_move_store_after_loop(n);
1733
1734 // Remove multiple allocations of the same inline type
1735 if (n->is_InlineType()) {
1736 n->as_InlineType()->remove_redundant_allocations(this);
1737 }
1738 }
1739
1740 // Transform:
1741 //
1742 // if (some_condition) {
1743 // // body 1
1744 // } else {
1745 // // body 2
1746 // }
1747 // if (some_condition) {
1748 // // body 3
1749 // } else {
1750 // // body 4
1751 // }
1752 //
1753 // into:
1754 //
1755 //
1756 // if (some_condition) {
1757 // // body 1
1758 // // body 3
1759 // } else {
1760 // // body 2
1761 // // body 4
1762 // }
1763 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1764 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1765 Node *n_ctrl = n->in(0);
1766 IfNode* dom_if = idom(n_ctrl)->as_If();
1767 if (n->in(1) != dom_if->in(1)) {
1768 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1769 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1770 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1771 _igvn.replace_input_of(n, 1, dom_if->in(1));
1772 }
1773 IfTrueNode* dom_proj_true = dom_if->true_proj();
1774 IfFalseNode* dom_proj_false = dom_if->false_proj();
1775
1776 // Now split the IF
1777 RegionNode* new_false_region;
1778 RegionNode* new_true_region;
1779 #ifndef PRODUCT
1780 if (TraceLoopOpts || TraceSplitIf) {
1781 tty->print_cr("Split-If Merging Identical Ifs: Dom-If: %d %s, If: %d %s", dom_if->_idx, dom_if->Name(), n->_idx, n->Name());
1782 }
1783 #endif
1784 do_split_if(n, &new_false_region, &new_true_region);
1785 assert(new_false_region->req() == new_true_region->req(), "");
1786 #ifdef ASSERT
1787 for (uint i = 1; i < new_false_region->req(); ++i) {
1788 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1789 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");
1790 }
1791 #endif
1792 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1793
1794 // We now have:
1795 // if (some_condition) {
1796 // // body 1
1797 // if (some_condition) {
1798 // body3: // new_true_region
1799 // // body3
1800 // } else {
1801 // goto body4;
1802 // }
1803 // } else {
1804 // // body 2
1805 // if (some_condition) {
1806 // goto body3;
1807 // } else {
1808 // body4: // new_false_region
1809 // // body4;
1810 // }
1811 // }
1812 //
1813
1814 // clone pinned nodes thru the resulting regions
1815 push_pinned_nodes_thru_region(dom_if, new_true_region);
1816 push_pinned_nodes_thru_region(dom_if, new_false_region);
1817
1818 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1819 // 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
1820 // unrelated control dependency.
1821 for (uint i = 1; i < new_false_region->req(); i++) {
1822 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1823 dominated_by(dom_proj_true, new_false_region->in(i)->in(0)->as_If());
1824 } else {
1825 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1826 dominated_by(dom_proj_false, new_false_region->in(i)->in(0)->as_If());
1827 }
1828 }
1829 return true;
1830 }
1831 return false;
1832 }
1833
1834 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1835 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1836 Node* u = region->out(i);
1837 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test()) {
1838 continue;
1839 }
1840 assert(u->in(0) == region, "not a control dependent node?");
1841 uint j = 1;
1842 for (; j < u->req(); ++j) {
1843 Node* in = u->in(j);
1844 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1845 break;
1846 }
1847 }
1848 if (j == u->req()) {
1849 Node *phi = PhiNode::make_blank(region, u);
1850 for (uint k = 1; k < region->req(); ++k) {
1851 Node* clone = u->clone();
1852 clone->set_req(0, region->in(k));
1853 register_new_node(clone, region->in(k));
1854 phi->init_req(k, clone);
1855 }
1856 register_new_node(phi, region);
1857 _igvn.replace_node(u, phi);
1858 --i;
1859 }
1860 }
1861 }
1862
1863 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1864 if (!dom->is_CountedLoopEnd()) {
1865 return true;
1866 }
1867 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1868 CountedLoopNode* cl = le->loopnode();
1869 if (cl == nullptr) {
1870 return true;
1871 }
1872 if (!cl->is_main_loop()) {
1873 return true;
1874 }
1875 if (cl->is_canonical_loop_entry() == nullptr) {
1876 return true;
1877 }
1878 // Further unrolling is possible so loop exit condition might change
1879 return false;
1880 }
1881
1882 // See if a shared loop-varying computation has no loop-varying uses.
1883 // Happens if something is only used for JVM state in uncommon trap exits,
1884 // like various versions of induction variable+offset. Clone the
1885 // computation per usage to allow it to sink out of the loop.
1886 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1887 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1888 n->in(1)->bottom_type()->isa_rawptr() &&
1889 !n->bottom_type()->isa_rawptr();
1890
1891 if (has_ctrl(n) &&
1892 !n->is_Phi() &&
1893 !n->is_Bool() &&
1894 !n->is_Proj() &&
1895 !n->is_MergeMem() &&
1896 !n->is_CMove() &&
1897 !n->is_OpaqueConstantBool() &&
1898 !n->is_OpaqueInitializedAssertionPredicate() &&
1899 !n->is_OpaqueTemplateAssertionPredicate() &&
1900 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1901 n->Opcode() != Op_CreateEx &&
1902 (KillPathsReachableByDeadTypeNode || !n->is_Type())
1903 ) {
1904 Node *n_ctrl = get_ctrl(n);
1905 IdealLoopTree *n_loop = get_loop(n_ctrl);
1906
1907 if (n->in(0) != nullptr) {
1908 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1909 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1910 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1911 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1912 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1913 Node* maybe_pinned_n = n;
1914 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1915 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) {
1916 if (n->depends_only_on_test()) {
1917 // If this node depends_only_on_test, it will be rewired to a control input that is not
1918 // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1919 // rewired to a dominating test equivalent to the new control.
1920 Node* pinned_clone = n->pin_node_under_control();
1921 if (pinned_clone != nullptr) {
1922 register_new_node(pinned_clone, n_ctrl);
1923 maybe_pinned_n = pinned_clone;
1924 _igvn.replace_node(n, pinned_clone);
1925 }
1926 }
1927 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1928 }
1929 }
1930 }
1931 if (n_loop != _ltree_root && n->outcnt() > 1) {
1932 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1933 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1934 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1935 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1936 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1937 if (n->is_Store() || n->is_LoadStore()) {
1938 assert(false, "no node with a side effect");
1939 C->record_failure("no node with a side effect");
1940 return;
1941 }
1942 Node* outer_loop_clone = nullptr;
1943 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1944 Node* u = n->last_out(j); // Clone private computation per use
1945 _igvn.rehash_node_delayed(u);
1946 Node* x = nullptr;
1947 if (n->in(0) != nullptr && n->depends_only_on_test()) {
1948 // If this node depends_only_on_test, it will be rewired to a control input that is not
1949 // the correct test. As a result, it must be pinned otherwise it can be incorrectly
1950 // rewired to a dominating test equivalent to the new control.
1951 x = n->pin_node_under_control();
1952 }
1953 if (x == nullptr) {
1954 x = n->clone();
1955 }
1956 Node* x_ctrl = nullptr;
1957 if (u->is_Phi()) {
1958 // Replace all uses of normal nodes. Replace Phi uses
1959 // individually, so the separate Nodes can sink down
1960 // different paths.
1961 uint k = 1;
1962 while (u->in(k) != n) k++;
1963 u->set_req(k, x);
1964 // x goes next to Phi input path
1965 x_ctrl = u->in(0)->in(k);
1966 // Find control for 'x' next to use but not inside inner loops.
1967 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1968 --j;
1969 } else { // Normal use
1970 if (has_ctrl(u)) {
1971 x_ctrl = get_ctrl(u);
1972 } else {
1973 x_ctrl = u->in(0);
1974 }
1975 // Find control for 'x' next to use but not inside inner loops.
1976 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1977 // Replace all uses
1978 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1979 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1980 // anymore now that we're going to pin n as well
1981 _igvn.replace_node(u, x);
1982 --j;
1983 } else {
1984 int nb = u->replace_edge(n, x, &_igvn);
1985 j -= nb;
1986 }
1987 }
1988
1989 if (n->is_Load()) {
1990 // For loads, add a control edge to a CFG node outside of the loop
1991 // to force them to not combine and return back inside the loop
1992 // during GVN optimization (4641526).
1993 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1994
1995 IdealLoopTree* x_loop = get_loop(x_ctrl);
1996 Node* x_head = x_loop->_head;
1997 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1998 // Do not add duplicate LoadNodes to the outer strip mined loop
1999 if (outer_loop_clone != nullptr) {
2000 _igvn.replace_node(x, outer_loop_clone);
2001 continue;
2002 }
2003 outer_loop_clone = x;
2004 }
2005 x->set_req(0, x_ctrl);
2006 } else if (n->in(0) != nullptr){
2007 x->set_req(0, x_ctrl);
2008 }
2009 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
2010 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
2011 register_new_node(x, x_ctrl);
2012
2013 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
2014 // All AddP nodes must keep the same base after sinking so:
2015 // 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,
2016 // their bases remain the same.
2017 // (see 2- below)
2018 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
2019 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
2020 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
2021 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
2022 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
2023 assert(!x->is_Load(), "load should be pinned");
2024 // Use a cast node to pin clone out of loop
2025 Node* cast = nullptr;
2026 for (uint k = 0; k < x->req(); k++) {
2027 Node* in = x->in(k);
2028 if (in != nullptr && ctrl_is_member(n_loop, in)) {
2029 const Type* in_t = _igvn.type(in);
2030 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
2031 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr);
2032 }
2033 if (cast != nullptr) {
2034 Node* prev = _igvn.hash_find_insert(cast);
2035 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
2036 cast->destruct(&_igvn);
2037 cast = prev;
2038 } else {
2039 register_new_node(cast, x_ctrl);
2040 }
2041 x->replace_edge(in, cast);
2042 // Chain of AddP nodes:
2043 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
2044 if (x->is_AddP() && k == AddPNode::Base) {
2045 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
2046 }
2047 break;
2048 }
2049 }
2050 assert(cast != nullptr, "must have added a cast to pin the node");
2051 }
2052 }
2053 _igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph);
2054 }
2055 _dom_lca_tags_round = 0;
2056 }
2057 }
2058 }
2059
2060 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
2061 ResourceMark rm;
2062 Node_List wq;
2063 wq.push(x);
2064 while (wq.size() != 0) {
2065 Node* n = wq.pop();
2066 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2067 Node* u = n->fast_out(i);
2068 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
2069 _igvn.replace_input_of(u, AddPNode::Base, new_base);
2070 wq.push(u);
2071 }
2072 }
2073 }
2074 }
2075
2076 // Compute the early control of a node by following its inputs until we reach
2077 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
2078 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
2079 Node* early_ctrl = nullptr;
2080 ResourceMark rm;
2081 Unique_Node_List wq;
2082 wq.push(n);
2083 for (uint i = 0; i < wq.size(); i++) {
2084 Node* m = wq.at(i);
2085 Node* c = nullptr;
2086 if (m->is_CFG()) {
2087 c = m;
2088 } else if (m->pinned()) {
2089 c = m->in(0);
2090 } else {
2091 for (uint j = 0; j < m->req(); j++) {
2092 Node* in = m->in(j);
2093 if (in != nullptr) {
2094 wq.push(in);
2095 }
2096 }
2097 }
2098 if (c != nullptr) {
2099 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
2100 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
2101 early_ctrl = c;
2102 }
2103 }
2104 }
2105 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
2106 return early_ctrl;
2107 }
2108
2109 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
2110 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2111 Node* u = n->fast_out(i);
2112 if (u->is_Opaque1()) {
2113 return false; // Found loop limit, bugfix for 4677003
2114 }
2115 if (u->is_Phi()) {
2116 for (uint j = 1; j < u->req(); ++j) {
2117 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
2118 return false;
2119 }
2120 }
2121 } else {
2122 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
2123 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
2124 return false;
2125 }
2126 }
2127 }
2128 return true;
2129 }
2130
2131 // 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
2132 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
2133 // test of the pre loop above the point in the graph where it's pinned. This results in a broken graph. One way to avoid
2134 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code
2135 // is generated for the main loop.
2136 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) {
2137 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
2138 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
2139 if (is_dominator(pre_loop->loopexit(), ctrl)) {
2140 return true;
2141 }
2142 }
2143 return false;
2144 }
2145
2146 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
2147 if (n->is_Load()) {
2148 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure each call to
2149 // get_late_ctrl_with_anti_dep() uses its own tag
2150 _dom_lca_tags_round++;
2151 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
2152
2153 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
2154 }
2155 IdealLoopTree *u_loop = get_loop(ctrl);
2156 if (u_loop == n_loop) {
2157 return false; // Found loop-varying use
2158 }
2159 if (n_loop->is_member(u_loop)) {
2160 return false; // Found use in inner loop
2161 }
2162 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) {
2163 return false;
2164 }
2165 return true;
2166 }
2167
2168 //------------------------------split_if_with_blocks---------------------------
2169 // Check for aggressive application of 'split-if' optimization,
2170 // using basic block level info.
2171 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
2172 Node* root = C->root();
2173 visited.set(root->_idx); // first, mark root as visited
2174 // Do pre-visit work for root
2175 Node* n = split_if_with_blocks_pre(root);
2176 uint cnt = n->outcnt();
2177 uint i = 0;
2178
2179 while (true) {
2180 // Visit all children
2181 if (i < cnt) {
2182 Node* use = n->raw_out(i);
2183 ++i;
2184 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
2185 // Now do pre-visit work for this use
2186 use = split_if_with_blocks_pre(use);
2187 nstack.push(n, i); // Save parent and next use's index.
2188 n = use; // Process all children of current use.
2189 cnt = use->outcnt();
2190 i = 0;
2191 }
2192 }
2193 else {
2194 // All of n's children have been processed, complete post-processing.
2195 if (cnt != 0 && !n->is_Con()) {
2196 assert(has_node(n), "no dead nodes");
2197 split_if_with_blocks_post(n);
2198 if (C->failing()) {
2199 return;
2200 }
2201 }
2202 if (must_throttle_split_if()) {
2203 nstack.clear();
2204 }
2205 if (nstack.is_empty()) {
2206 // Finished all nodes on stack.
2207 break;
2208 }
2209 // Get saved parent node and next use's index. Visit the rest of uses.
2210 n = nstack.node();
2211 cnt = n->outcnt();
2212 i = nstack.index();
2213 nstack.pop();
2214 }
2215 }
2216 }
2217
2218
2219 //=============================================================================
2220 //
2221 // C L O N E A L O O P B O D Y
2222 //
2223
2224 //------------------------------clone_iff--------------------------------------
2225 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2226 // "Nearly" because all Nodes have been cloned from the original in the loop,
2227 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2228 // through the Phi recursively, and return a Bool.
2229 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2230
2231 // Convert this Phi into a Phi merging Bools
2232 uint i;
2233 for (i = 1; i < phi->req(); i++) {
2234 Node* b = phi->in(i);
2235 if (b->is_Phi()) {
2236 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2237 } else {
2238 assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(),
2239 "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node");
2240 }
2241 }
2242 Node* n = phi->in(1);
2243 Node* sample_opaque = nullptr;
2244 Node *sample_bool = nullptr;
2245 if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) {
2246 sample_opaque = n;
2247 sample_bool = n->in(1);
2248 assert(sample_bool->is_Bool(), "wrong type");
2249 } else {
2250 sample_bool = n;
2251 }
2252 Node* sample_cmp = sample_bool->in(1);
2253 const Type* t = Type::TOP;
2254 const TypePtr* at = nullptr;
2255 if (sample_cmp->is_FlatArrayCheck()) {
2256 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2257 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2258 t = Type::MEMORY;
2259 at = TypeRawPtr::BOTTOM;
2260 }
2261
2262 // Make Phis to merge the Cmp's inputs.
2263 PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2264 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2265 for (i = 1; i < phi->req(); i++) {
2266 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2267 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2268 phi1->set_req(i, n1);
2269 phi2->set_req(i, n2);
2270 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2271 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2272 }
2273 // See if these Phis have been made before.
2274 // Register with optimizer
2275 Node *hit1 = _igvn.hash_find_insert(phi1);
2276 if (hit1) { // Hit, toss just made Phi
2277 _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2278 assert(hit1->is_Phi(), "" );
2279 phi1 = (PhiNode*)hit1; // Use existing phi
2280 } else { // Miss
2281 _igvn.register_new_node_with_optimizer(phi1);
2282 }
2283 Node *hit2 = _igvn.hash_find_insert(phi2);
2284 if (hit2) { // Hit, toss just made Phi
2285 _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2286 assert(hit2->is_Phi(), "" );
2287 phi2 = (PhiNode*)hit2; // Use existing phi
2288 } else { // Miss
2289 _igvn.register_new_node_with_optimizer(phi2);
2290 }
2291 // Register Phis with loop/block info
2292 set_ctrl(phi1, phi->in(0));
2293 set_ctrl(phi2, phi->in(0));
2294 // Make a new Cmp
2295 Node *cmp = sample_cmp->clone();
2296 cmp->set_req(1, phi1);
2297 cmp->set_req(2, phi2);
2298 _igvn.register_new_node_with_optimizer(cmp);
2299 set_ctrl(cmp, phi->in(0));
2300
2301 // Make a new Bool
2302 Node *b = sample_bool->clone();
2303 b->set_req(1,cmp);
2304 _igvn.register_new_node_with_optimizer(b);
2305 set_ctrl(b, phi->in(0));
2306
2307 if (sample_opaque != nullptr) {
2308 Node* opaque = sample_opaque->clone();
2309 opaque->set_req(1, b);
2310 _igvn.register_new_node_with_optimizer(opaque);
2311 set_ctrl(opaque, phi->in(0));
2312 return opaque;
2313 }
2314
2315 assert(b->is_Bool(), "");
2316 return b;
2317 }
2318
2319 //------------------------------clone_bool-------------------------------------
2320 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2321 // "Nearly" because all Nodes have been cloned from the original in the loop,
2322 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2323 // through the Phi recursively, and return a Bool.
2324 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2325 uint i;
2326 // Convert this Phi into a Phi merging Bools
2327 for( i = 1; i < phi->req(); i++ ) {
2328 Node *b = phi->in(i);
2329 if( b->is_Phi() ) {
2330 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2331 } else {
2332 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2333 }
2334 }
2335
2336 Node *sample_cmp = phi->in(1);
2337
2338 // Make Phis to merge the Cmp's inputs.
2339 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2340 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2341 for( uint j = 1; j < phi->req(); j++ ) {
2342 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2343 Node *n1, *n2;
2344 if( cmp_top->is_Cmp() ) {
2345 n1 = cmp_top->in(1);
2346 n2 = cmp_top->in(2);
2347 } else {
2348 n1 = n2 = cmp_top;
2349 }
2350 phi1->set_req( j, n1 );
2351 phi2->set_req( j, n2 );
2352 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2353 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2354 }
2355
2356 // See if these Phis have been made before.
2357 // Register with optimizer
2358 Node *hit1 = _igvn.hash_find_insert(phi1);
2359 if( hit1 ) { // Hit, toss just made Phi
2360 _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2361 assert( hit1->is_Phi(), "" );
2362 phi1 = (PhiNode*)hit1; // Use existing phi
2363 } else { // Miss
2364 _igvn.register_new_node_with_optimizer(phi1);
2365 }
2366 Node *hit2 = _igvn.hash_find_insert(phi2);
2367 if( hit2 ) { // Hit, toss just made Phi
2368 _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi
2369 assert( hit2->is_Phi(), "" );
2370 phi2 = (PhiNode*)hit2; // Use existing phi
2371 } else { // Miss
2372 _igvn.register_new_node_with_optimizer(phi2);
2373 }
2374 // Register Phis with loop/block info
2375 set_ctrl(phi1, phi->in(0));
2376 set_ctrl(phi2, phi->in(0));
2377 // Make a new Cmp
2378 Node *cmp = sample_cmp->clone();
2379 cmp->set_req( 1, phi1 );
2380 cmp->set_req( 2, phi2 );
2381 _igvn.register_new_node_with_optimizer(cmp);
2382 set_ctrl(cmp, phi->in(0));
2383
2384 assert( cmp->is_Cmp(), "" );
2385 return (CmpNode*)cmp;
2386 }
2387
2388 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2389 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2390 Node_List*& split_if_set, Node_List*& split_bool_set,
2391 Node_List*& split_cex_set, Node_List& worklist,
2392 uint new_counter, CloneLoopMode mode) {
2393 Node* nnn = old_new[old->_idx];
2394 // Copy uses to a worklist, so I can munge the def-use info
2395 // with impunity.
2396 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2397 worklist.push(old->fast_out(j));
2398
2399 while( worklist.size() ) {
2400 Node *use = worklist.pop();
2401 if (!has_node(use)) continue; // Ignore dead nodes
2402 if (use->in(0) == C->top()) continue;
2403 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2404 // Check for data-use outside of loop - at least one of OLD or USE
2405 // must not be a CFG node.
2406 #ifdef ASSERT
2407 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2408 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2409 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2410 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2411 }
2412 #endif
2413 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2414
2415 // If the Data use is an IF, that means we have an IF outside the
2416 // loop that is switching on a condition that is set inside the
2417 // loop. Happens if people set a loop-exit flag; then test the flag
2418 // in the loop to break the loop, then test is again outside the
2419 // loop to determine which way the loop exited.
2420 //
2421 // For several uses we need to make sure that there is no phi between,
2422 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2423 // to avoid such a phi in between.
2424 // For example, it is unexpected that there is a Phi between an
2425 // AllocateArray node and its ValidLengthTest input that could cause
2426 // split if to break.
2427 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2428 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2429 if (use->is_If() || use->is_CMove() || use->is_OpaqueConstantBool() || use->is_OpaqueInitializedAssertionPredicate() ||
2430 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2431 // Since this code is highly unlikely, we lazily build the worklist
2432 // of such Nodes to go split.
2433 if (!split_if_set) {
2434 split_if_set = new Node_List();
2435 }
2436 split_if_set->push(use);
2437 }
2438 if (use->is_Bool()) {
2439 if (!split_bool_set) {
2440 split_bool_set = new Node_List();
2441 }
2442 split_bool_set->push(use);
2443 }
2444 if (use->Opcode() == Op_CreateEx) {
2445 if (!split_cex_set) {
2446 split_cex_set = new Node_List();
2447 }
2448 split_cex_set->push(use);
2449 }
2450
2451
2452 // Get "block" use is in
2453 uint idx = 0;
2454 while( use->in(idx) != old ) idx++;
2455 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2456 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2457 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2458 ? prev->in(2)
2459 : idom(prev);
2460 if( use->is_Phi() ) // Phi use is in prior block
2461 cfg = prev->in(idx); // NOT in block of Phi itself
2462 if (cfg->is_top()) { // Use is dead?
2463 _igvn.replace_input_of(use, idx, C->top());
2464 continue;
2465 }
2466
2467 // If use is referenced through control edge... (idx == 0)
2468 if (mode == IgnoreStripMined && idx == 0) {
2469 LoopNode *head = loop->_head->as_Loop();
2470 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2471 // That node is outside the inner loop, leave it outside the
2472 // outer loop as well to not confuse verification code.
2473 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2474 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2475 continue;
2476 }
2477 }
2478
2479 while(!outer_loop->is_member(get_loop(cfg))) {
2480 prev = cfg;
2481 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2482 }
2483 // If the use occurs after merging several exits from the loop, then
2484 // old value must have dominated all those exits. Since the same old
2485 // value was used on all those exits we did not need a Phi at this
2486 // merge point. NOW we do need a Phi here. Each loop exit value
2487 // is now merged with the peeled body exit; each exit gets its own
2488 // private Phi and those Phis need to be merged here.
2489 Node *phi;
2490 if( prev->is_Region() ) {
2491 if( idx == 0 ) { // Updating control edge?
2492 phi = prev; // Just use existing control
2493 } else { // Else need a new Phi
2494 phi = PhiNode::make( prev, old );
2495 // Now recursively fix up the new uses of old!
2496 for( uint i = 1; i < prev->req(); i++ ) {
2497 worklist.push(phi); // Onto worklist once for each 'old' input
2498 }
2499 }
2500 } else {
2501 // Get new RegionNode merging old and new loop exits
2502 prev = old_new[prev->_idx];
2503 assert( prev, "just made this in step 7" );
2504 if( idx == 0) { // Updating control edge?
2505 phi = prev; // Just use existing control
2506 } else { // Else need a new Phi
2507 // Make a new Phi merging data values properly
2508 phi = PhiNode::make( prev, old );
2509 phi->set_req( 1, nnn );
2510 }
2511 }
2512 // If inserting a new Phi, check for prior hits
2513 if( idx != 0 ) {
2514 Node *hit = _igvn.hash_find_insert(phi);
2515 if( hit == nullptr ) {
2516 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2517 } else { // or
2518 // Remove the new phi from the graph and use the hit
2519 _igvn.remove_dead_node(phi, phi == prev ? PhaseIterGVN::NodeOrigin::Graph : PhaseIterGVN::NodeOrigin::Speculative);
2520 phi = hit; // Use existing phi
2521 }
2522 set_ctrl(phi, prev);
2523 }
2524 // Make 'use' use the Phi instead of the old loop body exit value
2525 assert(use->in(idx) == old, "old is still input of use");
2526 // We notify all uses of old, including use, and the indirect uses,
2527 // that may now be optimized because we have replaced old with phi.
2528 _igvn.add_users_to_worklist(old);
2529 if (idx == 0 && use->depends_only_on_test()) {
2530 // If this node depends_only_on_test, it will be rewired to a control input that is not the
2531 // correct test. As a result, it must be pinned otherwise it can be incorrectly rewired to
2532 // a dominating test equivalent to the new control.
2533 Node* pinned_clone = use->pin_node_under_control();
2534 if (pinned_clone != nullptr) {
2535 pinned_clone->set_req(0, phi);
2536 register_new_node_with_ctrl_of(pinned_clone, use);
2537 _igvn.replace_node(use, pinned_clone);
2538 continue;
2539 }
2540 }
2541 _igvn.replace_input_of(use, idx, phi);
2542 if( use->_idx >= new_counter ) { // If updating new phis
2543 // Not needed for correctness, but prevents a weak assert
2544 // in AddPNode from tripping (when we end up with different
2545 // base & derived Phis that will become the same after
2546 // IGVN does CSE).
2547 Node *hit = _igvn.hash_find_insert(use);
2548 if( hit ) // Go ahead and re-hash for hits.
2549 _igvn.replace_node( use, hit );
2550 }
2551 }
2552 }
2553 }
2554
2555 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2556 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2557 bool check_old_new) {
2558 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2559 Node* u = n->fast_out(j);
2560 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2561 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2562 assert(!phase->ctrl_is_member(loop, u) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2563 if (!phase->ctrl_is_member(loop, u)) {
2564 if (phase->ctrl_is_member(outer_loop, u)) {
2565 wq.push(u);
2566 } else {
2567 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2568 // the outer loop too
2569 Node* u_c = u->in(0);
2570 if (u_c != nullptr) {
2571 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2572 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2573 wq.push(u);
2574 }
2575 }
2576 }
2577 }
2578 }
2579 }
2580 }
2581
2582 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2583 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2584 Node_List& extra_data_nodes) {
2585 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2586 CountedLoopNode* cl = head->as_CountedLoop();
2587 Node* l = cl->outer_loop();
2588 Node* tail = cl->outer_loop_tail();
2589 IfNode* le = cl->outer_loop_end();
2590 Node* sfpt = cl->outer_safepoint();
2591 CountedLoopEndNode* cle = cl->loopexit();
2592 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2593 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2594 IfFalseNode* cle_out = cle->false_proj();
2595
2596 Node* new_sfpt = nullptr;
2597 Node* new_cle_out = cle_out->clone();
2598 old_new.map(cle_out->_idx, new_cle_out);
2599 if (mode == CloneIncludesStripMined) {
2600 // clone outer loop body
2601 Node* new_l = l->clone();
2602 Node* new_tail = tail->clone();
2603 IfNode* new_le = le->clone()->as_If();
2604 new_sfpt = sfpt->clone();
2605
2606 set_loop(new_l, outer_loop->_parent);
2607 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2608 set_loop(new_cle_out, outer_loop->_parent);
2609 set_idom(new_cle_out, new_cle, dd);
2610 set_loop(new_sfpt, outer_loop->_parent);
2611 set_idom(new_sfpt, new_cle_out, dd);
2612 set_loop(new_le, outer_loop->_parent);
2613 set_idom(new_le, new_sfpt, dd);
2614 set_loop(new_tail, outer_loop->_parent);
2615 set_idom(new_tail, new_le, dd);
2616 set_idom(new_cl, new_l, dd);
2617
2618 old_new.map(l->_idx, new_l);
2619 old_new.map(tail->_idx, new_tail);
2620 old_new.map(le->_idx, new_le);
2621 old_new.map(sfpt->_idx, new_sfpt);
2622
2623 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2624 new_l->set_req(0, new_l);
2625 new_tail->set_req(0, new_le);
2626 new_le->set_req(0, new_sfpt);
2627 new_sfpt->set_req(0, new_cle_out);
2628 new_cle_out->set_req(0, new_cle);
2629 new_cl->set_req(LoopNode::EntryControl, new_l);
2630
2631 _igvn.register_new_node_with_optimizer(new_l);
2632 _igvn.register_new_node_with_optimizer(new_tail);
2633 _igvn.register_new_node_with_optimizer(new_le);
2634 } else {
2635 Node *newhead = old_new[loop->_head->_idx];
2636 newhead->as_Loop()->clear_strip_mined();
2637 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2638 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2639 }
2640 // Look at data node that were assigned a control in the outer
2641 // loop: they are kept in the outer loop by the safepoint so start
2642 // from the safepoint node's inputs.
2643 IdealLoopTree* outer_loop = get_loop(l);
2644 Node_Stack stack(2);
2645 stack.push(sfpt, 1);
2646 uint new_counter = C->unique();
2647 while (stack.size() > 0) {
2648 Node* n = stack.node();
2649 uint i = stack.index();
2650 while (i < n->req() &&
2651 (n->in(i) == nullptr ||
2652 !has_ctrl(n->in(i)) ||
2653 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2654 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2655 i++;
2656 }
2657 if (i < n->req()) {
2658 stack.set_index(i+1);
2659 stack.push(n->in(i), 0);
2660 } else {
2661 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2662 Node* m = n == sfpt ? new_sfpt : n->clone();
2663 if (m != nullptr) {
2664 for (uint i = 0; i < n->req(); i++) {
2665 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2666 m->set_req(i, old_new[m->in(i)->_idx]);
2667 }
2668 }
2669 } else {
2670 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2671 }
2672 if (n != sfpt) {
2673 extra_data_nodes.push(n);
2674 _igvn.register_new_node_with_optimizer(m);
2675 assert(get_ctrl(n) == cle_out, "what other control?");
2676 set_ctrl(m, new_cle_out);
2677 old_new.map(n->_idx, m);
2678 }
2679 stack.pop();
2680 }
2681 }
2682 if (mode == CloneIncludesStripMined) {
2683 _igvn.register_new_node_with_optimizer(new_sfpt);
2684 _igvn.register_new_node_with_optimizer(new_cle_out);
2685 }
2686 // Some other transformation may have pessimistically assigned some
2687 // data nodes to the outer loop. Set their control so they are out
2688 // of the outer loop.
2689 ResourceMark rm;
2690 Unique_Node_List wq;
2691 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2692 Node* old = extra_data_nodes.at(i);
2693 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2694 }
2695
2696 for (uint i = 0; i < loop->_body.size(); i++) {
2697 Node* old = loop->_body.at(i);
2698 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2699 }
2700
2701 Node* inner_out = sfpt->in(0);
2702 if (inner_out->outcnt() > 1) {
2703 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2704 }
2705
2706 Node* new_ctrl = cl->outer_loop_exit();
2707 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2708 for (uint i = 0; i < wq.size(); i++) {
2709 Node* n = wq.at(i);
2710 set_ctrl(n, new_ctrl);
2711 if (n->in(0) != nullptr) {
2712 _igvn.replace_input_of(n, 0, new_ctrl);
2713 }
2714 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2715 }
2716 } else {
2717 Node *newhead = old_new[loop->_head->_idx];
2718 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2719 }
2720 }
2721
2722 //------------------------------clone_loop-------------------------------------
2723 //
2724 // C L O N E A L O O P B O D Y
2725 //
2726 // This is the basic building block of the loop optimizations. It clones an
2727 // entire loop body. It makes an old_new loop body mapping; with this mapping
2728 // you can find the new-loop equivalent to an old-loop node. All new-loop
2729 // nodes are exactly equal to their old-loop counterparts, all edges are the
2730 // same. All exits from the old-loop now have a RegionNode that merges the
2731 // equivalent new-loop path. This is true even for the normal "loop-exit"
2732 // condition. All uses of loop-invariant old-loop values now come from (one
2733 // or more) Phis that merge their new-loop equivalents.
2734 //
2735 // This operation leaves the graph in an illegal state: there are two valid
2736 // control edges coming from the loop pre-header to both loop bodies. I'll
2737 // definitely have to hack the graph after running this transform.
2738 //
2739 // From this building block I will further edit edges to perform loop peeling
2740 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2741 //
2742 // Parameter side_by_size_idom:
2743 // When side_by_size_idom is null, the dominator tree is constructed for
2744 // the clone loop to dominate the original. Used in construction of
2745 // pre-main-post loop sequence.
2746 // When nonnull, the clone and original are side-by-side, both are
2747 // dominated by the side_by_side_idom node. Used in construction of
2748 // unswitched loops.
2749 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2750 CloneLoopMode mode, Node* side_by_side_idom) {
2751
2752 LoopNode* head = loop->_head->as_Loop();
2753 head->verify_strip_mined(1);
2754
2755 if (C->do_vector_loop() && PrintOpto) {
2756 const char* mname = C->method()->name()->as_quoted_ascii();
2757 if (mname != nullptr) {
2758 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2759 }
2760 }
2761
2762 CloneMap& cm = C->clone_map();
2763 if (C->do_vector_loop()) {
2764 cm.set_clone_idx(cm.max_gen()+1);
2765 #ifndef PRODUCT
2766 if (PrintOpto) {
2767 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2768 loop->dump_head();
2769 }
2770 #endif
2771 }
2772
2773 // Step 1: Clone the loop body. Make the old->new mapping.
2774 clone_loop_body(loop->_body, old_new, &cm);
2775
2776 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2777
2778 // Step 2: Fix the edges in the new body. If the old input is outside the
2779 // loop use it. If the old input is INside the loop, use the corresponding
2780 // new node instead.
2781 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2782
2783 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2784 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2785
2786 // Step 3: Now fix control uses. Loop varying control uses have already
2787 // been fixed up (as part of all input edges in Step 2). Loop invariant
2788 // control uses must be either an IfFalse or an IfTrue. Make a merge
2789 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2790 // refer to this.
2791 Node_List worklist;
2792 uint new_counter = C->unique();
2793 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2794
2795 // Step 4: If loop-invariant use is not control, it must be dominated by a
2796 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2797 // there if needed. Make a Phi there merging old and new used values.
2798 Node_List *split_if_set = nullptr;
2799 Node_List *split_bool_set = nullptr;
2800 Node_List *split_cex_set = nullptr;
2801 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2802
2803 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2804 Node* old = extra_data_nodes.at(i);
2805 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2806 split_bool_set, split_cex_set, worklist, new_counter,
2807 mode);
2808 }
2809
2810 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2811 // the loop uses a condition set in the loop. The original IF probably
2812 // takes control from one or more OLD Regions (which in turn get from NEW
2813 // Regions). In any case, there will be a set of Phis for each merge point
2814 // from the IF up to where the original BOOL def exists the loop.
2815 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2816
2817 }
2818
2819 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2820 if (split_if_set) {
2821 while (split_if_set->size()) {
2822 Node *iff = split_if_set->pop();
2823 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2824 if (iff->in(input)->is_Phi()) {
2825 Node *b = clone_iff(iff->in(input)->as_Phi());
2826 _igvn.replace_input_of(iff, input, b);
2827 }
2828 }
2829 }
2830 if (split_bool_set) {
2831 while (split_bool_set->size()) {
2832 Node *b = split_bool_set->pop();
2833 Node *phi = b->in(1);
2834 assert(phi->is_Phi(), "");
2835 CmpNode *cmp = clone_bool((PhiNode*) phi);
2836 _igvn.replace_input_of(b, 1, cmp);
2837 }
2838 }
2839 if (split_cex_set) {
2840 while (split_cex_set->size()) {
2841 Node *b = split_cex_set->pop();
2842 assert(b->in(0)->is_Region(), "");
2843 assert(b->in(1)->is_Phi(), "");
2844 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2845 split_up(b, b->in(0), nullptr);
2846 }
2847 }
2848 }
2849
2850 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2851 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2852 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2853 for(uint i = 0; i < body.size(); i++ ) {
2854 Node* old = body.at(i);
2855 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2856 split_bool_set, split_cex_set, worklist, new_counter,
2857 mode);
2858 }
2859 }
2860
2861 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2862 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2863 LoopNode* head = loop->_head->as_Loop();
2864 for(uint i = 0; i < body.size(); i++ ) {
2865 Node* old = body.at(i);
2866 if( !old->is_CFG() ) continue;
2867
2868 // Copy uses to a worklist, so I can munge the def-use info
2869 // with impunity.
2870 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2871 worklist.push(old->fast_out(j));
2872 }
2873
2874 while (worklist.size()) { // Visit all uses
2875 Node *use = worklist.pop();
2876 if (!has_node(use)) continue; // Ignore dead nodes
2877 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2878 if (!loop->is_member(use_loop) && use->is_CFG()) {
2879 // Both OLD and USE are CFG nodes here.
2880 assert(use->is_Proj(), "" );
2881 Node* nnn = old_new[old->_idx];
2882
2883 Node* newuse = nullptr;
2884 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2885 CountedLoopNode* cl = head->as_CountedLoop();
2886 CountedLoopEndNode* cle = cl->loopexit();
2887 // is use the projection that exits the loop from the CountedLoopEndNode?
2888 if (use->in(0) == cle) {
2889 IfFalseNode* cle_out = use->as_IfFalse();
2890 IfNode* le = cl->outer_loop_end();
2891 use = le->false_proj();
2892 use_loop = get_loop(use);
2893 if (mode == CloneIncludesStripMined) {
2894 nnn = old_new[le->_idx];
2895 } else {
2896 newuse = old_new[cle_out->_idx];
2897 }
2898 }
2899 }
2900 if (newuse == nullptr) {
2901 newuse = use->clone();
2902 }
2903
2904 // Clone the loop exit control projection
2905 if (C->do_vector_loop() && cm != nullptr) {
2906 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2907 }
2908 newuse->set_req(0,nnn);
2909 _igvn.register_new_node_with_optimizer(newuse);
2910 set_loop(newuse, use_loop);
2911 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2912
2913 // We need a Region to merge the exit from the peeled body and the
2914 // exit from the old loop body.
2915 RegionNode *r = new RegionNode(3);
2916 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2917 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2918
2919 // The original user of 'use' uses 'r' instead.
2920 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2921 Node* useuse = use->last_out(l);
2922 _igvn.rehash_node_delayed(useuse);
2923 uint uses_found = 0;
2924 if (useuse->in(0) == use) {
2925 useuse->set_req(0, r);
2926 uses_found++;
2927 if (useuse->is_CFG()) {
2928 // This is not a dom_depth > dd_r because when new
2929 // control flow is constructed by a loop opt, a node and
2930 // its dominator can end up at the same dom_depth
2931 assert(dom_depth(useuse) >= dd_r, "");
2932 set_idom(useuse, r, dom_depth(useuse));
2933 }
2934 }
2935 for (uint k = 1; k < useuse->req(); k++) {
2936 if( useuse->in(k) == use ) {
2937 useuse->set_req(k, r);
2938 uses_found++;
2939 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2940 // This is not a dom_depth > dd_r because when new
2941 // control flow is constructed by a loop opt, a node
2942 // and its dominator can end up at the same dom_depth
2943 assert(dom_depth(useuse) >= dd_r , "");
2944 set_idom(useuse, r, dom_depth(useuse));
2945 }
2946 }
2947 }
2948 l -= uses_found; // we deleted 1 or more copies of this edge
2949 }
2950
2951 assert(use->is_Proj(), "loop exit should be projection");
2952 // replace_node_and_forward_ctrl() below moves all nodes that are:
2953 // - control dependent on the loop exit or
2954 // - have control set to the loop exit
2955 // below the post-loop merge point.
2956 // replace_node_and_forward_ctrl() takes a dead control as first input.
2957 // To make it possible to use it, the loop exit projection is cloned and becomes the
2958 // new exit projection. The initial one becomes dead and is "replaced" by the region.
2959 Node* use_clone = use->clone();
2960 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2961 // Now finish up 'r'
2962 r->set_req(1, newuse);
2963 r->set_req(2, use_clone);
2964 _igvn.register_new_node_with_optimizer(r);
2965 set_loop(r, use_loop);
2966 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2967 replace_node_and_forward_ctrl(use, r);
2968 // Map the (cloned) old use to the new merge point
2969 old_new.map(use_clone->_idx, r);
2970 } // End of if a loop-exit test
2971 }
2972 }
2973 }
2974
2975 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2976 IdealLoopTree* parent, bool partial) {
2977 for(uint i = 0; i < body.size(); i++ ) {
2978 Node *old = body.at(i);
2979 Node *nnn = old_new[old->_idx];
2980 // Fix CFG/Loop controlling the new node
2981 if (has_ctrl(old)) {
2982 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2983 } else {
2984 set_loop(nnn, parent);
2985 if (old->outcnt() > 0) {
2986 Node* dom = idom(old);
2987 if (old_new[dom->_idx] != nullptr) {
2988 dom = old_new[dom->_idx];
2989 set_idom(nnn, dom, dd );
2990 }
2991 }
2992 }
2993 // Correct edges to the new node
2994 for (uint j = 0; j < nnn->req(); j++) {
2995 Node *n = nnn->in(j);
2996 if (n != nullptr) {
2997 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2998 if (loop->is_member(old_in_loop)) {
2999 if (old_new[n->_idx] != nullptr) {
3000 nnn->set_req(j, old_new[n->_idx]);
3001 } else {
3002 assert(!body.contains(n), "");
3003 assert(partial, "node not cloned");
3004 }
3005 }
3006 }
3007 }
3008 _igvn.hash_find_insert(nnn);
3009 }
3010 }
3011
3012 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
3013 for (uint i = 0; i < body.size(); i++) {
3014 Node* old = body.at(i);
3015 Node* nnn = old->clone();
3016 old_new.map(old->_idx, nnn);
3017 if (C->do_vector_loop() && cm != nullptr) {
3018 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
3019 }
3020 _igvn.register_new_node_with_optimizer(nnn);
3021 }
3022 }
3023
3024
3025 //---------------------- stride_of_possible_iv -------------------------------------
3026 // Looks for an iff/bool/comp with one operand of the compare
3027 // being a cycle involving an add and a phi,
3028 // with an optional truncation (left-shift followed by a right-shift)
3029 // of the add. Returns zero if not an iv.
3030 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
3031 const TypeInteger* ttype = nullptr;
3032 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
3033 return 0;
3034 }
3035 BoolNode* bl = iff->in(1)->as_Bool();
3036 Node* cmp = bl->in(1);
3037 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
3038 return 0;
3039 }
3040 // Must have an invariant operand
3041 if (ctrl_is_member(get_loop(iff), cmp->in(2))) {
3042 return 0;
3043 }
3044 Node* add2 = nullptr;
3045 Node* cmp1 = cmp->in(1);
3046 if (cmp1->is_Phi()) {
3047 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
3048 Node* phi = cmp1;
3049 for (uint i = 1; i < phi->req(); i++) {
3050 Node* in = phi->in(i);
3051 CountedLoopConverter::TruncatedIncrement add(T_INT);
3052 add.build(in);
3053 if (add.is_valid() && add.incr()->in(1) == phi) {
3054 add2 = add.incr()->in(2);
3055 break;
3056 }
3057 }
3058 } else {
3059 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
3060 Node* addtrunc = cmp1;
3061 CountedLoopConverter::TruncatedIncrement add(T_INT);
3062 add.build(addtrunc);
3063 if (add.is_valid() && add.incr()->in(1)->is_Phi()) {
3064 Node* phi = add.incr()->in(1);
3065 for (uint i = 1; i < phi->req(); i++) {
3066 if (phi->in(i) == addtrunc) {
3067 add2 = add.incr()->in(2);
3068 break;
3069 }
3070 }
3071 }
3072 }
3073 if (add2 != nullptr) {
3074 const TypeInt* add2t = _igvn.type(add2)->is_int();
3075 if (add2t->is_con()) {
3076 return add2t->get_con();
3077 }
3078 }
3079 return 0;
3080 }
3081
3082
3083 //---------------------- stay_in_loop -------------------------------------
3084 // Return the (unique) control output node that's in the loop (if it exists.)
3085 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
3086 Node* unique = nullptr;
3087 if (!n) return nullptr;
3088 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3089 Node* use = n->fast_out(i);
3090 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
3091 if (unique != nullptr) {
3092 return nullptr;
3093 }
3094 unique = use;
3095 }
3096 }
3097 return unique;
3098 }
3099
3100 //------------------------------ register_node -------------------------------------
3101 // Utility to register node "n" with PhaseIdealLoop
3102 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
3103 _igvn.register_new_node_with_optimizer(n);
3104 loop->_body.push(n);
3105 if (n->is_CFG()) {
3106 set_loop(n, loop);
3107 set_idom(n, pred, ddepth);
3108 } else {
3109 set_ctrl(n, pred);
3110 }
3111 }
3112
3113 //------------------------------ proj_clone -------------------------------------
3114 // Utility to create an if-projection
3115 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
3116 ProjNode* c = p->clone()->as_Proj();
3117 c->set_req(0, iff);
3118 return c;
3119 }
3120
3121 //------------------------------ short_circuit_if -------------------------------------
3122 // Force the iff control output to be the live_proj
3123 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
3124 guarantee(live_proj != nullptr, "null projection");
3125 int proj_con = live_proj->_con;
3126 assert(proj_con == 0 || proj_con == 1, "false or true projection");
3127 Node* con = intcon(proj_con);
3128 if (iff) {
3129 iff->set_req(1, con);
3130 }
3131 return con;
3132 }
3133
3134 //------------------------------ insert_if_before_proj -------------------------------------
3135 // Insert a new if before an if projection (* - new node)
3136 //
3137 // before
3138 // if(test)
3139 // / \
3140 // v v
3141 // other-proj proj (arg)
3142 //
3143 // after
3144 // if(test)
3145 // / \
3146 // / v
3147 // | * proj-clone
3148 // v |
3149 // other-proj v
3150 // * new_if(relop(cmp[IU](left,right)))
3151 // / \
3152 // v v
3153 // * new-proj proj
3154 // (returned)
3155 //
3156 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
3157 IfNode* iff = proj->in(0)->as_If();
3158 IdealLoopTree *loop = get_loop(proj);
3159 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3160 uint ddepth = dom_depth(proj);
3161
3162 _igvn.rehash_node_delayed(iff);
3163 _igvn.rehash_node_delayed(proj);
3164
3165 proj->set_req(0, nullptr); // temporary disconnect
3166 ProjNode* proj2 = proj_clone(proj, iff);
3167 register_node(proj2, loop, iff, ddepth);
3168
3169 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
3170 register_node(cmp, loop, proj2, ddepth);
3171
3172 BoolNode* bol = new BoolNode(cmp, relop);
3173 register_node(bol, loop, proj2, ddepth);
3174
3175 int opcode = iff->Opcode();
3176 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
3177 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
3178 register_node(new_if, loop, proj2, ddepth);
3179
3180 proj->set_req(0, new_if); // reattach
3181 set_idom(proj, new_if, ddepth);
3182
3183 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
3184 guarantee(new_exit != nullptr, "null exit node");
3185 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
3186
3187 return new_exit;
3188 }
3189
3190 //------------------------------ insert_region_before_proj -------------------------------------
3191 // Insert a region before an if projection (* - new node)
3192 //
3193 // before
3194 // if(test)
3195 // / |
3196 // v |
3197 // proj v
3198 // other-proj
3199 //
3200 // after
3201 // if(test)
3202 // / |
3203 // v |
3204 // * proj-clone v
3205 // | other-proj
3206 // v
3207 // * new-region
3208 // |
3209 // v
3210 // * dum_if
3211 // / \
3212 // v \
3213 // * dum-proj v
3214 // proj
3215 //
3216 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3217 IfNode* iff = proj->in(0)->as_If();
3218 IdealLoopTree *loop = get_loop(proj);
3219 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3220 uint ddepth = dom_depth(proj);
3221
3222 _igvn.rehash_node_delayed(iff);
3223 _igvn.rehash_node_delayed(proj);
3224
3225 proj->set_req(0, nullptr); // temporary disconnect
3226 ProjNode* proj2 = proj_clone(proj, iff);
3227 register_node(proj2, loop, iff, ddepth);
3228
3229 RegionNode* reg = new RegionNode(2);
3230 reg->set_req(1, proj2);
3231 register_node(reg, loop, iff, ddepth);
3232
3233 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3234 register_node(dum_if, loop, reg, ddepth);
3235
3236 proj->set_req(0, dum_if); // reattach
3237 set_idom(proj, dum_if, ddepth);
3238
3239 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3240 register_node(dum_proj, loop, dum_if, ddepth);
3241
3242 return reg;
3243 }
3244
3245 // Idea
3246 // ----
3247 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3248 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3249 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3250 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3251 // test alone instead:
3252 //
3253 // Before Partial Peeling:
3254 // Loop:
3255 // <peeled section>
3256 // Split off signed loop exit test
3257 // <-- CUT HERE -->
3258 // Unchanged unsigned loop exit test
3259 // <rest of unpeeled section>
3260 // goto Loop
3261 //
3262 // After Partial Peeling:
3263 // <cloned peeled section>
3264 // Cloned split off signed loop exit test
3265 // Loop:
3266 // Unchanged unsigned loop exit test
3267 // <rest of unpeeled section>
3268 // <peeled section>
3269 // Split off signed loop exit test
3270 // goto Loop
3271 //
3272 // Details
3273 // -------
3274 // Before:
3275 // if (i <u limit) Unsigned loop exit condition
3276 // / |
3277 // v v
3278 // exit-proj stay-in-loop-proj
3279 //
3280 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3281 // before the CmpU on the stay-in-loop path and keep both tests:
3282 //
3283 // if (i <u limit) Signed loop exit test
3284 // / |
3285 // / if (i <u limit) Unsigned loop exit test
3286 // / / |
3287 // v v v
3288 // exit-region stay-in-loop-proj
3289 //
3290 // Implementation
3291 // --------------
3292 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3293 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3294 // exit tests is preserved, and their loop nesting is correct.
3295 //
3296 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3297 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3298 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3299 //
3300 // if (stay-in-loop-const) Killed original unsigned loop exit test
3301 // / |
3302 // / v
3303 // / if (i < limit) Split off signed loop exit test
3304 // / / |
3305 // / / v
3306 // / / if (i <u limit) Cloned unsigned loop exit test
3307 // / / / |
3308 // v v v |
3309 // exit-region |
3310 // | |
3311 // dummy-if |
3312 // / | |
3313 // dead | |
3314 // v v
3315 // exit-proj stay-in-loop-proj
3316 //
3317 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3318 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3319 //
3320 // Requirements
3321 // ------------
3322 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3323 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3324 // Otherwise, we need to bail out (see comments in the code below).
3325 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3326 const bool Signed = true;
3327 const bool Unsigned = false;
3328
3329 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3330 if (bol->_test._test != BoolTest::lt) {
3331 return nullptr;
3332 }
3333 CmpNode* cmpu = bol->in(1)->as_Cmp();
3334 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3335
3336 int stride = stride_of_possible_iv(if_cmpu);
3337 if (stride == 0) {
3338 return nullptr;
3339 }
3340
3341 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3342 guarantee(lp_proj != nullptr, "null loop node");
3343
3344 ProjNode* lp_continue = lp_proj->as_Proj();
3345 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3346 if (!lp_exit->is_IfFalse()) {
3347 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3348 // We therefore can't add a single exit condition.
3349 return nullptr;
3350 }
3351 // The unsigned loop exit condition is
3352 // !(i <u limit)
3353 // = i >=u limit
3354 //
3355 // First, we note that for any x for which
3356 // 0 <= x <= INT_MAX
3357 // we can convert x to an unsigned int and still get the same guarantee:
3358 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3359 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3360 //
3361 // With that in mind, if
3362 // limit >= 0 (COND)
3363 // then the unsigned loop exit condition
3364 // i >=u limit (ULE)
3365 // is equivalent to
3366 // i < 0 || i >= limit (SLE-full)
3367 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3368 // (uint) i >=u MAX_INT >= limit >= 0
3369 // or otherwise
3370 // i >= limit >= 0
3371 // holds due to (LEMMA).
3372 //
3373 // For completeness, a counterexample with limit < 0:
3374 // Assume i = -3 and limit = -2:
3375 // i < 0
3376 // -2 < 0
3377 // is true and thus also "i < 0 || i >= limit". But
3378 // i >=u limit
3379 // -3 >=u -2
3380 // is false.
3381 Node* limit = cmpu->in(2);
3382 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3383 if (type_limit->_lo < 0) {
3384 return nullptr;
3385 }
3386
3387 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3388 // stride < 0:
3389 // i < 0 (SLE = SLE-negative)
3390 // stride > 0:
3391 // i >= limit (SLE = SLE-positive)
3392 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3393 //
3394 // Loop:
3395 // <peeled section>
3396 // i >= limit (SLE-positive)
3397 // <-- CUT HERE -->
3398 // i >=u limit (ULE)
3399 // <rest of unpeeled section>
3400 // goto Loop
3401 //
3402 // We exit the loop if:
3403 // (SLE) is true OR (ULE) is true
3404 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3405 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3406 // (SLE) IMPLIES (ULE)
3407 // This indeed holds when (COND) is given:
3408 // - stride > 0:
3409 // i >= limit // (SLE = SLE-positive)
3410 // i >= limit >= 0 // (COND)
3411 // i >=u limit >= 0 // (LEMMA)
3412 // which is the unsigned loop exit condition (ULE).
3413 // - stride < 0:
3414 // i < 0 // (SLE = SLE-negative)
3415 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3416 // MAX_INT >= limit >= 0 // (COND)
3417 // MAX_INT >=u limit >= 0 // (LEMMA)
3418 // and thus from (NEG) and (LEMMA):
3419 // i >=u limit
3420 // which is the unsigned loop exit condition (ULE).
3421 //
3422 //
3423 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3424 // <cloned peeled section>
3425 // i >= limit (SLE-positive)
3426 // Loop:
3427 // i >=u limit (ULE)
3428 // <rest of unpeeled section>
3429 // <peeled section>
3430 // i >= limit (SLE-positive)
3431 // goto Loop
3432 Node* rhs_cmpi;
3433 if (stride > 0) {
3434 rhs_cmpi = limit; // For i >= limit
3435 } else {
3436 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3437 }
3438 // Create a new region on the exit path
3439 RegionNode* reg = insert_region_before_proj(lp_exit);
3440 guarantee(reg != nullptr, "null region node");
3441
3442 // Clone the if-cmpu-true-false using a signed compare
3443 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3444 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3445 reg->add_req(cmpi_exit);
3446
3447 // Clone the if-cmpu-true-false
3448 BoolTest::mask rel_u = bol->_test._test;
3449 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3450 reg->add_req(cmpu_exit);
3451
3452 // Force original if to stay in loop.
3453 short_circuit_if(if_cmpu, lp_continue);
3454
3455 return cmpi_exit->in(0)->as_If();
3456 }
3457
3458 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3459 // Remove a previously inserted signed compare loop exit.
3460 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3461 Node* lp_proj = stay_in_loop(if_cmp, loop);
3462 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3463 stay_in_loop(lp_proj, loop)->is_If() &&
3464 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3465 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3466 if_cmp->set_req(1, con);
3467 }
3468
3469 //------------------------------ scheduled_nodelist -------------------------------------
3470 // Create a post order schedule of nodes that are in the
3471 // "member" set. The list is returned in "sched".
3472 // The first node in "sched" is the loop head, followed by
3473 // nodes which have no inputs in the "member" set, and then
3474 // followed by the nodes that have an immediate input dependence
3475 // on a node in "sched".
3476 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3477
3478 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3479 VectorSet visited;
3480 Node_Stack nstack(loop->_body.size());
3481
3482 Node* n = loop->_head; // top of stack is cached in "n"
3483 uint idx = 0;
3484 visited.set(n->_idx);
3485
3486 // Initially push all with no inputs from within member set
3487 for(uint i = 0; i < loop->_body.size(); i++ ) {
3488 Node *elt = loop->_body.at(i);
3489 if (member.test(elt->_idx)) {
3490 bool found = false;
3491 for (uint j = 0; j < elt->req(); j++) {
3492 Node* def = elt->in(j);
3493 if (def && member.test(def->_idx) && def != elt) {
3494 found = true;
3495 break;
3496 }
3497 }
3498 if (!found && elt != loop->_head) {
3499 nstack.push(n, idx);
3500 n = elt;
3501 assert(!visited.test(n->_idx), "not seen yet");
3502 visited.set(n->_idx);
3503 }
3504 }
3505 }
3506
3507 // traverse out's that are in the member set
3508 while (true) {
3509 if (idx < n->outcnt()) {
3510 Node* use = n->raw_out(idx);
3511 idx++;
3512 if (!visited.test_set(use->_idx)) {
3513 if (member.test(use->_idx)) {
3514 nstack.push(n, idx);
3515 n = use;
3516 idx = 0;
3517 }
3518 }
3519 } else {
3520 // All outputs processed
3521 sched.push(n);
3522 if (nstack.is_empty()) break;
3523 n = nstack.node();
3524 idx = nstack.index();
3525 nstack.pop();
3526 }
3527 }
3528 }
3529
3530
3531 //------------------------------ has_use_in_set -------------------------------------
3532 // Has a use in the vector set
3533 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3534 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3535 Node* use = n->fast_out(j);
3536 if (vset.test(use->_idx)) {
3537 return true;
3538 }
3539 }
3540 return false;
3541 }
3542
3543
3544 //------------------------------ has_use_internal_to_set -------------------------------------
3545 // Has use internal to the vector set (ie. not in a phi at the loop head)
3546 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3547 Node* head = loop->_head;
3548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3549 Node* use = n->fast_out(j);
3550 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3551 return true;
3552 }
3553 }
3554 return false;
3555 }
3556
3557
3558 //------------------------------ clone_for_use_outside_loop -------------------------------------
3559 // clone "n" for uses that are outside of loop
3560 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3561 int cloned = 0;
3562 assert(worklist.size() == 0, "should be empty");
3563 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3564 Node* use = n->fast_out(j);
3565 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3566 worklist.push(use);
3567 }
3568 }
3569
3570 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3571 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3572 return -1;
3573 }
3574
3575 while( worklist.size() ) {
3576 Node *use = worklist.pop();
3577 if (!has_node(use) || use->in(0) == C->top()) continue;
3578 uint j;
3579 for (j = 0; j < use->req(); j++) {
3580 if (use->in(j) == n) break;
3581 }
3582 assert(j < use->req(), "must be there");
3583
3584 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3585 Node* n_clone = n->clone();
3586 _igvn.replace_input_of(use, j, n_clone);
3587 cloned++;
3588 Node* use_c;
3589 if (!use->is_Phi()) {
3590 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3591 } else {
3592 // Use in a phi is considered a use in the associated predecessor block
3593 use_c = use->in(0)->in(j);
3594 }
3595 set_ctrl(n_clone, use_c);
3596 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3597 get_loop(use_c)->_body.push(n_clone);
3598 _igvn.register_new_node_with_optimizer(n_clone);
3599 #ifndef PRODUCT
3600 if (TracePartialPeeling) {
3601 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3602 }
3603 #endif
3604 }
3605 return cloned;
3606 }
3607
3608
3609 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3610 // clone "n" for special uses that are in the not_peeled region.
3611 // If these def-uses occur in separate blocks, the code generator
3612 // marks the method as not compilable. For example, if a "BoolNode"
3613 // is in a different basic block than the "IfNode" that uses it, then
3614 // the compilation is aborted in the code generator.
3615 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3616 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3617 if (n->is_Phi() || n->is_Load()) {
3618 return;
3619 }
3620 assert(worklist.size() == 0, "should be empty");
3621 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3622 Node* use = n->fast_out(j);
3623 if ( not_peel.test(use->_idx) &&
3624 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3625 use->in(1) == n) {
3626 worklist.push(use);
3627 }
3628 }
3629 if (worklist.size() > 0) {
3630 // clone "n" and insert it between inputs of "n" and the use
3631 Node* n_clone = n->clone();
3632 loop->_body.push(n_clone);
3633 _igvn.register_new_node_with_optimizer(n_clone);
3634 set_ctrl(n_clone, get_ctrl(n));
3635 sink_list.push(n_clone);
3636 not_peel.set(n_clone->_idx);
3637 #ifndef PRODUCT
3638 if (TracePartialPeeling) {
3639 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3640 }
3641 #endif
3642 while( worklist.size() ) {
3643 Node *use = worklist.pop();
3644 _igvn.rehash_node_delayed(use);
3645 for (uint j = 1; j < use->req(); j++) {
3646 if (use->in(j) == n) {
3647 use->set_req(j, n_clone);
3648 }
3649 }
3650 }
3651 }
3652 }
3653
3654
3655 //------------------------------ insert_phi_for_loop -------------------------------------
3656 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3657 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3658 Node *phi = PhiNode::make(lp, back_edge_val);
3659 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3660 // Use existing phi if it already exists
3661 Node *hit = _igvn.hash_find_insert(phi);
3662 if( hit == nullptr ) {
3663 _igvn.register_new_node_with_optimizer(phi);
3664 set_ctrl(phi, lp);
3665 } else {
3666 // Remove the new phi from the graph and use the hit
3667 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Speculative);
3668 phi = hit;
3669 }
3670 _igvn.replace_input_of(use, idx, phi);
3671 }
3672
3673 #ifdef ASSERT
3674 //------------------------------ is_valid_loop_partition -------------------------------------
3675 // Validate the loop partition sets: peel and not_peel
3676 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3677 VectorSet& not_peel ) {
3678 uint i;
3679 // Check that peel_list entries are in the peel set
3680 for (i = 0; i < peel_list.size(); i++) {
3681 if (!peel.test(peel_list.at(i)->_idx)) {
3682 return false;
3683 }
3684 }
3685 // Check at loop members are in one of peel set or not_peel set
3686 for (i = 0; i < loop->_body.size(); i++ ) {
3687 Node *def = loop->_body.at(i);
3688 uint di = def->_idx;
3689 // Check that peel set elements are in peel_list
3690 if (peel.test(di)) {
3691 if (not_peel.test(di)) {
3692 return false;
3693 }
3694 // Must be in peel_list also
3695 bool found = false;
3696 for (uint j = 0; j < peel_list.size(); j++) {
3697 if (peel_list.at(j)->_idx == di) {
3698 found = true;
3699 break;
3700 }
3701 }
3702 if (!found) {
3703 return false;
3704 }
3705 } else if (not_peel.test(di)) {
3706 if (peel.test(di)) {
3707 return false;
3708 }
3709 } else {
3710 return false;
3711 }
3712 }
3713 return true;
3714 }
3715
3716 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3717 // Ensure a use outside of loop is of the right form
3718 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3719 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3720 return (use->is_Phi() &&
3721 use_c->is_Region() && use_c->req() == 3 &&
3722 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3723 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3724 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3725 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3726 }
3727
3728 //------------------------------ is_valid_clone_loop_form -------------------------------------
3729 // Ensure that all uses outside of loop are of the right form
3730 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3731 uint orig_exit_idx, uint clone_exit_idx) {
3732 uint len = peel_list.size();
3733 for (uint i = 0; i < len; i++) {
3734 Node *def = peel_list.at(i);
3735
3736 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3737 Node *use = def->fast_out(j);
3738 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3739 if (!loop->is_member(get_loop(use_c))) {
3740 // use is not in the loop, check for correct structure
3741 if (use->in(0) == def) {
3742 // Okay
3743 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3744 return false;
3745 }
3746 }
3747 }
3748 }
3749 return true;
3750 }
3751 #endif
3752
3753 //------------------------------ partial_peel -------------------------------------
3754 // Partially peel (aka loop rotation) the top portion of a loop (called
3755 // the peel section below) by cloning it and placing one copy just before
3756 // the new loop head and the other copy at the bottom of the new loop.
3757 //
3758 // before after where it came from
3759 //
3760 // stmt1 stmt1
3761 // loop: stmt2 clone
3762 // stmt2 if condA goto exitA clone
3763 // if condA goto exitA new_loop: new
3764 // stmt3 stmt3 clone
3765 // if !condB goto loop if condB goto exitB clone
3766 // exitB: stmt2 orig
3767 // stmt4 if !condA goto new_loop orig
3768 // exitA: goto exitA
3769 // exitB:
3770 // stmt4
3771 // exitA:
3772 //
3773 // Step 1: find the cut point: an exit test on probable
3774 // induction variable.
3775 // Step 2: schedule (with cloning) operations in the peel
3776 // section that can be executed after the cut into
3777 // the section that is not peeled. This may need
3778 // to clone operations into exit blocks. For
3779 // instance, a reference to A[i] in the not-peel
3780 // section and a reference to B[i] in an exit block
3781 // may cause a left-shift of i by 2 to be placed
3782 // in the peel block. This step will clone the left
3783 // shift into the exit block and sink the left shift
3784 // from the peel to the not-peel section.
3785 // Step 3: clone the loop, retarget the control, and insert
3786 // phis for values that are live across the new loop
3787 // head. This is very dependent on the graph structure
3788 // from clone_loop. It creates region nodes for
3789 // exit control and associated phi nodes for values
3790 // flow out of the loop through that exit. The region
3791 // node is dominated by the clone's control projection.
3792 // So the clone's peel section is placed before the
3793 // new loop head, and the clone's not-peel section is
3794 // forms the top part of the new loop. The original
3795 // peel section forms the tail of the new loop.
3796 // Step 4: update the dominator tree and recompute the
3797 // dominator depth.
3798 //
3799 // orig
3800 //
3801 // stmt1
3802 // |
3803 // v
3804 // predicates
3805 // |
3806 // v
3807 // loop<----+
3808 // | |
3809 // stmt2 |
3810 // | |
3811 // v |
3812 // ifA |
3813 // / | |
3814 // v v |
3815 // false true ^ <-- last_peel
3816 // / | |
3817 // / ===|==cut |
3818 // / stmt3 | <-- first_not_peel
3819 // / | |
3820 // | v |
3821 // v ifB |
3822 // exitA: / \ |
3823 // / \ |
3824 // v v |
3825 // false true |
3826 // / \ |
3827 // / ----+
3828 // |
3829 // v
3830 // exitB:
3831 // stmt4
3832 //
3833 //
3834 // after clone loop
3835 //
3836 // stmt1
3837 // |
3838 // v
3839 // predicates
3840 // / \
3841 // clone / \ orig
3842 // / \
3843 // / \
3844 // v v
3845 // +---->loop loop<----+
3846 // | | | |
3847 // | stmt2 stmt2 |
3848 // | | | |
3849 // | v v |
3850 // | ifA ifA |
3851 // | | \ / | |
3852 // | v v v v |
3853 // ^ true false false true ^ <-- last_peel
3854 // | | ^ \ / | |
3855 // | cut==|== \ \ / ===|==cut |
3856 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3857 // | | dom | | | |
3858 // | v \ 1v v2 v |
3859 // | ifB regionA ifB |
3860 // | / \ | / \ |
3861 // | / \ v / \ |
3862 // | v v exitA: v v |
3863 // | true false false true |
3864 // | / ^ \ / \ |
3865 // +---- \ \ / ----+
3866 // dom \ /
3867 // \ 1v v2
3868 // regionB
3869 // |
3870 // v
3871 // exitB:
3872 // stmt4
3873 //
3874 //
3875 // after partial peel
3876 //
3877 // stmt1
3878 // |
3879 // v
3880 // predicates
3881 // /
3882 // clone / orig
3883 // / TOP
3884 // / \
3885 // v v
3886 // TOP->loop loop----+
3887 // | | |
3888 // stmt2 stmt2 |
3889 // | | |
3890 // v v |
3891 // ifA ifA |
3892 // | \ / | |
3893 // v v v v |
3894 // true false false true | <-- last_peel
3895 // | ^ \ / +------|---+
3896 // +->newloop \ \ / === ==cut | |
3897 // | stmt3 \ \ / TOP | |
3898 // | | dom | | stmt3 | | <-- first_not_peel
3899 // | v \ 1v v2 v | |
3900 // | ifB regionA ifB ^ v
3901 // | / \ | / \ | |
3902 // | / \ v / \ | |
3903 // | v v exitA: v v | |
3904 // | true false false true | |
3905 // | / ^ \ / \ | |
3906 // | | \ \ / v | |
3907 // | | dom \ / TOP | |
3908 // | | \ 1v v2 | |
3909 // ^ v regionB | |
3910 // | | | | |
3911 // | | v ^ v
3912 // | | exitB: | |
3913 // | | stmt4 | |
3914 // | +------------>-----------------+ |
3915 // | |
3916 // +-----------------<---------------------+
3917 //
3918 //
3919 // final graph
3920 //
3921 // stmt1
3922 // |
3923 // v
3924 // predicates
3925 // |
3926 // v
3927 // stmt2 clone
3928 // |
3929 // v
3930 // ........> ifA clone
3931 // : / |
3932 // dom / |
3933 // : v v
3934 // : false true
3935 // : | |
3936 // : | v
3937 // : | newloop<-----+
3938 // : | | |
3939 // : | stmt3 clone |
3940 // : | | |
3941 // : | v |
3942 // : | ifB |
3943 // : | / \ |
3944 // : | v v |
3945 // : | false true |
3946 // : | | | |
3947 // : | v stmt2 |
3948 // : | exitB: | |
3949 // : | stmt4 v |
3950 // : | ifA orig |
3951 // : | / \ |
3952 // : | / \ |
3953 // : | v v |
3954 // : | false true |
3955 // : | / \ |
3956 // : v v -----+
3957 // RegionA
3958 // |
3959 // v
3960 // exitA
3961 //
3962 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3963
3964 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3965 if (!loop->_head->is_Loop()) {
3966 return false;
3967 }
3968 LoopNode *head = loop->_head->as_Loop();
3969
3970 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3971 return false;
3972 }
3973
3974 // Check for complex exit control
3975 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3976 Node *n = loop->_body.at(ii);
3977 int opc = n->Opcode();
3978 if (n->is_Call() ||
3979 opc == Op_Catch ||
3980 opc == Op_CatchProj ||
3981 opc == Op_Jump ||
3982 opc == Op_JumpProj) {
3983 #ifndef PRODUCT
3984 if (TracePartialPeeling) {
3985 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3986 }
3987 #endif
3988 return false;
3989 }
3990 }
3991
3992 int dd = dom_depth(head);
3993
3994 // Step 1: find cut point
3995
3996 // Walk up dominators to loop head looking for first loop exit
3997 // which is executed on every path thru loop.
3998 IfNode *peel_if = nullptr;
3999 IfNode *peel_if_cmpu = nullptr;
4000
4001 Node *iff = loop->tail();
4002 while (iff != head) {
4003 if (iff->is_If()) {
4004 Node *ctrl = get_ctrl(iff->in(1));
4005 if (ctrl->is_top()) return false; // Dead test on live IF.
4006 // If loop-varying exit-test, check for induction variable
4007 if (loop->is_member(get_loop(ctrl)) &&
4008 loop->is_loop_exit(iff) &&
4009 is_possible_iv_test(iff)) {
4010 Node* cmp = iff->in(1)->in(1);
4011 if (cmp->Opcode() == Op_CmpI) {
4012 peel_if = iff->as_If();
4013 } else {
4014 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
4015 peel_if_cmpu = iff->as_If();
4016 }
4017 }
4018 }
4019 iff = idom(iff);
4020 }
4021
4022 // Prefer signed compare over unsigned compare.
4023 IfNode* new_peel_if = nullptr;
4024 if (peel_if == nullptr) {
4025 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
4026 return false; // No peel point found
4027 }
4028 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
4029 if (new_peel_if == nullptr) {
4030 return false; // No peel point found
4031 }
4032 peel_if = new_peel_if;
4033 }
4034 Node* last_peel = stay_in_loop(peel_if, loop);
4035 Node* first_not_peeled = stay_in_loop(last_peel, loop);
4036 if (first_not_peeled == nullptr || first_not_peeled == head) {
4037 return false;
4038 }
4039
4040 #ifndef PRODUCT
4041 if (TraceLoopOpts) {
4042 tty->print("PartialPeel ");
4043 loop->dump_head();
4044 }
4045
4046 if (TracePartialPeeling) {
4047 tty->print_cr("before partial peel one iteration");
4048 Node_List wl;
4049 Node* t = head->in(2);
4050 while (true) {
4051 wl.push(t);
4052 if (t == head) break;
4053 t = idom(t);
4054 }
4055 while (wl.size() > 0) {
4056 Node* tt = wl.pop();
4057 tt->dump();
4058 if (tt == last_peel) tty->print_cr("-- cut --");
4059 }
4060 }
4061 #endif
4062
4063 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
4064
4065 VectorSet peel;
4066 VectorSet not_peel;
4067 Node_List peel_list;
4068 Node_List worklist;
4069 Node_List sink_list;
4070
4071 uint estimate = loop->est_loop_clone_sz(1);
4072 if (exceeding_node_budget(estimate)) {
4073 return false;
4074 }
4075
4076 // Set of cfg nodes to peel are those that are executable from
4077 // the head through last_peel.
4078 assert(worklist.size() == 0, "should be empty");
4079 worklist.push(head);
4080 peel.set(head->_idx);
4081 while (worklist.size() > 0) {
4082 Node *n = worklist.pop();
4083 if (n != last_peel) {
4084 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
4085 Node* use = n->fast_out(j);
4086 if (use->is_CFG() &&
4087 loop->is_member(get_loop(use)) &&
4088 !peel.test_set(use->_idx)) {
4089 worklist.push(use);
4090 }
4091 }
4092 }
4093 }
4094
4095 // Set of non-cfg nodes to peel are those that are control
4096 // dependent on the cfg nodes.
4097 for (uint i = 0; i < loop->_body.size(); i++) {
4098 Node *n = loop->_body.at(i);
4099 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
4100 if (peel.test(n_c->_idx)) {
4101 peel.set(n->_idx);
4102 } else {
4103 not_peel.set(n->_idx);
4104 }
4105 }
4106
4107 // Step 2: move operations from the peeled section down into the
4108 // not-peeled section
4109
4110 // Get a post order schedule of nodes in the peel region
4111 // Result in right-most operand.
4112 scheduled_nodelist(loop, peel, peel_list);
4113
4114 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4115
4116 // For future check for too many new phis
4117 uint old_phi_cnt = 0;
4118 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
4119 Node* use = head->fast_out(j);
4120 if (use->is_Phi()) old_phi_cnt++;
4121 }
4122
4123 #ifndef PRODUCT
4124 if (TracePartialPeeling) {
4125 tty->print_cr("\npeeled list");
4126 }
4127 #endif
4128
4129 // Evacuate nodes in peel region into the not_peeled region if possible
4130 bool too_many_clones = false;
4131 uint new_phi_cnt = 0;
4132 uint cloned_for_outside_use = 0;
4133 for (uint i = 0; i < peel_list.size();) {
4134 Node* n = peel_list.at(i);
4135 #ifndef PRODUCT
4136 if (TracePartialPeeling) n->dump();
4137 #endif
4138 bool incr = true;
4139 if (!n->is_CFG()) {
4140 if (has_use_in_set(n, not_peel)) {
4141 // If not used internal to the peeled region,
4142 // move "n" from peeled to not_peeled region.
4143 if (!has_use_internal_to_set(n, peel, loop)) {
4144 // if not pinned and not a load (which maybe anti-dependent on a store)
4145 // and not a CMove (Matcher expects only bool->cmove).
4146 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
4147 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
4148 if (C->failing()) return false;
4149 if (new_clones == -1) {
4150 too_many_clones = true;
4151 break;
4152 }
4153 cloned_for_outside_use += new_clones;
4154 sink_list.push(n);
4155 peel.remove(n->_idx);
4156 not_peel.set(n->_idx);
4157 peel_list.remove(i);
4158 incr = false;
4159 #ifndef PRODUCT
4160 if (TracePartialPeeling) {
4161 tty->print_cr("sink to not_peeled region: %d newbb: %d",
4162 n->_idx, get_ctrl(n)->_idx);
4163 }
4164 #endif
4165 }
4166 } else {
4167 // Otherwise check for special def-use cases that span
4168 // the peel/not_peel boundary such as bool->if
4169 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
4170 new_phi_cnt++;
4171 }
4172 }
4173 }
4174 if (incr) i++;
4175 }
4176
4177 estimate += cloned_for_outside_use + new_phi_cnt;
4178 bool exceed_node_budget = !may_require_nodes(estimate);
4179 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
4180
4181 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
4182 #ifndef PRODUCT
4183 if (TracePartialPeeling && exceed_phi_limit) {
4184 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
4185 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
4186 }
4187 #endif
4188 if (new_peel_if != nullptr) {
4189 remove_cmpi_loop_exit(new_peel_if, loop);
4190 }
4191 // Inhibit more partial peeling on this loop
4192 assert(!head->is_partial_peel_loop(), "not partial peeled");
4193 head->mark_partial_peel_failed();
4194 if (cloned_for_outside_use > 0) {
4195 // Terminate this round of loop opts because
4196 // the graph outside this loop was changed.
4197 C->set_major_progress();
4198 return true;
4199 }
4200 return false;
4201 }
4202
4203 // Step 3: clone loop, retarget control, and insert new phis
4204
4205 // Create new loop head for new phis and to hang
4206 // the nodes being moved (sinked) from the peel region.
4207 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4208 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4209 _igvn.register_new_node_with_optimizer(new_head);
4210 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4211 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4212 set_loop(new_head, loop);
4213 loop->_body.push(new_head);
4214 not_peel.set(new_head->_idx);
4215 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4216 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4217
4218 while (sink_list.size() > 0) {
4219 Node* n = sink_list.pop();
4220 set_ctrl(n, new_head);
4221 }
4222
4223 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4224
4225 clone_loop(loop, old_new, dd, IgnoreStripMined);
4226
4227 const uint clone_exit_idx = 1;
4228 const uint orig_exit_idx = 2;
4229 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4230
4231 Node* head_clone = old_new[head->_idx];
4232 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4233 Node* orig_tail_clone = head_clone->in(2);
4234
4235 // Add phi if "def" node is in peel set and "use" is not
4236
4237 for (uint i = 0; i < peel_list.size(); i++) {
4238 Node *def = peel_list.at(i);
4239 if (!def->is_CFG()) {
4240 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4241 Node *use = def->fast_out(j);
4242 if (has_node(use) && use->in(0) != C->top() &&
4243 (!peel.test(use->_idx) ||
4244 (use->is_Phi() && use->in(0) == head)) ) {
4245 worklist.push(use);
4246 }
4247 }
4248 while( worklist.size() ) {
4249 Node *use = worklist.pop();
4250 for (uint j = 1; j < use->req(); j++) {
4251 Node* n = use->in(j);
4252 if (n == def) {
4253
4254 // "def" is in peel set, "use" is not in peel set
4255 // or "use" is in the entry boundary (a phi) of the peel set
4256
4257 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4258
4259 if ( loop->is_member(get_loop( use_c )) ) {
4260 // use is in loop
4261 if (old_new[use->_idx] != nullptr) { // null for dead code
4262 Node* use_clone = old_new[use->_idx];
4263 _igvn.replace_input_of(use, j, C->top());
4264 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4265 }
4266 } else {
4267 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4268 // use is not in the loop, check if the live range includes the cut
4269 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4270 if (not_peel.test(lp_if->_idx)) {
4271 assert(j == orig_exit_idx, "use from original loop");
4272 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4273 }
4274 }
4275 }
4276 }
4277 }
4278 }
4279 }
4280
4281 // Step 3b: retarget control
4282
4283 // Redirect control to the new loop head if a cloned node in
4284 // the not_peeled region has control that points into the peeled region.
4285 // This necessary because the cloned peeled region will be outside
4286 // the loop.
4287 // from to
4288 // cloned-peeled <---+
4289 // new_head_clone: | <--+
4290 // cloned-not_peeled in(0) in(0)
4291 // orig-peeled
4292
4293 for (uint i = 0; i < loop->_body.size(); i++) {
4294 Node *n = loop->_body.at(i);
4295 if (!n->is_CFG() && n->in(0) != nullptr &&
4296 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4297 Node* n_clone = old_new[n->_idx];
4298 if (n_clone->depends_only_on_test()) {
4299 // If this node depends_only_on_test, it will be rewire to the loop head, which is not the
4300 // correct test
4301 Node* pinned_clone = n_clone->pin_node_under_control();
4302 if (pinned_clone != nullptr) {
4303 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4304 old_new.map(n->_idx, pinned_clone);
4305 _igvn.replace_node(n_clone, pinned_clone);
4306 n_clone = pinned_clone;
4307 }
4308 }
4309 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4310 }
4311 }
4312
4313 // Backedge of the surviving new_head (the clone) is original last_peel
4314 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4315
4316 // Cut first node in original not_peel set
4317 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4318 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4319 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4320
4321 // Copy head_clone back-branch info to original head
4322 // and remove original head's loop entry and
4323 // clone head's back-branch
4324 _igvn.rehash_node_delayed(head); // Multiple edge updates
4325 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4326 head->set_req(LoopNode::LoopBackControl, C->top());
4327 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4328
4329 // Similarly modify the phis
4330 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4331 Node* use = head->fast_out(k);
4332 if (use->is_Phi() && use->outcnt() > 0) {
4333 Node* use_clone = old_new[use->_idx];
4334 _igvn.rehash_node_delayed(use); // Multiple edge updates
4335 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4336 use->set_req(LoopNode::LoopBackControl, C->top());
4337 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4338 }
4339 }
4340
4341 // Step 4: update dominator tree and dominator depth
4342
4343 set_idom(head, orig_tail_clone, dd);
4344 recompute_dom_depth();
4345
4346 // Inhibit more partial peeling on this loop
4347 new_head_clone->set_partial_peel_loop();
4348 C->set_major_progress();
4349 loop->record_for_igvn();
4350
4351 #ifndef PRODUCT
4352 if (TracePartialPeeling) {
4353 tty->print_cr("\nafter partial peel one iteration");
4354 Node_List wl;
4355 Node* t = last_peel;
4356 while (true) {
4357 wl.push(t);
4358 if (t == head_clone) break;
4359 t = idom(t);
4360 }
4361 while (wl.size() > 0) {
4362 Node* tt = wl.pop();
4363 if (tt == head) tty->print_cr("orig head");
4364 else if (tt == new_head_clone) tty->print_cr("new head");
4365 else if (tt == head_clone) tty->print_cr("clone head");
4366 tt->dump();
4367 }
4368 }
4369 #endif
4370
4371 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4372
4373 return true;
4374 }
4375
4376 #ifdef ASSERT
4377
4378 // Moves Template Assertion Predicates to a target loop by cloning and killing the old ones. The target loop is the
4379 // original, not-cloned loop. This is currently only used with StressLoopBackedge which is a develop flag only and
4380 // false with product builds. We can therefore guard it with an ifdef. More details can be found at the use-site.
4381 class MoveAssertionPredicatesVisitor : public PredicateVisitor {
4382 ClonePredicateToTargetLoop _clone_predicate_to_loop;
4383 PhaseIdealLoop* const _phase;
4384
4385 public:
4386 MoveAssertionPredicatesVisitor(LoopNode* target_loop_head,
4387 const NodeInSingleLoopBody &node_in_loop_body,
4388 PhaseIdealLoop* phase)
4389 : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
4390 _phase(phase) {
4391 }
4392 NONCOPYABLE(MoveAssertionPredicatesVisitor);
4393
4394 using PredicateVisitor::visit;
4395
4396 void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
4397 _clone_predicate_to_loop.clone_template_assertion_predicate(template_assertion_predicate);
4398 template_assertion_predicate.kill(_phase->igvn());
4399 }
4400 };
4401 #endif // ASSERT
4402
4403 // Transform:
4404 //
4405 // loop<-----------------+
4406 // | |
4407 // stmt1 stmt2 .. stmtn |
4408 // | | | |
4409 // \ | / |
4410 // v v v |
4411 // region |
4412 // | |
4413 // shared_stmt |
4414 // | |
4415 // v |
4416 // if |
4417 // / \ |
4418 // | -----------+
4419 // v
4420 //
4421 // into:
4422 //
4423 // loop<-------------------+
4424 // | |
4425 // v |
4426 // +->loop |
4427 // | | |
4428 // | stmt1 stmt2 .. stmtn |
4429 // | | | | |
4430 // | | \ / |
4431 // | | v v |
4432 // | | region1 |
4433 // | | | |
4434 // | shared_stmt shared_stmt |
4435 // | | | |
4436 // | v v |
4437 // | if if |
4438 // | /\ / \ |
4439 // +-- | | -------+
4440 // \ /
4441 // v v
4442 // region2
4443 //
4444 // (region2 is shown to merge mirrored projections of the loop exit
4445 // ifs to make the diagram clearer but they really merge the same
4446 // projection)
4447 //
4448 // Conditions for this transformation to trigger:
4449 // - the path through stmt1 is frequent enough
4450 // - the inner loop will be turned into a counted loop after transformation
4451 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4452 if (!DuplicateBackedge) {
4453 return false;
4454 }
4455 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4456 if (!loop->_head->is_Loop()) {
4457 return false;
4458 }
4459
4460 uint estimate = loop->est_loop_clone_sz(1);
4461 if (exceeding_node_budget(estimate)) {
4462 return false;
4463 }
4464
4465 LoopNode *head = loop->_head->as_Loop();
4466
4467 Node* region = nullptr;
4468 IfNode* exit_test = nullptr;
4469 uint inner;
4470 float f;
4471 #ifdef ASSERT
4472 if (StressDuplicateBackedge) {
4473 if (head->is_strip_mined()) {
4474 return false;
4475 }
4476 Node* c = head->in(LoopNode::LoopBackControl);
4477
4478 while (c != head) {
4479 if (c->is_Region()) {
4480 region = c;
4481 }
4482 c = idom(c);
4483 }
4484
4485 if (region == nullptr) {
4486 return false;
4487 }
4488
4489 inner = 1;
4490 } else
4491 #endif //ASSERT
4492 {
4493 // Is the shape of the loop that of a counted loop...
4494 Node* back_control = loop_exit_control(loop);
4495 if (back_control == nullptr) {
4496 return false;
4497 }
4498
4499 LoopExitTest loop_exit(back_control, loop, this);
4500 loop_exit.build();
4501 if (!loop_exit.is_valid_with_bt(T_INT)) {
4502 return false;
4503 }
4504
4505 const Node* loop_incr = loop_exit.incr();
4506
4507 // With an extra phi for the candidate iv?
4508 // Or the region node is the loop head
4509 if (!loop_incr->is_Phi() || loop_incr->in(0) == head) {
4510 return false;
4511 }
4512
4513 PathFrequency pf(head, this);
4514 region = loop_incr->in(0);
4515
4516 // Go over all paths for the extra phi's region and see if that
4517 // path is frequent enough and would match the expected iv shape
4518 // if the extra phi is removed
4519 inner = 0;
4520 for (uint i = 1; i < loop_incr->req(); ++i) {
4521 CountedLoopConverter::TruncatedIncrement increment(T_INT);
4522 increment.build(loop_incr->in(i));
4523 if (!increment.is_valid()) {
4524 continue;
4525 }
4526 assert(increment.incr()->Opcode() == Op_AddI, "wrong increment code");
4527
4528 LoopIVStride stride = LoopIVStride(T_INT);
4529 stride.build(increment.incr());
4530 if (!stride.is_valid()) {
4531 continue;
4532 }
4533
4534 PhiNode* phi = loop_iv_phi(stride.xphi(), nullptr, head);
4535 if (phi == nullptr ||
4536 (increment.outer_trunc() == nullptr && phi->in(LoopNode::LoopBackControl) != loop_exit.incr()) ||
4537 (increment.outer_trunc() != nullptr && phi->in(LoopNode::LoopBackControl) != increment.outer_trunc())) {
4538 return false;
4539 }
4540
4541 f = pf.to(region->in(i));
4542 if (f > 0.5) {
4543 inner = i;
4544 break;
4545 }
4546 }
4547
4548 if (inner == 0) {
4549 return false;
4550 }
4551
4552 exit_test = back_control->in(0)->as_If();
4553 }
4554
4555 if (idom(region)->is_Catch()) {
4556 return false;
4557 }
4558
4559 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4560 Unique_Node_List wq;
4561 wq.push(head->in(LoopNode::LoopBackControl));
4562 for (uint i = 0; i < wq.size(); i++) {
4563 Node* c = wq.at(i);
4564 assert(get_loop(c) == loop, "not in the right loop?");
4565 if (c->is_Region()) {
4566 if (c != region) {
4567 for (uint j = 1; j < c->req(); ++j) {
4568 wq.push(c->in(j));
4569 }
4570 }
4571 } else {
4572 wq.push(c->in(0));
4573 }
4574 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4575 }
4576
4577 Node* region_dom = idom(region);
4578
4579 // Can't do the transformation if this would cause a membar pair to
4580 // be split
4581 for (uint i = 0; i < wq.size(); i++) {
4582 Node* c = wq.at(i);
4583 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4584 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4585 if (!wq.member(c->as_MemBar()->leading_membar())) {
4586 return false;
4587 }
4588 }
4589 }
4590 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head);
4591
4592 // Collect data nodes that need to be clones as well
4593 int dd = dom_depth(head);
4594
4595 for (uint i = 0; i < loop->_body.size(); ++i) {
4596 Node* n = loop->_body.at(i);
4597 if (has_ctrl(n)) {
4598 Node* c = get_ctrl(n);
4599 if (wq.member(c)) {
4600 wq.push(n);
4601 }
4602 } else {
4603 set_idom(n, idom(n), dd);
4604 }
4605 }
4606
4607 // clone shared_stmt
4608 clone_loop_body(wq, old_new, nullptr);
4609
4610 Node* region_clone = old_new[region->_idx];
4611 region_clone->set_req(inner, C->top());
4612 set_idom(region, region->in(inner), dd);
4613
4614 // Prepare the outer loop
4615 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4616 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4617 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4618 set_idom(head, outer_head, dd);
4619
4620 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4621
4622 // Make one of the shared_stmt copies only reachable from stmt1, the
4623 // other only from stmt2..stmtn.
4624 Node* dom = nullptr;
4625 for (uint i = 1; i < region->req(); ++i) {
4626 if (i != inner) {
4627 _igvn.replace_input_of(region, i, C->top());
4628 }
4629 Node* in = region_clone->in(i);
4630 if (in->is_top()) {
4631 continue;
4632 }
4633 if (dom == nullptr) {
4634 dom = in;
4635 } else {
4636 dom = dom_lca(dom, in);
4637 }
4638 }
4639
4640 set_idom(region_clone, dom, dd);
4641
4642 // Set up the outer loop
4643 for (uint i = 0; i < head->outcnt(); i++) {
4644 Node* u = head->raw_out(i);
4645 if (u->is_Phi()) {
4646 Node* outer_phi = u->clone();
4647 outer_phi->set_req(0, outer_head);
4648 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4649 if (backedge == nullptr) {
4650 backedge = u->in(LoopNode::LoopBackControl);
4651 }
4652 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4653 register_new_node(outer_phi, outer_head);
4654 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4655 }
4656 }
4657
4658 // create control and data nodes for out of loop uses (including region2)
4659 Node_List worklist;
4660 uint new_counter = C->unique();
4661 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4662
4663 Node_List *split_if_set = nullptr;
4664 Node_List *split_bool_set = nullptr;
4665 Node_List *split_cex_set = nullptr;
4666 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4667 split_if_set, split_bool_set, split_cex_set);
4668
4669 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4670
4671 if (exit_test != nullptr) {
4672 float cnt = exit_test->_fcnt;
4673 if (cnt != COUNT_UNKNOWN) {
4674 exit_test->_fcnt = cnt * f;
4675 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4676 }
4677 }
4678
4679 #ifdef ASSERT
4680 if (StressDuplicateBackedge && head->is_CountedLoop()) {
4681 // The Template Assertion Predicates from the old counted loop are now at the new outer loop - clone them to
4682 // the inner counted loop and kill the old ones. We only need to do this with debug builds because
4683 // StressDuplicateBackedge is a devlop flag and false by default. Without StressDuplicateBackedge 'head' will be a
4684 // non-counted loop, and thus we have no Template Assertion Predicates above the old loop to move down.
4685 PredicateIterator predicate_iterator(outer_head->in(LoopNode::EntryControl));
4686 NodeInSingleLoopBody node_in_body(this, loop);
4687 MoveAssertionPredicatesVisitor move_assertion_predicates_visitor(head, node_in_body, this);
4688 predicate_iterator.for_each(move_assertion_predicates_visitor);
4689 }
4690 #endif // ASSERT
4691
4692 C->set_major_progress();
4693
4694 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head);
4695
4696 return true;
4697 }
4698
4699 // AutoVectorize the loop: replace scalar ops with vector ops.
4700 PhaseIdealLoop::AutoVectorizeStatus
4701 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4702 // Counted loop only
4703 if (!lpt->is_counted()) {
4704 return AutoVectorizeStatus::Impossible;
4705 }
4706
4707 // Main-loop only
4708 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4709 if (!cl->is_main_loop()) {
4710 return AutoVectorizeStatus::Impossible;
4711 }
4712
4713 VLoop vloop(lpt, false);
4714 if (!vloop.check_preconditions()) {
4715 return AutoVectorizeStatus::TriedAndFailed;
4716 }
4717
4718 // Ensure the shared data is cleared before each use
4719 vshared.clear();
4720
4721 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4722 if (!vloop_analyzer.success()) {
4723 return AutoVectorizeStatus::TriedAndFailed;
4724 }
4725
4726 SuperWord sw(vloop_analyzer);
4727 if (!sw.transform_loop()) {
4728 return AutoVectorizeStatus::TriedAndFailed;
4729 }
4730
4731 return AutoVectorizeStatus::Success;
4732 }
4733
4734 // Just before insert_pre_post_loops, we can multiversion the loop:
4735 //
4736 // multiversion_if
4737 // | |
4738 // fast_loop slow_loop
4739 //
4740 // In the fast_loop we can make speculative assumptions, and put the
4741 // conditions into the multiversion_if. If the conditions hold at runtime,
4742 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4743 // instead which does not make any of the speculative assumptions.
4744 //
4745 // Note: we only multiversion the loop if the loop does not have any
4746 // auto vectorization check Predicate. If we have that predicate,
4747 // then we can simply add the speculative assumption checks to
4748 // that Predicate. This means we do not need to duplicate the
4749 // loop - we have a smaller graph and save compile time. Should
4750 // the conditions ever fail, then we deopt / trap at the Predicate
4751 // and recompile without that Predicate. At that point we will
4752 // multiversion the loop, so that we can still have speculative
4753 // runtime checks.
4754 //
4755 // We perform the multiversioning when the loop is still in its single
4756 // iteration form, even before we insert pre and post loops. This makes
4757 // the cloning much simpler. However, this means that both the fast
4758 // and the slow loop have to be optimized independently (adding pre
4759 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4760 // we may end up not needing any speculative assumptions in the fast_loop
4761 // and then rejecting the slow_loop by constant folding the multiversion_if.
4762 //
4763 // Therefore, we "delay" the optimization of the slow_loop until we add
4764 // at least one speculative assumption for the fast_loop. If we never
4765 // add such a speculative runtime check, the OpaqueMultiversioningNode
4766 // of the multiversion_if constant folds to true after loop opts, and the
4767 // multiversion_if folds away the "delayed" slow_loop. If we add any
4768 // speculative assumption, then we notify the OpaqueMultiversioningNode
4769 // with "notify_slow_loop_that_it_can_resume_optimizations".
4770 //
4771 // Note: new runtime checks can be added to the multiversion_if with
4772 // PhaseIdealLoop::create_new_if_for_multiversion
4773 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4774 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4775 LoopNode* outer_loop = cl->skip_strip_mined();
4776 Node* entry = outer_loop->in(LoopNode::EntryControl);
4777
4778 // Check we have multiversioning enabled, and are not already multiversioned.
4779 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4780
4781 // Check that we do not have a parse-predicate where we can add the runtime checks
4782 // during auto-vectorization.
4783 const Predicates predicates(entry);
4784 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4785 if (predicate_block->has_parse_predicate()) { return; }
4786
4787 // Check node budget.
4788 uint estimate = lpt->est_loop_clone_sz(2);
4789 if (!may_require_nodes(estimate)) { return; }
4790
4791 do_multiversioning(lpt, old_new);
4792 }
4793
4794 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4795 for (uint i = 0; i < _data_nodes.size(); i++) {
4796 clone(_data_nodes[i], new_ctrl);
4797 }
4798 }
4799
4800 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4801 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4802 Node* clone = node->clone();
4803 _phase->igvn().register_new_node_with_optimizer(clone);
4804 _orig_to_new.put(node, clone);
4805 _phase->set_ctrl(clone, new_ctrl);
4806 if (node->is_CastII()) {
4807 clone->set_req(0, new_ctrl);
4808 }
4809 }
4810
4811 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4812 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4813 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4814 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4815 for (uint i = 1; i < node->req(); i++) {
4816 Node** cloned_input = _orig_to_new.get(node->in(i));
4817 if (cloned_input != nullptr) {
4818 // Input was also cloned -> rewire clone to the cloned input.
4819 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4820 }
4821 }
4822 });
4823 }
4824
4825 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4826 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4827 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4828 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4829 Node* new_ctrl) {
4830 for (uint i = 0; i < _data_nodes.size(); i++) {
4831 Node* data_node = _data_nodes[i];
4832 if (data_node->is_Opaque1()) {
4833 transform_opaque_node(transform_strategy, data_node);
4834 } else {
4835 clone(data_node, new_ctrl);
4836 }
4837 }
4838 }
4839
4840 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4841 Node* transformed_node;
4842 if (node->is_OpaqueLoopInit()) {
4843 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4844 } else {
4845 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4846 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4847 }
4848 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4849 _orig_to_new.put(node, transformed_node);
4850 }