1 /*
2 * Copyright (c) 1999, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/divnode.hpp"
35 #include "opto/inlinetypenode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "opto/superword.hpp"
45 #include "opto/vectornode.hpp"
46 #include "utilities/macros.hpp"
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
52 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
53 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
54 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
55 // so disable this for now
56 return nullptr;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
63 n->in(1) == region->as_CountedLoop()->phi()) {
64 return nullptr;
65 }
66
67 // 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 int wins = 0;
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++;
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++;
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++;
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);
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 = 0;
177 break;
178 }
179 }
180 // Too few wins?
181 if (wins <= policy) {
182 _igvn.remove_dead_node(phi);
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 return phi;
239 }
240
241 // Test whether node 'x' can move into an inner loop relative to node 'n'.
242 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
243 // BUT it can also return true and 'x' is in the outer loop
244 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
245 IdealLoopTree* n_loop_tree = get_loop(n_loop);
246 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
247 // x_loop_tree should be outer or same loop as n_loop_tree
248 return !x_loop_tree->is_member(n_loop_tree);
249 }
250
251 // Subtype checks that carry profile data don't common so look for a replacement by following edges
252 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
253 if (x->is_SubTypeCheck()) {
254 Node* in1 = x->in(1);
255 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
256 Node* u = in1->fast_out(i);
257 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
258 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
259 Node* bol = u->fast_out(j);
260 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
261 Node* iff = bol->fast_out(k);
262 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
263 // unrelated profile
264 if (iff->is_If() && is_dominator(iff, r_in)) {
265 return u;
266 }
267 }
268 }
269 }
270 }
271 }
272 return nullptr;
273 }
274
275 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
276 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
277 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
278 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
279 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
280 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
281 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
282 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
283 const Type* zero;
284 switch (n->Opcode()) {
285 case Op_DivI:
286 case Op_ModI:
287 case Op_UDivI:
288 case Op_UModI:
289 zero = TypeInt::ZERO;
290 break;
291 case Op_DivL:
292 case Op_ModL:
293 case Op_UDivL:
294 case Op_UModL:
295 zero = TypeLong::ZERO;
296 break;
297 default:
298 return false;
299 }
300
301 if (n->in(0) != nullptr) {
302 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
303 return true;
304 }
305
306 Node* divisor = n->in(2);
307 return is_divisor_loop_phi(divisor, region) &&
308 loop_phi_backedge_type_contains_zero(divisor, zero);
309 }
310
311 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
312 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
313 }
314
315 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
316 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
317 }
318
319 //------------------------------dominated_by------------------------------------
320 // Replace the dominated test with an obvious true or false. Place it on the
321 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
322 // live path up to the dominating control.
323 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
324 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
325
326 // prevdom is the dominating projection of the dominating test.
327 assert(iff->Opcode() == Op_If ||
328 iff->Opcode() == Op_CountedLoopEnd ||
329 iff->Opcode() == Op_LongCountedLoopEnd ||
330 iff->Opcode() == Op_RangeCheck ||
331 iff->Opcode() == Op_ParsePredicate,
332 "Check this code when new subtype is added");
333
334 int pop = prevdom->Opcode();
335 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
336 if (flip) {
337 if (pop == Op_IfTrue)
338 pop = Op_IfFalse;
339 else
340 pop = Op_IfTrue;
341 }
342 // 'con' is set to true or false to kill the dominated test.
343 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
344 // Hack the dominated test
345 _igvn.replace_input_of(iff, 1, con);
346
347 // If I don't have a reachable TRUE and FALSE path following the IfNode then
348 // I can assume this path reaches an infinite loop. In this case it's not
349 // important to optimize the data Nodes - either the whole compilation will
350 // be tossed or this path (and all data Nodes) will go dead.
351 if (iff->outcnt() != 2) {
352 return;
353 }
354
355 // Make control-dependent data Nodes on the live path (path that will remain
356 // once the dominated IF is removed) become control-dependent on the
357 // dominating projection.
358 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
359
360 if (dp == nullptr) {
361 return;
362 }
363
364 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
365 }
366
367 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
368 IdealLoopTree* old_loop = get_loop(source);
369
370 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
371 Node* out = source->fast_out(i); // Control-dependent node
372 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
373 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
374 assert(out->in(0) == source, "must be control dependent on source");
375 _igvn.replace_input_of(out, 0, dominator);
376 if (pin_array_access_nodes) {
377 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
378 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
379 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
380 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
381 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
382 // dominating check.
383 Node* clone = out->pin_array_access_node();
384 if (clone != nullptr) {
385 clone = _igvn.register_new_node_with_optimizer(clone, out);
386 _igvn.replace_node(out, clone);
387 out = clone;
388 }
389 }
390 set_early_ctrl(out, false);
391 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
392 if (old_loop != new_loop) {
393 if (!old_loop->_child) {
394 old_loop->_body.yank(out);
395 }
396 if (!new_loop->_child) {
397 new_loop->_body.push(out);
398 }
399 }
400 --i;
401 --imax;
402 }
403 }
404 }
405
406 //------------------------------has_local_phi_input----------------------------
407 // Return TRUE if 'n' has Phi inputs from its local block and no other
408 // block-local inputs (all non-local-phi inputs come from earlier blocks)
409 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
410 Node *n_ctrl = get_ctrl(n);
411 // See if some inputs come from a Phi in this block, or from before
412 // this block.
413 uint i;
414 for( i = 1; i < n->req(); i++ ) {
415 Node *phi = n->in(i);
416 if( phi->is_Phi() && phi->in(0) == n_ctrl )
417 break;
418 }
419 if( i >= n->req() )
420 return nullptr; // No Phi inputs; nowhere to clone thru
421
422 // Check for inputs created between 'n' and the Phi input. These
423 // must split as well; they have already been given the chance
424 // (courtesy of a post-order visit) and since they did not we must
425 // recover the 'cost' of splitting them by being very profitable
426 // when splitting 'n'. Since this is unlikely we simply give up.
427 for( i = 1; i < n->req(); i++ ) {
428 Node *m = n->in(i);
429 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
430 // We allow the special case of AddP's with no local inputs.
431 // This allows us to split-up address expressions.
432 if (m->is_AddP() &&
433 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
434 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
435 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
436 // Move the AddP up to the dominating point. That's fine because control of m's inputs
437 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
438 Node* c = find_non_split_ctrl(idom(n_ctrl));
439 if (c->is_OuterStripMinedLoop()) {
440 c->as_Loop()->verify_strip_mined(1);
441 c = c->in(LoopNode::EntryControl);
442 }
443 set_ctrl_and_loop(m, c);
444 continue;
445 }
446 return nullptr;
447 }
448 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
449 }
450
451 return n_ctrl;
452 }
453
454 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
455 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
456 assert(bt == T_INT || bt == T_LONG, "only for integers");
457 int n_op = n->Opcode();
458
459 if (n_op == Op_LShift(bt)) {
460 // Scale is loop invariant
461 Node* scale = n->in(2);
462 Node* scale_ctrl = get_ctrl(scale);
463 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
464 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
465 return nullptr;
466 }
467 const TypeInt* scale_t = scale->bottom_type()->isa_int();
468 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
469 return nullptr; // Dont bother with byte/short masking
470 }
471 // Add must vary with loop (else shift would be loop-invariant)
472 Node* add = n->in(1);
473 Node* add_ctrl = get_ctrl(add);
474 IdealLoopTree* add_loop = get_loop(add_ctrl);
475 if (n_loop != add_loop) {
476 return nullptr; // happens w/ evil ZKM loops
477 }
478
479 // Convert I-V into I+ (0-V); same for V-I
480 if (add->Opcode() == Op_Sub(bt) &&
481 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
482 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
483 Node* zero = integercon(0, bt);
484 Node* neg = SubNode::make(zero, add->in(2), bt);
485 register_new_node_with_ctrl_of(neg, add->in(2));
486 add = AddNode::make(add->in(1), neg, bt);
487 register_new_node(add, add_ctrl);
488 }
489 if (add->Opcode() != Op_Add(bt)) return nullptr;
490 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
491 // See if one add input is loop invariant
492 Node* add_var = add->in(1);
493 Node* add_var_ctrl = get_ctrl(add_var);
494 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
495 Node* add_invar = add->in(2);
496 Node* add_invar_ctrl = get_ctrl(add_invar);
497 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
498 if (add_invar_loop == n_loop) {
499 // Swap to find the invariant part
500 add_invar = add_var;
501 add_invar_ctrl = add_var_ctrl;
502 add_invar_loop = add_var_loop;
503 add_var = add->in(2);
504 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
505 return nullptr;
506 }
507 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
508 return nullptr; // No invariant part of the add?
509 }
510
511 // Yes! Reshape address expression!
512 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
513 Node* inv_scale_ctrl =
514 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
515 add_invar_ctrl : scale_ctrl;
516 register_new_node(inv_scale, inv_scale_ctrl);
517 Node* var_scale = LShiftNode::make(add_var, scale, bt);
518 register_new_node(var_scale, n_ctrl);
519 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
520 register_new_node(var_add, n_ctrl);
521 _igvn.replace_node(n, var_add);
522 return var_add;
523 }
524 return nullptr;
525 }
526
527 //------------------------------remix_address_expressions----------------------
528 // Rework addressing expressions to get the most loop-invariant stuff
529 // moved out. We'd like to do all associative operators, but it's especially
530 // important (common) to do address expressions.
531 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
532 if (!has_ctrl(n)) return nullptr;
533 Node* n_ctrl = get_ctrl(n);
534 IdealLoopTree* n_loop = get_loop(n_ctrl);
535
536 // See if 'n' mixes loop-varying and loop-invariant inputs and
537 // itself is loop-varying.
538
539 // Only interested in binary ops (and AddP)
540 if (n->req() < 3 || n->req() > 4) return nullptr;
541
542 Node* n1_ctrl = get_ctrl(n->in( 1));
543 Node* n2_ctrl = get_ctrl(n->in( 2));
544 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
545 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
546 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
547 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
548
549 // Does one of my inputs spin in a tighter loop than self?
550 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
551 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
552 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
553 return nullptr; // Leave well enough alone
554 }
555
556 // Is at least one of my inputs loop-invariant?
557 if (n1_loop == n_loop &&
558 n2_loop == n_loop &&
559 n3_loop == n_loop) {
560 return nullptr; // No loop-invariant inputs
561 }
562
563 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
564 if (res != nullptr) {
565 return res;
566 }
567 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
568 if (res != nullptr) {
569 return res;
570 }
571
572 int n_op = n->Opcode();
573 // Replace (I+V) with (V+I)
574 if (n_op == Op_AddI ||
575 n_op == Op_AddL ||
576 n_op == Op_AddF ||
577 n_op == Op_AddD ||
578 n_op == Op_MulI ||
579 n_op == Op_MulL ||
580 n_op == Op_MulF ||
581 n_op == Op_MulD) {
582 if (n2_loop == n_loop) {
583 assert(n1_loop != n_loop, "");
584 n->swap_edges(1, 2);
585 }
586 }
587
588 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
589 // but not if I2 is a constant. Skip for irreducible loops.
590 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
591 if (n2_loop == n_loop && n3_loop != n_loop) {
592 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
593 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
594 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
595 IdealLoopTree* n22loop = get_loop(n22_ctrl);
596 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
597 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
598 n23_loop == n_loop) {
599 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
600 // Stuff new AddP in the loop preheader
601 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
602 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
603 register_new_node(add2, n_ctrl);
604 _igvn.replace_node(n, add2);
605 return add2;
606 }
607 }
608 }
609
610 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
611 if (n2_loop != n_loop && n3_loop == n_loop) {
612 if (n->in(3)->Opcode() == Op_AddX) {
613 Node* V = n->in(3)->in(1);
614 Node* I = n->in(3)->in(2);
615 if (is_member(n_loop,get_ctrl(V))) {
616 } else {
617 Node *tmp = V; V = I; I = tmp;
618 }
619 if (!is_member(n_loop,get_ctrl(I))) {
620 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
621 // Stuff new AddP in the loop preheader
622 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
623 Node* add2 = new AddPNode(n->in(1), add1, V);
624 register_new_node(add2, n_ctrl);
625 _igvn.replace_node(n, add2);
626 return add2;
627 }
628 }
629 }
630 }
631
632 return nullptr;
633 }
634
635 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
636 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
637 assert(n->Opcode() == Op_AddI, "sanity");
638 Node * nn = nullptr;
639 Node * in1 = n->in(1);
640 Node * in2 = n->in(2);
641 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
642 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
643 if (loop_n->is_counted() &&
644 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
645 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
646 Matcher::match_rule_supported(Op_MulAddS2I)) {
647 Node* mul_in1 = in1->in(1);
648 Node* mul_in2 = in1->in(2);
649 Node* mul_in3 = in2->in(1);
650 Node* mul_in4 = in2->in(2);
651 if (mul_in1->Opcode() == Op_LoadS &&
652 mul_in2->Opcode() == Op_LoadS &&
653 mul_in3->Opcode() == Op_LoadS &&
654 mul_in4->Opcode() == Op_LoadS) {
655 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
656 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
657 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
658 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
659 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
660 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
661 // All nodes should be in the same counted loop.
662 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
663 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
664 Node* adr1 = mul_in1->in(MemNode::Address);
665 Node* adr2 = mul_in2->in(MemNode::Address);
666 Node* adr3 = mul_in3->in(MemNode::Address);
667 Node* adr4 = mul_in4->in(MemNode::Address);
668 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
669 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
670 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
671 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
672 register_new_node_with_ctrl_of(nn, n);
673 _igvn.replace_node(n, nn);
674 return nn;
675 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
676 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
677 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
678 register_new_node_with_ctrl_of(nn, n);
679 _igvn.replace_node(n, nn);
680 return nn;
681 }
682 }
683 }
684 }
685 }
686 }
687 return nn;
688 }
689
690 //------------------------------conditional_move-------------------------------
691 // Attempt to replace a Phi with a conditional move. We have some pretty
692 // strict profitability requirements. All Phis at the merge point must
693 // be converted, so we can remove the control flow. We need to limit the
694 // number of c-moves to a small handful. All code that was in the side-arms
695 // of the CFG diamond is now speculatively executed. This code has to be
696 // "cheap enough". We are pretty much limited to CFG diamonds that merge
697 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
698 Node *PhaseIdealLoop::conditional_move( Node *region ) {
699
700 assert(region->is_Region(), "sanity check");
701 if (region->req() != 3) return nullptr;
702
703 // Check for CFG diamond
704 Node *lp = region->in(1);
705 Node *rp = region->in(2);
706 if (!lp || !rp) return nullptr;
707 Node *lp_c = lp->in(0);
708 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
709 IfNode *iff = lp_c->as_If();
710
711 // Check for ops pinned in an arm of the diamond.
712 // Can't remove the control flow in this case
713 if (lp->outcnt() > 1) return nullptr;
714 if (rp->outcnt() > 1) return nullptr;
715
716 IdealLoopTree* r_loop = get_loop(region);
717 assert(r_loop == get_loop(iff), "sanity");
718 // Always convert to CMOVE if all results are used only outside this loop.
719 bool used_inside_loop = (r_loop == _ltree_root);
720
721 // Check profitability
722 int cost = 0;
723 int phis = 0;
724 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
725 Node *out = region->fast_out(i);
726 if (!out->is_Phi()) continue; // Ignore other control edges, etc
727 phis++;
728 PhiNode* phi = out->as_Phi();
729 BasicType bt = phi->type()->basic_type();
730 switch (bt) {
731 case T_DOUBLE:
732 case T_FLOAT:
733 if (C->use_cmove()) {
734 continue; //TODO: maybe we want to add some cost
735 }
736 cost += Matcher::float_cmove_cost(); // Could be very expensive
737 break;
738 case T_LONG: {
739 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
740 }
741 case T_INT: // These all CMOV fine
742 case T_ADDRESS: { // (RawPtr)
743 cost++;
744 break;
745 }
746 case T_NARROWOOP: // Fall through
747 case T_OBJECT: { // Base oops are OK, but not derived oops
748 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
749 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
750 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
751 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
752 // have a Phi for the base here that we convert to a CMOVE all is well
753 // and good. But if the base is dead, we'll not make a CMOVE. Later
754 // the allocator will have to produce a base by creating a CMOVE of the
755 // relevant bases. This puts the allocator in the business of
756 // manufacturing expensive instructions, generally a bad plan.
757 // Just Say No to Conditionally-Moved Derived Pointers.
758 if (tp && tp->offset() != 0)
759 return nullptr;
760 cost++;
761 break;
762 }
763 default:
764 return nullptr; // In particular, can't do memory or I/O
765 }
766 // Add in cost any speculative ops
767 for (uint j = 1; j < region->req(); j++) {
768 Node *proj = region->in(j);
769 Node *inp = phi->in(j);
770 if (inp->isa_InlineType()) {
771 // TODO 8302217 This prevents PhiNode::push_inline_types_through
772 return nullptr;
773 }
774 if (get_ctrl(inp) == proj) { // Found local op
775 cost++;
776 // Check for a chain of dependent ops; these will all become
777 // speculative in a CMOV.
778 for (uint k = 1; k < inp->req(); k++)
779 if (get_ctrl(inp->in(k)) == proj)
780 cost += ConditionalMoveLimit; // Too much speculative goo
781 }
782 }
783 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
784 // This will likely Split-If, a higher-payoff operation.
785 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
786 Node* use = phi->fast_out(k);
787 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
788 cost += ConditionalMoveLimit;
789 // Is there a use inside the loop?
790 // Note: check only basic types since CMoveP is pinned.
791 if (!used_inside_loop && is_java_primitive(bt)) {
792 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
793 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
794 used_inside_loop = true;
795 }
796 }
797 }
798 }//for
799 Node* bol = iff->in(1);
800 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
801 if (bol->is_OpaqueTemplateAssertionPredicate()) {
802 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
803 return nullptr;
804 }
805 if (bol->is_OpaqueMultiversioning()) {
806 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
807 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
808 // and will make the multiversion_if constant fold in the next IGVN round.
809 return nullptr;
810 }
811 if (!bol->is_Bool()) {
812 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
813 return nullptr;
814 }
815 int cmp_op = bol->in(1)->Opcode();
816 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
817 return nullptr;
818 }
819 // It is expensive to generate flags from a float compare.
820 // Avoid duplicated float compare.
821 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
822
823 float infrequent_prob = PROB_UNLIKELY_MAG(3);
824 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
825 if (used_inside_loop) {
826 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
827
828 // BlockLayoutByFrequency optimization moves infrequent branch
829 // from hot path. No point in CMOV'ing in such case (110 is used
830 // instead of 100 to take into account not exactness of float value).
831 if (BlockLayoutByFrequency) {
832 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
833 }
834 }
835 // Check for highly predictable branch. No point in CMOV'ing if
836 // we are going to predict accurately all the time.
837 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
838 //keep going
839 } else if (iff->_prob < infrequent_prob ||
840 iff->_prob > (1.0f - infrequent_prob))
841 return nullptr;
842
843 // --------------
844 // Now replace all Phis with CMOV's
845 Node *cmov_ctrl = iff->in(0);
846 uint flip = (lp->Opcode() == Op_IfTrue);
847 Node_List wq;
848 while (1) {
849 PhiNode* phi = nullptr;
850 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
851 Node *out = region->fast_out(i);
852 if (out->is_Phi()) {
853 phi = out->as_Phi();
854 break;
855 }
856 }
857 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
858 break;
859 }
860 // Move speculative ops
861 wq.push(phi);
862 while (wq.size() > 0) {
863 Node *n = wq.pop();
864 for (uint j = 1; j < n->req(); j++) {
865 Node* m = n->in(j);
866 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
867 set_ctrl(m, cmov_ctrl);
868 wq.push(m);
869 }
870 }
871 }
872 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
873 register_new_node(cmov, cmov_ctrl);
874 _igvn.replace_node(phi, cmov);
875 #ifndef PRODUCT
876 if (TraceLoopOpts) {
877 tty->print("CMOV ");
878 r_loop->dump_head();
879 if (Verbose) {
880 bol->in(1)->dump(1);
881 cmov->dump(1);
882 }
883 }
884 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
885 #endif
886 }
887
888 // The useless CFG diamond will fold up later; see the optimization in
889 // RegionNode::Ideal.
890 _igvn._worklist.push(region);
891
892 return iff->in(1);
893 }
894
895 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
896 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
897 Node* u = m->fast_out(i);
898 if (u->is_CFG()) {
899 if (u->is_NeverBranch()) {
900 u = u->as_NeverBranch()->proj_out(0);
901 enqueue_cfg_uses(u, wq);
902 } else {
903 wq.push(u);
904 }
905 }
906 }
907 }
908
909 // Try moving a store out of a loop, right before the loop
910 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
911 // Store has to be first in the loop body
912 IdealLoopTree *n_loop = get_loop(n_ctrl);
913 if (n->is_Store() && n_loop != _ltree_root &&
914 n_loop->is_loop() && n_loop->_head->is_Loop() &&
915 n->in(0) != nullptr) {
916 Node* address = n->in(MemNode::Address);
917 Node* value = n->in(MemNode::ValueIn);
918 Node* mem = n->in(MemNode::Memory);
919 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
920 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
921
922 // - address and value must be loop invariant
923 // - memory must be a memory Phi for the loop
924 // - Store must be the only store on this memory slice in the
925 // loop: if there's another store following this one then value
926 // written at iteration i by the second store could be overwritten
927 // at iteration i+n by the first store: it's not safe to move the
928 // first store out of the loop
929 // - nothing must observe the memory Phi: it guarantees no read
930 // before the store, we are also guaranteed the store post
931 // dominates the loop head (ignoring a possible early
932 // exit). Otherwise there would be extra Phi involved between the
933 // loop's Phi and the store.
934 // - there must be no early exit from the loop before the Store
935 // (such an exit most of the time would be an extra use of the
936 // memory Phi but sometimes is a bottom memory Phi that takes the
937 // store as input).
938
939 if (!n_loop->is_member(address_loop) &&
940 !n_loop->is_member(value_loop) &&
941 mem->is_Phi() && mem->in(0) == n_loop->_head &&
942 mem->outcnt() == 1 &&
943 mem->in(LoopNode::LoopBackControl) == n) {
944
945 assert(n_loop->_tail != nullptr, "need a tail");
946 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
947
948 // Verify that there's no early exit of the loop before the store.
949 bool ctrl_ok = false;
950 {
951 // Follow control from loop head until n, we exit the loop or
952 // we reach the tail
953 ResourceMark rm;
954 Unique_Node_List wq;
955 wq.push(n_loop->_head);
956
957 for (uint next = 0; next < wq.size(); ++next) {
958 Node *m = wq.at(next);
959 if (m == n->in(0)) {
960 ctrl_ok = true;
961 continue;
962 }
963 assert(!has_ctrl(m), "should be CFG");
964 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
965 ctrl_ok = false;
966 break;
967 }
968 enqueue_cfg_uses(m, wq);
969 if (wq.size() > 10) {
970 ctrl_ok = false;
971 break;
972 }
973 }
974 }
975 if (ctrl_ok) {
976 // move the Store
977 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
978 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
979 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
980 // Disconnect the phi now. An empty phi can confuse other
981 // optimizations in this pass of loop opts.
982 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
983 n_loop->_body.yank(mem);
984
985 set_ctrl_and_loop(n, n->in(0));
986
987 return n;
988 }
989 }
990 }
991 return nullptr;
992 }
993
994 // Try moving a store out of a loop, right after the loop
995 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
996 if (n->is_Store() && n->in(0) != nullptr) {
997 Node *n_ctrl = get_ctrl(n);
998 IdealLoopTree *n_loop = get_loop(n_ctrl);
999 // Store must be in a loop
1000 if (n_loop != _ltree_root && !n_loop->_irreducible) {
1001 Node* address = n->in(MemNode::Address);
1002 Node* value = n->in(MemNode::ValueIn);
1003 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
1004 // address must be loop invariant
1005 if (!n_loop->is_member(address_loop)) {
1006 // Store must be last on this memory slice in the loop and
1007 // nothing in the loop must observe it
1008 Node* phi = nullptr;
1009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1010 Node* u = n->fast_out(i);
1011 if (has_ctrl(u)) { // control use?
1012 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1013 if (!n_loop->is_member(u_loop)) {
1014 continue;
1015 }
1016 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1017 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1018 // multiple phis on the same slice are possible
1019 if (phi != nullptr) {
1020 return;
1021 }
1022 phi = u;
1023 continue;
1024 }
1025 }
1026 return;
1027 }
1028 if (phi != nullptr) {
1029 // Nothing in the loop before the store (next iteration)
1030 // must observe the stored value
1031 bool mem_ok = true;
1032 {
1033 ResourceMark rm;
1034 Unique_Node_List wq;
1035 wq.push(phi);
1036 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1037 Node *m = wq.at(next);
1038 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1039 Node* u = m->fast_out(i);
1040 if (u->is_Store() || u->is_Phi()) {
1041 if (u != n) {
1042 wq.push(u);
1043 mem_ok = (wq.size() <= 10);
1044 }
1045 } else {
1046 mem_ok = false;
1047 break;
1048 }
1049 }
1050 }
1051 }
1052 if (mem_ok) {
1053 // Move the store out of the loop if the LCA of all
1054 // users (except for the phi) is outside the loop.
1055 Node* hook = new Node(1);
1056 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1057 _igvn.rehash_node_delayed(phi);
1058 int count = phi->replace_edge(n, hook, &_igvn);
1059 assert(count > 0, "inconsistent phi");
1060
1061 // Compute latest point this store can go
1062 Node* lca = get_late_ctrl(n, get_ctrl(n));
1063 if (lca->is_OuterStripMinedLoop()) {
1064 lca = lca->in(LoopNode::EntryControl);
1065 }
1066 if (n_loop->is_member(get_loop(lca))) {
1067 // LCA is in the loop - bail out
1068 _igvn.replace_node(hook, n);
1069 return;
1070 }
1071 #ifdef ASSERT
1072 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1073 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1074 n_loop->_head->as_Loop()->verify_strip_mined(1);
1075 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1076 IdealLoopTree* outer_loop = get_loop(outer);
1077 assert(n_loop->_parent == outer_loop, "broken loop tree");
1078 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1079 }
1080 #endif
1081 lca = place_outside_loop(lca, n_loop);
1082 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1083 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");
1084
1085 // Move store out of the loop
1086 _igvn.replace_node(hook, n->in(MemNode::Memory));
1087 _igvn.replace_input_of(n, 0, lca);
1088 set_ctrl_and_loop(n, lca);
1089
1090 // Disconnect the phi now. An empty phi can confuse other
1091 // optimizations in this pass of loop opts..
1092 if (phi->in(LoopNode::LoopBackControl) == phi) {
1093 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1094 n_loop->_body.yank(phi);
1095 }
1096 }
1097 }
1098 }
1099 }
1100 }
1101 }
1102
1103 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1104 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1105 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1106 // flat array checks out of loops, mainly to enable loop unswitching.
1107 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1108 // Skip checks for more than one array
1109 if (n->req() > 3) {
1110 return;
1111 }
1112 Node* mem = n->in(FlatArrayCheckNode::Memory);
1113 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1114 IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1115 IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1116
1117 // Check if array is loop invariant
1118 if (!check_loop->is_member(ary_loop)) {
1119 // Walk up memory graph from the check until we leave the loop
1120 VectorSet wq;
1121 wq.set(mem->_idx);
1122 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1123 if (mem->is_Phi()) {
1124 mem = mem->in(1);
1125 } else if (mem->is_MergeMem()) {
1126 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1127 } else if (mem->is_Proj()) {
1128 mem = mem->in(0);
1129 } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1130 mem = mem->in(TypeFunc::Memory);
1131 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1132 mem = mem->in(MemNode::Memory);
1133 } else {
1134 #ifdef ASSERT
1135 mem->dump();
1136 #endif
1137 ShouldNotReachHere();
1138 }
1139 if (wq.test_set(mem->_idx)) {
1140 return;
1141 }
1142 }
1143 // Replace memory input and re-compute ctrl to move the check out of the loop
1144 _igvn.replace_input_of(n, 1, mem);
1145 set_ctrl_and_loop(n, get_early_ctrl(n));
1146 Node* bol = n->unique_out();
1147 set_ctrl_and_loop(bol, get_early_ctrl(bol));
1148 }
1149 }
1150
1151 //------------------------------split_if_with_blocks_pre-----------------------
1152 // Do the real work in a non-recursive function. Data nodes want to be
1153 // cloned in the pre-order so they can feed each other nicely.
1154 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1155 // Cloning these guys is unlikely to win
1156 int n_op = n->Opcode();
1157 if (n_op == Op_MergeMem) {
1158 return n;
1159 }
1160 if (n->is_Proj()) {
1161 return n;
1162 }
1163
1164 if (n->isa_FlatArrayCheck()) {
1165 move_flat_array_check_out_of_loop(n);
1166 return n;
1167 }
1168
1169 // Do not clone-up CmpFXXX variations, as these are always
1170 // followed by a CmpI
1171 if (n->is_Cmp()) {
1172 return n;
1173 }
1174 // Attempt to use a conditional move instead of a phi/branch
1175 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1176 Node *cmov = conditional_move( n );
1177 if (cmov) {
1178 return cmov;
1179 }
1180 }
1181 if (n->is_CFG() || n->is_LoadStore()) {
1182 return n;
1183 }
1184 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1185 if (!C->major_progress()) { // If chance of no more loop opts...
1186 _igvn._worklist.push(n); // maybe we'll remove them
1187 }
1188 return n;
1189 }
1190
1191 if (n->is_Con()) {
1192 return n; // No cloning for Con nodes
1193 }
1194
1195 Node *n_ctrl = get_ctrl(n);
1196 if (!n_ctrl) {
1197 return n; // Dead node
1198 }
1199
1200 Node* res = try_move_store_before_loop(n, n_ctrl);
1201 if (res != nullptr) {
1202 return n;
1203 }
1204
1205 // Attempt to remix address expressions for loop invariants
1206 Node *m = remix_address_expressions( n );
1207 if( m ) return m;
1208
1209 if (n_op == Op_AddI) {
1210 Node *nn = convert_add_to_muladd( n );
1211 if ( nn ) return nn;
1212 }
1213
1214 if (n->is_ConstraintCast()) {
1215 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1216 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1217 // Node control inputs don't necessarily agree with loop control info (due to
1218 // transformations happened in between), thus additional dominance check is needed
1219 // to keep loop info valid.
1220 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1221 _igvn.replace_node(n, dom_cast);
1222 return dom_cast;
1223 }
1224 }
1225
1226 // Determine if the Node has inputs from some local Phi.
1227 // Returns the block to clone thru.
1228 Node *n_blk = has_local_phi_input( n );
1229 if( !n_blk ) return n;
1230
1231 // Do not clone the trip counter through on a CountedLoop
1232 // (messes up the canonical shape).
1233 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1234 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1235 return n;
1236 }
1237 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1238 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1239 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1240 return n;
1241 }
1242
1243 // Check for having no control input; not pinned. Allow
1244 // dominating control.
1245 if (n->in(0)) {
1246 Node *dom = idom(n_blk);
1247 if (dom_lca(n->in(0), dom) != n->in(0)) {
1248 return n;
1249 }
1250 }
1251 // Policy: when is it profitable. You must get more wins than
1252 // policy before it is considered profitable. Policy is usually 0,
1253 // so 1 win is considered profitable. Big merges will require big
1254 // cloning, so get a larger policy.
1255 int policy = n_blk->req() >> 2;
1256
1257 // If the loop is a candidate for range check elimination,
1258 // delay splitting through it's phi until a later loop optimization
1259 if (n_blk->is_BaseCountedLoop()) {
1260 IdealLoopTree *lp = get_loop(n_blk);
1261 if (lp && lp->_rce_candidate) {
1262 return n;
1263 }
1264 }
1265
1266 if (must_throttle_split_if()) return n;
1267
1268 // Split 'n' through the merge point if it is profitable
1269 Node *phi = split_thru_phi( n, n_blk, policy );
1270 if (!phi) return n;
1271
1272 // Found a Phi to split thru!
1273 // Replace 'n' with the new phi
1274 _igvn.replace_node( n, phi );
1275 // Moved a load around the loop, 'en-registering' something.
1276 if (n_blk->is_Loop() && n->is_Load() &&
1277 !phi->in(LoopNode::LoopBackControl)->is_Load())
1278 C->set_major_progress();
1279
1280 return phi;
1281 }
1282
1283 static bool merge_point_too_heavy(Compile* C, Node* region) {
1284 // Bail out if the region and its phis have too many users.
1285 int weight = 0;
1286 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1287 weight += region->fast_out(i)->outcnt();
1288 }
1289 int nodes_left = C->max_node_limit() - C->live_nodes();
1290 if (weight * 8 > nodes_left) {
1291 if (PrintOpto) {
1292 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1293 }
1294 return true;
1295 } else {
1296 return false;
1297 }
1298 }
1299
1300 static bool merge_point_safe(Node* region) {
1301 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1302 // having a PhiNode input. This sidesteps the dangerous case where the split
1303 // ConvI2LNode may become TOP if the input Value() does not
1304 // overlap the ConvI2L range, leaving a node which may not dominate its
1305 // uses.
1306 // A better fix for this problem can be found in the BugTraq entry, but
1307 // expediency for Mantis demands this hack.
1308 #ifdef _LP64
1309 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1310 Node* n = region->fast_out(i);
1311 if (n->is_Phi()) {
1312 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1313 Node* m = n->fast_out(j);
1314 if (m->Opcode() == Op_ConvI2L)
1315 return false;
1316 if (m->is_CastII()) {
1317 return false;
1318 }
1319 }
1320 }
1321 }
1322 #endif
1323 return true;
1324 }
1325
1326
1327 //------------------------------place_outside_loop---------------------------------
1328 // Place some computation outside of this loop on the path to the use passed as argument
1329 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1330 Node* head = loop->_head;
1331 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1332 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1333 loop = loop->_parent;
1334 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1335 }
1336
1337 // Pick control right outside the loop
1338 for (;;) {
1339 Node* dom = idom(useblock);
1340 if (loop->is_member(get_loop(dom))) {
1341 break;
1342 }
1343 useblock = dom;
1344 }
1345 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1346 return useblock;
1347 }
1348
1349
1350 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1351 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1352 return false;
1353 }
1354 if (!n->in(0)->is_Region()) {
1355 return false;
1356 }
1357
1358 Node* region = n->in(0);
1359 Node* dom = idom(region);
1360 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1361 return false;
1362 }
1363 IfNode* dom_if = dom->as_If();
1364 Node* proj_true = dom_if->proj_out(1);
1365 Node* proj_false = dom_if->proj_out(0);
1366
1367 for (uint i = 1; i < region->req(); i++) {
1368 if (is_dominator(proj_true, region->in(i))) {
1369 continue;
1370 }
1371 if (is_dominator(proj_false, region->in(i))) {
1372 continue;
1373 }
1374 return false;
1375 }
1376
1377 return true;
1378 }
1379
1380
1381 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1382 if (must_throttle_split_if()) {
1383 return false;
1384 }
1385
1386 // Do not do 'split-if' if irreducible loops are present.
1387 if (_has_irreducible_loops) {
1388 return false;
1389 }
1390
1391 if (merge_point_too_heavy(C, n_ctrl)) {
1392 return false;
1393 }
1394
1395 // Do not do 'split-if' if some paths are dead. First do dead code
1396 // elimination and then see if its still profitable.
1397 for (uint i = 1; i < n_ctrl->req(); i++) {
1398 if (n_ctrl->in(i) == C->top()) {
1399 return false;
1400 }
1401 }
1402
1403 // If trying to do a 'Split-If' at the loop head, it is only
1404 // profitable if the cmp folds up on BOTH paths. Otherwise we
1405 // risk peeling a loop forever.
1406
1407 // CNC - Disabled for now. Requires careful handling of loop
1408 // body selection for the cloned code. Also, make sure we check
1409 // for any input path not being in the same loop as n_ctrl. For
1410 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1411 // because the alternative loop entry points won't be converted
1412 // into LoopNodes.
1413 IdealLoopTree *n_loop = get_loop(n_ctrl);
1414 for (uint j = 1; j < n_ctrl->req(); j++) {
1415 if (get_loop(n_ctrl->in(j)) != n_loop) {
1416 return false;
1417 }
1418 }
1419
1420 // Check for safety of the merge point.
1421 if (!merge_point_safe(n_ctrl)) {
1422 return false;
1423 }
1424
1425 return true;
1426 }
1427
1428 // Detect if the node is the inner strip-mined loop
1429 // Return: null if it's not the case, or the exit of outer strip-mined loop
1430 static Node* is_inner_of_stripmined_loop(const Node* out) {
1431 Node* out_le = nullptr;
1432
1433 if (out->is_CountedLoopEnd()) {
1434 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1435
1436 if (loop != nullptr && loop->is_strip_mined()) {
1437 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1438 }
1439 }
1440
1441 return out_le;
1442 }
1443
1444 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) {
1445 // If the CmpP is a subtype check for a value that has just been
1446 // loaded from an array, the subtype check guarantees the value
1447 // can't be stored in a flat array and the load of the value
1448 // happens with a flat array check then: push the type check
1449 // through the phi of the flat array check. This needs special
1450 // logic because the subtype check's input is not a phi but a
1451 // LoadKlass that must first be cloned through the phi.
1452 if (n->Opcode() != Op_CmpP) {
1453 return false;
1454 }
1455
1456 Node* klassptr = n->in(1);
1457 Node* klasscon = n->in(2);
1458
1459 if (klassptr->is_DecodeNarrowPtr()) {
1460 klassptr = klassptr->in(1);
1461 }
1462
1463 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1464 return false;
1465 }
1466
1467 if (!klasscon->is_Con()) {
1468 return false;
1469 }
1470
1471 Node* addr = klassptr->in(MemNode::Address);
1472
1473 if (!addr->is_AddP()) {
1474 return false;
1475 }
1476
1477 intptr_t offset;
1478 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1479
1480 if (obj == nullptr) {
1481 return false;
1482 }
1483
1484 assert(obj != nullptr && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?");
1485 if (obj->Opcode() == Op_CastPP) {
1486 obj = obj->in(1);
1487 }
1488
1489 if (!obj->is_Phi()) {
1490 return false;
1491 }
1492
1493 Node* region = obj->in(0);
1494
1495 Node* phi = PhiNode::make_blank(region, n->in(1));
1496 for (uint i = 1; i < region->req(); i++) {
1497 Node* in = obj->in(i);
1498 Node* ctrl = region->in(i);
1499 if (addr->in(AddPNode::Base) != obj) {
1500 Node* cast = addr->in(AddPNode::Base);
1501 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph");
1502 Node* cast_clone = cast->clone();
1503 cast_clone->set_req(0, ctrl);
1504 cast_clone->set_req(1, in);
1505 register_new_node(cast_clone, ctrl);
1506 const Type* tcast = cast_clone->Value(&_igvn);
1507 _igvn.set_type(cast_clone, tcast);
1508 cast_clone->as_Type()->set_type(tcast);
1509 in = cast_clone;
1510 }
1511 Node* addr_clone = addr->clone();
1512 addr_clone->set_req(AddPNode::Base, in);
1513 addr_clone->set_req(AddPNode::Address, in);
1514 register_new_node(addr_clone, ctrl);
1515 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1516 Node* klassptr_clone = klassptr->clone();
1517 klassptr_clone->set_req(2, addr_clone);
1518 register_new_node(klassptr_clone, ctrl);
1519 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1520 if (klassptr != n->in(1)) {
1521 Node* decode = n->in(1);
1522 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1523 Node* decode_clone = decode->clone();
1524 decode_clone->set_req(1, klassptr_clone);
1525 register_new_node(decode_clone, ctrl);
1526 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1527 klassptr_clone = decode_clone;
1528 }
1529 phi->set_req(i, klassptr_clone);
1530 }
1531 register_new_node(phi, region);
1532 Node* orig = n->in(1);
1533 _igvn.replace_input_of(n, 1, phi);
1534 split_if_with_blocks_post(n);
1535 if (n->outcnt() != 0) {
1536 _igvn.replace_input_of(n, 1, orig);
1537 _igvn.remove_dead_node(phi);
1538 }
1539 return true;
1540 }
1541
1542 //------------------------------split_if_with_blocks_post----------------------
1543 // Do the real work in a non-recursive function. CFG hackery wants to be
1544 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1545 // info.
1546 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1547
1548 if (flat_array_element_type_check(n)) {
1549 return;
1550 }
1551
1552 // Cloning Cmp through Phi's involves the split-if transform.
1553 // FastLock is not used by an If
1554 if (n->is_Cmp() && !n->is_FastLock()) {
1555 Node *n_ctrl = get_ctrl(n);
1556 // Determine if the Node has inputs from some local Phi.
1557 // Returns the block to clone thru.
1558 Node *n_blk = has_local_phi_input(n);
1559 if (n_blk != n_ctrl) {
1560 return;
1561 }
1562
1563 if (!can_split_if(n_ctrl)) {
1564 return;
1565 }
1566
1567 if (n->outcnt() != 1) {
1568 return; // Multiple bool's from 1 compare?
1569 }
1570 Node *bol = n->unique_out();
1571 assert(bol->is_Bool(), "expect a bool here");
1572 if (bol->outcnt() != 1) {
1573 return;// Multiple branches from 1 compare?
1574 }
1575 Node *iff = bol->unique_out();
1576
1577 // Check some safety conditions
1578 if (iff->is_If()) { // Classic split-if?
1579 if (iff->in(0) != n_ctrl) {
1580 return; // Compare must be in same blk as if
1581 }
1582 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1583 // Can't split CMove with different control.
1584 if (get_ctrl(iff) != n_ctrl) {
1585 return;
1586 }
1587 if (get_ctrl(iff->in(2)) == n_ctrl ||
1588 get_ctrl(iff->in(3)) == n_ctrl) {
1589 return; // Inputs not yet split-up
1590 }
1591 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1592 return; // Loop-invar test gates loop-varying CMOVE
1593 }
1594 } else {
1595 return; // some other kind of node, such as an Allocate
1596 }
1597
1598 // When is split-if profitable? Every 'win' on means some control flow
1599 // goes dead, so it's almost always a win.
1600 int policy = 0;
1601 // Split compare 'n' through the merge point if it is profitable
1602 Node *phi = split_thru_phi( n, n_ctrl, policy);
1603 if (!phi) {
1604 return;
1605 }
1606
1607 // Found a Phi to split thru!
1608 // Replace 'n' with the new phi
1609 _igvn.replace_node(n, phi);
1610
1611 // Now split the bool up thru the phi
1612 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1613 guarantee(bolphi != nullptr, "null boolean phi node");
1614
1615 _igvn.replace_node(bol, bolphi);
1616 assert(iff->in(1) == bolphi, "");
1617
1618 if (bolphi->Value(&_igvn)->singleton()) {
1619 return;
1620 }
1621
1622 // Conditional-move? Must split up now
1623 if (!iff->is_If()) {
1624 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1625 _igvn.replace_node(iff, cmovphi);
1626 return;
1627 }
1628
1629 // Now split the IF
1630 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1631 if (TraceLoopOpts) {
1632 tty->print_cr("Split-If");
1633 }
1634 do_split_if(iff);
1635 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1636 return;
1637 }
1638
1639 // Two identical ifs back to back can be merged
1640 if (try_merge_identical_ifs(n)) {
1641 return;
1642 }
1643
1644 // Check for an IF ready to split; one that has its
1645 // condition codes input coming from a Phi at the block start.
1646 int n_op = n->Opcode();
1647
1648 // Check for an IF being dominated by another IF same test
1649 if (n_op == Op_If ||
1650 n_op == Op_RangeCheck) {
1651 Node *bol = n->in(1);
1652 uint max = bol->outcnt();
1653 // Check for same test used more than once?
1654 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1655 // Search up IDOMs to see if this IF is dominated.
1656 Node* cmp = bol->in(1);
1657 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1658
1659 // Now search up IDOMs till cutoff, looking for a dominating test
1660 Node *prevdom = n;
1661 Node *dom = idom(prevdom);
1662 while (dom != cutoff) {
1663 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1664 safe_for_if_replacement(dom)) {
1665 // It's invalid to move control dependent data nodes in the inner
1666 // strip-mined loop, because:
1667 // 1) break validation of LoopNode::verify_strip_mined()
1668 // 2) move code with side-effect in strip-mined loop
1669 // Move to the exit of outer strip-mined loop in that case.
1670 Node* out_le = is_inner_of_stripmined_loop(dom);
1671 if (out_le != nullptr) {
1672 prevdom = out_le;
1673 }
1674 // Replace the dominated test with an obvious true or false.
1675 // Place it on the IGVN worklist for later cleanup.
1676 C->set_major_progress();
1677 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1678 // to prevent an array load from floating above its range check. There are three cases:
1679 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1680 // all its array accesses at that point.
1681 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1682 // accesses would start to float, since we don't pin at that point.
1683 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1684 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1685 prevdom->in(0)->Opcode() != Op_RangeCheck;
1686 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1687 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1688 return;
1689 }
1690 prevdom = dom;
1691 dom = idom(prevdom);
1692 }
1693 }
1694 }
1695
1696 try_sink_out_of_loop(n);
1697 if (C->failing()) {
1698 return;
1699 }
1700
1701 try_move_store_after_loop(n);
1702
1703 // Remove multiple allocations of the same inline type
1704 if (n->is_InlineType()) {
1705 n->as_InlineType()->remove_redundant_allocations(this);
1706 }
1707 }
1708
1709 // Transform:
1710 //
1711 // if (some_condition) {
1712 // // body 1
1713 // } else {
1714 // // body 2
1715 // }
1716 // if (some_condition) {
1717 // // body 3
1718 // } else {
1719 // // body 4
1720 // }
1721 //
1722 // into:
1723 //
1724 //
1725 // if (some_condition) {
1726 // // body 1
1727 // // body 3
1728 // } else {
1729 // // body 2
1730 // // body 4
1731 // }
1732 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1733 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1734 Node *n_ctrl = n->in(0);
1735 IfNode* dom_if = idom(n_ctrl)->as_If();
1736 if (n->in(1) != dom_if->in(1)) {
1737 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1738 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1739 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1740 _igvn.replace_input_of(n, 1, dom_if->in(1));
1741 }
1742 ProjNode* dom_proj_true = dom_if->proj_out(1);
1743 ProjNode* dom_proj_false = dom_if->proj_out(0);
1744
1745 // Now split the IF
1746 RegionNode* new_false_region;
1747 RegionNode* new_true_region;
1748 do_split_if(n, &new_false_region, &new_true_region);
1749 assert(new_false_region->req() == new_true_region->req(), "");
1750 #ifdef ASSERT
1751 for (uint i = 1; i < new_false_region->req(); ++i) {
1752 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1753 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");
1754 }
1755 #endif
1756 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1757
1758 // We now have:
1759 // if (some_condition) {
1760 // // body 1
1761 // if (some_condition) {
1762 // body3: // new_true_region
1763 // // body3
1764 // } else {
1765 // goto body4;
1766 // }
1767 // } else {
1768 // // body 2
1769 // if (some_condition) {
1770 // goto body3;
1771 // } else {
1772 // body4: // new_false_region
1773 // // body4;
1774 // }
1775 // }
1776 //
1777
1778 // clone pinned nodes thru the resulting regions
1779 push_pinned_nodes_thru_region(dom_if, new_true_region);
1780 push_pinned_nodes_thru_region(dom_if, new_false_region);
1781
1782 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1783 // 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
1784 // unrelated control dependency.
1785 for (uint i = 1; i < new_false_region->req(); i++) {
1786 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1787 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1788 } else {
1789 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1790 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1791 }
1792 }
1793 return true;
1794 }
1795 return false;
1796 }
1797
1798 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1799 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1800 Node* u = region->out(i);
1801 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1802 continue;
1803 }
1804 assert(u->in(0) == region, "not a control dependent node?");
1805 uint j = 1;
1806 for (; j < u->req(); ++j) {
1807 Node* in = u->in(j);
1808 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1809 break;
1810 }
1811 }
1812 if (j == u->req()) {
1813 Node *phi = PhiNode::make_blank(region, u);
1814 for (uint k = 1; k < region->req(); ++k) {
1815 Node* clone = u->clone();
1816 clone->set_req(0, region->in(k));
1817 register_new_node(clone, region->in(k));
1818 phi->init_req(k, clone);
1819 }
1820 register_new_node(phi, region);
1821 _igvn.replace_node(u, phi);
1822 --i;
1823 }
1824 }
1825 }
1826
1827 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1828 if (!dom->is_CountedLoopEnd()) {
1829 return true;
1830 }
1831 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1832 CountedLoopNode* cl = le->loopnode();
1833 if (cl == nullptr) {
1834 return true;
1835 }
1836 if (!cl->is_main_loop()) {
1837 return true;
1838 }
1839 if (cl->is_canonical_loop_entry() == nullptr) {
1840 return true;
1841 }
1842 // Further unrolling is possible so loop exit condition might change
1843 return false;
1844 }
1845
1846 // See if a shared loop-varying computation has no loop-varying uses.
1847 // Happens if something is only used for JVM state in uncommon trap exits,
1848 // like various versions of induction variable+offset. Clone the
1849 // computation per usage to allow it to sink out of the loop.
1850 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1851 if (has_ctrl(n) &&
1852 !n->is_Phi() &&
1853 !n->is_Bool() &&
1854 !n->is_Proj() &&
1855 !n->is_MergeMem() &&
1856 !n->is_CMove() &&
1857 !n->is_OpaqueNotNull() &&
1858 !n->is_OpaqueInitializedAssertionPredicate() &&
1859 !n->is_OpaqueTemplateAssertionPredicate() &&
1860 !n->is_Type()) {
1861 Node *n_ctrl = get_ctrl(n);
1862 IdealLoopTree *n_loop = get_loop(n_ctrl);
1863
1864 if (n->in(0) != nullptr) {
1865 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1866 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1867 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1868 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1869 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1870 Node* maybe_pinned_n = n;
1871 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1872 if (n->depends_only_on_test()) {
1873 Node* pinned_clone = n->pin_array_access_node();
1874 if (pinned_clone != nullptr) {
1875 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1876 // range check for that access. If that condition is replaced by an identical dominating one, then an
1877 // unpinned load would risk floating above its range check.
1878 register_new_node(pinned_clone, n_ctrl);
1879 maybe_pinned_n = pinned_clone;
1880 _igvn.replace_node(n, pinned_clone);
1881 }
1882 }
1883 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1884 }
1885 }
1886 if (n_loop != _ltree_root && n->outcnt() > 1) {
1887 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1888 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1889 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1890 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1891 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1892 if (n->is_Store() || n->is_LoadStore()) {
1893 assert(false, "no node with a side effect");
1894 C->record_failure("no node with a side effect");
1895 return;
1896 }
1897 Node* outer_loop_clone = nullptr;
1898 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1899 Node* u = n->last_out(j); // Clone private computation per use
1900 _igvn.rehash_node_delayed(u);
1901 Node* x = nullptr;
1902 if (n->depends_only_on_test()) {
1903 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1904 // range check for that access. If that condition is replaced by an identical dominating one, then an
1905 // unpinned load would risk floating above its range check.
1906 x = n->pin_array_access_node();
1907 }
1908 if (x == nullptr) {
1909 x = n->clone();
1910 }
1911 Node* x_ctrl = nullptr;
1912 if (u->is_Phi()) {
1913 // Replace all uses of normal nodes. Replace Phi uses
1914 // individually, so the separate Nodes can sink down
1915 // different paths.
1916 uint k = 1;
1917 while (u->in(k) != n) k++;
1918 u->set_req(k, x);
1919 // x goes next to Phi input path
1920 x_ctrl = u->in(0)->in(k);
1921 // Find control for 'x' next to use but not inside inner loops.
1922 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1923 --j;
1924 } else { // Normal use
1925 if (has_ctrl(u)) {
1926 x_ctrl = get_ctrl(u);
1927 } else {
1928 x_ctrl = u->in(0);
1929 }
1930 // Find control for 'x' next to use but not inside inner loops.
1931 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1932 // Replace all uses
1933 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1934 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1935 // anymore now that we're going to pin n as well
1936 _igvn.replace_node(u, x);
1937 --j;
1938 } else {
1939 int nb = u->replace_edge(n, x, &_igvn);
1940 j -= nb;
1941 }
1942 }
1943
1944 if (n->is_Load()) {
1945 // For loads, add a control edge to a CFG node outside of the loop
1946 // to force them to not combine and return back inside the loop
1947 // during GVN optimization (4641526).
1948 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1949
1950 IdealLoopTree* x_loop = get_loop(x_ctrl);
1951 Node* x_head = x_loop->_head;
1952 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1953 // Do not add duplicate LoadNodes to the outer strip mined loop
1954 if (outer_loop_clone != nullptr) {
1955 _igvn.replace_node(x, outer_loop_clone);
1956 continue;
1957 }
1958 outer_loop_clone = x;
1959 }
1960 x->set_req(0, x_ctrl);
1961 } else if (n->in(0) != nullptr){
1962 x->set_req(0, x_ctrl);
1963 }
1964 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1965 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1966 register_new_node(x, x_ctrl);
1967
1968 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1969 // All AddP nodes must keep the same base after sinking so:
1970 // 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,
1971 // their bases remain the same.
1972 // (see 2- below)
1973 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1974 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1975 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1976 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1977 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1978 assert(!x->is_Load(), "load should be pinned");
1979 // Use a cast node to pin clone out of loop
1980 Node* cast = nullptr;
1981 for (uint k = 0; k < x->req(); k++) {
1982 Node* in = x->in(k);
1983 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1984 const Type* in_t = _igvn.type(in);
1985 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1986 ConstraintCastNode::UnconditionalDependency, nullptr);
1987 }
1988 if (cast != nullptr) {
1989 Node* prev = _igvn.hash_find_insert(cast);
1990 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1991 cast->destruct(&_igvn);
1992 cast = prev;
1993 } else {
1994 register_new_node(cast, x_ctrl);
1995 }
1996 x->replace_edge(in, cast);
1997 // Chain of AddP nodes:
1998 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1999 if (x->is_AddP() && k == AddPNode::Base) {
2000 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
2001 }
2002 break;
2003 }
2004 }
2005 assert(cast != nullptr, "must have added a cast to pin the node");
2006 }
2007 }
2008 _igvn.remove_dead_node(n);
2009 }
2010 _dom_lca_tags_round = 0;
2011 }
2012 }
2013 }
2014
2015 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
2016 ResourceMark rm;
2017 Node_List wq;
2018 wq.push(x);
2019 while (wq.size() != 0) {
2020 Node* n = wq.pop();
2021 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2022 Node* u = n->fast_out(i);
2023 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
2024 _igvn.replace_input_of(u, AddPNode::Base, new_base);
2025 wq.push(u);
2026 }
2027 }
2028 }
2029 }
2030
2031 // Compute the early control of a node by following its inputs until we reach
2032 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
2033 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
2034 Node* early_ctrl = nullptr;
2035 ResourceMark rm;
2036 Unique_Node_List wq;
2037 wq.push(n);
2038 for (uint i = 0; i < wq.size(); i++) {
2039 Node* m = wq.at(i);
2040 Node* c = nullptr;
2041 if (m->is_CFG()) {
2042 c = m;
2043 } else if (m->pinned()) {
2044 c = m->in(0);
2045 } else {
2046 for (uint j = 0; j < m->req(); j++) {
2047 Node* in = m->in(j);
2048 if (in != nullptr) {
2049 wq.push(in);
2050 }
2051 }
2052 }
2053 if (c != nullptr) {
2054 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
2055 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
2056 early_ctrl = c;
2057 }
2058 }
2059 }
2060 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
2061 return early_ctrl;
2062 }
2063
2064 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
2065 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2066 Node* u = n->fast_out(i);
2067 if (u->is_Opaque1()) {
2068 return false; // Found loop limit, bugfix for 4677003
2069 }
2070 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
2071 // get_late_ctrl_with_anti_dep() use their own tag
2072 _dom_lca_tags_round++;
2073 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
2074
2075 if (u->is_Phi()) {
2076 for (uint j = 1; j < u->req(); ++j) {
2077 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
2078 return false;
2079 }
2080 }
2081 } else {
2082 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
2083 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
2084 return false;
2085 }
2086 }
2087 }
2088 return true;
2089 }
2090
2091 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
2092 if (n->is_Load()) {
2093 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
2094 }
2095 IdealLoopTree *u_loop = get_loop(ctrl);
2096 if (u_loop == n_loop) {
2097 return false; // Found loop-varying use
2098 }
2099 if (n_loop->is_member(u_loop)) {
2100 return false; // Found use in inner loop
2101 }
2102 // 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
2103 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
2104 // test of the pre loop above the point in the graph where it's pinned.
2105 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
2106 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
2107 if (is_dominator(pre_loop->loopexit(), ctrl)) {
2108 return false;
2109 }
2110 }
2111 return true;
2112 }
2113
2114 //------------------------------split_if_with_blocks---------------------------
2115 // Check for aggressive application of 'split-if' optimization,
2116 // using basic block level info.
2117 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
2118 Node* root = C->root();
2119 visited.set(root->_idx); // first, mark root as visited
2120 // Do pre-visit work for root
2121 Node* n = split_if_with_blocks_pre(root);
2122 uint cnt = n->outcnt();
2123 uint i = 0;
2124
2125 while (true) {
2126 // Visit all children
2127 if (i < cnt) {
2128 Node* use = n->raw_out(i);
2129 ++i;
2130 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
2131 // Now do pre-visit work for this use
2132 use = split_if_with_blocks_pre(use);
2133 nstack.push(n, i); // Save parent and next use's index.
2134 n = use; // Process all children of current use.
2135 cnt = use->outcnt();
2136 i = 0;
2137 }
2138 }
2139 else {
2140 // All of n's children have been processed, complete post-processing.
2141 if (cnt != 0 && !n->is_Con()) {
2142 assert(has_node(n), "no dead nodes");
2143 split_if_with_blocks_post(n);
2144 if (C->failing()) {
2145 return;
2146 }
2147 }
2148 if (must_throttle_split_if()) {
2149 nstack.clear();
2150 }
2151 if (nstack.is_empty()) {
2152 // Finished all nodes on stack.
2153 break;
2154 }
2155 // Get saved parent node and next use's index. Visit the rest of uses.
2156 n = nstack.node();
2157 cnt = n->outcnt();
2158 i = nstack.index();
2159 nstack.pop();
2160 }
2161 }
2162 }
2163
2164
2165 //=============================================================================
2166 //
2167 // C L O N E A L O O P B O D Y
2168 //
2169
2170 //------------------------------clone_iff--------------------------------------
2171 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2172 // "Nearly" because all Nodes have been cloned from the original in the loop,
2173 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2174 // through the Phi recursively, and return a Bool.
2175 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2176
2177 // Convert this Phi into a Phi merging Bools
2178 uint i;
2179 for (i = 1; i < phi->req(); i++) {
2180 Node* b = phi->in(i);
2181 if (b->is_Phi()) {
2182 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2183 } else {
2184 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2185 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2186 }
2187 }
2188 Node* n = phi->in(1);
2189 Node* sample_opaque = nullptr;
2190 Node *sample_bool = nullptr;
2191 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2192 sample_opaque = n;
2193 sample_bool = n->in(1);
2194 assert(sample_bool->is_Bool(), "wrong type");
2195 } else {
2196 sample_bool = n;
2197 }
2198 Node* sample_cmp = sample_bool->in(1);
2199 const Type* t = Type::TOP;
2200 const TypePtr* at = nullptr;
2201 if (sample_cmp->is_FlatArrayCheck()) {
2202 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2203 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2204 t = Type::MEMORY;
2205 at = TypeRawPtr::BOTTOM;
2206 }
2207
2208 // Make Phis to merge the Cmp's inputs.
2209 PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2210 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2211 for (i = 1; i < phi->req(); i++) {
2212 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2213 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2214 phi1->set_req(i, n1);
2215 phi2->set_req(i, n2);
2216 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2217 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2218 }
2219 // See if these Phis have been made before.
2220 // Register with optimizer
2221 Node *hit1 = _igvn.hash_find_insert(phi1);
2222 if (hit1) { // Hit, toss just made Phi
2223 _igvn.remove_dead_node(phi1); // Remove new phi
2224 assert(hit1->is_Phi(), "" );
2225 phi1 = (PhiNode*)hit1; // Use existing phi
2226 } else { // Miss
2227 _igvn.register_new_node_with_optimizer(phi1);
2228 }
2229 Node *hit2 = _igvn.hash_find_insert(phi2);
2230 if (hit2) { // Hit, toss just made Phi
2231 _igvn.remove_dead_node(phi2); // Remove new phi
2232 assert(hit2->is_Phi(), "" );
2233 phi2 = (PhiNode*)hit2; // Use existing phi
2234 } else { // Miss
2235 _igvn.register_new_node_with_optimizer(phi2);
2236 }
2237 // Register Phis with loop/block info
2238 set_ctrl(phi1, phi->in(0));
2239 set_ctrl(phi2, phi->in(0));
2240 // Make a new Cmp
2241 Node *cmp = sample_cmp->clone();
2242 cmp->set_req(1, phi1);
2243 cmp->set_req(2, phi2);
2244 _igvn.register_new_node_with_optimizer(cmp);
2245 set_ctrl(cmp, phi->in(0));
2246
2247 // Make a new Bool
2248 Node *b = sample_bool->clone();
2249 b->set_req(1,cmp);
2250 _igvn.register_new_node_with_optimizer(b);
2251 set_ctrl(b, phi->in(0));
2252
2253 if (sample_opaque != nullptr) {
2254 Node* opaque = sample_opaque->clone();
2255 opaque->set_req(1, b);
2256 _igvn.register_new_node_with_optimizer(opaque);
2257 set_ctrl(opaque, phi->in(0));
2258 return opaque;
2259 }
2260
2261 assert(b->is_Bool(), "");
2262 return b;
2263 }
2264
2265 //------------------------------clone_bool-------------------------------------
2266 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2267 // "Nearly" because all Nodes have been cloned from the original in the loop,
2268 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2269 // through the Phi recursively, and return a Bool.
2270 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2271 uint i;
2272 // Convert this Phi into a Phi merging Bools
2273 for( i = 1; i < phi->req(); i++ ) {
2274 Node *b = phi->in(i);
2275 if( b->is_Phi() ) {
2276 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2277 } else {
2278 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2279 }
2280 }
2281
2282 Node *sample_cmp = phi->in(1);
2283
2284 // Make Phis to merge the Cmp's inputs.
2285 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2286 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2287 for( uint j = 1; j < phi->req(); j++ ) {
2288 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2289 Node *n1, *n2;
2290 if( cmp_top->is_Cmp() ) {
2291 n1 = cmp_top->in(1);
2292 n2 = cmp_top->in(2);
2293 } else {
2294 n1 = n2 = cmp_top;
2295 }
2296 phi1->set_req( j, n1 );
2297 phi2->set_req( j, n2 );
2298 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2299 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2300 }
2301
2302 // See if these Phis have been made before.
2303 // Register with optimizer
2304 Node *hit1 = _igvn.hash_find_insert(phi1);
2305 if( hit1 ) { // Hit, toss just made Phi
2306 _igvn.remove_dead_node(phi1); // Remove new phi
2307 assert( hit1->is_Phi(), "" );
2308 phi1 = (PhiNode*)hit1; // Use existing phi
2309 } else { // Miss
2310 _igvn.register_new_node_with_optimizer(phi1);
2311 }
2312 Node *hit2 = _igvn.hash_find_insert(phi2);
2313 if( hit2 ) { // Hit, toss just made Phi
2314 _igvn.remove_dead_node(phi2); // Remove new phi
2315 assert( hit2->is_Phi(), "" );
2316 phi2 = (PhiNode*)hit2; // Use existing phi
2317 } else { // Miss
2318 _igvn.register_new_node_with_optimizer(phi2);
2319 }
2320 // Register Phis with loop/block info
2321 set_ctrl(phi1, phi->in(0));
2322 set_ctrl(phi2, phi->in(0));
2323 // Make a new Cmp
2324 Node *cmp = sample_cmp->clone();
2325 cmp->set_req( 1, phi1 );
2326 cmp->set_req( 2, phi2 );
2327 _igvn.register_new_node_with_optimizer(cmp);
2328 set_ctrl(cmp, phi->in(0));
2329
2330 assert( cmp->is_Cmp(), "" );
2331 return (CmpNode*)cmp;
2332 }
2333
2334 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2335 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2336 Node_List*& split_if_set, Node_List*& split_bool_set,
2337 Node_List*& split_cex_set, Node_List& worklist,
2338 uint new_counter, CloneLoopMode mode) {
2339 Node* nnn = old_new[old->_idx];
2340 // Copy uses to a worklist, so I can munge the def-use info
2341 // with impunity.
2342 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2343 worklist.push(old->fast_out(j));
2344
2345 while( worklist.size() ) {
2346 Node *use = worklist.pop();
2347 if (!has_node(use)) continue; // Ignore dead nodes
2348 if (use->in(0) == C->top()) continue;
2349 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2350 // Check for data-use outside of loop - at least one of OLD or USE
2351 // must not be a CFG node.
2352 #ifdef ASSERT
2353 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2354 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2355 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2356 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2357 }
2358 #endif
2359 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2360
2361 // If the Data use is an IF, that means we have an IF outside the
2362 // loop that is switching on a condition that is set inside the
2363 // loop. Happens if people set a loop-exit flag; then test the flag
2364 // in the loop to break the loop, then test is again outside the
2365 // loop to determine which way the loop exited.
2366 //
2367 // For several uses we need to make sure that there is no phi between,
2368 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2369 // to avoid such a phi in between.
2370 // For example, it is unexpected that there is a Phi between an
2371 // AllocateArray node and its ValidLengthTest input that could cause
2372 // split if to break.
2373 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2374 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2375 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2376 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2377 // Since this code is highly unlikely, we lazily build the worklist
2378 // of such Nodes to go split.
2379 if (!split_if_set) {
2380 split_if_set = new Node_List();
2381 }
2382 split_if_set->push(use);
2383 }
2384 if (use->is_Bool()) {
2385 if (!split_bool_set) {
2386 split_bool_set = new Node_List();
2387 }
2388 split_bool_set->push(use);
2389 }
2390 if (use->Opcode() == Op_CreateEx) {
2391 if (!split_cex_set) {
2392 split_cex_set = new Node_List();
2393 }
2394 split_cex_set->push(use);
2395 }
2396
2397
2398 // Get "block" use is in
2399 uint idx = 0;
2400 while( use->in(idx) != old ) idx++;
2401 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2402 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2403 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2404 ? prev->in(2)
2405 : idom(prev);
2406 if( use->is_Phi() ) // Phi use is in prior block
2407 cfg = prev->in(idx); // NOT in block of Phi itself
2408 if (cfg->is_top()) { // Use is dead?
2409 _igvn.replace_input_of(use, idx, C->top());
2410 continue;
2411 }
2412
2413 // If use is referenced through control edge... (idx == 0)
2414 if (mode == IgnoreStripMined && idx == 0) {
2415 LoopNode *head = loop->_head->as_Loop();
2416 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2417 // That node is outside the inner loop, leave it outside the
2418 // outer loop as well to not confuse verification code.
2419 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2420 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2421 continue;
2422 }
2423 }
2424
2425 while(!outer_loop->is_member(get_loop(cfg))) {
2426 prev = cfg;
2427 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2428 }
2429 // If the use occurs after merging several exits from the loop, then
2430 // old value must have dominated all those exits. Since the same old
2431 // value was used on all those exits we did not need a Phi at this
2432 // merge point. NOW we do need a Phi here. Each loop exit value
2433 // is now merged with the peeled body exit; each exit gets its own
2434 // private Phi and those Phis need to be merged here.
2435 Node *phi;
2436 if( prev->is_Region() ) {
2437 if( idx == 0 ) { // Updating control edge?
2438 phi = prev; // Just use existing control
2439 } else { // Else need a new Phi
2440 phi = PhiNode::make( prev, old );
2441 // Now recursively fix up the new uses of old!
2442 for( uint i = 1; i < prev->req(); i++ ) {
2443 worklist.push(phi); // Onto worklist once for each 'old' input
2444 }
2445 }
2446 } else {
2447 // Get new RegionNode merging old and new loop exits
2448 prev = old_new[prev->_idx];
2449 assert( prev, "just made this in step 7" );
2450 if( idx == 0) { // Updating control edge?
2451 phi = prev; // Just use existing control
2452 } else { // Else need a new Phi
2453 // Make a new Phi merging data values properly
2454 phi = PhiNode::make( prev, old );
2455 phi->set_req( 1, nnn );
2456 }
2457 }
2458 // If inserting a new Phi, check for prior hits
2459 if( idx != 0 ) {
2460 Node *hit = _igvn.hash_find_insert(phi);
2461 if( hit == nullptr ) {
2462 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2463 } else { // or
2464 // Remove the new phi from the graph and use the hit
2465 _igvn.remove_dead_node(phi);
2466 phi = hit; // Use existing phi
2467 }
2468 set_ctrl(phi, prev);
2469 }
2470 // Make 'use' use the Phi instead of the old loop body exit value
2471 assert(use->in(idx) == old, "old is still input of use");
2472 // We notify all uses of old, including use, and the indirect uses,
2473 // that may now be optimized because we have replaced old with phi.
2474 _igvn.add_users_to_worklist(old);
2475 if (idx == 0 &&
2476 use->depends_only_on_test()) {
2477 Node* pinned_clone = use->pin_array_access_node();
2478 if (pinned_clone != nullptr) {
2479 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2480 // into the region is left, an array load could become dependent on a condition that's not a range check for
2481 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2482 // floating above its range check.
2483 pinned_clone->set_req(0, phi);
2484 register_new_node_with_ctrl_of(pinned_clone, use);
2485 _igvn.replace_node(use, pinned_clone);
2486 continue;
2487 }
2488 }
2489 _igvn.replace_input_of(use, idx, phi);
2490 if( use->_idx >= new_counter ) { // If updating new phis
2491 // Not needed for correctness, but prevents a weak assert
2492 // in AddPNode from tripping (when we end up with different
2493 // base & derived Phis that will become the same after
2494 // IGVN does CSE).
2495 Node *hit = _igvn.hash_find_insert(use);
2496 if( hit ) // Go ahead and re-hash for hits.
2497 _igvn.replace_node( use, hit );
2498 }
2499 }
2500 }
2501 }
2502
2503 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2504 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2505 bool check_old_new) {
2506 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2507 Node* u = n->fast_out(j);
2508 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2509 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2510 Node* c = phase->get_ctrl(u);
2511 IdealLoopTree* u_loop = phase->get_loop(c);
2512 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2513 if (!loop->is_member(u_loop)) {
2514 if (outer_loop->is_member(u_loop)) {
2515 wq.push(u);
2516 } else {
2517 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2518 // the outer loop too
2519 Node* u_c = u->in(0);
2520 if (u_c != nullptr) {
2521 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2522 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2523 wq.push(u);
2524 }
2525 }
2526 }
2527 }
2528 }
2529 }
2530 }
2531
2532 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2533 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2534 Node_List& extra_data_nodes) {
2535 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2536 CountedLoopNode* cl = head->as_CountedLoop();
2537 Node* l = cl->outer_loop();
2538 Node* tail = cl->outer_loop_tail();
2539 IfNode* le = cl->outer_loop_end();
2540 Node* sfpt = cl->outer_safepoint();
2541 CountedLoopEndNode* cle = cl->loopexit();
2542 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2543 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2544 Node* cle_out = cle->proj_out(false);
2545
2546 Node* new_sfpt = nullptr;
2547 Node* new_cle_out = cle_out->clone();
2548 old_new.map(cle_out->_idx, new_cle_out);
2549 if (mode == CloneIncludesStripMined) {
2550 // clone outer loop body
2551 Node* new_l = l->clone();
2552 Node* new_tail = tail->clone();
2553 IfNode* new_le = le->clone()->as_If();
2554 new_sfpt = sfpt->clone();
2555
2556 set_loop(new_l, outer_loop->_parent);
2557 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2558 set_loop(new_cle_out, outer_loop->_parent);
2559 set_idom(new_cle_out, new_cle, dd);
2560 set_loop(new_sfpt, outer_loop->_parent);
2561 set_idom(new_sfpt, new_cle_out, dd);
2562 set_loop(new_le, outer_loop->_parent);
2563 set_idom(new_le, new_sfpt, dd);
2564 set_loop(new_tail, outer_loop->_parent);
2565 set_idom(new_tail, new_le, dd);
2566 set_idom(new_cl, new_l, dd);
2567
2568 old_new.map(l->_idx, new_l);
2569 old_new.map(tail->_idx, new_tail);
2570 old_new.map(le->_idx, new_le);
2571 old_new.map(sfpt->_idx, new_sfpt);
2572
2573 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2574 new_l->set_req(0, new_l);
2575 new_tail->set_req(0, new_le);
2576 new_le->set_req(0, new_sfpt);
2577 new_sfpt->set_req(0, new_cle_out);
2578 new_cle_out->set_req(0, new_cle);
2579 new_cl->set_req(LoopNode::EntryControl, new_l);
2580
2581 _igvn.register_new_node_with_optimizer(new_l);
2582 _igvn.register_new_node_with_optimizer(new_tail);
2583 _igvn.register_new_node_with_optimizer(new_le);
2584 } else {
2585 Node *newhead = old_new[loop->_head->_idx];
2586 newhead->as_Loop()->clear_strip_mined();
2587 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2588 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2589 }
2590 // Look at data node that were assigned a control in the outer
2591 // loop: they are kept in the outer loop by the safepoint so start
2592 // from the safepoint node's inputs.
2593 IdealLoopTree* outer_loop = get_loop(l);
2594 Node_Stack stack(2);
2595 stack.push(sfpt, 1);
2596 uint new_counter = C->unique();
2597 while (stack.size() > 0) {
2598 Node* n = stack.node();
2599 uint i = stack.index();
2600 while (i < n->req() &&
2601 (n->in(i) == nullptr ||
2602 !has_ctrl(n->in(i)) ||
2603 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2604 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2605 i++;
2606 }
2607 if (i < n->req()) {
2608 stack.set_index(i+1);
2609 stack.push(n->in(i), 0);
2610 } else {
2611 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2612 Node* m = n == sfpt ? new_sfpt : n->clone();
2613 if (m != nullptr) {
2614 for (uint i = 0; i < n->req(); i++) {
2615 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2616 m->set_req(i, old_new[m->in(i)->_idx]);
2617 }
2618 }
2619 } else {
2620 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2621 }
2622 if (n != sfpt) {
2623 extra_data_nodes.push(n);
2624 _igvn.register_new_node_with_optimizer(m);
2625 assert(get_ctrl(n) == cle_out, "what other control?");
2626 set_ctrl(m, new_cle_out);
2627 old_new.map(n->_idx, m);
2628 }
2629 stack.pop();
2630 }
2631 }
2632 if (mode == CloneIncludesStripMined) {
2633 _igvn.register_new_node_with_optimizer(new_sfpt);
2634 _igvn.register_new_node_with_optimizer(new_cle_out);
2635 }
2636 // Some other transformation may have pessimistically assigned some
2637 // data nodes to the outer loop. Set their control so they are out
2638 // of the outer loop.
2639 ResourceMark rm;
2640 Unique_Node_List wq;
2641 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2642 Node* old = extra_data_nodes.at(i);
2643 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2644 }
2645
2646 for (uint i = 0; i < loop->_body.size(); i++) {
2647 Node* old = loop->_body.at(i);
2648 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2649 }
2650
2651 Node* inner_out = sfpt->in(0);
2652 if (inner_out->outcnt() > 1) {
2653 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2654 }
2655
2656 Node* new_ctrl = cl->outer_loop_exit();
2657 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2658 for (uint i = 0; i < wq.size(); i++) {
2659 Node* n = wq.at(i);
2660 set_ctrl(n, new_ctrl);
2661 if (n->in(0) != nullptr) {
2662 _igvn.replace_input_of(n, 0, new_ctrl);
2663 }
2664 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2665 }
2666 } else {
2667 Node *newhead = old_new[loop->_head->_idx];
2668 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2669 }
2670 }
2671
2672 //------------------------------clone_loop-------------------------------------
2673 //
2674 // C L O N E A L O O P B O D Y
2675 //
2676 // This is the basic building block of the loop optimizations. It clones an
2677 // entire loop body. It makes an old_new loop body mapping; with this mapping
2678 // you can find the new-loop equivalent to an old-loop node. All new-loop
2679 // nodes are exactly equal to their old-loop counterparts, all edges are the
2680 // same. All exits from the old-loop now have a RegionNode that merges the
2681 // equivalent new-loop path. This is true even for the normal "loop-exit"
2682 // condition. All uses of loop-invariant old-loop values now come from (one
2683 // or more) Phis that merge their new-loop equivalents.
2684 //
2685 // This operation leaves the graph in an illegal state: there are two valid
2686 // control edges coming from the loop pre-header to both loop bodies. I'll
2687 // definitely have to hack the graph after running this transform.
2688 //
2689 // From this building block I will further edit edges to perform loop peeling
2690 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2691 //
2692 // Parameter side_by_size_idom:
2693 // When side_by_size_idom is null, the dominator tree is constructed for
2694 // the clone loop to dominate the original. Used in construction of
2695 // pre-main-post loop sequence.
2696 // When nonnull, the clone and original are side-by-side, both are
2697 // dominated by the side_by_side_idom node. Used in construction of
2698 // unswitched loops.
2699 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2700 CloneLoopMode mode, Node* side_by_side_idom) {
2701
2702 LoopNode* head = loop->_head->as_Loop();
2703 head->verify_strip_mined(1);
2704
2705 if (C->do_vector_loop() && PrintOpto) {
2706 const char* mname = C->method()->name()->as_quoted_ascii();
2707 if (mname != nullptr) {
2708 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2709 }
2710 }
2711
2712 CloneMap& cm = C->clone_map();
2713 if (C->do_vector_loop()) {
2714 cm.set_clone_idx(cm.max_gen()+1);
2715 #ifndef PRODUCT
2716 if (PrintOpto) {
2717 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2718 loop->dump_head();
2719 }
2720 #endif
2721 }
2722
2723 // Step 1: Clone the loop body. Make the old->new mapping.
2724 clone_loop_body(loop->_body, old_new, &cm);
2725
2726 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2727
2728 // Step 2: Fix the edges in the new body. If the old input is outside the
2729 // loop use it. If the old input is INside the loop, use the corresponding
2730 // new node instead.
2731 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2732
2733 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2734 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2735
2736 // Step 3: Now fix control uses. Loop varying control uses have already
2737 // been fixed up (as part of all input edges in Step 2). Loop invariant
2738 // control uses must be either an IfFalse or an IfTrue. Make a merge
2739 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2740 // refer to this.
2741 Node_List worklist;
2742 uint new_counter = C->unique();
2743 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2744
2745 // Step 4: If loop-invariant use is not control, it must be dominated by a
2746 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2747 // there if needed. Make a Phi there merging old and new used values.
2748 Node_List *split_if_set = nullptr;
2749 Node_List *split_bool_set = nullptr;
2750 Node_List *split_cex_set = nullptr;
2751 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2752
2753 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2754 Node* old = extra_data_nodes.at(i);
2755 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2756 split_bool_set, split_cex_set, worklist, new_counter,
2757 mode);
2758 }
2759
2760 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2761 // the loop uses a condition set in the loop. The original IF probably
2762 // takes control from one or more OLD Regions (which in turn get from NEW
2763 // Regions). In any case, there will be a set of Phis for each merge point
2764 // from the IF up to where the original BOOL def exists the loop.
2765 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2766
2767 }
2768
2769 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2770 if (split_if_set) {
2771 while (split_if_set->size()) {
2772 Node *iff = split_if_set->pop();
2773 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2774 if (iff->in(input)->is_Phi()) {
2775 Node *b = clone_iff(iff->in(input)->as_Phi());
2776 _igvn.replace_input_of(iff, input, b);
2777 }
2778 }
2779 }
2780 if (split_bool_set) {
2781 while (split_bool_set->size()) {
2782 Node *b = split_bool_set->pop();
2783 Node *phi = b->in(1);
2784 assert(phi->is_Phi(), "");
2785 CmpNode *cmp = clone_bool((PhiNode*) phi);
2786 _igvn.replace_input_of(b, 1, cmp);
2787 }
2788 }
2789 if (split_cex_set) {
2790 while (split_cex_set->size()) {
2791 Node *b = split_cex_set->pop();
2792 assert(b->in(0)->is_Region(), "");
2793 assert(b->in(1)->is_Phi(), "");
2794 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2795 split_up(b, b->in(0), nullptr);
2796 }
2797 }
2798 }
2799
2800 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2801 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2802 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2803 for(uint i = 0; i < body.size(); i++ ) {
2804 Node* old = body.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
2811 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2812 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2813 LoopNode* head = loop->_head->as_Loop();
2814 for(uint i = 0; i < body.size(); i++ ) {
2815 Node* old = body.at(i);
2816 if( !old->is_CFG() ) continue;
2817
2818 // Copy uses to a worklist, so I can munge the def-use info
2819 // with impunity.
2820 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2821 worklist.push(old->fast_out(j));
2822 }
2823
2824 while (worklist.size()) { // Visit all uses
2825 Node *use = worklist.pop();
2826 if (!has_node(use)) continue; // Ignore dead nodes
2827 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2828 if (!loop->is_member(use_loop) && use->is_CFG()) {
2829 // Both OLD and USE are CFG nodes here.
2830 assert(use->is_Proj(), "" );
2831 Node* nnn = old_new[old->_idx];
2832
2833 Node* newuse = nullptr;
2834 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2835 CountedLoopNode* cl = head->as_CountedLoop();
2836 CountedLoopEndNode* cle = cl->loopexit();
2837 Node* cle_out = cle->proj_out_or_null(false);
2838 if (use == cle_out) {
2839 IfNode* le = cl->outer_loop_end();
2840 use = le->proj_out(false);
2841 use_loop = get_loop(use);
2842 if (mode == CloneIncludesStripMined) {
2843 nnn = old_new[le->_idx];
2844 } else {
2845 newuse = old_new[cle_out->_idx];
2846 }
2847 }
2848 }
2849 if (newuse == nullptr) {
2850 newuse = use->clone();
2851 }
2852
2853 // Clone the loop exit control projection
2854 if (C->do_vector_loop() && cm != nullptr) {
2855 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2856 }
2857 newuse->set_req(0,nnn);
2858 _igvn.register_new_node_with_optimizer(newuse);
2859 set_loop(newuse, use_loop);
2860 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2861
2862 // We need a Region to merge the exit from the peeled body and the
2863 // exit from the old loop body.
2864 RegionNode *r = new RegionNode(3);
2865 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2866 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2867
2868 // The original user of 'use' uses 'r' instead.
2869 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2870 Node* useuse = use->last_out(l);
2871 _igvn.rehash_node_delayed(useuse);
2872 uint uses_found = 0;
2873 if (useuse->in(0) == use) {
2874 useuse->set_req(0, r);
2875 uses_found++;
2876 if (useuse->is_CFG()) {
2877 // This is not a dom_depth > dd_r because when new
2878 // control flow is constructed by a loop opt, a node and
2879 // its dominator can end up at the same dom_depth
2880 assert(dom_depth(useuse) >= dd_r, "");
2881 set_idom(useuse, r, dom_depth(useuse));
2882 }
2883 }
2884 for (uint k = 1; k < useuse->req(); k++) {
2885 if( useuse->in(k) == use ) {
2886 useuse->set_req(k, r);
2887 uses_found++;
2888 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2889 // This is not a dom_depth > dd_r because when new
2890 // control flow is constructed by a loop opt, a node
2891 // and its dominator can end up at the same dom_depth
2892 assert(dom_depth(useuse) >= dd_r , "");
2893 set_idom(useuse, r, dom_depth(useuse));
2894 }
2895 }
2896 }
2897 l -= uses_found; // we deleted 1 or more copies of this edge
2898 }
2899
2900 assert(use->is_Proj(), "loop exit should be projection");
2901 // lazy_replace() below moves all nodes that are:
2902 // - control dependent on the loop exit or
2903 // - have control set to the loop exit
2904 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2905 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2906 // becomes dead and is "replaced" by the region.
2907 Node* use_clone = use->clone();
2908 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2909 // Now finish up 'r'
2910 r->set_req(1, newuse);
2911 r->set_req(2, use_clone);
2912 _igvn.register_new_node_with_optimizer(r);
2913 set_loop(r, use_loop);
2914 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2915 lazy_replace(use, r);
2916 // Map the (cloned) old use to the new merge point
2917 old_new.map(use_clone->_idx, r);
2918 } // End of if a loop-exit test
2919 }
2920 }
2921 }
2922
2923 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2924 IdealLoopTree* parent, bool partial) {
2925 for(uint i = 0; i < body.size(); i++ ) {
2926 Node *old = body.at(i);
2927 Node *nnn = old_new[old->_idx];
2928 // Fix CFG/Loop controlling the new node
2929 if (has_ctrl(old)) {
2930 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2931 } else {
2932 set_loop(nnn, parent);
2933 if (old->outcnt() > 0) {
2934 Node* dom = idom(old);
2935 if (old_new[dom->_idx] != nullptr) {
2936 dom = old_new[dom->_idx];
2937 set_idom(nnn, dom, dd );
2938 }
2939 }
2940 }
2941 // Correct edges to the new node
2942 for (uint j = 0; j < nnn->req(); j++) {
2943 Node *n = nnn->in(j);
2944 if (n != nullptr) {
2945 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2946 if (loop->is_member(old_in_loop)) {
2947 if (old_new[n->_idx] != nullptr) {
2948 nnn->set_req(j, old_new[n->_idx]);
2949 } else {
2950 assert(!body.contains(n), "");
2951 assert(partial, "node not cloned");
2952 }
2953 }
2954 }
2955 }
2956 _igvn.hash_find_insert(nnn);
2957 }
2958 }
2959
2960 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2961 for (uint i = 0; i < body.size(); i++) {
2962 Node* old = body.at(i);
2963 Node* nnn = old->clone();
2964 old_new.map(old->_idx, nnn);
2965 if (C->do_vector_loop() && cm != nullptr) {
2966 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2967 }
2968 _igvn.register_new_node_with_optimizer(nnn);
2969 }
2970 }
2971
2972
2973 //---------------------- stride_of_possible_iv -------------------------------------
2974 // Looks for an iff/bool/comp with one operand of the compare
2975 // being a cycle involving an add and a phi,
2976 // with an optional truncation (left-shift followed by a right-shift)
2977 // of the add. Returns zero if not an iv.
2978 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2979 Node* trunc1 = nullptr;
2980 Node* trunc2 = nullptr;
2981 const TypeInteger* ttype = nullptr;
2982 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2983 return 0;
2984 }
2985 BoolNode* bl = iff->in(1)->as_Bool();
2986 Node* cmp = bl->in(1);
2987 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2988 return 0;
2989 }
2990 // Must have an invariant operand
2991 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2992 return 0;
2993 }
2994 Node* add2 = nullptr;
2995 Node* cmp1 = cmp->in(1);
2996 if (cmp1->is_Phi()) {
2997 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2998 Node* phi = cmp1;
2999 for (uint i = 1; i < phi->req(); i++) {
3000 Node* in = phi->in(i);
3001 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
3002 &trunc1, &trunc2, &ttype, T_INT);
3003 if (add && add->in(1) == phi) {
3004 add2 = add->in(2);
3005 break;
3006 }
3007 }
3008 } else {
3009 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
3010 Node* addtrunc = cmp1;
3011 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
3012 &trunc1, &trunc2, &ttype, T_INT);
3013 if (add && add->in(1)->is_Phi()) {
3014 Node* phi = add->in(1);
3015 for (uint i = 1; i < phi->req(); i++) {
3016 if (phi->in(i) == addtrunc) {
3017 add2 = add->in(2);
3018 break;
3019 }
3020 }
3021 }
3022 }
3023 if (add2 != nullptr) {
3024 const TypeInt* add2t = _igvn.type(add2)->is_int();
3025 if (add2t->is_con()) {
3026 return add2t->get_con();
3027 }
3028 }
3029 return 0;
3030 }
3031
3032
3033 //---------------------- stay_in_loop -------------------------------------
3034 // Return the (unique) control output node that's in the loop (if it exists.)
3035 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
3036 Node* unique = nullptr;
3037 if (!n) return nullptr;
3038 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3039 Node* use = n->fast_out(i);
3040 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
3041 if (unique != nullptr) {
3042 return nullptr;
3043 }
3044 unique = use;
3045 }
3046 }
3047 return unique;
3048 }
3049
3050 //------------------------------ register_node -------------------------------------
3051 // Utility to register node "n" with PhaseIdealLoop
3052 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
3053 _igvn.register_new_node_with_optimizer(n);
3054 loop->_body.push(n);
3055 if (n->is_CFG()) {
3056 set_loop(n, loop);
3057 set_idom(n, pred, ddepth);
3058 } else {
3059 set_ctrl(n, pred);
3060 }
3061 }
3062
3063 //------------------------------ proj_clone -------------------------------------
3064 // Utility to create an if-projection
3065 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
3066 ProjNode* c = p->clone()->as_Proj();
3067 c->set_req(0, iff);
3068 return c;
3069 }
3070
3071 //------------------------------ short_circuit_if -------------------------------------
3072 // Force the iff control output to be the live_proj
3073 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
3074 guarantee(live_proj != nullptr, "null projection");
3075 int proj_con = live_proj->_con;
3076 assert(proj_con == 0 || proj_con == 1, "false or true projection");
3077 Node* con = intcon(proj_con);
3078 if (iff) {
3079 iff->set_req(1, con);
3080 }
3081 return con;
3082 }
3083
3084 //------------------------------ insert_if_before_proj -------------------------------------
3085 // Insert a new if before an if projection (* - new node)
3086 //
3087 // before
3088 // if(test)
3089 // / \
3090 // v v
3091 // other-proj proj (arg)
3092 //
3093 // after
3094 // if(test)
3095 // / \
3096 // / v
3097 // | * proj-clone
3098 // v |
3099 // other-proj v
3100 // * new_if(relop(cmp[IU](left,right)))
3101 // / \
3102 // v v
3103 // * new-proj proj
3104 // (returned)
3105 //
3106 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
3107 IfNode* iff = proj->in(0)->as_If();
3108 IdealLoopTree *loop = get_loop(proj);
3109 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3110 uint ddepth = dom_depth(proj);
3111
3112 _igvn.rehash_node_delayed(iff);
3113 _igvn.rehash_node_delayed(proj);
3114
3115 proj->set_req(0, nullptr); // temporary disconnect
3116 ProjNode* proj2 = proj_clone(proj, iff);
3117 register_node(proj2, loop, iff, ddepth);
3118
3119 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
3120 register_node(cmp, loop, proj2, ddepth);
3121
3122 BoolNode* bol = new BoolNode(cmp, relop);
3123 register_node(bol, loop, proj2, ddepth);
3124
3125 int opcode = iff->Opcode();
3126 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
3127 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
3128 register_node(new_if, loop, proj2, ddepth);
3129
3130 proj->set_req(0, new_if); // reattach
3131 set_idom(proj, new_if, ddepth);
3132
3133 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
3134 guarantee(new_exit != nullptr, "null exit node");
3135 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
3136
3137 return new_exit;
3138 }
3139
3140 //------------------------------ insert_region_before_proj -------------------------------------
3141 // Insert a region before an if projection (* - new node)
3142 //
3143 // before
3144 // if(test)
3145 // / |
3146 // v |
3147 // proj v
3148 // other-proj
3149 //
3150 // after
3151 // if(test)
3152 // / |
3153 // v |
3154 // * proj-clone v
3155 // | other-proj
3156 // v
3157 // * new-region
3158 // |
3159 // v
3160 // * dum_if
3161 // / \
3162 // v \
3163 // * dum-proj v
3164 // proj
3165 //
3166 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3167 IfNode* iff = proj->in(0)->as_If();
3168 IdealLoopTree *loop = get_loop(proj);
3169 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3170 uint ddepth = dom_depth(proj);
3171
3172 _igvn.rehash_node_delayed(iff);
3173 _igvn.rehash_node_delayed(proj);
3174
3175 proj->set_req(0, nullptr); // temporary disconnect
3176 ProjNode* proj2 = proj_clone(proj, iff);
3177 register_node(proj2, loop, iff, ddepth);
3178
3179 RegionNode* reg = new RegionNode(2);
3180 reg->set_req(1, proj2);
3181 register_node(reg, loop, iff, ddepth);
3182
3183 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3184 register_node(dum_if, loop, reg, ddepth);
3185
3186 proj->set_req(0, dum_if); // reattach
3187 set_idom(proj, dum_if, ddepth);
3188
3189 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3190 register_node(dum_proj, loop, dum_if, ddepth);
3191
3192 return reg;
3193 }
3194
3195 // Idea
3196 // ----
3197 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3198 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3199 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3200 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3201 // test alone instead:
3202 //
3203 // Before Partial Peeling:
3204 // Loop:
3205 // <peeled section>
3206 // Split off signed loop exit test
3207 // <-- CUT HERE -->
3208 // Unchanged unsigned loop exit test
3209 // <rest of unpeeled section>
3210 // goto Loop
3211 //
3212 // After Partial Peeling:
3213 // <cloned peeled section>
3214 // Cloned split off signed loop exit test
3215 // Loop:
3216 // Unchanged unsigned loop exit test
3217 // <rest of unpeeled section>
3218 // <peeled section>
3219 // Split off signed loop exit test
3220 // goto Loop
3221 //
3222 // Details
3223 // -------
3224 // Before:
3225 // if (i <u limit) Unsigned loop exit condition
3226 // / |
3227 // v v
3228 // exit-proj stay-in-loop-proj
3229 //
3230 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3231 // before the CmpU on the stay-in-loop path and keep both tests:
3232 //
3233 // if (i <u limit) Signed loop exit test
3234 // / |
3235 // / if (i <u limit) Unsigned loop exit test
3236 // / / |
3237 // v v v
3238 // exit-region stay-in-loop-proj
3239 //
3240 // Implementation
3241 // --------------
3242 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3243 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3244 // exit tests is preserved, and their loop nesting is correct.
3245 //
3246 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3247 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3248 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3249 //
3250 // if (stay-in-loop-const) Killed original unsigned loop exit test
3251 // / |
3252 // / v
3253 // / if (i < limit) Split off signed loop exit test
3254 // / / |
3255 // / / v
3256 // / / if (i <u limit) Cloned unsigned loop exit test
3257 // / / / |
3258 // v v v |
3259 // exit-region |
3260 // | |
3261 // dummy-if |
3262 // / | |
3263 // dead | |
3264 // v v
3265 // exit-proj stay-in-loop-proj
3266 //
3267 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3268 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3269 //
3270 // Requirements
3271 // ------------
3272 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3273 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3274 // Otherwise, we need to bail out (see comments in the code below).
3275 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3276 const bool Signed = true;
3277 const bool Unsigned = false;
3278
3279 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3280 if (bol->_test._test != BoolTest::lt) {
3281 return nullptr;
3282 }
3283 CmpNode* cmpu = bol->in(1)->as_Cmp();
3284 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3285
3286 int stride = stride_of_possible_iv(if_cmpu);
3287 if (stride == 0) {
3288 return nullptr;
3289 }
3290
3291 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3292 guarantee(lp_proj != nullptr, "null loop node");
3293
3294 ProjNode* lp_continue = lp_proj->as_Proj();
3295 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3296 if (!lp_exit->is_IfFalse()) {
3297 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3298 // We therefore can't add a single exit condition.
3299 return nullptr;
3300 }
3301 // The unsigned loop exit condition is
3302 // !(i <u limit)
3303 // = i >=u limit
3304 //
3305 // First, we note that for any x for which
3306 // 0 <= x <= INT_MAX
3307 // we can convert x to an unsigned int and still get the same guarantee:
3308 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3309 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3310 //
3311 // With that in mind, if
3312 // limit >= 0 (COND)
3313 // then the unsigned loop exit condition
3314 // i >=u limit (ULE)
3315 // is equivalent to
3316 // i < 0 || i >= limit (SLE-full)
3317 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3318 // (uint) i >=u MAX_INT >= limit >= 0
3319 // or otherwise
3320 // i >= limit >= 0
3321 // holds due to (LEMMA).
3322 //
3323 // For completeness, a counterexample with limit < 0:
3324 // Assume i = -3 and limit = -2:
3325 // i < 0
3326 // -2 < 0
3327 // is true and thus also "i < 0 || i >= limit". But
3328 // i >=u limit
3329 // -3 >=u -2
3330 // is false.
3331 Node* limit = cmpu->in(2);
3332 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3333 if (type_limit->_lo < 0) {
3334 return nullptr;
3335 }
3336
3337 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3338 // stride < 0:
3339 // i < 0 (SLE = SLE-negative)
3340 // stride > 0:
3341 // i >= limit (SLE = SLE-positive)
3342 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3343 //
3344 // Loop:
3345 // <peeled section>
3346 // i >= limit (SLE-positive)
3347 // <-- CUT HERE -->
3348 // i >=u limit (ULE)
3349 // <rest of unpeeled section>
3350 // goto Loop
3351 //
3352 // We exit the loop if:
3353 // (SLE) is true OR (ULE) is true
3354 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3355 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3356 // (SLE) IMPLIES (ULE)
3357 // This indeed holds when (COND) is given:
3358 // - stride > 0:
3359 // i >= limit // (SLE = SLE-positive)
3360 // i >= limit >= 0 // (COND)
3361 // i >=u limit >= 0 // (LEMMA)
3362 // which is the unsigned loop exit condition (ULE).
3363 // - stride < 0:
3364 // i < 0 // (SLE = SLE-negative)
3365 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3366 // MAX_INT >= limit >= 0 // (COND)
3367 // MAX_INT >=u limit >= 0 // (LEMMA)
3368 // and thus from (NEG) and (LEMMA):
3369 // i >=u limit
3370 // which is the unsigned loop exit condition (ULE).
3371 //
3372 //
3373 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3374 // <cloned peeled section>
3375 // i >= limit (SLE-positive)
3376 // Loop:
3377 // i >=u limit (ULE)
3378 // <rest of unpeeled section>
3379 // <peeled section>
3380 // i >= limit (SLE-positive)
3381 // goto Loop
3382 Node* rhs_cmpi;
3383 if (stride > 0) {
3384 rhs_cmpi = limit; // For i >= limit
3385 } else {
3386 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3387 }
3388 // Create a new region on the exit path
3389 RegionNode* reg = insert_region_before_proj(lp_exit);
3390 guarantee(reg != nullptr, "null region node");
3391
3392 // Clone the if-cmpu-true-false using a signed compare
3393 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3394 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3395 reg->add_req(cmpi_exit);
3396
3397 // Clone the if-cmpu-true-false
3398 BoolTest::mask rel_u = bol->_test._test;
3399 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3400 reg->add_req(cmpu_exit);
3401
3402 // Force original if to stay in loop.
3403 short_circuit_if(if_cmpu, lp_continue);
3404
3405 return cmpi_exit->in(0)->as_If();
3406 }
3407
3408 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3409 // Remove a previously inserted signed compare loop exit.
3410 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3411 Node* lp_proj = stay_in_loop(if_cmp, loop);
3412 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3413 stay_in_loop(lp_proj, loop)->is_If() &&
3414 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3415 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3416 if_cmp->set_req(1, con);
3417 }
3418
3419 //------------------------------ scheduled_nodelist -------------------------------------
3420 // Create a post order schedule of nodes that are in the
3421 // "member" set. The list is returned in "sched".
3422 // The first node in "sched" is the loop head, followed by
3423 // nodes which have no inputs in the "member" set, and then
3424 // followed by the nodes that have an immediate input dependence
3425 // on a node in "sched".
3426 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3427
3428 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3429 VectorSet visited;
3430 Node_Stack nstack(loop->_body.size());
3431
3432 Node* n = loop->_head; // top of stack is cached in "n"
3433 uint idx = 0;
3434 visited.set(n->_idx);
3435
3436 // Initially push all with no inputs from within member set
3437 for(uint i = 0; i < loop->_body.size(); i++ ) {
3438 Node *elt = loop->_body.at(i);
3439 if (member.test(elt->_idx)) {
3440 bool found = false;
3441 for (uint j = 0; j < elt->req(); j++) {
3442 Node* def = elt->in(j);
3443 if (def && member.test(def->_idx) && def != elt) {
3444 found = true;
3445 break;
3446 }
3447 }
3448 if (!found && elt != loop->_head) {
3449 nstack.push(n, idx);
3450 n = elt;
3451 assert(!visited.test(n->_idx), "not seen yet");
3452 visited.set(n->_idx);
3453 }
3454 }
3455 }
3456
3457 // traverse out's that are in the member set
3458 while (true) {
3459 if (idx < n->outcnt()) {
3460 Node* use = n->raw_out(idx);
3461 idx++;
3462 if (!visited.test_set(use->_idx)) {
3463 if (member.test(use->_idx)) {
3464 nstack.push(n, idx);
3465 n = use;
3466 idx = 0;
3467 }
3468 }
3469 } else {
3470 // All outputs processed
3471 sched.push(n);
3472 if (nstack.is_empty()) break;
3473 n = nstack.node();
3474 idx = nstack.index();
3475 nstack.pop();
3476 }
3477 }
3478 }
3479
3480
3481 //------------------------------ has_use_in_set -------------------------------------
3482 // Has a use in the vector set
3483 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3484 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3485 Node* use = n->fast_out(j);
3486 if (vset.test(use->_idx)) {
3487 return true;
3488 }
3489 }
3490 return false;
3491 }
3492
3493
3494 //------------------------------ has_use_internal_to_set -------------------------------------
3495 // Has use internal to the vector set (ie. not in a phi at the loop head)
3496 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3497 Node* head = loop->_head;
3498 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3499 Node* use = n->fast_out(j);
3500 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3501 return true;
3502 }
3503 }
3504 return false;
3505 }
3506
3507
3508 //------------------------------ clone_for_use_outside_loop -------------------------------------
3509 // clone "n" for uses that are outside of loop
3510 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3511 int cloned = 0;
3512 assert(worklist.size() == 0, "should be empty");
3513 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3514 Node* use = n->fast_out(j);
3515 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3516 worklist.push(use);
3517 }
3518 }
3519
3520 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3521 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3522 return -1;
3523 }
3524
3525 while( worklist.size() ) {
3526 Node *use = worklist.pop();
3527 if (!has_node(use) || use->in(0) == C->top()) continue;
3528 uint j;
3529 for (j = 0; j < use->req(); j++) {
3530 if (use->in(j) == n) break;
3531 }
3532 assert(j < use->req(), "must be there");
3533
3534 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3535 Node* n_clone = n->clone();
3536 _igvn.replace_input_of(use, j, n_clone);
3537 cloned++;
3538 Node* use_c;
3539 if (!use->is_Phi()) {
3540 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3541 } else {
3542 // Use in a phi is considered a use in the associated predecessor block
3543 use_c = use->in(0)->in(j);
3544 }
3545 set_ctrl(n_clone, use_c);
3546 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3547 get_loop(use_c)->_body.push(n_clone);
3548 _igvn.register_new_node_with_optimizer(n_clone);
3549 #ifndef PRODUCT
3550 if (TracePartialPeeling) {
3551 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3552 }
3553 #endif
3554 }
3555 return cloned;
3556 }
3557
3558
3559 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3560 // clone "n" for special uses that are in the not_peeled region.
3561 // If these def-uses occur in separate blocks, the code generator
3562 // marks the method as not compilable. For example, if a "BoolNode"
3563 // is in a different basic block than the "IfNode" that uses it, then
3564 // the compilation is aborted in the code generator.
3565 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3566 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3567 if (n->is_Phi() || n->is_Load()) {
3568 return;
3569 }
3570 assert(worklist.size() == 0, "should be empty");
3571 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3572 Node* use = n->fast_out(j);
3573 if ( not_peel.test(use->_idx) &&
3574 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3575 use->in(1) == n) {
3576 worklist.push(use);
3577 }
3578 }
3579 if (worklist.size() > 0) {
3580 // clone "n" and insert it between inputs of "n" and the use
3581 Node* n_clone = n->clone();
3582 loop->_body.push(n_clone);
3583 _igvn.register_new_node_with_optimizer(n_clone);
3584 set_ctrl(n_clone, get_ctrl(n));
3585 sink_list.push(n_clone);
3586 not_peel.set(n_clone->_idx);
3587 #ifndef PRODUCT
3588 if (TracePartialPeeling) {
3589 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3590 }
3591 #endif
3592 while( worklist.size() ) {
3593 Node *use = worklist.pop();
3594 _igvn.rehash_node_delayed(use);
3595 for (uint j = 1; j < use->req(); j++) {
3596 if (use->in(j) == n) {
3597 use->set_req(j, n_clone);
3598 }
3599 }
3600 }
3601 }
3602 }
3603
3604
3605 //------------------------------ insert_phi_for_loop -------------------------------------
3606 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3607 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3608 Node *phi = PhiNode::make(lp, back_edge_val);
3609 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3610 // Use existing phi if it already exists
3611 Node *hit = _igvn.hash_find_insert(phi);
3612 if( hit == nullptr ) {
3613 _igvn.register_new_node_with_optimizer(phi);
3614 set_ctrl(phi, lp);
3615 } else {
3616 // Remove the new phi from the graph and use the hit
3617 _igvn.remove_dead_node(phi);
3618 phi = hit;
3619 }
3620 _igvn.replace_input_of(use, idx, phi);
3621 }
3622
3623 #ifdef ASSERT
3624 //------------------------------ is_valid_loop_partition -------------------------------------
3625 // Validate the loop partition sets: peel and not_peel
3626 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3627 VectorSet& not_peel ) {
3628 uint i;
3629 // Check that peel_list entries are in the peel set
3630 for (i = 0; i < peel_list.size(); i++) {
3631 if (!peel.test(peel_list.at(i)->_idx)) {
3632 return false;
3633 }
3634 }
3635 // Check at loop members are in one of peel set or not_peel set
3636 for (i = 0; i < loop->_body.size(); i++ ) {
3637 Node *def = loop->_body.at(i);
3638 uint di = def->_idx;
3639 // Check that peel set elements are in peel_list
3640 if (peel.test(di)) {
3641 if (not_peel.test(di)) {
3642 return false;
3643 }
3644 // Must be in peel_list also
3645 bool found = false;
3646 for (uint j = 0; j < peel_list.size(); j++) {
3647 if (peel_list.at(j)->_idx == di) {
3648 found = true;
3649 break;
3650 }
3651 }
3652 if (!found) {
3653 return false;
3654 }
3655 } else if (not_peel.test(di)) {
3656 if (peel.test(di)) {
3657 return false;
3658 }
3659 } else {
3660 return false;
3661 }
3662 }
3663 return true;
3664 }
3665
3666 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3667 // Ensure a use outside of loop is of the right form
3668 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3669 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3670 return (use->is_Phi() &&
3671 use_c->is_Region() && use_c->req() == 3 &&
3672 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3673 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3674 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3675 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3676 }
3677
3678 //------------------------------ is_valid_clone_loop_form -------------------------------------
3679 // Ensure that all uses outside of loop are of the right form
3680 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3681 uint orig_exit_idx, uint clone_exit_idx) {
3682 uint len = peel_list.size();
3683 for (uint i = 0; i < len; i++) {
3684 Node *def = peel_list.at(i);
3685
3686 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3687 Node *use = def->fast_out(j);
3688 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3689 if (!loop->is_member(get_loop(use_c))) {
3690 // use is not in the loop, check for correct structure
3691 if (use->in(0) == def) {
3692 // Okay
3693 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3694 return false;
3695 }
3696 }
3697 }
3698 }
3699 return true;
3700 }
3701 #endif
3702
3703 //------------------------------ partial_peel -------------------------------------
3704 // Partially peel (aka loop rotation) the top portion of a loop (called
3705 // the peel section below) by cloning it and placing one copy just before
3706 // the new loop head and the other copy at the bottom of the new loop.
3707 //
3708 // before after where it came from
3709 //
3710 // stmt1 stmt1
3711 // loop: stmt2 clone
3712 // stmt2 if condA goto exitA clone
3713 // if condA goto exitA new_loop: new
3714 // stmt3 stmt3 clone
3715 // if !condB goto loop if condB goto exitB clone
3716 // exitB: stmt2 orig
3717 // stmt4 if !condA goto new_loop orig
3718 // exitA: goto exitA
3719 // exitB:
3720 // stmt4
3721 // exitA:
3722 //
3723 // Step 1: find the cut point: an exit test on probable
3724 // induction variable.
3725 // Step 2: schedule (with cloning) operations in the peel
3726 // section that can be executed after the cut into
3727 // the section that is not peeled. This may need
3728 // to clone operations into exit blocks. For
3729 // instance, a reference to A[i] in the not-peel
3730 // section and a reference to B[i] in an exit block
3731 // may cause a left-shift of i by 2 to be placed
3732 // in the peel block. This step will clone the left
3733 // shift into the exit block and sink the left shift
3734 // from the peel to the not-peel section.
3735 // Step 3: clone the loop, retarget the control, and insert
3736 // phis for values that are live across the new loop
3737 // head. This is very dependent on the graph structure
3738 // from clone_loop. It creates region nodes for
3739 // exit control and associated phi nodes for values
3740 // flow out of the loop through that exit. The region
3741 // node is dominated by the clone's control projection.
3742 // So the clone's peel section is placed before the
3743 // new loop head, and the clone's not-peel section is
3744 // forms the top part of the new loop. The original
3745 // peel section forms the tail of the new loop.
3746 // Step 4: update the dominator tree and recompute the
3747 // dominator depth.
3748 //
3749 // orig
3750 //
3751 // stmt1
3752 // |
3753 // v
3754 // predicates
3755 // |
3756 // v
3757 // loop<----+
3758 // | |
3759 // stmt2 |
3760 // | |
3761 // v |
3762 // ifA |
3763 // / | |
3764 // v v |
3765 // false true ^ <-- last_peel
3766 // / | |
3767 // / ===|==cut |
3768 // / stmt3 | <-- first_not_peel
3769 // / | |
3770 // | v |
3771 // v ifB |
3772 // exitA: / \ |
3773 // / \ |
3774 // v v |
3775 // false true |
3776 // / \ |
3777 // / ----+
3778 // |
3779 // v
3780 // exitB:
3781 // stmt4
3782 //
3783 //
3784 // after clone loop
3785 //
3786 // stmt1
3787 // |
3788 // v
3789 // predicates
3790 // / \
3791 // clone / \ orig
3792 // / \
3793 // / \
3794 // v v
3795 // +---->loop loop<----+
3796 // | | | |
3797 // | stmt2 stmt2 |
3798 // | | | |
3799 // | v v |
3800 // | ifA ifA |
3801 // | | \ / | |
3802 // | v v v v |
3803 // ^ true false false true ^ <-- last_peel
3804 // | | ^ \ / | |
3805 // | cut==|== \ \ / ===|==cut |
3806 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3807 // | | dom | | | |
3808 // | v \ 1v v2 v |
3809 // | ifB regionA ifB |
3810 // | / \ | / \ |
3811 // | / \ v / \ |
3812 // | v v exitA: v v |
3813 // | true false false true |
3814 // | / ^ \ / \ |
3815 // +---- \ \ / ----+
3816 // dom \ /
3817 // \ 1v v2
3818 // regionB
3819 // |
3820 // v
3821 // exitB:
3822 // stmt4
3823 //
3824 //
3825 // after partial peel
3826 //
3827 // stmt1
3828 // |
3829 // v
3830 // predicates
3831 // /
3832 // clone / orig
3833 // / TOP
3834 // / \
3835 // v v
3836 // TOP->loop loop----+
3837 // | | |
3838 // stmt2 stmt2 |
3839 // | | |
3840 // v v |
3841 // ifA ifA |
3842 // | \ / | |
3843 // v v v v |
3844 // true false false true | <-- last_peel
3845 // | ^ \ / +------|---+
3846 // +->newloop \ \ / === ==cut | |
3847 // | stmt3 \ \ / TOP | |
3848 // | | dom | | stmt3 | | <-- first_not_peel
3849 // | v \ 1v v2 v | |
3850 // | ifB regionA ifB ^ v
3851 // | / \ | / \ | |
3852 // | / \ v / \ | |
3853 // | v v exitA: v v | |
3854 // | true false false true | |
3855 // | / ^ \ / \ | |
3856 // | | \ \ / v | |
3857 // | | dom \ / TOP | |
3858 // | | \ 1v v2 | |
3859 // ^ v regionB | |
3860 // | | | | |
3861 // | | v ^ v
3862 // | | exitB: | |
3863 // | | stmt4 | |
3864 // | +------------>-----------------+ |
3865 // | |
3866 // +-----------------<---------------------+
3867 //
3868 //
3869 // final graph
3870 //
3871 // stmt1
3872 // |
3873 // v
3874 // predicates
3875 // |
3876 // v
3877 // stmt2 clone
3878 // |
3879 // v
3880 // ........> ifA clone
3881 // : / |
3882 // dom / |
3883 // : v v
3884 // : false true
3885 // : | |
3886 // : | v
3887 // : | newloop<-----+
3888 // : | | |
3889 // : | stmt3 clone |
3890 // : | | |
3891 // : | v |
3892 // : | ifB |
3893 // : | / \ |
3894 // : | v v |
3895 // : | false true |
3896 // : | | | |
3897 // : | v stmt2 |
3898 // : | exitB: | |
3899 // : | stmt4 v |
3900 // : | ifA orig |
3901 // : | / \ |
3902 // : | / \ |
3903 // : | v v |
3904 // : | false true |
3905 // : | / \ |
3906 // : v v -----+
3907 // RegionA
3908 // |
3909 // v
3910 // exitA
3911 //
3912 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3913
3914 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3915 if (!loop->_head->is_Loop()) {
3916 return false;
3917 }
3918 LoopNode *head = loop->_head->as_Loop();
3919
3920 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3921 return false;
3922 }
3923
3924 // Check for complex exit control
3925 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3926 Node *n = loop->_body.at(ii);
3927 int opc = n->Opcode();
3928 if (n->is_Call() ||
3929 opc == Op_Catch ||
3930 opc == Op_CatchProj ||
3931 opc == Op_Jump ||
3932 opc == Op_JumpProj) {
3933 #ifndef PRODUCT
3934 if (TracePartialPeeling) {
3935 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3936 }
3937 #endif
3938 return false;
3939 }
3940 }
3941
3942 int dd = dom_depth(head);
3943
3944 // Step 1: find cut point
3945
3946 // Walk up dominators to loop head looking for first loop exit
3947 // which is executed on every path thru loop.
3948 IfNode *peel_if = nullptr;
3949 IfNode *peel_if_cmpu = nullptr;
3950
3951 Node *iff = loop->tail();
3952 while (iff != head) {
3953 if (iff->is_If()) {
3954 Node *ctrl = get_ctrl(iff->in(1));
3955 if (ctrl->is_top()) return false; // Dead test on live IF.
3956 // If loop-varying exit-test, check for induction variable
3957 if (loop->is_member(get_loop(ctrl)) &&
3958 loop->is_loop_exit(iff) &&
3959 is_possible_iv_test(iff)) {
3960 Node* cmp = iff->in(1)->in(1);
3961 if (cmp->Opcode() == Op_CmpI) {
3962 peel_if = iff->as_If();
3963 } else {
3964 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3965 peel_if_cmpu = iff->as_If();
3966 }
3967 }
3968 }
3969 iff = idom(iff);
3970 }
3971
3972 // Prefer signed compare over unsigned compare.
3973 IfNode* new_peel_if = nullptr;
3974 if (peel_if == nullptr) {
3975 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3976 return false; // No peel point found
3977 }
3978 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3979 if (new_peel_if == nullptr) {
3980 return false; // No peel point found
3981 }
3982 peel_if = new_peel_if;
3983 }
3984 Node* last_peel = stay_in_loop(peel_if, loop);
3985 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3986 if (first_not_peeled == nullptr || first_not_peeled == head) {
3987 return false;
3988 }
3989
3990 #ifndef PRODUCT
3991 if (TraceLoopOpts) {
3992 tty->print("PartialPeel ");
3993 loop->dump_head();
3994 }
3995
3996 if (TracePartialPeeling) {
3997 tty->print_cr("before partial peel one iteration");
3998 Node_List wl;
3999 Node* t = head->in(2);
4000 while (true) {
4001 wl.push(t);
4002 if (t == head) break;
4003 t = idom(t);
4004 }
4005 while (wl.size() > 0) {
4006 Node* tt = wl.pop();
4007 tt->dump();
4008 if (tt == last_peel) tty->print_cr("-- cut --");
4009 }
4010 }
4011 #endif
4012
4013 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
4014
4015 VectorSet peel;
4016 VectorSet not_peel;
4017 Node_List peel_list;
4018 Node_List worklist;
4019 Node_List sink_list;
4020
4021 uint estimate = loop->est_loop_clone_sz(1);
4022 if (exceeding_node_budget(estimate)) {
4023 return false;
4024 }
4025
4026 // Set of cfg nodes to peel are those that are executable from
4027 // the head through last_peel.
4028 assert(worklist.size() == 0, "should be empty");
4029 worklist.push(head);
4030 peel.set(head->_idx);
4031 while (worklist.size() > 0) {
4032 Node *n = worklist.pop();
4033 if (n != last_peel) {
4034 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
4035 Node* use = n->fast_out(j);
4036 if (use->is_CFG() &&
4037 loop->is_member(get_loop(use)) &&
4038 !peel.test_set(use->_idx)) {
4039 worklist.push(use);
4040 }
4041 }
4042 }
4043 }
4044
4045 // Set of non-cfg nodes to peel are those that are control
4046 // dependent on the cfg nodes.
4047 for (uint i = 0; i < loop->_body.size(); i++) {
4048 Node *n = loop->_body.at(i);
4049 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
4050 if (peel.test(n_c->_idx)) {
4051 peel.set(n->_idx);
4052 } else {
4053 not_peel.set(n->_idx);
4054 }
4055 }
4056
4057 // Step 2: move operations from the peeled section down into the
4058 // not-peeled section
4059
4060 // Get a post order schedule of nodes in the peel region
4061 // Result in right-most operand.
4062 scheduled_nodelist(loop, peel, peel_list);
4063
4064 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4065
4066 // For future check for too many new phis
4067 uint old_phi_cnt = 0;
4068 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
4069 Node* use = head->fast_out(j);
4070 if (use->is_Phi()) old_phi_cnt++;
4071 }
4072
4073 #ifndef PRODUCT
4074 if (TracePartialPeeling) {
4075 tty->print_cr("\npeeled list");
4076 }
4077 #endif
4078
4079 // Evacuate nodes in peel region into the not_peeled region if possible
4080 bool too_many_clones = false;
4081 uint new_phi_cnt = 0;
4082 uint cloned_for_outside_use = 0;
4083 for (uint i = 0; i < peel_list.size();) {
4084 Node* n = peel_list.at(i);
4085 #ifndef PRODUCT
4086 if (TracePartialPeeling) n->dump();
4087 #endif
4088 bool incr = true;
4089 if (!n->is_CFG()) {
4090 if (has_use_in_set(n, not_peel)) {
4091 // If not used internal to the peeled region,
4092 // move "n" from peeled to not_peeled region.
4093 if (!has_use_internal_to_set(n, peel, loop)) {
4094 // if not pinned and not a load (which maybe anti-dependent on a store)
4095 // and not a CMove (Matcher expects only bool->cmove).
4096 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
4097 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
4098 if (C->failing()) return false;
4099 if (new_clones == -1) {
4100 too_many_clones = true;
4101 break;
4102 }
4103 cloned_for_outside_use += new_clones;
4104 sink_list.push(n);
4105 peel.remove(n->_idx);
4106 not_peel.set(n->_idx);
4107 peel_list.remove(i);
4108 incr = false;
4109 #ifndef PRODUCT
4110 if (TracePartialPeeling) {
4111 tty->print_cr("sink to not_peeled region: %d newbb: %d",
4112 n->_idx, get_ctrl(n)->_idx);
4113 }
4114 #endif
4115 }
4116 } else {
4117 // Otherwise check for special def-use cases that span
4118 // the peel/not_peel boundary such as bool->if
4119 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
4120 new_phi_cnt++;
4121 }
4122 }
4123 }
4124 if (incr) i++;
4125 }
4126
4127 estimate += cloned_for_outside_use + new_phi_cnt;
4128 bool exceed_node_budget = !may_require_nodes(estimate);
4129 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
4130
4131 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
4132 #ifndef PRODUCT
4133 if (TracePartialPeeling && exceed_phi_limit) {
4134 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
4135 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
4136 }
4137 #endif
4138 if (new_peel_if != nullptr) {
4139 remove_cmpi_loop_exit(new_peel_if, loop);
4140 }
4141 // Inhibit more partial peeling on this loop
4142 assert(!head->is_partial_peel_loop(), "not partial peeled");
4143 head->mark_partial_peel_failed();
4144 if (cloned_for_outside_use > 0) {
4145 // Terminate this round of loop opts because
4146 // the graph outside this loop was changed.
4147 C->set_major_progress();
4148 return true;
4149 }
4150 return false;
4151 }
4152
4153 // Step 3: clone loop, retarget control, and insert new phis
4154
4155 // Create new loop head for new phis and to hang
4156 // the nodes being moved (sinked) from the peel region.
4157 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4158 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4159 _igvn.register_new_node_with_optimizer(new_head);
4160 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4161 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4162 set_loop(new_head, loop);
4163 loop->_body.push(new_head);
4164 not_peel.set(new_head->_idx);
4165 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4166 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4167
4168 while (sink_list.size() > 0) {
4169 Node* n = sink_list.pop();
4170 set_ctrl(n, new_head);
4171 }
4172
4173 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4174
4175 clone_loop(loop, old_new, dd, IgnoreStripMined);
4176
4177 const uint clone_exit_idx = 1;
4178 const uint orig_exit_idx = 2;
4179 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4180
4181 Node* head_clone = old_new[head->_idx];
4182 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4183 Node* orig_tail_clone = head_clone->in(2);
4184
4185 // Add phi if "def" node is in peel set and "use" is not
4186
4187 for (uint i = 0; i < peel_list.size(); i++) {
4188 Node *def = peel_list.at(i);
4189 if (!def->is_CFG()) {
4190 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4191 Node *use = def->fast_out(j);
4192 if (has_node(use) && use->in(0) != C->top() &&
4193 (!peel.test(use->_idx) ||
4194 (use->is_Phi() && use->in(0) == head)) ) {
4195 worklist.push(use);
4196 }
4197 }
4198 while( worklist.size() ) {
4199 Node *use = worklist.pop();
4200 for (uint j = 1; j < use->req(); j++) {
4201 Node* n = use->in(j);
4202 if (n == def) {
4203
4204 // "def" is in peel set, "use" is not in peel set
4205 // or "use" is in the entry boundary (a phi) of the peel set
4206
4207 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4208
4209 if ( loop->is_member(get_loop( use_c )) ) {
4210 // use is in loop
4211 if (old_new[use->_idx] != nullptr) { // null for dead code
4212 Node* use_clone = old_new[use->_idx];
4213 _igvn.replace_input_of(use, j, C->top());
4214 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4215 }
4216 } else {
4217 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4218 // use is not in the loop, check if the live range includes the cut
4219 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4220 if (not_peel.test(lp_if->_idx)) {
4221 assert(j == orig_exit_idx, "use from original loop");
4222 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4223 }
4224 }
4225 }
4226 }
4227 }
4228 }
4229 }
4230
4231 // Step 3b: retarget control
4232
4233 // Redirect control to the new loop head if a cloned node in
4234 // the not_peeled region has control that points into the peeled region.
4235 // This necessary because the cloned peeled region will be outside
4236 // the loop.
4237 // from to
4238 // cloned-peeled <---+
4239 // new_head_clone: | <--+
4240 // cloned-not_peeled in(0) in(0)
4241 // orig-peeled
4242
4243 for (uint i = 0; i < loop->_body.size(); i++) {
4244 Node *n = loop->_body.at(i);
4245 if (!n->is_CFG() && n->in(0) != nullptr &&
4246 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4247 Node* n_clone = old_new[n->_idx];
4248 if (n_clone->depends_only_on_test()) {
4249 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4250 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4251 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4252 // floating above its range check.
4253 Node* pinned_clone = n_clone->pin_array_access_node();
4254 if (pinned_clone != nullptr) {
4255 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4256 old_new.map(n->_idx, pinned_clone);
4257 _igvn.replace_node(n_clone, pinned_clone);
4258 n_clone = pinned_clone;
4259 }
4260 }
4261 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4262 }
4263 }
4264
4265 // Backedge of the surviving new_head (the clone) is original last_peel
4266 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4267
4268 // Cut first node in original not_peel set
4269 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4270 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4271 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4272
4273 // Copy head_clone back-branch info to original head
4274 // and remove original head's loop entry and
4275 // clone head's back-branch
4276 _igvn.rehash_node_delayed(head); // Multiple edge updates
4277 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4278 head->set_req(LoopNode::LoopBackControl, C->top());
4279 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4280
4281 // Similarly modify the phis
4282 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4283 Node* use = head->fast_out(k);
4284 if (use->is_Phi() && use->outcnt() > 0) {
4285 Node* use_clone = old_new[use->_idx];
4286 _igvn.rehash_node_delayed(use); // Multiple edge updates
4287 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4288 use->set_req(LoopNode::LoopBackControl, C->top());
4289 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4290 }
4291 }
4292
4293 // Step 4: update dominator tree and dominator depth
4294
4295 set_idom(head, orig_tail_clone, dd);
4296 recompute_dom_depth();
4297
4298 // Inhibit more partial peeling on this loop
4299 new_head_clone->set_partial_peel_loop();
4300 C->set_major_progress();
4301 loop->record_for_igvn();
4302
4303 #ifndef PRODUCT
4304 if (TracePartialPeeling) {
4305 tty->print_cr("\nafter partial peel one iteration");
4306 Node_List wl;
4307 Node* t = last_peel;
4308 while (true) {
4309 wl.push(t);
4310 if (t == head_clone) break;
4311 t = idom(t);
4312 }
4313 while (wl.size() > 0) {
4314 Node* tt = wl.pop();
4315 if (tt == head) tty->print_cr("orig head");
4316 else if (tt == new_head_clone) tty->print_cr("new head");
4317 else if (tt == head_clone) tty->print_cr("clone head");
4318 tt->dump();
4319 }
4320 }
4321 #endif
4322
4323 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4324
4325 return true;
4326 }
4327
4328 // Transform:
4329 //
4330 // loop<-----------------+
4331 // | |
4332 // stmt1 stmt2 .. stmtn |
4333 // | | | |
4334 // \ | / |
4335 // v v v |
4336 // region |
4337 // | |
4338 // shared_stmt |
4339 // | |
4340 // v |
4341 // if |
4342 // / \ |
4343 // | -----------+
4344 // v
4345 //
4346 // into:
4347 //
4348 // loop<-------------------+
4349 // | |
4350 // v |
4351 // +->loop |
4352 // | | |
4353 // | stmt1 stmt2 .. stmtn |
4354 // | | | | |
4355 // | | \ / |
4356 // | | v v |
4357 // | | region1 |
4358 // | | | |
4359 // | shared_stmt shared_stmt |
4360 // | | | |
4361 // | v v |
4362 // | if if |
4363 // | /\ / \ |
4364 // +-- | | -------+
4365 // \ /
4366 // v v
4367 // region2
4368 //
4369 // (region2 is shown to merge mirrored projections of the loop exit
4370 // ifs to make the diagram clearer but they really merge the same
4371 // projection)
4372 //
4373 // Conditions for this transformation to trigger:
4374 // - the path through stmt1 is frequent enough
4375 // - the inner loop will be turned into a counted loop after transformation
4376 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4377 if (!DuplicateBackedge) {
4378 return false;
4379 }
4380 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4381 if (!loop->_head->is_Loop()) {
4382 return false;
4383 }
4384
4385 uint estimate = loop->est_loop_clone_sz(1);
4386 if (exceeding_node_budget(estimate)) {
4387 return false;
4388 }
4389
4390 LoopNode *head = loop->_head->as_Loop();
4391
4392 Node* region = nullptr;
4393 IfNode* exit_test = nullptr;
4394 uint inner;
4395 float f;
4396 if (StressDuplicateBackedge) {
4397 if (head->is_strip_mined()) {
4398 return false;
4399 }
4400 Node* c = head->in(LoopNode::LoopBackControl);
4401
4402 while (c != head) {
4403 if (c->is_Region()) {
4404 region = c;
4405 }
4406 c = idom(c);
4407 }
4408
4409 if (region == nullptr) {
4410 return false;
4411 }
4412
4413 inner = 1;
4414 } else {
4415 // Is the shape of the loop that of a counted loop...
4416 Node* back_control = loop_exit_control(head, loop);
4417 if (back_control == nullptr) {
4418 return false;
4419 }
4420
4421 BoolTest::mask bt = BoolTest::illegal;
4422 float cl_prob = 0;
4423 Node* incr = nullptr;
4424 Node* limit = nullptr;
4425 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4426 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4427 return false;
4428 }
4429
4430 // With an extra phi for the candidate iv?
4431 // Or the region node is the loop head
4432 if (!incr->is_Phi() || incr->in(0) == head) {
4433 return false;
4434 }
4435
4436 PathFrequency pf(head, this);
4437 region = incr->in(0);
4438
4439 // Go over all paths for the extra phi's region and see if that
4440 // path is frequent enough and would match the expected iv shape
4441 // if the extra phi is removed
4442 inner = 0;
4443 for (uint i = 1; i < incr->req(); ++i) {
4444 Node* in = incr->in(i);
4445 Node* trunc1 = nullptr;
4446 Node* trunc2 = nullptr;
4447 const TypeInteger* iv_trunc_t = nullptr;
4448 Node* orig_in = in;
4449 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4450 continue;
4451 }
4452 assert(in->Opcode() == Op_AddI, "wrong increment code");
4453 Node* xphi = nullptr;
4454 Node* stride = loop_iv_stride(in, loop, xphi);
4455
4456 if (stride == nullptr) {
4457 continue;
4458 }
4459
4460 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4461 if (phi == nullptr ||
4462 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4463 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4464 return false;
4465 }
4466
4467 f = pf.to(region->in(i));
4468 if (f > 0.5) {
4469 inner = i;
4470 break;
4471 }
4472 }
4473
4474 if (inner == 0) {
4475 return false;
4476 }
4477
4478 exit_test = back_control->in(0)->as_If();
4479 }
4480
4481 if (idom(region)->is_Catch()) {
4482 return false;
4483 }
4484
4485 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4486 Unique_Node_List wq;
4487 wq.push(head->in(LoopNode::LoopBackControl));
4488 for (uint i = 0; i < wq.size(); i++) {
4489 Node* c = wq.at(i);
4490 assert(get_loop(c) == loop, "not in the right loop?");
4491 if (c->is_Region()) {
4492 if (c != region) {
4493 for (uint j = 1; j < c->req(); ++j) {
4494 wq.push(c->in(j));
4495 }
4496 }
4497 } else {
4498 wq.push(c->in(0));
4499 }
4500 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4501 }
4502
4503 Node* region_dom = idom(region);
4504
4505 // Can't do the transformation if this would cause a membar pair to
4506 // be split
4507 for (uint i = 0; i < wq.size(); i++) {
4508 Node* c = wq.at(i);
4509 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4510 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4511 if (!wq.member(c->as_MemBar()->leading_membar())) {
4512 return false;
4513 }
4514 }
4515 }
4516
4517 // Collect data nodes that need to be clones as well
4518 int dd = dom_depth(head);
4519
4520 for (uint i = 0; i < loop->_body.size(); ++i) {
4521 Node* n = loop->_body.at(i);
4522 if (has_ctrl(n)) {
4523 Node* c = get_ctrl(n);
4524 if (wq.member(c)) {
4525 wq.push(n);
4526 }
4527 } else {
4528 set_idom(n, idom(n), dd);
4529 }
4530 }
4531
4532 // clone shared_stmt
4533 clone_loop_body(wq, old_new, nullptr);
4534
4535 Node* region_clone = old_new[region->_idx];
4536 region_clone->set_req(inner, C->top());
4537 set_idom(region, region->in(inner), dd);
4538
4539 // Prepare the outer loop
4540 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4541 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4542 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4543 set_idom(head, outer_head, dd);
4544
4545 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4546
4547 // Make one of the shared_stmt copies only reachable from stmt1, the
4548 // other only from stmt2..stmtn.
4549 Node* dom = nullptr;
4550 for (uint i = 1; i < region->req(); ++i) {
4551 if (i != inner) {
4552 _igvn.replace_input_of(region, i, C->top());
4553 }
4554 Node* in = region_clone->in(i);
4555 if (in->is_top()) {
4556 continue;
4557 }
4558 if (dom == nullptr) {
4559 dom = in;
4560 } else {
4561 dom = dom_lca(dom, in);
4562 }
4563 }
4564
4565 set_idom(region_clone, dom, dd);
4566
4567 // Set up the outer loop
4568 for (uint i = 0; i < head->outcnt(); i++) {
4569 Node* u = head->raw_out(i);
4570 if (u->is_Phi()) {
4571 Node* outer_phi = u->clone();
4572 outer_phi->set_req(0, outer_head);
4573 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4574 if (backedge == nullptr) {
4575 backedge = u->in(LoopNode::LoopBackControl);
4576 }
4577 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4578 register_new_node(outer_phi, outer_head);
4579 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4580 }
4581 }
4582
4583 // create control and data nodes for out of loop uses (including region2)
4584 Node_List worklist;
4585 uint new_counter = C->unique();
4586 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4587
4588 Node_List *split_if_set = nullptr;
4589 Node_List *split_bool_set = nullptr;
4590 Node_List *split_cex_set = nullptr;
4591 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4592 split_if_set, split_bool_set, split_cex_set);
4593
4594 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4595
4596 if (exit_test != nullptr) {
4597 float cnt = exit_test->_fcnt;
4598 if (cnt != COUNT_UNKNOWN) {
4599 exit_test->_fcnt = cnt * f;
4600 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4601 }
4602 }
4603
4604 C->set_major_progress();
4605
4606 return true;
4607 }
4608
4609 // AutoVectorize the loop: replace scalar ops with vector ops.
4610 PhaseIdealLoop::AutoVectorizeStatus
4611 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4612 // Counted loop only
4613 if (!lpt->is_counted()) {
4614 return AutoVectorizeStatus::Impossible;
4615 }
4616
4617 // Main-loop only
4618 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4619 if (!cl->is_main_loop()) {
4620 return AutoVectorizeStatus::Impossible;
4621 }
4622
4623 VLoop vloop(lpt, false);
4624 if (!vloop.check_preconditions()) {
4625 return AutoVectorizeStatus::TriedAndFailed;
4626 }
4627
4628 // Ensure the shared data is cleared before each use
4629 vshared.clear();
4630
4631 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4632 if (!vloop_analyzer.success()) {
4633 return AutoVectorizeStatus::TriedAndFailed;
4634 }
4635
4636 SuperWord sw(vloop_analyzer);
4637 if (!sw.transform_loop()) {
4638 return AutoVectorizeStatus::TriedAndFailed;
4639 }
4640
4641 return AutoVectorizeStatus::Success;
4642 }
4643
4644 // Just before insert_pre_post_loops, we can multiversion the loop:
4645 //
4646 // multiversion_if
4647 // | |
4648 // fast_loop slow_loop
4649 //
4650 // In the fast_loop we can make speculative assumptions, and put the
4651 // conditions into the multiversion_if. If the conditions hold at runtime,
4652 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4653 // instead which does not make any of the speculative assumptions.
4654 //
4655 // Note: we only multiversion the loop if the loop does not have any
4656 // auto vectorization check Predicate. If we have that predicate,
4657 // then we can simply add the speculative assumption checks to
4658 // that Predicate. This means we do not need to duplicate the
4659 // loop - we have a smaller graph and save compile time. Should
4660 // the conditions ever fail, then we deopt / trap at the Predicate
4661 // and recompile without that Predicate. At that point we will
4662 // multiversion the loop, so that we can still have speculative
4663 // runtime checks.
4664 //
4665 // We perform the multiversioning when the loop is still in its single
4666 // iteration form, even before we insert pre and post loops. This makes
4667 // the cloning much simpler. However, this means that both the fast
4668 // and the slow loop have to be optimized independently (adding pre
4669 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4670 // we may end up not needing any speculative assumptions in the fast_loop
4671 // and then rejecting the slow_loop by constant folding the multiversion_if.
4672 //
4673 // Therefore, we "delay" the optimization of the slow_loop until we add
4674 // at least one speculative assumption for the fast_loop. If we never
4675 // add such a speculative runtime check, the OpaqueMultiversioningNode
4676 // of the multiversion_if constant folds to true after loop opts, and the
4677 // multiversion_if folds away the "delayed" slow_loop. If we add any
4678 // speculative assumption, then we notify the OpaqueMultiversioningNode
4679 // with "notify_slow_loop_that_it_can_resume_optimizations".
4680 //
4681 // Note: new runtime checks can be added to the multiversion_if with
4682 // PhaseIdealLoop::create_new_if_for_multiversion
4683 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4684 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4685 LoopNode* outer_loop = cl->skip_strip_mined();
4686 Node* entry = outer_loop->in(LoopNode::EntryControl);
4687
4688 // Check we have multiversioning enabled, and are not already multiversioned.
4689 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4690
4691 // Check that we do not have a parse-predicate where we can add the runtime checks
4692 // during auto-vectorization.
4693 const Predicates predicates(entry);
4694 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4695 if (predicate_block->has_parse_predicate()) { return; }
4696
4697 // Check node budget.
4698 uint estimate = lpt->est_loop_clone_sz(2);
4699 if (!may_require_nodes(estimate)) { return; }
4700
4701 do_multiversioning(lpt, old_new);
4702 }
4703
4704 // Returns true if the Reduction node is unordered.
4705 static bool is_unordered_reduction(Node* n) {
4706 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4707 }
4708
4709 // Having ReductionNodes in the loop is expensive. They need to recursively
4710 // fold together the vector values, for every vectorized loop iteration. If
4711 // we encounter the following pattern, we can vector accumulate the values
4712 // inside the loop, and only have a single UnorderedReduction after the loop.
4713 //
4714 // Note: UnorderedReduction represents a ReductionNode which does not require
4715 // calculating in strict order.
4716 //
4717 // CountedLoop init
4718 // | |
4719 // +------+ | +-----------------------+
4720 // | | | |
4721 // PhiNode (s) |
4722 // | |
4723 // | Vector |
4724 // | | |
4725 // UnorderedReduction (first_ur) |
4726 // | |
4727 // ... Vector |
4728 // | | |
4729 // UnorderedReduction (last_ur) |
4730 // | |
4731 // +---------------------+
4732 //
4733 // We patch the graph to look like this:
4734 //
4735 // CountedLoop identity_vector
4736 // | |
4737 // +-------+ | +---------------+
4738 // | | | |
4739 // PhiNode (v) |
4740 // | |
4741 // | Vector |
4742 // | | |
4743 // VectorAccumulator |
4744 // | |
4745 // ... Vector |
4746 // | | |
4747 // init VectorAccumulator |
4748 // | | | |
4749 // UnorderedReduction +-----------+
4750 //
4751 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4752 // use vector_accumulators, which do the same reductions, but only element
4753 // wise. This is a single operation per vector_accumulator, rather than many
4754 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4755 // after the loop, and also reduce the init value into it.
4756 //
4757 // We can not do this with all reductions. Some reductions do not allow the
4758 // reordering of operations (for example float addition/multiplication require
4759 // strict order).
4760 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4761 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4762
4763 // Find all Phi nodes with an unordered Reduction on backedge.
4764 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4765 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4766 Node* phi = cl->fast_out(j);
4767 // We have a phi with a single use, and an unordered Reduction on the backedge.
4768 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4769 continue;
4770 }
4771
4772 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4773 assert(!last_ur->requires_strict_order(), "must be");
4774
4775 // Determine types
4776 const TypeVect* vec_t = last_ur->vect_type();
4777 uint vector_length = vec_t->length();
4778 BasicType bt = vec_t->element_basic_type();
4779
4780 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4781 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4782 const int vopc = VectorNode::opcode(sopc, bt);
4783 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4784 DEBUG_ONLY( last_ur->dump(); )
4785 assert(false, "do not have normal vector op for this reduction");
4786 continue; // not implemented -> fails
4787 }
4788
4789 // Traverse up the chain of unordered Reductions, checking that it loops back to
4790 // the phi. Check that all unordered Reductions only have a single use, except for
4791 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4792 // are outside the loop.
4793 ReductionNode* current = last_ur;
4794 ReductionNode* first_ur = nullptr;
4795 while (true) {
4796 assert(!current->requires_strict_order(), "sanity");
4797
4798 // Expect no ctrl and a vector_input from within the loop.
4799 Node* ctrl = current->in(0);
4800 Node* vector_input = current->in(2);
4801 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4802 DEBUG_ONLY( current->dump(1); )
4803 assert(false, "reduction has ctrl or bad vector_input");
4804 break; // Chain traversal fails.
4805 }
4806
4807 assert(current->vect_type() != nullptr, "must have vector type");
4808 if (current->vect_type() != last_ur->vect_type()) {
4809 // Reductions do not have the same vector type (length and element type).
4810 break; // Chain traversal fails.
4811 }
4812
4813 // Expect single use of an unordered Reduction, except for last_ur.
4814 if (current == last_ur) {
4815 // Expect all uses to be outside the loop, except phi.
4816 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4817 Node* use = current->fast_out(k);
4818 if (use != phi && ctrl_or_self(use) == cl) {
4819 DEBUG_ONLY( current->dump(-1); )
4820 assert(false, "reduction has use inside loop");
4821 // Should not be allowed by SuperWord::mark_reductions
4822 return; // bail out of optimization
4823 }
4824 }
4825 } else {
4826 if (current->outcnt() != 1) {
4827 break; // Chain traversal fails.
4828 }
4829 }
4830
4831 // Expect another unordered Reduction or phi as the scalar input.
4832 Node* scalar_input = current->in(1);
4833 if (is_unordered_reduction(scalar_input) &&
4834 scalar_input->Opcode() == current->Opcode()) {
4835 // Move up the unordered Reduction chain.
4836 current = scalar_input->as_Reduction();
4837 assert(!current->requires_strict_order(), "must be");
4838 } else if (scalar_input == phi) {
4839 // Chain terminates at phi.
4840 first_ur = current;
4841 current = nullptr;
4842 break; // Success.
4843 } else {
4844 // scalar_input is neither phi nor a matching reduction
4845 // Can for example be scalar reduction when we have
4846 // partial vectorization.
4847 break; // Chain traversal fails.
4848 }
4849 }
4850 if (current != nullptr) {
4851 // Chain traversal was not successful.
4852 continue;
4853 }
4854 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4855
4856 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4857 set_root_as_ctrl(identity_scalar);
4858 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4859 register_new_node(identity_vector, C->root());
4860 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4861 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4862
4863 // Turn the scalar phi into a vector phi.
4864 _igvn.rehash_node_delayed(phi);
4865 Node* init = phi->in(1); // Remember init before replacing it.
4866 phi->set_req_X(1, identity_vector, &_igvn);
4867 phi->as_Type()->set_type(vec_t);
4868 _igvn.set_type(phi, vec_t);
4869
4870 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4871 current = first_ur;
4872 while (true) {
4873 // Create vector_accumulator to replace current.
4874 Node* last_vector_accumulator = current->in(1);
4875 Node* vector_input = current->in(2);
4876 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4877 register_new_node(vector_accumulator, cl);
4878 _igvn.replace_node(current, vector_accumulator);
4879 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4880 if (current == last_ur) {
4881 break;
4882 }
4883 current = vector_accumulator->unique_out()->as_Reduction();
4884 assert(!current->requires_strict_order(), "must be");
4885 }
4886
4887 // Create post-loop reduction.
4888 Node* last_accumulator = phi->in(2);
4889 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4890
4891 // Take over uses of last_accumulator that are not in the loop.
4892 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4893 Node* use = last_accumulator->out(i);
4894 if (use != phi && use != post_loop_reduction) {
4895 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4896 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4897 --i;
4898 }
4899 }
4900 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4901 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4902
4903 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4904 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4905 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4906 }
4907 }
4908
4909 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4910 for (uint i = 0; i < _data_nodes.size(); i++) {
4911 clone(_data_nodes[i], new_ctrl);
4912 }
4913 }
4914
4915 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4916 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4917 Node* clone = node->clone();
4918 _phase->igvn().register_new_node_with_optimizer(clone);
4919 _orig_to_new.put(node, clone);
4920 _phase->set_ctrl(clone, new_ctrl);
4921 if (node->is_CastII()) {
4922 clone->set_req(0, new_ctrl);
4923 }
4924 }
4925
4926 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4927 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4928 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4929 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4930 for (uint i = 1; i < node->req(); i++) {
4931 Node** cloned_input = _orig_to_new.get(node->in(i));
4932 if (cloned_input != nullptr) {
4933 // Input was also cloned -> rewire clone to the cloned input.
4934 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4935 }
4936 }
4937 });
4938 }
4939
4940 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4941 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4942 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4943 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4944 Node* new_ctrl) {
4945 for (uint i = 0; i < _data_nodes.size(); i++) {
4946 Node* data_node = _data_nodes[i];
4947 if (data_node->is_Opaque1()) {
4948 transform_opaque_node(transform_strategy, data_node);
4949 } else {
4950 clone(data_node, new_ctrl);
4951 }
4952 }
4953 }
4954
4955 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4956 Node* transformed_node;
4957 if (node->is_OpaqueLoopInit()) {
4958 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4959 } else {
4960 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4961 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4962 }
4963 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4964 _orig_to_new.put(node, transformed_node);
4965 }