1 /*
2 * Copyright (c) 1999, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/divnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "opto/vectornode.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
52 // ConvI2L may have type information on it which is unsafe to push up
53 // so disable this for now
54 return nullptr;
55 }
56
57 // Splitting range check CastIIs through a loop induction Phi can
58 // cause new Phis to be created that are left unrelated to the loop
59 // induction Phi and prevent optimizations (vectorization)
60 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
61 n->in(1) == region->as_CountedLoop()->phi()) {
62 return nullptr;
63 }
64
65 if (cannot_split_division(n, region)) {
66 return nullptr;
67 }
68
69 int wins = 0;
70 assert(!n->is_CFG(), "");
71 assert(region->is_Region(), "");
72
73 const Type* type = n->bottom_type();
74 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
75 Node* phi;
76 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
77 int iid = t_oop->instance_id();
78 int index = C->get_alias_index(t_oop);
79 int offset = t_oop->offset();
80 phi = new PhiNode(region, type, nullptr, iid, index, offset);
81 } else {
82 phi = PhiNode::make_blank(region, n);
83 }
84 uint old_unique = C->unique();
85 for (uint i = 1; i < region->req(); i++) {
86 Node* x;
87 Node* the_clone = nullptr;
88 if (region->in(i) == C->top()) {
89 x = C->top(); // Dead path? Use a dead data op
90 } else {
91 x = n->clone(); // Else clone up the data op
92 the_clone = x; // Remember for possible deletion.
93 // Alter data node to use pre-phi inputs
94 if (n->in(0) == region)
95 x->set_req( 0, region->in(i) );
96 for (uint j = 1; j < n->req(); j++) {
97 Node* in = n->in(j);
98 if (in->is_Phi() && in->in(0) == region)
99 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
100 }
101 }
102 // Check for a 'win' on some paths
103 const Type* t = x->Value(&_igvn);
104
105 bool singleton = t->singleton();
106
107 // A TOP singleton indicates that there are no possible values incoming
108 // along a particular edge. In most cases, this is OK, and the Phi will
109 // be eliminated later in an Ideal call. However, we can't allow this to
110 // happen if the singleton occurs on loop entry, as the elimination of
111 // the PhiNode may cause the resulting node to migrate back to a previous
112 // loop iteration.
113 if (singleton && t == Type::TOP) {
114 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
115 // irreducible loop may not be indicated by an affirmative is_Loop());
116 // therefore, the only top we can split thru a phi is on a backedge of
117 // a loop.
118 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
119 }
120
121 if (singleton) {
122 wins++;
123 x = ((PhaseGVN&)_igvn).makecon(t);
124 } else {
125 // We now call Identity to try to simplify the cloned node.
126 // Note that some Identity methods call phase->type(this).
127 // Make sure that the type array is big enough for
128 // our new node, even though we may throw the node away.
129 // (Note: This tweaking with igvn only works because x is a new node.)
130 _igvn.set_type(x, t);
131 // If x is a TypeNode, capture any more-precise type permanently into Node
132 // otherwise it will be not updated during igvn->transform since
133 // igvn->type(x) is set to x->Value() already.
134 x->raise_bottom_type(t);
135 Node* y = x->Identity(&_igvn);
136 if (y != x) {
137 wins++;
138 x = y;
139 } else {
140 y = _igvn.hash_find(x);
141 if (y) {
142 wins++;
143 x = y;
144 } else {
145 // Else x is a new node we are keeping
146 // We do not need register_new_node_with_optimizer
147 // because set_type has already been called.
148 _igvn._worklist.push(x);
149 }
150 }
151 }
152 if (x != the_clone && the_clone != nullptr)
153 _igvn.remove_dead_node(the_clone);
154 phi->set_req( i, x );
155 }
156 // Too few wins?
157 if (wins <= policy) {
158 _igvn.remove_dead_node(phi);
159 return nullptr;
160 }
161
162 // Record Phi
163 register_new_node( phi, region );
164
165 for (uint i2 = 1; i2 < phi->req(); i2++) {
166 Node *x = phi->in(i2);
167 // If we commoned up the cloned 'x' with another existing Node,
168 // the existing Node picks up a new use. We need to make the
169 // existing Node occur higher up so it dominates its uses.
170 Node *old_ctrl;
171 IdealLoopTree *old_loop;
172
173 if (x->is_Con()) {
174 // Constant's control is always root.
175 set_ctrl(x, C->root());
176 continue;
177 }
178 // The occasional new node
179 if (x->_idx >= old_unique) { // Found a new, unplaced node?
180 old_ctrl = nullptr;
181 old_loop = nullptr; // Not in any prior loop
182 } else {
183 old_ctrl = get_ctrl(x);
184 old_loop = get_loop(old_ctrl); // Get prior loop
185 }
186 // New late point must dominate new use
187 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
188 if (new_ctrl == old_ctrl) // Nothing is changed
189 continue;
190
191 IdealLoopTree *new_loop = get_loop(new_ctrl);
192
193 // Don't move x into a loop if its uses are
194 // outside of loop. Otherwise x will be cloned
195 // for each use outside of this loop.
196 IdealLoopTree *use_loop = get_loop(region);
197 if (!new_loop->is_member(use_loop) &&
198 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
199 // Take early control, later control will be recalculated
200 // during next iteration of loop optimizations.
201 new_ctrl = get_early_ctrl(x);
202 new_loop = get_loop(new_ctrl);
203 }
204 // Set new location
205 set_ctrl(x, new_ctrl);
206 // If changing loop bodies, see if we need to collect into new body
207 if (old_loop != new_loop) {
208 if (old_loop && !old_loop->_child)
209 old_loop->_body.yank(x);
210 if (!new_loop->_child)
211 new_loop->_body.push(x); // Collect body info
212 }
213 }
214
215 return phi;
216 }
217
218 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
219 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
220 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
221 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
222 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
223 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
224 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
225 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
226 const Type* zero;
227 switch (n->Opcode()) {
228 case Op_DivI:
229 case Op_ModI:
230 zero = TypeInt::ZERO;
231 break;
232 case Op_DivL:
233 case Op_ModL:
234 zero = TypeLong::ZERO;
235 break;
236 default:
237 return false;
238 }
239
240 assert(n->in(0) == nullptr, "divisions with zero check should already have bailed out earlier in split-if");
241 Node* divisor = n->in(2);
242 return is_divisor_counted_loop_phi(divisor, region) &&
243 loop_phi_backedge_type_contains_zero(divisor, zero);
244 }
245
246 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) {
247 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop;
248 }
249
250 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
251 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
252 }
253
254 //------------------------------dominated_by------------------------------------
255 // Replace the dominated test with an obvious true or false. Place it on the
256 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
257 // live path up to the dominating control.
258 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) {
259 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
260
261 // prevdom is the dominating projection of the dominating test.
262 assert(iff->Opcode() == Op_If ||
263 iff->Opcode() == Op_CountedLoopEnd ||
264 iff->Opcode() == Op_LongCountedLoopEnd ||
265 iff->Opcode() == Op_RangeCheck ||
266 iff->Opcode() == Op_ParsePredicate,
267 "Check this code when new subtype is added");
268
269 int pop = prevdom->Opcode();
270 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
271 if (flip) {
272 if (pop == Op_IfTrue)
273 pop = Op_IfFalse;
274 else
275 pop = Op_IfTrue;
276 }
277 // 'con' is set to true or false to kill the dominated test.
278 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
279 set_ctrl(con, C->root()); // Constant gets a new use
280 // Hack the dominated test
281 _igvn.replace_input_of(iff, 1, con);
282
283 // If I dont have a reachable TRUE and FALSE path following the IfNode then
284 // I can assume this path reaches an infinite loop. In this case it's not
285 // important to optimize the data Nodes - either the whole compilation will
286 // be tossed or this path (and all data Nodes) will go dead.
287 if (iff->outcnt() != 2) return;
288
289 // Make control-dependent data Nodes on the live path (path that will remain
290 // once the dominated IF is removed) become control-dependent on the
291 // dominating projection.
292 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
293
294 // Loop predicates may have depending checks which should not
295 // be skipped. For example, range check predicate has two checks
296 // for lower and upper bounds.
297 if (dp == nullptr)
298 return;
299
300 ProjNode* dp_proj = dp->as_Proj();
301 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj();
302 if (exclude_loop_predicate &&
303 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr ||
304 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr ||
305 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != nullptr)) {
306 // If this is a range check (IfNode::is_range_check), do not
307 // reorder because Compile::allow_range_check_smearing might have
308 // changed the check.
309 return; // Let IGVN transformation change control dependence.
310 }
311
312 IdealLoopTree* old_loop = get_loop(dp);
313
314 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
315 Node* cd = dp->fast_out(i); // Control-dependent node
316 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
317 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) {
318 assert(cd->in(0) == dp, "");
319 _igvn.replace_input_of(cd, 0, prevdom);
320 set_early_ctrl(cd, false);
321 IdealLoopTree* new_loop = get_loop(get_ctrl(cd));
322 if (old_loop != new_loop) {
323 if (!old_loop->_child) {
324 old_loop->_body.yank(cd);
325 }
326 if (!new_loop->_child) {
327 new_loop->_body.push(cd);
328 }
329 }
330 --i;
331 --imax;
332 }
333 }
334 }
335
336 //------------------------------has_local_phi_input----------------------------
337 // Return TRUE if 'n' has Phi inputs from its local block and no other
338 // block-local inputs (all non-local-phi inputs come from earlier blocks)
339 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
340 Node *n_ctrl = get_ctrl(n);
341 // See if some inputs come from a Phi in this block, or from before
342 // this block.
343 uint i;
344 for( i = 1; i < n->req(); i++ ) {
345 Node *phi = n->in(i);
346 if( phi->is_Phi() && phi->in(0) == n_ctrl )
347 break;
348 }
349 if( i >= n->req() )
350 return nullptr; // No Phi inputs; nowhere to clone thru
351
352 // Check for inputs created between 'n' and the Phi input. These
353 // must split as well; they have already been given the chance
354 // (courtesy of a post-order visit) and since they did not we must
355 // recover the 'cost' of splitting them by being very profitable
356 // when splitting 'n'. Since this is unlikely we simply give up.
357 for( i = 1; i < n->req(); i++ ) {
358 Node *m = n->in(i);
359 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
360 // We allow the special case of AddP's with no local inputs.
361 // This allows us to split-up address expressions.
362 if (m->is_AddP() &&
363 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
364 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
365 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
366 // Move the AddP up to the dominating point. That's fine because control of m's inputs
367 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
368 Node* c = find_non_split_ctrl(idom(n_ctrl));
369 if (c->is_OuterStripMinedLoop()) {
370 c->as_Loop()->verify_strip_mined(1);
371 c = c->in(LoopNode::EntryControl);
372 }
373 set_ctrl_and_loop(m, c);
374 continue;
375 }
376 return nullptr;
377 }
378 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
379 }
380
381 return n_ctrl;
382 }
383
384 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
385 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
386 assert(bt == T_INT || bt == T_LONG, "only for integers");
387 int n_op = n->Opcode();
388
389 if (n_op == Op_LShift(bt)) {
390 // Scale is loop invariant
391 Node* scale = n->in(2);
392 Node* scale_ctrl = get_ctrl(scale);
393 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
394 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
395 return nullptr;
396 }
397 const TypeInt* scale_t = scale->bottom_type()->isa_int();
398 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
399 return nullptr; // Dont bother with byte/short masking
400 }
401 // Add must vary with loop (else shift would be loop-invariant)
402 Node* add = n->in(1);
403 Node* add_ctrl = get_ctrl(add);
404 IdealLoopTree* add_loop = get_loop(add_ctrl);
405 if (n_loop != add_loop) {
406 return nullptr; // happens w/ evil ZKM loops
407 }
408
409 // Convert I-V into I+ (0-V); same for V-I
410 if (add->Opcode() == Op_Sub(bt) &&
411 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
412 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
413 Node* zero = _igvn.integercon(0, bt);
414 set_ctrl(zero, C->root());
415 Node* neg = SubNode::make(zero, add->in(2), bt);
416 register_new_node(neg, get_ctrl(add->in(2)));
417 add = AddNode::make(add->in(1), neg, bt);
418 register_new_node(add, add_ctrl);
419 }
420 if (add->Opcode() != Op_Add(bt)) return nullptr;
421 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
422 // See if one add input is loop invariant
423 Node* add_var = add->in(1);
424 Node* add_var_ctrl = get_ctrl(add_var);
425 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
426 Node* add_invar = add->in(2);
427 Node* add_invar_ctrl = get_ctrl(add_invar);
428 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
429 if (add_invar_loop == n_loop) {
430 // Swap to find the invariant part
431 add_invar = add_var;
432 add_invar_ctrl = add_var_ctrl;
433 add_invar_loop = add_var_loop;
434 add_var = add->in(2);
435 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
436 return nullptr;
437 }
438 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
439 return nullptr; // No invariant part of the add?
440 }
441
442 // Yes! Reshape address expression!
443 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
444 Node* inv_scale_ctrl =
445 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
446 add_invar_ctrl : scale_ctrl;
447 register_new_node(inv_scale, inv_scale_ctrl);
448 Node* var_scale = LShiftNode::make(add_var, scale, bt);
449 register_new_node(var_scale, n_ctrl);
450 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
451 register_new_node(var_add, n_ctrl);
452 _igvn.replace_node(n, var_add);
453 return var_add;
454 }
455 return nullptr;
456 }
457
458 //------------------------------remix_address_expressions----------------------
459 // Rework addressing expressions to get the most loop-invariant stuff
460 // moved out. We'd like to do all associative operators, but it's especially
461 // important (common) to do address expressions.
462 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
463 if (!has_ctrl(n)) return nullptr;
464 Node* n_ctrl = get_ctrl(n);
465 IdealLoopTree* n_loop = get_loop(n_ctrl);
466
467 // See if 'n' mixes loop-varying and loop-invariant inputs and
468 // itself is loop-varying.
469
470 // Only interested in binary ops (and AddP)
471 if (n->req() < 3 || n->req() > 4) return nullptr;
472
473 Node* n1_ctrl = get_ctrl(n->in( 1));
474 Node* n2_ctrl = get_ctrl(n->in( 2));
475 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
476 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
477 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
478 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
479
480 // Does one of my inputs spin in a tighter loop than self?
481 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
482 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
483 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
484 return nullptr; // Leave well enough alone
485 }
486
487 // Is at least one of my inputs loop-invariant?
488 if (n1_loop == n_loop &&
489 n2_loop == n_loop &&
490 n3_loop == n_loop) {
491 return nullptr; // No loop-invariant inputs
492 }
493
494 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
495 if (res != nullptr) {
496 return res;
497 }
498 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
499 if (res != nullptr) {
500 return res;
501 }
502
503 int n_op = n->Opcode();
504 // Replace (I+V) with (V+I)
505 if (n_op == Op_AddI ||
506 n_op == Op_AddL ||
507 n_op == Op_AddF ||
508 n_op == Op_AddD ||
509 n_op == Op_MulI ||
510 n_op == Op_MulL ||
511 n_op == Op_MulF ||
512 n_op == Op_MulD) {
513 if (n2_loop == n_loop) {
514 assert(n1_loop != n_loop, "");
515 n->swap_edges(1, 2);
516 }
517 }
518
519 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
520 // but not if I2 is a constant.
521 if (n_op == Op_AddP) {
522 if (n2_loop == n_loop && n3_loop != n_loop) {
523 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
524 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
525 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
526 IdealLoopTree* n22loop = get_loop(n22_ctrl);
527 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
528 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
529 n23_loop == n_loop) {
530 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
531 // Stuff new AddP in the loop preheader
532 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
533 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
534 register_new_node(add2, n_ctrl);
535 _igvn.replace_node(n, add2);
536 return add2;
537 }
538 }
539 }
540
541 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
542 if (n2_loop != n_loop && n3_loop == n_loop) {
543 if (n->in(3)->Opcode() == Op_AddX) {
544 Node* V = n->in(3)->in(1);
545 Node* I = n->in(3)->in(2);
546 if (is_member(n_loop,get_ctrl(V))) {
547 } else {
548 Node *tmp = V; V = I; I = tmp;
549 }
550 if (!is_member(n_loop,get_ctrl(I))) {
551 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
552 // Stuff new AddP in the loop preheader
553 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
554 Node* add2 = new AddPNode(n->in(1), add1, V);
555 register_new_node(add2, n_ctrl);
556 _igvn.replace_node(n, add2);
557 return add2;
558 }
559 }
560 }
561 }
562
563 return nullptr;
564 }
565
566 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
567 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
568 assert(n->Opcode() == Op_AddI, "sanity");
569 Node * nn = nullptr;
570 Node * in1 = n->in(1);
571 Node * in2 = n->in(2);
572 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
573 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
574 if (loop_n->is_counted() &&
575 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
576 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
577 Matcher::match_rule_supported(Op_MulAddS2I)) {
578 Node* mul_in1 = in1->in(1);
579 Node* mul_in2 = in1->in(2);
580 Node* mul_in3 = in2->in(1);
581 Node* mul_in4 = in2->in(2);
582 if (mul_in1->Opcode() == Op_LoadS &&
583 mul_in2->Opcode() == Op_LoadS &&
584 mul_in3->Opcode() == Op_LoadS &&
585 mul_in4->Opcode() == Op_LoadS) {
586 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
587 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
588 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
589 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
590 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
591 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
592 // All nodes should be in the same counted loop.
593 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
594 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
595 Node* adr1 = mul_in1->in(MemNode::Address);
596 Node* adr2 = mul_in2->in(MemNode::Address);
597 Node* adr3 = mul_in3->in(MemNode::Address);
598 Node* adr4 = mul_in4->in(MemNode::Address);
599 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
600 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
601 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
602 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
603 register_new_node(nn, get_ctrl(n));
604 _igvn.replace_node(n, nn);
605 return nn;
606 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
607 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
608 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
609 register_new_node(nn, get_ctrl(n));
610 _igvn.replace_node(n, nn);
611 return nn;
612 }
613 }
614 }
615 }
616 }
617 }
618 return nn;
619 }
620
621 //------------------------------conditional_move-------------------------------
622 // Attempt to replace a Phi with a conditional move. We have some pretty
623 // strict profitability requirements. All Phis at the merge point must
624 // be converted, so we can remove the control flow. We need to limit the
625 // number of c-moves to a small handful. All code that was in the side-arms
626 // of the CFG diamond is now speculatively executed. This code has to be
627 // "cheap enough". We are pretty much limited to CFG diamonds that merge
628 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
629 Node *PhaseIdealLoop::conditional_move( Node *region ) {
630
631 assert(region->is_Region(), "sanity check");
632 if (region->req() != 3) return nullptr;
633
634 // Check for CFG diamond
635 Node *lp = region->in(1);
636 Node *rp = region->in(2);
637 if (!lp || !rp) return nullptr;
638 Node *lp_c = lp->in(0);
639 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
640 IfNode *iff = lp_c->as_If();
641
642 // Check for ops pinned in an arm of the diamond.
643 // Can't remove the control flow in this case
644 if (lp->outcnt() > 1) return nullptr;
645 if (rp->outcnt() > 1) return nullptr;
646
647 IdealLoopTree* r_loop = get_loop(region);
648 assert(r_loop == get_loop(iff), "sanity");
649 // Always convert to CMOVE if all results are used only outside this loop.
650 bool used_inside_loop = (r_loop == _ltree_root);
651
652 // Check profitability
653 int cost = 0;
654 int phis = 0;
655 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
656 Node *out = region->fast_out(i);
657 if (!out->is_Phi()) continue; // Ignore other control edges, etc
658 phis++;
659 PhiNode* phi = out->as_Phi();
660 BasicType bt = phi->type()->basic_type();
661 switch (bt) {
662 case T_DOUBLE:
663 case T_FLOAT:
664 if (C->use_cmove()) {
665 continue; //TODO: maybe we want to add some cost
666 }
667 cost += Matcher::float_cmove_cost(); // Could be very expensive
668 break;
669 case T_LONG: {
670 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
671 }
672 case T_INT: // These all CMOV fine
673 case T_ADDRESS: { // (RawPtr)
674 cost++;
675 break;
676 }
677 case T_NARROWOOP: // Fall through
678 case T_OBJECT: { // Base oops are OK, but not derived oops
679 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
680 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
681 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
682 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
683 // have a Phi for the base here that we convert to a CMOVE all is well
684 // and good. But if the base is dead, we'll not make a CMOVE. Later
685 // the allocator will have to produce a base by creating a CMOVE of the
686 // relevant bases. This puts the allocator in the business of
687 // manufacturing expensive instructions, generally a bad plan.
688 // Just Say No to Conditionally-Moved Derived Pointers.
689 if (tp && tp->offset() != 0)
690 return nullptr;
691 cost++;
692 break;
693 }
694 default:
695 return nullptr; // In particular, can't do memory or I/O
696 }
697 // Add in cost any speculative ops
698 for (uint j = 1; j < region->req(); j++) {
699 Node *proj = region->in(j);
700 Node *inp = phi->in(j);
701 if (get_ctrl(inp) == proj) { // Found local op
702 cost++;
703 // Check for a chain of dependent ops; these will all become
704 // speculative in a CMOV.
705 for (uint k = 1; k < inp->req(); k++)
706 if (get_ctrl(inp->in(k)) == proj)
707 cost += ConditionalMoveLimit; // Too much speculative goo
708 }
709 }
710 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
711 // This will likely Split-If, a higher-payoff operation.
712 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
713 Node* use = phi->fast_out(k);
714 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
715 cost += ConditionalMoveLimit;
716 // Is there a use inside the loop?
717 // Note: check only basic types since CMoveP is pinned.
718 if (!used_inside_loop && is_java_primitive(bt)) {
719 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
720 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
721 used_inside_loop = true;
722 }
723 }
724 }
725 }//for
726 Node* bol = iff->in(1);
727 if (bol->Opcode() == Op_Opaque4) {
728 return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
729 }
730 assert(bol->Opcode() == Op_Bool, "Unexpected node");
731 int cmp_op = bol->in(1)->Opcode();
732 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
733 return nullptr;
734 }
735 // It is expensive to generate flags from a float compare.
736 // Avoid duplicated float compare.
737 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
738
739 float infrequent_prob = PROB_UNLIKELY_MAG(3);
740 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
741 if (used_inside_loop) {
742 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
743
744 // BlockLayoutByFrequency optimization moves infrequent branch
745 // from hot path. No point in CMOV'ing in such case (110 is used
746 // instead of 100 to take into account not exactness of float value).
747 if (BlockLayoutByFrequency) {
748 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
749 }
750 }
751 // Check for highly predictable branch. No point in CMOV'ing if
752 // we are going to predict accurately all the time.
753 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
754 //keep going
755 } else if (iff->_prob < infrequent_prob ||
756 iff->_prob > (1.0f - infrequent_prob))
757 return nullptr;
758
759 // --------------
760 // Now replace all Phis with CMOV's
761 Node *cmov_ctrl = iff->in(0);
762 uint flip = (lp->Opcode() == Op_IfTrue);
763 Node_List wq;
764 while (1) {
765 PhiNode* phi = nullptr;
766 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
767 Node *out = region->fast_out(i);
768 if (out->is_Phi()) {
769 phi = out->as_Phi();
770 break;
771 }
772 }
773 if (phi == nullptr || _igvn.type(phi) == Type::TOP) {
774 break;
775 }
776 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
777 // Move speculative ops
778 wq.push(phi);
779 while (wq.size() > 0) {
780 Node *n = wq.pop();
781 for (uint j = 1; j < n->req(); j++) {
782 Node* m = n->in(j);
783 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
784 #ifndef PRODUCT
785 if (PrintOpto && VerifyLoopOptimizations) {
786 tty->print(" speculate: ");
787 m->dump();
788 }
789 #endif
790 set_ctrl(m, cmov_ctrl);
791 wq.push(m);
792 }
793 }
794 }
795 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
796 register_new_node( cmov, cmov_ctrl );
797 _igvn.replace_node( phi, cmov );
798 #ifndef PRODUCT
799 if (TraceLoopOpts) {
800 tty->print("CMOV ");
801 r_loop->dump_head();
802 if (Verbose) {
803 bol->in(1)->dump(1);
804 cmov->dump(1);
805 }
806 }
807 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
808 #endif
809 }
810
811 // The useless CFG diamond will fold up later; see the optimization in
812 // RegionNode::Ideal.
813 _igvn._worklist.push(region);
814
815 return iff->in(1);
816 }
817
818 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
819 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
820 Node* u = m->fast_out(i);
821 if (u->is_CFG()) {
822 if (u->is_NeverBranch()) {
823 u = u->as_NeverBranch()->proj_out(0);
824 enqueue_cfg_uses(u, wq);
825 } else {
826 wq.push(u);
827 }
828 }
829 }
830 }
831
832 // Try moving a store out of a loop, right before the loop
833 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
834 // Store has to be first in the loop body
835 IdealLoopTree *n_loop = get_loop(n_ctrl);
836 if (n->is_Store() && n_loop != _ltree_root &&
837 n_loop->is_loop() && n_loop->_head->is_Loop() &&
838 n->in(0) != nullptr) {
839 Node* address = n->in(MemNode::Address);
840 Node* value = n->in(MemNode::ValueIn);
841 Node* mem = n->in(MemNode::Memory);
842 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
843 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
844
845 // - address and value must be loop invariant
846 // - memory must be a memory Phi for the loop
847 // - Store must be the only store on this memory slice in the
848 // loop: if there's another store following this one then value
849 // written at iteration i by the second store could be overwritten
850 // at iteration i+n by the first store: it's not safe to move the
851 // first store out of the loop
852 // - nothing must observe the memory Phi: it guarantees no read
853 // before the store, we are also guaranteed the store post
854 // dominates the loop head (ignoring a possible early
855 // exit). Otherwise there would be extra Phi involved between the
856 // loop's Phi and the store.
857 // - there must be no early exit from the loop before the Store
858 // (such an exit most of the time would be an extra use of the
859 // memory Phi but sometimes is a bottom memory Phi that takes the
860 // store as input).
861
862 if (!n_loop->is_member(address_loop) &&
863 !n_loop->is_member(value_loop) &&
864 mem->is_Phi() && mem->in(0) == n_loop->_head &&
865 mem->outcnt() == 1 &&
866 mem->in(LoopNode::LoopBackControl) == n) {
867
868 assert(n_loop->_tail != nullptr, "need a tail");
869 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
870
871 // Verify that there's no early exit of the loop before the store.
872 bool ctrl_ok = false;
873 {
874 // Follow control from loop head until n, we exit the loop or
875 // we reach the tail
876 ResourceMark rm;
877 Unique_Node_List wq;
878 wq.push(n_loop->_head);
879
880 for (uint next = 0; next < wq.size(); ++next) {
881 Node *m = wq.at(next);
882 if (m == n->in(0)) {
883 ctrl_ok = true;
884 continue;
885 }
886 assert(!has_ctrl(m), "should be CFG");
887 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
888 ctrl_ok = false;
889 break;
890 }
891 enqueue_cfg_uses(m, wq);
892 if (wq.size() > 10) {
893 ctrl_ok = false;
894 break;
895 }
896 }
897 }
898 if (ctrl_ok) {
899 // move the Store
900 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
901 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
902 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
903 // Disconnect the phi now. An empty phi can confuse other
904 // optimizations in this pass of loop opts.
905 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
906 n_loop->_body.yank(mem);
907
908 set_ctrl_and_loop(n, n->in(0));
909
910 return n;
911 }
912 }
913 }
914 return nullptr;
915 }
916
917 // Try moving a store out of a loop, right after the loop
918 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
919 if (n->is_Store() && n->in(0) != nullptr) {
920 Node *n_ctrl = get_ctrl(n);
921 IdealLoopTree *n_loop = get_loop(n_ctrl);
922 // Store must be in a loop
923 if (n_loop != _ltree_root && !n_loop->_irreducible) {
924 Node* address = n->in(MemNode::Address);
925 Node* value = n->in(MemNode::ValueIn);
926 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
927 // address must be loop invariant
928 if (!n_loop->is_member(address_loop)) {
929 // Store must be last on this memory slice in the loop and
930 // nothing in the loop must observe it
931 Node* phi = nullptr;
932 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
933 Node* u = n->fast_out(i);
934 if (has_ctrl(u)) { // control use?
935 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
936 if (!n_loop->is_member(u_loop)) {
937 continue;
938 }
939 if (u->is_Phi() && u->in(0) == n_loop->_head) {
940 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
941 // multiple phis on the same slice are possible
942 if (phi != nullptr) {
943 return;
944 }
945 phi = u;
946 continue;
947 }
948 }
949 return;
950 }
951 if (phi != nullptr) {
952 // Nothing in the loop before the store (next iteration)
953 // must observe the stored value
954 bool mem_ok = true;
955 {
956 ResourceMark rm;
957 Unique_Node_List wq;
958 wq.push(phi);
959 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
960 Node *m = wq.at(next);
961 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
962 Node* u = m->fast_out(i);
963 if (u->is_Store() || u->is_Phi()) {
964 if (u != n) {
965 wq.push(u);
966 mem_ok = (wq.size() <= 10);
967 }
968 } else {
969 mem_ok = false;
970 break;
971 }
972 }
973 }
974 }
975 if (mem_ok) {
976 // Move the store out of the loop if the LCA of all
977 // users (except for the phi) is outside the loop.
978 Node* hook = new Node(1);
979 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
980 _igvn.rehash_node_delayed(phi);
981 int count = phi->replace_edge(n, hook, &_igvn);
982 assert(count > 0, "inconsistent phi");
983
984 // Compute latest point this store can go
985 Node* lca = get_late_ctrl(n, get_ctrl(n));
986 if (lca->is_OuterStripMinedLoop()) {
987 lca = lca->in(LoopNode::EntryControl);
988 }
989 if (n_loop->is_member(get_loop(lca))) {
990 // LCA is in the loop - bail out
991 _igvn.replace_node(hook, n);
992 return;
993 }
994 #ifdef ASSERT
995 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
996 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
997 n_loop->_head->as_Loop()->verify_strip_mined(1);
998 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
999 IdealLoopTree* outer_loop = get_loop(outer);
1000 assert(n_loop->_parent == outer_loop, "broken loop tree");
1001 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1002 }
1003 #endif
1004 lca = place_outside_loop(lca, n_loop);
1005 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1006 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop");
1007
1008 // Move store out of the loop
1009 _igvn.replace_node(hook, n->in(MemNode::Memory));
1010 _igvn.replace_input_of(n, 0, lca);
1011 set_ctrl_and_loop(n, lca);
1012
1013 // Disconnect the phi now. An empty phi can confuse other
1014 // optimizations in this pass of loop opts..
1015 if (phi->in(LoopNode::LoopBackControl) == phi) {
1016 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1017 n_loop->_body.yank(phi);
1018 }
1019 }
1020 }
1021 }
1022 }
1023 }
1024 }
1025
1026 //------------------------------split_if_with_blocks_pre-----------------------
1027 // Do the real work in a non-recursive function. Data nodes want to be
1028 // cloned in the pre-order so they can feed each other nicely.
1029 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1030 // Cloning these guys is unlikely to win
1031 int n_op = n->Opcode();
1032 if (n_op == Op_MergeMem) {
1033 return n;
1034 }
1035 if (n->is_Proj()) {
1036 return n;
1037 }
1038 // Do not clone-up CmpFXXX variations, as these are always
1039 // followed by a CmpI
1040 if (n->is_Cmp()) {
1041 return n;
1042 }
1043 // Attempt to use a conditional move instead of a phi/branch
1044 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1045 Node *cmov = conditional_move( n );
1046 if (cmov) {
1047 return cmov;
1048 }
1049 }
1050 if (n->is_CFG() || n->is_LoadStore()) {
1051 return n;
1052 }
1053 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1054 if (!C->major_progress()) { // If chance of no more loop opts...
1055 _igvn._worklist.push(n); // maybe we'll remove them
1056 }
1057 return n;
1058 }
1059
1060 if (n->is_Con()) {
1061 return n; // No cloning for Con nodes
1062 }
1063
1064 Node *n_ctrl = get_ctrl(n);
1065 if (!n_ctrl) {
1066 return n; // Dead node
1067 }
1068
1069 Node* res = try_move_store_before_loop(n, n_ctrl);
1070 if (res != nullptr) {
1071 return n;
1072 }
1073
1074 // Attempt to remix address expressions for loop invariants
1075 Node *m = remix_address_expressions( n );
1076 if( m ) return m;
1077
1078 if (n_op == Op_AddI) {
1079 Node *nn = convert_add_to_muladd( n );
1080 if ( nn ) return nn;
1081 }
1082
1083 if (n->is_ConstraintCast()) {
1084 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1085 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1086 // Node control inputs don't necessarily agree with loop control info (due to
1087 // transformations happened in between), thus additional dominance check is needed
1088 // to keep loop info valid.
1089 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1090 _igvn.replace_node(n, dom_cast);
1091 return dom_cast;
1092 }
1093 }
1094
1095 // Determine if the Node has inputs from some local Phi.
1096 // Returns the block to clone thru.
1097 Node *n_blk = has_local_phi_input( n );
1098 if( !n_blk ) return n;
1099
1100 // Do not clone the trip counter through on a CountedLoop
1101 // (messes up the canonical shape).
1102 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1103 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1104 return n;
1105 }
1106 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1107 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1108 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1109 return n;
1110 }
1111
1112 // Check for having no control input; not pinned. Allow
1113 // dominating control.
1114 if (n->in(0)) {
1115 Node *dom = idom(n_blk);
1116 if (dom_lca(n->in(0), dom) != n->in(0)) {
1117 return n;
1118 }
1119 }
1120 // Policy: when is it profitable. You must get more wins than
1121 // policy before it is considered profitable. Policy is usually 0,
1122 // so 1 win is considered profitable. Big merges will require big
1123 // cloning, so get a larger policy.
1124 int policy = n_blk->req() >> 2;
1125
1126 // If the loop is a candidate for range check elimination,
1127 // delay splitting through it's phi until a later loop optimization
1128 if (n_blk->is_BaseCountedLoop()) {
1129 IdealLoopTree *lp = get_loop(n_blk);
1130 if (lp && lp->_rce_candidate) {
1131 return n;
1132 }
1133 }
1134
1135 if (must_throttle_split_if()) return n;
1136
1137 // Split 'n' through the merge point if it is profitable
1138 Node *phi = split_thru_phi( n, n_blk, policy );
1139 if (!phi) return n;
1140
1141 // Found a Phi to split thru!
1142 // Replace 'n' with the new phi
1143 _igvn.replace_node( n, phi );
1144 // Moved a load around the loop, 'en-registering' something.
1145 if (n_blk->is_Loop() && n->is_Load() &&
1146 !phi->in(LoopNode::LoopBackControl)->is_Load())
1147 C->set_major_progress();
1148
1149 return phi;
1150 }
1151
1152 static bool merge_point_too_heavy(Compile* C, Node* region) {
1153 // Bail out if the region and its phis have too many users.
1154 int weight = 0;
1155 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1156 weight += region->fast_out(i)->outcnt();
1157 }
1158 int nodes_left = C->max_node_limit() - C->live_nodes();
1159 if (weight * 8 > nodes_left) {
1160 if (PrintOpto) {
1161 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1162 }
1163 return true;
1164 } else {
1165 return false;
1166 }
1167 }
1168
1169 static bool merge_point_safe(Node* region) {
1170 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1171 // having a PhiNode input. This sidesteps the dangerous case where the split
1172 // ConvI2LNode may become TOP if the input Value() does not
1173 // overlap the ConvI2L range, leaving a node which may not dominate its
1174 // uses.
1175 // A better fix for this problem can be found in the BugTraq entry, but
1176 // expediency for Mantis demands this hack.
1177 #ifdef _LP64
1178 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1179 Node* n = region->fast_out(i);
1180 if (n->is_Phi()) {
1181 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1182 Node* m = n->fast_out(j);
1183 if (m->Opcode() == Op_ConvI2L)
1184 return false;
1185 if (m->is_CastII()) {
1186 return false;
1187 }
1188 }
1189 }
1190 }
1191 #endif
1192 return true;
1193 }
1194
1195
1196 //------------------------------place_outside_loop---------------------------------
1197 // Place some computation outside of this loop on the path to the use passed as argument
1198 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1199 Node* head = loop->_head;
1200 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1201 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1202 loop = loop->_parent;
1203 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1204 }
1205
1206 // Pick control right outside the loop
1207 for (;;) {
1208 Node* dom = idom(useblock);
1209 if (loop->is_member(get_loop(dom)) ||
1210 // NeverBranch nodes are not assigned to the loop when constructed
1211 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) {
1212 break;
1213 }
1214 useblock = dom;
1215 }
1216 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1217 return useblock;
1218 }
1219
1220
1221 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1222 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1223 return false;
1224 }
1225 if (!n->in(0)->is_Region()) {
1226 return false;
1227 }
1228
1229 Node* region = n->in(0);
1230 Node* dom = idom(region);
1231 if (!dom->is_If() || dom->in(1) != n->in(1)) {
1232 return false;
1233 }
1234 IfNode* dom_if = dom->as_If();
1235 Node* proj_true = dom_if->proj_out(1);
1236 Node* proj_false = dom_if->proj_out(0);
1237
1238 for (uint i = 1; i < region->req(); i++) {
1239 if (is_dominator(proj_true, region->in(i))) {
1240 continue;
1241 }
1242 if (is_dominator(proj_false, region->in(i))) {
1243 continue;
1244 }
1245 return false;
1246 }
1247
1248 return true;
1249 }
1250
1251
1252 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1253 if (must_throttle_split_if()) {
1254 return false;
1255 }
1256
1257 // Do not do 'split-if' if irreducible loops are present.
1258 if (_has_irreducible_loops) {
1259 return false;
1260 }
1261
1262 if (merge_point_too_heavy(C, n_ctrl)) {
1263 return false;
1264 }
1265
1266 // Do not do 'split-if' if some paths are dead. First do dead code
1267 // elimination and then see if its still profitable.
1268 for (uint i = 1; i < n_ctrl->req(); i++) {
1269 if (n_ctrl->in(i) == C->top()) {
1270 return false;
1271 }
1272 }
1273
1274 // If trying to do a 'Split-If' at the loop head, it is only
1275 // profitable if the cmp folds up on BOTH paths. Otherwise we
1276 // risk peeling a loop forever.
1277
1278 // CNC - Disabled for now. Requires careful handling of loop
1279 // body selection for the cloned code. Also, make sure we check
1280 // for any input path not being in the same loop as n_ctrl. For
1281 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1282 // because the alternative loop entry points won't be converted
1283 // into LoopNodes.
1284 IdealLoopTree *n_loop = get_loop(n_ctrl);
1285 for (uint j = 1; j < n_ctrl->req(); j++) {
1286 if (get_loop(n_ctrl->in(j)) != n_loop) {
1287 return false;
1288 }
1289 }
1290
1291 // Check for safety of the merge point.
1292 if (!merge_point_safe(n_ctrl)) {
1293 return false;
1294 }
1295
1296 return true;
1297 }
1298
1299 // Detect if the node is the inner strip-mined loop
1300 // Return: null if it's not the case, or the exit of outer strip-mined loop
1301 static Node* is_inner_of_stripmined_loop(const Node* out) {
1302 Node* out_le = nullptr;
1303
1304 if (out->is_CountedLoopEnd()) {
1305 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1306
1307 if (loop != nullptr && loop->is_strip_mined()) {
1308 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1309 }
1310 }
1311
1312 return out_le;
1313 }
1314
1315 //------------------------------split_if_with_blocks_post----------------------
1316 // Do the real work in a non-recursive function. CFG hackery wants to be
1317 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1318 // info.
1319 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1320
1321 // Cloning Cmp through Phi's involves the split-if transform.
1322 // FastLock is not used by an If
1323 if (n->is_Cmp() && !n->is_FastLock()) {
1324 Node *n_ctrl = get_ctrl(n);
1325 // Determine if the Node has inputs from some local Phi.
1326 // Returns the block to clone thru.
1327 Node *n_blk = has_local_phi_input(n);
1328 if (n_blk != n_ctrl) {
1329 return;
1330 }
1331
1332 if (!can_split_if(n_ctrl)) {
1333 return;
1334 }
1335
1336 if (n->outcnt() != 1) {
1337 return; // Multiple bool's from 1 compare?
1338 }
1339 Node *bol = n->unique_out();
1340 assert(bol->is_Bool(), "expect a bool here");
1341 if (bol->outcnt() != 1) {
1342 return;// Multiple branches from 1 compare?
1343 }
1344 Node *iff = bol->unique_out();
1345
1346 // Check some safety conditions
1347 if (iff->is_If()) { // Classic split-if?
1348 if (iff->in(0) != n_ctrl) {
1349 return; // Compare must be in same blk as if
1350 }
1351 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1352 // Can't split CMove with different control.
1353 if (get_ctrl(iff) != n_ctrl) {
1354 return;
1355 }
1356 if (get_ctrl(iff->in(2)) == n_ctrl ||
1357 get_ctrl(iff->in(3)) == n_ctrl) {
1358 return; // Inputs not yet split-up
1359 }
1360 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1361 return; // Loop-invar test gates loop-varying CMOVE
1362 }
1363 } else {
1364 return; // some other kind of node, such as an Allocate
1365 }
1366
1367 // When is split-if profitable? Every 'win' on means some control flow
1368 // goes dead, so it's almost always a win.
1369 int policy = 0;
1370 // Split compare 'n' through the merge point if it is profitable
1371 Node *phi = split_thru_phi( n, n_ctrl, policy);
1372 if (!phi) {
1373 return;
1374 }
1375
1376 // Found a Phi to split thru!
1377 // Replace 'n' with the new phi
1378 _igvn.replace_node(n, phi);
1379
1380 // Now split the bool up thru the phi
1381 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1382 guarantee(bolphi != nullptr, "null boolean phi node");
1383
1384 _igvn.replace_node(bol, bolphi);
1385 assert(iff->in(1) == bolphi, "");
1386
1387 if (bolphi->Value(&_igvn)->singleton()) {
1388 return;
1389 }
1390
1391 // Conditional-move? Must split up now
1392 if (!iff->is_If()) {
1393 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1394 _igvn.replace_node(iff, cmovphi);
1395 return;
1396 }
1397
1398 // Now split the IF
1399 do_split_if(iff);
1400 return;
1401 }
1402
1403 // Two identical ifs back to back can be merged
1404 if (try_merge_identical_ifs(n)) {
1405 return;
1406 }
1407
1408 // Check for an IF ready to split; one that has its
1409 // condition codes input coming from a Phi at the block start.
1410 int n_op = n->Opcode();
1411
1412 // Check for an IF being dominated by another IF same test
1413 if (n_op == Op_If ||
1414 n_op == Op_RangeCheck) {
1415 Node *bol = n->in(1);
1416 uint max = bol->outcnt();
1417 // Check for same test used more than once?
1418 if (max > 1 && bol->is_Bool()) {
1419 // Search up IDOMs to see if this IF is dominated.
1420 Node *cutoff = get_ctrl(bol);
1421
1422 // Now search up IDOMs till cutoff, looking for a dominating test
1423 Node *prevdom = n;
1424 Node *dom = idom(prevdom);
1425 while (dom != cutoff) {
1426 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom &&
1427 safe_for_if_replacement(dom)) {
1428 // It's invalid to move control dependent data nodes in the inner
1429 // strip-mined loop, because:
1430 // 1) break validation of LoopNode::verify_strip_mined()
1431 // 2) move code with side-effect in strip-mined loop
1432 // Move to the exit of outer strip-mined loop in that case.
1433 Node* out_le = is_inner_of_stripmined_loop(dom);
1434 if (out_le != nullptr) {
1435 prevdom = out_le;
1436 }
1437 // Replace the dominated test with an obvious true or false.
1438 // Place it on the IGVN worklist for later cleanup.
1439 C->set_major_progress();
1440 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true);
1441 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1442 return;
1443 }
1444 prevdom = dom;
1445 dom = idom(prevdom);
1446 }
1447 }
1448 }
1449
1450 try_sink_out_of_loop(n);
1451
1452 try_move_store_after_loop(n);
1453 }
1454
1455 // Transform:
1456 //
1457 // if (some_condition) {
1458 // // body 1
1459 // } else {
1460 // // body 2
1461 // }
1462 // if (some_condition) {
1463 // // body 3
1464 // } else {
1465 // // body 4
1466 // }
1467 //
1468 // into:
1469 //
1470 //
1471 // if (some_condition) {
1472 // // body 1
1473 // // body 3
1474 // } else {
1475 // // body 2
1476 // // body 4
1477 // }
1478 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1479 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1480 Node *n_ctrl = n->in(0);
1481 IfNode* dom_if = idom(n_ctrl)->as_If();
1482 ProjNode* dom_proj_true = dom_if->proj_out(1);
1483 ProjNode* dom_proj_false = dom_if->proj_out(0);
1484
1485 // Now split the IF
1486 RegionNode* new_false_region;
1487 RegionNode* new_true_region;
1488 do_split_if(n, &new_false_region, &new_true_region);
1489 assert(new_false_region->req() == new_true_region->req(), "");
1490 #ifdef ASSERT
1491 for (uint i = 1; i < new_false_region->req(); ++i) {
1492 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1493 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");
1494 }
1495 #endif
1496 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1497
1498 // We now have:
1499 // if (some_condition) {
1500 // // body 1
1501 // if (some_condition) {
1502 // body3: // new_true_region
1503 // // body3
1504 // } else {
1505 // goto body4;
1506 // }
1507 // } else {
1508 // // body 2
1509 // if (some_condition) {
1510 // goto body3;
1511 // } else {
1512 // body4: // new_false_region
1513 // // body4;
1514 // }
1515 // }
1516 //
1517
1518 // clone pinned nodes thru the resulting regions
1519 push_pinned_nodes_thru_region(dom_if, new_true_region);
1520 push_pinned_nodes_thru_region(dom_if, new_false_region);
1521
1522 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1523 // 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
1524 // unrelated control dependency.
1525 for (uint i = 1; i < new_false_region->req(); i++) {
1526 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1527 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1528 } else {
1529 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1530 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false);
1531 }
1532 }
1533 return true;
1534 }
1535 return false;
1536 }
1537
1538 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1539 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1540 Node* u = region->out(i);
1541 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1542 continue;
1543 }
1544 assert(u->in(0) == region, "not a control dependent node?");
1545 uint j = 1;
1546 for (; j < u->req(); ++j) {
1547 Node* in = u->in(j);
1548 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1549 break;
1550 }
1551 }
1552 if (j == u->req()) {
1553 Node *phi = PhiNode::make_blank(region, u);
1554 for (uint k = 1; k < region->req(); ++k) {
1555 Node* clone = u->clone();
1556 clone->set_req(0, region->in(k));
1557 register_new_node(clone, region->in(k));
1558 phi->init_req(k, clone);
1559 }
1560 register_new_node(phi, region);
1561 _igvn.replace_node(u, phi);
1562 --i;
1563 }
1564 }
1565 }
1566
1567 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1568 if (!dom->is_CountedLoopEnd()) {
1569 return true;
1570 }
1571 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1572 CountedLoopNode* cl = le->loopnode();
1573 if (cl == nullptr) {
1574 return true;
1575 }
1576 if (!cl->is_main_loop()) {
1577 return true;
1578 }
1579 if (cl->is_canonical_loop_entry() == nullptr) {
1580 return true;
1581 }
1582 // Further unrolling is possible so loop exit condition might change
1583 return false;
1584 }
1585
1586 // See if a shared loop-varying computation has no loop-varying uses.
1587 // Happens if something is only used for JVM state in uncommon trap exits,
1588 // like various versions of induction variable+offset. Clone the
1589 // computation per usage to allow it to sink out of the loop.
1590 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1591 if (has_ctrl(n) &&
1592 !n->is_Phi() &&
1593 !n->is_Bool() &&
1594 !n->is_Proj() &&
1595 !n->is_MergeMem() &&
1596 !n->is_CMove() &&
1597 n->Opcode() != Op_Opaque4 &&
1598 !n->is_Type()) {
1599 Node *n_ctrl = get_ctrl(n);
1600 IdealLoopTree *n_loop = get_loop(n_ctrl);
1601
1602 if (n->in(0) != nullptr) {
1603 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1604 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1605 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1606 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1607 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1608 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl));
1609 }
1610 }
1611 if (n_loop != _ltree_root && n->outcnt() > 1) {
1612 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1613 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1614 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1615 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1616 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1617 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect");
1618 Node* outer_loop_clone = nullptr;
1619 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1620 Node* u = n->last_out(j); // Clone private computation per use
1621 _igvn.rehash_node_delayed(u);
1622 Node* x = n->clone(); // Clone computation
1623 Node* x_ctrl = nullptr;
1624 if (u->is_Phi()) {
1625 // Replace all uses of normal nodes. Replace Phi uses
1626 // individually, so the separate Nodes can sink down
1627 // different paths.
1628 uint k = 1;
1629 while (u->in(k) != n) k++;
1630 u->set_req(k, x);
1631 // x goes next to Phi input path
1632 x_ctrl = u->in(0)->in(k);
1633 // Find control for 'x' next to use but not inside inner loops.
1634 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1635 --j;
1636 } else { // Normal use
1637 if (has_ctrl(u)) {
1638 x_ctrl = get_ctrl(u);
1639 } else {
1640 x_ctrl = u->in(0);
1641 }
1642 // Find control for 'x' next to use but not inside inner loops.
1643 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1644 // Replace all uses
1645 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) {
1646 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1647 // anymore now that we're going to pin n as well
1648 _igvn.replace_node(u, x);
1649 --j;
1650 } else {
1651 int nb = u->replace_edge(n, x, &_igvn);
1652 j -= nb;
1653 }
1654 }
1655
1656 if (n->is_Load()) {
1657 // For loads, add a control edge to a CFG node outside of the loop
1658 // to force them to not combine and return back inside the loop
1659 // during GVN optimization (4641526).
1660 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1661
1662 IdealLoopTree* x_loop = get_loop(x_ctrl);
1663 Node* x_head = x_loop->_head;
1664 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1665 // Do not add duplicate LoadNodes to the outer strip mined loop
1666 if (outer_loop_clone != nullptr) {
1667 _igvn.replace_node(x, outer_loop_clone);
1668 continue;
1669 }
1670 outer_loop_clone = x;
1671 }
1672 x->set_req(0, x_ctrl);
1673 } else if (n->in(0) != nullptr){
1674 x->set_req(0, x_ctrl);
1675 }
1676 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1677 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1678 register_new_node(x, x_ctrl);
1679
1680 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so:
1681 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain
1682 // the same.
1683 // (see 2- below)
1684 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1685 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1686 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1687 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1688 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1689 assert(!x->is_Load(), "load should be pinned");
1690 // Use a cast node to pin clone out of loop
1691 Node* cast = nullptr;
1692 for (uint k = 0; k < x->req(); k++) {
1693 Node* in = x->in(k);
1694 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1695 const Type* in_t = _igvn.type(in);
1696 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency);
1697 }
1698 if (cast != nullptr) {
1699 Node* prev = _igvn.hash_find_insert(cast);
1700 if (prev != nullptr) {
1701 cast->destruct(&_igvn);
1702 cast = prev;
1703 } else {
1704 register_new_node(cast, x_ctrl);
1705 }
1706 x->replace_edge(in, cast);
1707 // Chain of AddP:
1708 // 2- A CastPP of the base is only added now that both AddP nodes are sunk
1709 if (x->is_AddP() && k == AddPNode::Base) {
1710 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
1711 Node* u = x->fast_out(i);
1712 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) {
1713 _igvn.replace_input_of(u, AddPNode::Base, cast);
1714 assert(u->find_out_with(Op_AddP) == nullptr, "more than 2 chained AddP nodes?");
1715 }
1716 }
1717 }
1718 break;
1719 }
1720 }
1721 assert(cast != nullptr, "must have added a cast to pin the node");
1722 }
1723 }
1724 _igvn.remove_dead_node(n);
1725 }
1726 _dom_lca_tags_round = 0;
1727 }
1728 }
1729 }
1730
1731 // Compute the early control of a node by following its inputs until we reach
1732 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1733 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1734 Node* early_ctrl = nullptr;
1735 ResourceMark rm;
1736 Unique_Node_List wq;
1737 wq.push(n);
1738 for (uint i = 0; i < wq.size(); i++) {
1739 Node* m = wq.at(i);
1740 Node* c = nullptr;
1741 if (m->is_CFG()) {
1742 c = m;
1743 } else if (m->pinned()) {
1744 c = m->in(0);
1745 } else {
1746 for (uint j = 0; j < m->req(); j++) {
1747 Node* in = m->in(j);
1748 if (in != nullptr) {
1749 wq.push(in);
1750 }
1751 }
1752 }
1753 if (c != nullptr) {
1754 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1755 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1756 early_ctrl = c;
1757 }
1758 }
1759 }
1760 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1761 return early_ctrl;
1762 }
1763
1764 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1765 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1766 Node* u = n->fast_out(i);
1767 if (u->is_Opaque1()) {
1768 return false; // Found loop limit, bugfix for 4677003
1769 }
1770 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1771 // get_late_ctrl_with_anti_dep() use their own tag
1772 _dom_lca_tags_round++;
1773 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1774
1775 if (u->is_Phi()) {
1776 for (uint j = 1; j < u->req(); ++j) {
1777 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1778 return false;
1779 }
1780 }
1781 } else {
1782 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1783 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1784 return false;
1785 }
1786 }
1787 }
1788 return true;
1789 }
1790
1791 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1792 if (n->is_Load()) {
1793 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1794 }
1795 IdealLoopTree *u_loop = get_loop(ctrl);
1796 if (u_loop == n_loop) {
1797 return false; // Found loop-varying use
1798 }
1799 if (n_loop->is_member(u_loop)) {
1800 return false; // Found use in inner loop
1801 }
1802 return true;
1803 }
1804
1805 //------------------------------split_if_with_blocks---------------------------
1806 // Check for aggressive application of 'split-if' optimization,
1807 // using basic block level info.
1808 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1809 Node* root = C->root();
1810 visited.set(root->_idx); // first, mark root as visited
1811 // Do pre-visit work for root
1812 Node* n = split_if_with_blocks_pre(root);
1813 uint cnt = n->outcnt();
1814 uint i = 0;
1815
1816 while (true) {
1817 // Visit all children
1818 if (i < cnt) {
1819 Node* use = n->raw_out(i);
1820 ++i;
1821 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1822 // Now do pre-visit work for this use
1823 use = split_if_with_blocks_pre(use);
1824 nstack.push(n, i); // Save parent and next use's index.
1825 n = use; // Process all children of current use.
1826 cnt = use->outcnt();
1827 i = 0;
1828 }
1829 }
1830 else {
1831 // All of n's children have been processed, complete post-processing.
1832 if (cnt != 0 && !n->is_Con()) {
1833 assert(has_node(n), "no dead nodes");
1834 split_if_with_blocks_post(n);
1835 }
1836 if (must_throttle_split_if()) {
1837 nstack.clear();
1838 }
1839 if (nstack.is_empty()) {
1840 // Finished all nodes on stack.
1841 break;
1842 }
1843 // Get saved parent node and next use's index. Visit the rest of uses.
1844 n = nstack.node();
1845 cnt = n->outcnt();
1846 i = nstack.index();
1847 nstack.pop();
1848 }
1849 }
1850 }
1851
1852
1853 //=============================================================================
1854 //
1855 // C L O N E A L O O P B O D Y
1856 //
1857
1858 //------------------------------clone_iff--------------------------------------
1859 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1860 // "Nearly" because all Nodes have been cloned from the original in the loop,
1861 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1862 // through the Phi recursively, and return a Bool.
1863 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
1864
1865 // Convert this Phi into a Phi merging Bools
1866 uint i;
1867 for (i = 1; i < phi->req(); i++) {
1868 Node *b = phi->in(i);
1869 if (b->is_Phi()) {
1870 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
1871 } else {
1872 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1873 }
1874 }
1875
1876 Node* n = phi->in(1);
1877 Node* sample_opaque = nullptr;
1878 Node *sample_bool = nullptr;
1879 if (n->Opcode() == Op_Opaque4) {
1880 sample_opaque = n;
1881 sample_bool = n->in(1);
1882 assert(sample_bool->is_Bool(), "wrong type");
1883 } else {
1884 sample_bool = n;
1885 }
1886 Node *sample_cmp = sample_bool->in(1);
1887
1888 // Make Phis to merge the Cmp's inputs.
1889 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1890 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1891 for (i = 1; i < phi->req(); i++) {
1892 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1893 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1894 phi1->set_req(i, n1);
1895 phi2->set_req(i, n2);
1896 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1897 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1898 }
1899 // See if these Phis have been made before.
1900 // Register with optimizer
1901 Node *hit1 = _igvn.hash_find_insert(phi1);
1902 if (hit1) { // Hit, toss just made Phi
1903 _igvn.remove_dead_node(phi1); // Remove new phi
1904 assert(hit1->is_Phi(), "" );
1905 phi1 = (PhiNode*)hit1; // Use existing phi
1906 } else { // Miss
1907 _igvn.register_new_node_with_optimizer(phi1);
1908 }
1909 Node *hit2 = _igvn.hash_find_insert(phi2);
1910 if (hit2) { // Hit, toss just made Phi
1911 _igvn.remove_dead_node(phi2); // Remove new phi
1912 assert(hit2->is_Phi(), "" );
1913 phi2 = (PhiNode*)hit2; // Use existing phi
1914 } else { // Miss
1915 _igvn.register_new_node_with_optimizer(phi2);
1916 }
1917 // Register Phis with loop/block info
1918 set_ctrl(phi1, phi->in(0));
1919 set_ctrl(phi2, phi->in(0));
1920 // Make a new Cmp
1921 Node *cmp = sample_cmp->clone();
1922 cmp->set_req(1, phi1);
1923 cmp->set_req(2, phi2);
1924 _igvn.register_new_node_with_optimizer(cmp);
1925 set_ctrl(cmp, phi->in(0));
1926
1927 // Make a new Bool
1928 Node *b = sample_bool->clone();
1929 b->set_req(1,cmp);
1930 _igvn.register_new_node_with_optimizer(b);
1931 set_ctrl(b, phi->in(0));
1932
1933 if (sample_opaque != nullptr) {
1934 Node* opaque = sample_opaque->clone();
1935 opaque->set_req(1, b);
1936 _igvn.register_new_node_with_optimizer(opaque);
1937 set_ctrl(opaque, phi->in(0));
1938 return opaque;
1939 }
1940
1941 assert(b->is_Bool(), "");
1942 return b;
1943 }
1944
1945 //------------------------------clone_bool-------------------------------------
1946 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1947 // "Nearly" because all Nodes have been cloned from the original in the loop,
1948 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1949 // through the Phi recursively, and return a Bool.
1950 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
1951 uint i;
1952 // Convert this Phi into a Phi merging Bools
1953 for( i = 1; i < phi->req(); i++ ) {
1954 Node *b = phi->in(i);
1955 if( b->is_Phi() ) {
1956 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
1957 } else {
1958 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1959 }
1960 }
1961
1962 Node *sample_cmp = phi->in(1);
1963
1964 // Make Phis to merge the Cmp's inputs.
1965 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1966 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1967 for( uint j = 1; j < phi->req(); j++ ) {
1968 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1969 Node *n1, *n2;
1970 if( cmp_top->is_Cmp() ) {
1971 n1 = cmp_top->in(1);
1972 n2 = cmp_top->in(2);
1973 } else {
1974 n1 = n2 = cmp_top;
1975 }
1976 phi1->set_req( j, n1 );
1977 phi2->set_req( j, n2 );
1978 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1979 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1980 }
1981
1982 // See if these Phis have been made before.
1983 // Register with optimizer
1984 Node *hit1 = _igvn.hash_find_insert(phi1);
1985 if( hit1 ) { // Hit, toss just made Phi
1986 _igvn.remove_dead_node(phi1); // Remove new phi
1987 assert( hit1->is_Phi(), "" );
1988 phi1 = (PhiNode*)hit1; // Use existing phi
1989 } else { // Miss
1990 _igvn.register_new_node_with_optimizer(phi1);
1991 }
1992 Node *hit2 = _igvn.hash_find_insert(phi2);
1993 if( hit2 ) { // Hit, toss just made Phi
1994 _igvn.remove_dead_node(phi2); // Remove new phi
1995 assert( hit2->is_Phi(), "" );
1996 phi2 = (PhiNode*)hit2; // Use existing phi
1997 } else { // Miss
1998 _igvn.register_new_node_with_optimizer(phi2);
1999 }
2000 // Register Phis with loop/block info
2001 set_ctrl(phi1, phi->in(0));
2002 set_ctrl(phi2, phi->in(0));
2003 // Make a new Cmp
2004 Node *cmp = sample_cmp->clone();
2005 cmp->set_req( 1, phi1 );
2006 cmp->set_req( 2, phi2 );
2007 _igvn.register_new_node_with_optimizer(cmp);
2008 set_ctrl(cmp, phi->in(0));
2009
2010 assert( cmp->is_Cmp(), "" );
2011 return (CmpNode*)cmp;
2012 }
2013
2014 //------------------------------sink_use---------------------------------------
2015 // If 'use' was in the loop-exit block, it now needs to be sunk
2016 // below the post-loop merge point.
2017 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
2018 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
2019 set_ctrl(use, post_loop);
2020 for (DUIterator j = use->outs(); use->has_out(j); j++)
2021 sink_use(use->out(j), post_loop);
2022 }
2023 }
2024
2025 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2026 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2027 Node_List*& split_if_set, Node_List*& split_bool_set,
2028 Node_List*& split_cex_set, Node_List& worklist,
2029 uint new_counter, CloneLoopMode mode) {
2030 Node* nnn = old_new[old->_idx];
2031 // Copy uses to a worklist, so I can munge the def-use info
2032 // with impunity.
2033 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2034 worklist.push(old->fast_out(j));
2035
2036 while( worklist.size() ) {
2037 Node *use = worklist.pop();
2038 if (!has_node(use)) continue; // Ignore dead nodes
2039 if (use->in(0) == C->top()) continue;
2040 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2041 // Check for data-use outside of loop - at least one of OLD or USE
2042 // must not be a CFG node.
2043 #ifdef ASSERT
2044 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2045 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2046 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2047 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2048 }
2049 #endif
2050 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2051
2052 // If the Data use is an IF, that means we have an IF outside of the
2053 // loop that is switching on a condition that is set inside of the
2054 // loop. Happens if people set a loop-exit flag; then test the flag
2055 // in the loop to break the loop, then test is again outside of the
2056 // loop to determine which way the loop exited.
2057 // Loop predicate If node connects to Bool node through Opaque1 node.
2058 //
2059 // If the use is an AllocateArray through its ValidLengthTest input,
2060 // make sure the Bool/Cmp input is cloned down to avoid a Phi between
2061 // the AllocateArray node and its ValidLengthTest input that could cause
2062 // split if to break.
2063 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4 ||
2064 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2065 // Since this code is highly unlikely, we lazily build the worklist
2066 // of such Nodes to go split.
2067 if (!split_if_set) {
2068 split_if_set = new Node_List();
2069 }
2070 split_if_set->push(use);
2071 }
2072 if (use->is_Bool()) {
2073 if (!split_bool_set) {
2074 split_bool_set = new Node_List();
2075 }
2076 split_bool_set->push(use);
2077 }
2078 if (use->Opcode() == Op_CreateEx) {
2079 if (!split_cex_set) {
2080 split_cex_set = new Node_List();
2081 }
2082 split_cex_set->push(use);
2083 }
2084
2085
2086 // Get "block" use is in
2087 uint idx = 0;
2088 while( use->in(idx) != old ) idx++;
2089 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2090 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2091 Node *cfg = prev->_idx >= new_counter
2092 ? prev->in(2)
2093 : idom(prev);
2094 if( use->is_Phi() ) // Phi use is in prior block
2095 cfg = prev->in(idx); // NOT in block of Phi itself
2096 if (cfg->is_top()) { // Use is dead?
2097 _igvn.replace_input_of(use, idx, C->top());
2098 continue;
2099 }
2100
2101 // If use is referenced through control edge... (idx == 0)
2102 if (mode == IgnoreStripMined && idx == 0) {
2103 LoopNode *head = loop->_head->as_Loop();
2104 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2105 // That node is outside the inner loop, leave it outside the
2106 // outer loop as well to not confuse verification code.
2107 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2108 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2109 continue;
2110 }
2111 }
2112
2113 while(!outer_loop->is_member(get_loop(cfg))) {
2114 prev = cfg;
2115 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
2116 }
2117 // If the use occurs after merging several exits from the loop, then
2118 // old value must have dominated all those exits. Since the same old
2119 // value was used on all those exits we did not need a Phi at this
2120 // merge point. NOW we do need a Phi here. Each loop exit value
2121 // is now merged with the peeled body exit; each exit gets its own
2122 // private Phi and those Phis need to be merged here.
2123 Node *phi;
2124 if( prev->is_Region() ) {
2125 if( idx == 0 ) { // Updating control edge?
2126 phi = prev; // Just use existing control
2127 } else { // Else need a new Phi
2128 phi = PhiNode::make( prev, old );
2129 // Now recursively fix up the new uses of old!
2130 for( uint i = 1; i < prev->req(); i++ ) {
2131 worklist.push(phi); // Onto worklist once for each 'old' input
2132 }
2133 }
2134 } else {
2135 // Get new RegionNode merging old and new loop exits
2136 prev = old_new[prev->_idx];
2137 assert( prev, "just made this in step 7" );
2138 if( idx == 0) { // Updating control edge?
2139 phi = prev; // Just use existing control
2140 } else { // Else need a new Phi
2141 // Make a new Phi merging data values properly
2142 phi = PhiNode::make( prev, old );
2143 phi->set_req( 1, nnn );
2144 }
2145 }
2146 // If inserting a new Phi, check for prior hits
2147 if( idx != 0 ) {
2148 Node *hit = _igvn.hash_find_insert(phi);
2149 if( hit == nullptr ) {
2150 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2151 } else { // or
2152 // Remove the new phi from the graph and use the hit
2153 _igvn.remove_dead_node(phi);
2154 phi = hit; // Use existing phi
2155 }
2156 set_ctrl(phi, prev);
2157 }
2158 // Make 'use' use the Phi instead of the old loop body exit value
2159 assert(use->in(idx) == old, "old is still input of use");
2160 // We notify all uses of old, including use, and the indirect uses,
2161 // that may now be optimized because we have replaced old with phi.
2162 _igvn.add_users_to_worklist(old);
2163 _igvn.replace_input_of(use, idx, phi);
2164 if( use->_idx >= new_counter ) { // If updating new phis
2165 // Not needed for correctness, but prevents a weak assert
2166 // in AddPNode from tripping (when we end up with different
2167 // base & derived Phis that will become the same after
2168 // IGVN does CSE).
2169 Node *hit = _igvn.hash_find_insert(use);
2170 if( hit ) // Go ahead and re-hash for hits.
2171 _igvn.replace_node( use, hit );
2172 }
2173
2174 // If 'use' was in the loop-exit block, it now needs to be sunk
2175 // below the post-loop merge point.
2176 sink_use( use, prev );
2177 }
2178 }
2179 }
2180
2181 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2182 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2183 bool check_old_new) {
2184 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2185 Node* u = n->fast_out(j);
2186 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2187 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2188 Node* c = phase->get_ctrl(u);
2189 IdealLoopTree* u_loop = phase->get_loop(c);
2190 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2191 if (!loop->is_member(u_loop)) {
2192 if (outer_loop->is_member(u_loop)) {
2193 wq.push(u);
2194 } else {
2195 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2196 // the outer loop too
2197 Node* u_c = u->in(0);
2198 if (u_c != nullptr) {
2199 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2200 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2201 wq.push(u);
2202 }
2203 }
2204 }
2205 }
2206 }
2207 }
2208 }
2209
2210 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2211 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2212 Node_List& extra_data_nodes) {
2213 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2214 CountedLoopNode* cl = head->as_CountedLoop();
2215 Node* l = cl->outer_loop();
2216 Node* tail = cl->outer_loop_tail();
2217 IfNode* le = cl->outer_loop_end();
2218 Node* sfpt = cl->outer_safepoint();
2219 CountedLoopEndNode* cle = cl->loopexit();
2220 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2221 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2222 Node* cle_out = cle->proj_out(false);
2223
2224 Node* new_sfpt = nullptr;
2225 Node* new_cle_out = cle_out->clone();
2226 old_new.map(cle_out->_idx, new_cle_out);
2227 if (mode == CloneIncludesStripMined) {
2228 // clone outer loop body
2229 Node* new_l = l->clone();
2230 Node* new_tail = tail->clone();
2231 IfNode* new_le = le->clone()->as_If();
2232 new_sfpt = sfpt->clone();
2233
2234 set_loop(new_l, outer_loop->_parent);
2235 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2236 set_loop(new_cle_out, outer_loop->_parent);
2237 set_idom(new_cle_out, new_cle, dd);
2238 set_loop(new_sfpt, outer_loop->_parent);
2239 set_idom(new_sfpt, new_cle_out, dd);
2240 set_loop(new_le, outer_loop->_parent);
2241 set_idom(new_le, new_sfpt, dd);
2242 set_loop(new_tail, outer_loop->_parent);
2243 set_idom(new_tail, new_le, dd);
2244 set_idom(new_cl, new_l, dd);
2245
2246 old_new.map(l->_idx, new_l);
2247 old_new.map(tail->_idx, new_tail);
2248 old_new.map(le->_idx, new_le);
2249 old_new.map(sfpt->_idx, new_sfpt);
2250
2251 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2252 new_l->set_req(0, new_l);
2253 new_tail->set_req(0, new_le);
2254 new_le->set_req(0, new_sfpt);
2255 new_sfpt->set_req(0, new_cle_out);
2256 new_cle_out->set_req(0, new_cle);
2257 new_cl->set_req(LoopNode::EntryControl, new_l);
2258
2259 _igvn.register_new_node_with_optimizer(new_l);
2260 _igvn.register_new_node_with_optimizer(new_tail);
2261 _igvn.register_new_node_with_optimizer(new_le);
2262 } else {
2263 Node *newhead = old_new[loop->_head->_idx];
2264 newhead->as_Loop()->clear_strip_mined();
2265 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2266 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2267 }
2268 // Look at data node that were assigned a control in the outer
2269 // loop: they are kept in the outer loop by the safepoint so start
2270 // from the safepoint node's inputs.
2271 IdealLoopTree* outer_loop = get_loop(l);
2272 Node_Stack stack(2);
2273 stack.push(sfpt, 1);
2274 uint new_counter = C->unique();
2275 while (stack.size() > 0) {
2276 Node* n = stack.node();
2277 uint i = stack.index();
2278 while (i < n->req() &&
2279 (n->in(i) == nullptr ||
2280 !has_ctrl(n->in(i)) ||
2281 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2282 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2283 i++;
2284 }
2285 if (i < n->req()) {
2286 stack.set_index(i+1);
2287 stack.push(n->in(i), 0);
2288 } else {
2289 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2290 Node* m = n == sfpt ? new_sfpt : n->clone();
2291 if (m != nullptr) {
2292 for (uint i = 0; i < n->req(); i++) {
2293 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2294 m->set_req(i, old_new[m->in(i)->_idx]);
2295 }
2296 }
2297 } else {
2298 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2299 }
2300 if (n != sfpt) {
2301 extra_data_nodes.push(n);
2302 _igvn.register_new_node_with_optimizer(m);
2303 assert(get_ctrl(n) == cle_out, "what other control?");
2304 set_ctrl(m, new_cle_out);
2305 old_new.map(n->_idx, m);
2306 }
2307 stack.pop();
2308 }
2309 }
2310 if (mode == CloneIncludesStripMined) {
2311 _igvn.register_new_node_with_optimizer(new_sfpt);
2312 _igvn.register_new_node_with_optimizer(new_cle_out);
2313 }
2314 // Some other transformation may have pessimistically assigned some
2315 // data nodes to the outer loop. Set their control so they are out
2316 // of the outer loop.
2317 ResourceMark rm;
2318 Unique_Node_List wq;
2319 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2320 Node* old = extra_data_nodes.at(i);
2321 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2322 }
2323
2324 for (uint i = 0; i < loop->_body.size(); i++) {
2325 Node* old = loop->_body.at(i);
2326 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2327 }
2328
2329 Node* inner_out = sfpt->in(0);
2330 if (inner_out->outcnt() > 1) {
2331 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2332 }
2333
2334 Node* new_ctrl = cl->outer_loop_exit();
2335 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2336 for (uint i = 0; i < wq.size(); i++) {
2337 Node* n = wq.at(i);
2338 set_ctrl(n, new_ctrl);
2339 if (n->in(0) != nullptr) {
2340 _igvn.replace_input_of(n, 0, new_ctrl);
2341 }
2342 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2343 }
2344 } else {
2345 Node *newhead = old_new[loop->_head->_idx];
2346 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2347 }
2348 }
2349
2350 //------------------------------clone_loop-------------------------------------
2351 //
2352 // C L O N E A L O O P B O D Y
2353 //
2354 // This is the basic building block of the loop optimizations. It clones an
2355 // entire loop body. It makes an old_new loop body mapping; with this mapping
2356 // you can find the new-loop equivalent to an old-loop node. All new-loop
2357 // nodes are exactly equal to their old-loop counterparts, all edges are the
2358 // same. All exits from the old-loop now have a RegionNode that merges the
2359 // equivalent new-loop path. This is true even for the normal "loop-exit"
2360 // condition. All uses of loop-invariant old-loop values now come from (one
2361 // or more) Phis that merge their new-loop equivalents.
2362 //
2363 // This operation leaves the graph in an illegal state: there are two valid
2364 // control edges coming from the loop pre-header to both loop bodies. I'll
2365 // definitely have to hack the graph after running this transform.
2366 //
2367 // From this building block I will further edit edges to perform loop peeling
2368 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2369 //
2370 // Parameter side_by_size_idom:
2371 // When side_by_size_idom is null, the dominator tree is constructed for
2372 // the clone loop to dominate the original. Used in construction of
2373 // pre-main-post loop sequence.
2374 // When nonnull, the clone and original are side-by-side, both are
2375 // dominated by the side_by_side_idom node. Used in construction of
2376 // unswitched loops.
2377 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2378 CloneLoopMode mode, Node* side_by_side_idom) {
2379
2380 LoopNode* head = loop->_head->as_Loop();
2381 head->verify_strip_mined(1);
2382
2383 if (C->do_vector_loop() && PrintOpto) {
2384 const char* mname = C->method()->name()->as_quoted_ascii();
2385 if (mname != nullptr) {
2386 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2387 }
2388 }
2389
2390 CloneMap& cm = C->clone_map();
2391 if (C->do_vector_loop()) {
2392 cm.set_clone_idx(cm.max_gen()+1);
2393 #ifndef PRODUCT
2394 if (PrintOpto) {
2395 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2396 loop->dump_head();
2397 }
2398 #endif
2399 }
2400
2401 // Step 1: Clone the loop body. Make the old->new mapping.
2402 clone_loop_body(loop->_body, old_new, &cm);
2403
2404 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2405
2406 // Step 2: Fix the edges in the new body. If the old input is outside the
2407 // loop use it. If the old input is INside the loop, use the corresponding
2408 // new node instead.
2409 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2410
2411 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2412 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2413
2414 // Step 3: Now fix control uses. Loop varying control uses have already
2415 // been fixed up (as part of all input edges in Step 2). Loop invariant
2416 // control uses must be either an IfFalse or an IfTrue. Make a merge
2417 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2418 // refer to this.
2419 Node_List worklist;
2420 uint new_counter = C->unique();
2421 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2422
2423 // Step 4: If loop-invariant use is not control, it must be dominated by a
2424 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2425 // there if needed. Make a Phi there merging old and new used values.
2426 Node_List *split_if_set = nullptr;
2427 Node_List *split_bool_set = nullptr;
2428 Node_List *split_cex_set = nullptr;
2429 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2430
2431 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2432 Node* old = extra_data_nodes.at(i);
2433 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2434 split_bool_set, split_cex_set, worklist, new_counter,
2435 mode);
2436 }
2437
2438 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2439 // the loop uses a condition set in the loop. The original IF probably
2440 // takes control from one or more OLD Regions (which in turn get from NEW
2441 // Regions). In any case, there will be a set of Phis for each merge point
2442 // from the IF up to where the original BOOL def exists the loop.
2443 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2444
2445 }
2446
2447 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2448 if (split_if_set) {
2449 while (split_if_set->size()) {
2450 Node *iff = split_if_set->pop();
2451 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2452 if (iff->in(input)->is_Phi()) {
2453 Node *b = clone_iff(iff->in(input)->as_Phi());
2454 _igvn.replace_input_of(iff, input, b);
2455 }
2456 }
2457 }
2458 if (split_bool_set) {
2459 while (split_bool_set->size()) {
2460 Node *b = split_bool_set->pop();
2461 Node *phi = b->in(1);
2462 assert(phi->is_Phi(), "");
2463 CmpNode *cmp = clone_bool((PhiNode*) phi);
2464 _igvn.replace_input_of(b, 1, cmp);
2465 }
2466 }
2467 if (split_cex_set) {
2468 while (split_cex_set->size()) {
2469 Node *b = split_cex_set->pop();
2470 assert(b->in(0)->is_Region(), "");
2471 assert(b->in(1)->is_Phi(), "");
2472 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2473 split_up(b, b->in(0), nullptr);
2474 }
2475 }
2476 }
2477
2478 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2479 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2480 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2481 for(uint i = 0; i < body.size(); i++ ) {
2482 Node* old = body.at(i);
2483 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2484 split_bool_set, split_cex_set, worklist, new_counter,
2485 mode);
2486 }
2487 }
2488
2489 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2490 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2491 LoopNode* head = loop->_head->as_Loop();
2492 for(uint i = 0; i < body.size(); i++ ) {
2493 Node* old = body.at(i);
2494 if( !old->is_CFG() ) continue;
2495
2496 // Copy uses to a worklist, so I can munge the def-use info
2497 // with impunity.
2498 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2499 worklist.push(old->fast_out(j));
2500 }
2501
2502 while (worklist.size()) { // Visit all uses
2503 Node *use = worklist.pop();
2504 if (!has_node(use)) continue; // Ignore dead nodes
2505 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2506 if (!loop->is_member(use_loop) && use->is_CFG()) {
2507 // Both OLD and USE are CFG nodes here.
2508 assert(use->is_Proj(), "" );
2509 Node* nnn = old_new[old->_idx];
2510
2511 Node* newuse = nullptr;
2512 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2513 CountedLoopNode* cl = head->as_CountedLoop();
2514 CountedLoopEndNode* cle = cl->loopexit();
2515 Node* cle_out = cle->proj_out_or_null(false);
2516 if (use == cle_out) {
2517 IfNode* le = cl->outer_loop_end();
2518 use = le->proj_out(false);
2519 use_loop = get_loop(use);
2520 if (mode == CloneIncludesStripMined) {
2521 nnn = old_new[le->_idx];
2522 } else {
2523 newuse = old_new[cle_out->_idx];
2524 }
2525 }
2526 }
2527 if (newuse == nullptr) {
2528 newuse = use->clone();
2529 }
2530
2531 // Clone the loop exit control projection
2532 if (C->do_vector_loop() && cm != nullptr) {
2533 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2534 }
2535 newuse->set_req(0,nnn);
2536 _igvn.register_new_node_with_optimizer(newuse);
2537 set_loop(newuse, use_loop);
2538 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2539
2540 // We need a Region to merge the exit from the peeled body and the
2541 // exit from the old loop body.
2542 RegionNode *r = new RegionNode(3);
2543 // Map the old use to the new merge point
2544 old_new.map( use->_idx, r );
2545 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2546 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2547
2548 // The original user of 'use' uses 'r' instead.
2549 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2550 Node* useuse = use->last_out(l);
2551 _igvn.rehash_node_delayed(useuse);
2552 uint uses_found = 0;
2553 if (useuse->in(0) == use) {
2554 useuse->set_req(0, r);
2555 uses_found++;
2556 if (useuse->is_CFG()) {
2557 // This is not a dom_depth > dd_r because when new
2558 // control flow is constructed by a loop opt, a node and
2559 // its dominator can end up at the same dom_depth
2560 assert(dom_depth(useuse) >= dd_r, "");
2561 set_idom(useuse, r, dom_depth(useuse));
2562 }
2563 }
2564 for (uint k = 1; k < useuse->req(); k++) {
2565 if( useuse->in(k) == use ) {
2566 useuse->set_req(k, r);
2567 uses_found++;
2568 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2569 // This is not a dom_depth > dd_r because when new
2570 // control flow is constructed by a loop opt, a node
2571 // and its dominator can end up at the same dom_depth
2572 assert(dom_depth(useuse) >= dd_r , "");
2573 set_idom(useuse, r, dom_depth(useuse));
2574 }
2575 }
2576 }
2577 l -= uses_found; // we deleted 1 or more copies of this edge
2578 }
2579
2580 // Now finish up 'r'
2581 r->set_req(1, newuse);
2582 r->set_req(2, use);
2583 _igvn.register_new_node_with_optimizer(r);
2584 set_loop(r, use_loop);
2585 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2586 } // End of if a loop-exit test
2587 }
2588 }
2589 }
2590
2591 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2592 IdealLoopTree* parent, bool partial) {
2593 for(uint i = 0; i < body.size(); i++ ) {
2594 Node *old = body.at(i);
2595 Node *nnn = old_new[old->_idx];
2596 // Fix CFG/Loop controlling the new node
2597 if (has_ctrl(old)) {
2598 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2599 } else {
2600 set_loop(nnn, parent);
2601 if (old->outcnt() > 0) {
2602 Node* dom = idom(old);
2603 if (old_new[dom->_idx] != nullptr) {
2604 dom = old_new[dom->_idx];
2605 set_idom(nnn, dom, dd );
2606 }
2607 }
2608 }
2609 // Correct edges to the new node
2610 for (uint j = 0; j < nnn->req(); j++) {
2611 Node *n = nnn->in(j);
2612 if (n != nullptr) {
2613 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2614 if (loop->is_member(old_in_loop)) {
2615 if (old_new[n->_idx] != nullptr) {
2616 nnn->set_req(j, old_new[n->_idx]);
2617 } else {
2618 assert(!body.contains(n), "");
2619 assert(partial, "node not cloned");
2620 }
2621 }
2622 }
2623 }
2624 _igvn.hash_find_insert(nnn);
2625 }
2626 }
2627
2628 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2629 for (uint i = 0; i < body.size(); i++) {
2630 Node* old = body.at(i);
2631 Node* nnn = old->clone();
2632 old_new.map(old->_idx, nnn);
2633 if (C->do_vector_loop() && cm != nullptr) {
2634 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2635 }
2636 _igvn.register_new_node_with_optimizer(nnn);
2637 }
2638 }
2639
2640
2641 //---------------------- stride_of_possible_iv -------------------------------------
2642 // Looks for an iff/bool/comp with one operand of the compare
2643 // being a cycle involving an add and a phi,
2644 // with an optional truncation (left-shift followed by a right-shift)
2645 // of the add. Returns zero if not an iv.
2646 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2647 Node* trunc1 = nullptr;
2648 Node* trunc2 = nullptr;
2649 const TypeInteger* ttype = nullptr;
2650 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2651 return 0;
2652 }
2653 BoolNode* bl = iff->in(1)->as_Bool();
2654 Node* cmp = bl->in(1);
2655 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2656 return 0;
2657 }
2658 // Must have an invariant operand
2659 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2660 return 0;
2661 }
2662 Node* add2 = nullptr;
2663 Node* cmp1 = cmp->in(1);
2664 if (cmp1->is_Phi()) {
2665 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2666 Node* phi = cmp1;
2667 for (uint i = 1; i < phi->req(); i++) {
2668 Node* in = phi->in(i);
2669 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2670 &trunc1, &trunc2, &ttype, T_INT);
2671 if (add && add->in(1) == phi) {
2672 add2 = add->in(2);
2673 break;
2674 }
2675 }
2676 } else {
2677 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2678 Node* addtrunc = cmp1;
2679 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2680 &trunc1, &trunc2, &ttype, T_INT);
2681 if (add && add->in(1)->is_Phi()) {
2682 Node* phi = add->in(1);
2683 for (uint i = 1; i < phi->req(); i++) {
2684 if (phi->in(i) == addtrunc) {
2685 add2 = add->in(2);
2686 break;
2687 }
2688 }
2689 }
2690 }
2691 if (add2 != nullptr) {
2692 const TypeInt* add2t = _igvn.type(add2)->is_int();
2693 if (add2t->is_con()) {
2694 return add2t->get_con();
2695 }
2696 }
2697 return 0;
2698 }
2699
2700
2701 //---------------------- stay_in_loop -------------------------------------
2702 // Return the (unique) control output node that's in the loop (if it exists.)
2703 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2704 Node* unique = nullptr;
2705 if (!n) return nullptr;
2706 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2707 Node* use = n->fast_out(i);
2708 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2709 if (unique != nullptr) {
2710 return nullptr;
2711 }
2712 unique = use;
2713 }
2714 }
2715 return unique;
2716 }
2717
2718 //------------------------------ register_node -------------------------------------
2719 // Utility to register node "n" with PhaseIdealLoop
2720 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2721 _igvn.register_new_node_with_optimizer(n);
2722 loop->_body.push(n);
2723 if (n->is_CFG()) {
2724 set_loop(n, loop);
2725 set_idom(n, pred, ddepth);
2726 } else {
2727 set_ctrl(n, pred);
2728 }
2729 }
2730
2731 //------------------------------ proj_clone -------------------------------------
2732 // Utility to create an if-projection
2733 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2734 ProjNode* c = p->clone()->as_Proj();
2735 c->set_req(0, iff);
2736 return c;
2737 }
2738
2739 //------------------------------ short_circuit_if -------------------------------------
2740 // Force the iff control output to be the live_proj
2741 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2742 guarantee(live_proj != nullptr, "null projection");
2743 int proj_con = live_proj->_con;
2744 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2745 Node *con = _igvn.intcon(proj_con);
2746 set_ctrl(con, C->root());
2747 if (iff) {
2748 iff->set_req(1, con);
2749 }
2750 return con;
2751 }
2752
2753 //------------------------------ insert_if_before_proj -------------------------------------
2754 // Insert a new if before an if projection (* - new node)
2755 //
2756 // before
2757 // if(test)
2758 // / \
2759 // v v
2760 // other-proj proj (arg)
2761 //
2762 // after
2763 // if(test)
2764 // / \
2765 // / v
2766 // | * proj-clone
2767 // v |
2768 // other-proj v
2769 // * new_if(relop(cmp[IU](left,right)))
2770 // / \
2771 // v v
2772 // * new-proj proj
2773 // (returned)
2774 //
2775 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2776 IfNode* iff = proj->in(0)->as_If();
2777 IdealLoopTree *loop = get_loop(proj);
2778 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2779 uint ddepth = dom_depth(proj);
2780
2781 _igvn.rehash_node_delayed(iff);
2782 _igvn.rehash_node_delayed(proj);
2783
2784 proj->set_req(0, nullptr); // temporary disconnect
2785 ProjNode* proj2 = proj_clone(proj, iff);
2786 register_node(proj2, loop, iff, ddepth);
2787
2788 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2789 register_node(cmp, loop, proj2, ddepth);
2790
2791 BoolNode* bol = new BoolNode(cmp, relop);
2792 register_node(bol, loop, proj2, ddepth);
2793
2794 int opcode = iff->Opcode();
2795 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2796 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2797 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2798 register_node(new_if, loop, proj2, ddepth);
2799
2800 proj->set_req(0, new_if); // reattach
2801 set_idom(proj, new_if, ddepth);
2802
2803 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2804 guarantee(new_exit != nullptr, "null exit node");
2805 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2806
2807 return new_exit;
2808 }
2809
2810 //------------------------------ insert_region_before_proj -------------------------------------
2811 // Insert a region before an if projection (* - new node)
2812 //
2813 // before
2814 // if(test)
2815 // / |
2816 // v |
2817 // proj v
2818 // other-proj
2819 //
2820 // after
2821 // if(test)
2822 // / |
2823 // v |
2824 // * proj-clone v
2825 // | other-proj
2826 // v
2827 // * new-region
2828 // |
2829 // v
2830 // * dum_if
2831 // / \
2832 // v \
2833 // * dum-proj v
2834 // proj
2835 //
2836 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2837 IfNode* iff = proj->in(0)->as_If();
2838 IdealLoopTree *loop = get_loop(proj);
2839 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2840 uint ddepth = dom_depth(proj);
2841
2842 _igvn.rehash_node_delayed(iff);
2843 _igvn.rehash_node_delayed(proj);
2844
2845 proj->set_req(0, nullptr); // temporary disconnect
2846 ProjNode* proj2 = proj_clone(proj, iff);
2847 register_node(proj2, loop, iff, ddepth);
2848
2849 RegionNode* reg = new RegionNode(2);
2850 reg->set_req(1, proj2);
2851 register_node(reg, loop, iff, ddepth);
2852
2853 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
2854 register_node(dum_if, loop, reg, ddepth);
2855
2856 proj->set_req(0, dum_if); // reattach
2857 set_idom(proj, dum_if, ddepth);
2858
2859 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2860 register_node(dum_proj, loop, dum_if, ddepth);
2861
2862 return reg;
2863 }
2864
2865 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2866 // Clone a signed compare loop exit from an unsigned compare and
2867 // insert it before the unsigned cmp on the stay-in-loop path.
2868 // All new nodes inserted in the dominator tree between the original
2869 // if and it's projections. The original if test is replaced with
2870 // a constant to force the stay-in-loop path.
2871 //
2872 // This is done to make sure that the original if and it's projections
2873 // still dominate the same set of control nodes, that the ctrl() relation
2874 // from data nodes to them is preserved, and that their loop nesting is
2875 // preserved.
2876 //
2877 // before
2878 // if(i <u limit) unsigned compare loop exit
2879 // / |
2880 // v v
2881 // exit-proj stay-in-loop-proj
2882 //
2883 // after
2884 // if(stay-in-loop-const) original if
2885 // / |
2886 // / v
2887 // / if(i < limit) new signed test
2888 // / / |
2889 // / / v
2890 // / / if(i <u limit) new cloned unsigned test
2891 // / / / |
2892 // v v v |
2893 // region |
2894 // | |
2895 // dum-if |
2896 // / | |
2897 // ether | |
2898 // v v
2899 // exit-proj stay-in-loop-proj
2900 //
2901 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2902 const bool Signed = true;
2903 const bool Unsigned = false;
2904
2905 BoolNode* bol = if_cmpu->in(1)->as_Bool();
2906 if (bol->_test._test != BoolTest::lt) return nullptr;
2907 CmpNode* cmpu = bol->in(1)->as_Cmp();
2908 if (cmpu->Opcode() != Op_CmpU) return nullptr;
2909 int stride = stride_of_possible_iv(if_cmpu);
2910 if (stride == 0) return nullptr;
2911
2912 Node* lp_proj = stay_in_loop(if_cmpu, loop);
2913 guarantee(lp_proj != nullptr, "null loop node");
2914
2915 ProjNode* lp_continue = lp_proj->as_Proj();
2916 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2917 if (!lp_exit->is_IfFalse()) {
2918 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
2919 // We therefore can't add a single exit condition.
2920 return nullptr;
2921 }
2922 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit).
2923 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0.
2924 Node* limit = nullptr;
2925 if (stride > 0) {
2926 limit = cmpu->in(2);
2927 } else {
2928 limit = _igvn.makecon(TypeInt::ZERO);
2929 set_ctrl(limit, C->root());
2930 }
2931 // Create a new region on the exit path
2932 RegionNode* reg = insert_region_before_proj(lp_exit);
2933 guarantee(reg != nullptr, "null region node");
2934
2935 // Clone the if-cmpu-true-false using a signed compare
2936 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2937 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2938 reg->add_req(cmpi_exit);
2939
2940 // Clone the if-cmpu-true-false
2941 BoolTest::mask rel_u = bol->_test._test;
2942 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2943 reg->add_req(cmpu_exit);
2944
2945 // Force original if to stay in loop.
2946 short_circuit_if(if_cmpu, lp_continue);
2947
2948 return cmpi_exit->in(0)->as_If();
2949 }
2950
2951 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2952 // Remove a previously inserted signed compare loop exit.
2953 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2954 Node* lp_proj = stay_in_loop(if_cmp, loop);
2955 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2956 stay_in_loop(lp_proj, loop)->is_If() &&
2957 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2958 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2959 set_ctrl(con, C->root());
2960 if_cmp->set_req(1, con);
2961 }
2962
2963 //------------------------------ scheduled_nodelist -------------------------------------
2964 // Create a post order schedule of nodes that are in the
2965 // "member" set. The list is returned in "sched".
2966 // The first node in "sched" is the loop head, followed by
2967 // nodes which have no inputs in the "member" set, and then
2968 // followed by the nodes that have an immediate input dependence
2969 // on a node in "sched".
2970 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2971
2972 assert(member.test(loop->_head->_idx), "loop head must be in member set");
2973 VectorSet visited;
2974 Node_Stack nstack(loop->_body.size());
2975
2976 Node* n = loop->_head; // top of stack is cached in "n"
2977 uint idx = 0;
2978 visited.set(n->_idx);
2979
2980 // Initially push all with no inputs from within member set
2981 for(uint i = 0; i < loop->_body.size(); i++ ) {
2982 Node *elt = loop->_body.at(i);
2983 if (member.test(elt->_idx)) {
2984 bool found = false;
2985 for (uint j = 0; j < elt->req(); j++) {
2986 Node* def = elt->in(j);
2987 if (def && member.test(def->_idx) && def != elt) {
2988 found = true;
2989 break;
2990 }
2991 }
2992 if (!found && elt != loop->_head) {
2993 nstack.push(n, idx);
2994 n = elt;
2995 assert(!visited.test(n->_idx), "not seen yet");
2996 visited.set(n->_idx);
2997 }
2998 }
2999 }
3000
3001 // traverse out's that are in the member set
3002 while (true) {
3003 if (idx < n->outcnt()) {
3004 Node* use = n->raw_out(idx);
3005 idx++;
3006 if (!visited.test_set(use->_idx)) {
3007 if (member.test(use->_idx)) {
3008 nstack.push(n, idx);
3009 n = use;
3010 idx = 0;
3011 }
3012 }
3013 } else {
3014 // All outputs processed
3015 sched.push(n);
3016 if (nstack.is_empty()) break;
3017 n = nstack.node();
3018 idx = nstack.index();
3019 nstack.pop();
3020 }
3021 }
3022 }
3023
3024
3025 //------------------------------ has_use_in_set -------------------------------------
3026 // Has a use in the vector set
3027 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3028 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3029 Node* use = n->fast_out(j);
3030 if (vset.test(use->_idx)) {
3031 return true;
3032 }
3033 }
3034 return false;
3035 }
3036
3037
3038 //------------------------------ has_use_internal_to_set -------------------------------------
3039 // Has use internal to the vector set (ie. not in a phi at the loop head)
3040 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3041 Node* head = loop->_head;
3042 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3043 Node* use = n->fast_out(j);
3044 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3045 return true;
3046 }
3047 }
3048 return false;
3049 }
3050
3051
3052 //------------------------------ clone_for_use_outside_loop -------------------------------------
3053 // clone "n" for uses that are outside of loop
3054 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3055 int cloned = 0;
3056 assert(worklist.size() == 0, "should be empty");
3057 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3058 Node* use = n->fast_out(j);
3059 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3060 worklist.push(use);
3061 }
3062 }
3063
3064 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3065 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3066 return -1;
3067 }
3068
3069 while( worklist.size() ) {
3070 Node *use = worklist.pop();
3071 if (!has_node(use) || use->in(0) == C->top()) continue;
3072 uint j;
3073 for (j = 0; j < use->req(); j++) {
3074 if (use->in(j) == n) break;
3075 }
3076 assert(j < use->req(), "must be there");
3077
3078 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3079 Node* n_clone = n->clone();
3080 _igvn.replace_input_of(use, j, n_clone);
3081 cloned++;
3082 Node* use_c;
3083 if (!use->is_Phi()) {
3084 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3085 } else {
3086 // Use in a phi is considered a use in the associated predecessor block
3087 use_c = use->in(0)->in(j);
3088 }
3089 set_ctrl(n_clone, use_c);
3090 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3091 get_loop(use_c)->_body.push(n_clone);
3092 _igvn.register_new_node_with_optimizer(n_clone);
3093 #ifndef PRODUCT
3094 if (TracePartialPeeling) {
3095 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3096 }
3097 #endif
3098 }
3099 return cloned;
3100 }
3101
3102
3103 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3104 // clone "n" for special uses that are in the not_peeled region.
3105 // If these def-uses occur in separate blocks, the code generator
3106 // marks the method as not compilable. For example, if a "BoolNode"
3107 // is in a different basic block than the "IfNode" that uses it, then
3108 // the compilation is aborted in the code generator.
3109 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3110 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3111 if (n->is_Phi() || n->is_Load()) {
3112 return;
3113 }
3114 assert(worklist.size() == 0, "should be empty");
3115 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3116 Node* use = n->fast_out(j);
3117 if ( not_peel.test(use->_idx) &&
3118 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3119 use->in(1) == n) {
3120 worklist.push(use);
3121 }
3122 }
3123 if (worklist.size() > 0) {
3124 // clone "n" and insert it between inputs of "n" and the use
3125 Node* n_clone = n->clone();
3126 loop->_body.push(n_clone);
3127 _igvn.register_new_node_with_optimizer(n_clone);
3128 set_ctrl(n_clone, get_ctrl(n));
3129 sink_list.push(n_clone);
3130 not_peel.set(n_clone->_idx);
3131 #ifndef PRODUCT
3132 if (TracePartialPeeling) {
3133 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3134 }
3135 #endif
3136 while( worklist.size() ) {
3137 Node *use = worklist.pop();
3138 _igvn.rehash_node_delayed(use);
3139 for (uint j = 1; j < use->req(); j++) {
3140 if (use->in(j) == n) {
3141 use->set_req(j, n_clone);
3142 }
3143 }
3144 }
3145 }
3146 }
3147
3148
3149 //------------------------------ insert_phi_for_loop -------------------------------------
3150 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3151 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3152 Node *phi = PhiNode::make(lp, back_edge_val);
3153 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3154 // Use existing phi if it already exists
3155 Node *hit = _igvn.hash_find_insert(phi);
3156 if( hit == nullptr ) {
3157 _igvn.register_new_node_with_optimizer(phi);
3158 set_ctrl(phi, lp);
3159 } else {
3160 // Remove the new phi from the graph and use the hit
3161 _igvn.remove_dead_node(phi);
3162 phi = hit;
3163 }
3164 _igvn.replace_input_of(use, idx, phi);
3165 }
3166
3167 #ifdef ASSERT
3168 //------------------------------ is_valid_loop_partition -------------------------------------
3169 // Validate the loop partition sets: peel and not_peel
3170 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3171 VectorSet& not_peel ) {
3172 uint i;
3173 // Check that peel_list entries are in the peel set
3174 for (i = 0; i < peel_list.size(); i++) {
3175 if (!peel.test(peel_list.at(i)->_idx)) {
3176 return false;
3177 }
3178 }
3179 // Check at loop members are in one of peel set or not_peel set
3180 for (i = 0; i < loop->_body.size(); i++ ) {
3181 Node *def = loop->_body.at(i);
3182 uint di = def->_idx;
3183 // Check that peel set elements are in peel_list
3184 if (peel.test(di)) {
3185 if (not_peel.test(di)) {
3186 return false;
3187 }
3188 // Must be in peel_list also
3189 bool found = false;
3190 for (uint j = 0; j < peel_list.size(); j++) {
3191 if (peel_list.at(j)->_idx == di) {
3192 found = true;
3193 break;
3194 }
3195 }
3196 if (!found) {
3197 return false;
3198 }
3199 } else if (not_peel.test(di)) {
3200 if (peel.test(di)) {
3201 return false;
3202 }
3203 } else {
3204 return false;
3205 }
3206 }
3207 return true;
3208 }
3209
3210 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3211 // Ensure a use outside of loop is of the right form
3212 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3213 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3214 return (use->is_Phi() &&
3215 use_c->is_Region() && use_c->req() == 3 &&
3216 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3217 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3218 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3219 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3220 }
3221
3222 //------------------------------ is_valid_clone_loop_form -------------------------------------
3223 // Ensure that all uses outside of loop are of the right form
3224 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3225 uint orig_exit_idx, uint clone_exit_idx) {
3226 uint len = peel_list.size();
3227 for (uint i = 0; i < len; i++) {
3228 Node *def = peel_list.at(i);
3229
3230 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3231 Node *use = def->fast_out(j);
3232 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3233 if (!loop->is_member(get_loop(use_c))) {
3234 // use is not in the loop, check for correct structure
3235 if (use->in(0) == def) {
3236 // Okay
3237 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3238 return false;
3239 }
3240 }
3241 }
3242 }
3243 return true;
3244 }
3245 #endif
3246
3247 //------------------------------ partial_peel -------------------------------------
3248 // Partially peel (aka loop rotation) the top portion of a loop (called
3249 // the peel section below) by cloning it and placing one copy just before
3250 // the new loop head and the other copy at the bottom of the new loop.
3251 //
3252 // before after where it came from
3253 //
3254 // stmt1 stmt1
3255 // loop: stmt2 clone
3256 // stmt2 if condA goto exitA clone
3257 // if condA goto exitA new_loop: new
3258 // stmt3 stmt3 clone
3259 // if !condB goto loop if condB goto exitB clone
3260 // exitB: stmt2 orig
3261 // stmt4 if !condA goto new_loop orig
3262 // exitA: goto exitA
3263 // exitB:
3264 // stmt4
3265 // exitA:
3266 //
3267 // Step 1: find the cut point: an exit test on probable
3268 // induction variable.
3269 // Step 2: schedule (with cloning) operations in the peel
3270 // section that can be executed after the cut into
3271 // the section that is not peeled. This may need
3272 // to clone operations into exit blocks. For
3273 // instance, a reference to A[i] in the not-peel
3274 // section and a reference to B[i] in an exit block
3275 // may cause a left-shift of i by 2 to be placed
3276 // in the peel block. This step will clone the left
3277 // shift into the exit block and sink the left shift
3278 // from the peel to the not-peel section.
3279 // Step 3: clone the loop, retarget the control, and insert
3280 // phis for values that are live across the new loop
3281 // head. This is very dependent on the graph structure
3282 // from clone_loop. It creates region nodes for
3283 // exit control and associated phi nodes for values
3284 // flow out of the loop through that exit. The region
3285 // node is dominated by the clone's control projection.
3286 // So the clone's peel section is placed before the
3287 // new loop head, and the clone's not-peel section is
3288 // forms the top part of the new loop. The original
3289 // peel section forms the tail of the new loop.
3290 // Step 4: update the dominator tree and recompute the
3291 // dominator depth.
3292 //
3293 // orig
3294 //
3295 // stmt1
3296 // |
3297 // v
3298 // predicates
3299 // |
3300 // v
3301 // loop<----+
3302 // | |
3303 // stmt2 |
3304 // | |
3305 // v |
3306 // ifA |
3307 // / | |
3308 // v v |
3309 // false true ^ <-- last_peel
3310 // / | |
3311 // / ===|==cut |
3312 // / stmt3 | <-- first_not_peel
3313 // / | |
3314 // | v |
3315 // v ifB |
3316 // exitA: / \ |
3317 // / \ |
3318 // v v |
3319 // false true |
3320 // / \ |
3321 // / ----+
3322 // |
3323 // v
3324 // exitB:
3325 // stmt4
3326 //
3327 //
3328 // after clone loop
3329 //
3330 // stmt1
3331 // |
3332 // v
3333 // predicates
3334 // / \
3335 // clone / \ orig
3336 // / \
3337 // / \
3338 // v v
3339 // +---->loop loop<----+
3340 // | | | |
3341 // | stmt2 stmt2 |
3342 // | | | |
3343 // | v v |
3344 // | ifA ifA |
3345 // | | \ / | |
3346 // | v v v v |
3347 // ^ true false false true ^ <-- last_peel
3348 // | | ^ \ / | |
3349 // | cut==|== \ \ / ===|==cut |
3350 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3351 // | | dom | | | |
3352 // | v \ 1v v2 v |
3353 // | ifB regionA ifB |
3354 // | / \ | / \ |
3355 // | / \ v / \ |
3356 // | v v exitA: v v |
3357 // | true false false true |
3358 // | / ^ \ / \ |
3359 // +---- \ \ / ----+
3360 // dom \ /
3361 // \ 1v v2
3362 // regionB
3363 // |
3364 // v
3365 // exitB:
3366 // stmt4
3367 //
3368 //
3369 // after partial peel
3370 //
3371 // stmt1
3372 // |
3373 // v
3374 // predicates
3375 // /
3376 // clone / orig
3377 // / TOP
3378 // / \
3379 // v v
3380 // TOP->loop loop----+
3381 // | | |
3382 // stmt2 stmt2 |
3383 // | | |
3384 // v v |
3385 // ifA ifA |
3386 // | \ / | |
3387 // v v v v |
3388 // true false false true | <-- last_peel
3389 // | ^ \ / +------|---+
3390 // +->newloop \ \ / === ==cut | |
3391 // | stmt3 \ \ / TOP | |
3392 // | | dom | | stmt3 | | <-- first_not_peel
3393 // | v \ 1v v2 v | |
3394 // | ifB regionA ifB ^ v
3395 // | / \ | / \ | |
3396 // | / \ v / \ | |
3397 // | v v exitA: v v | |
3398 // | true false false true | |
3399 // | / ^ \ / \ | |
3400 // | | \ \ / v | |
3401 // | | dom \ / TOP | |
3402 // | | \ 1v v2 | |
3403 // ^ v regionB | |
3404 // | | | | |
3405 // | | v ^ v
3406 // | | exitB: | |
3407 // | | stmt4 | |
3408 // | +------------>-----------------+ |
3409 // | |
3410 // +-----------------<---------------------+
3411 //
3412 //
3413 // final graph
3414 //
3415 // stmt1
3416 // |
3417 // v
3418 // predicates
3419 // |
3420 // v
3421 // stmt2 clone
3422 // |
3423 // v
3424 // ........> ifA clone
3425 // : / |
3426 // dom / |
3427 // : v v
3428 // : false true
3429 // : | |
3430 // : | v
3431 // : | newloop<-----+
3432 // : | | |
3433 // : | stmt3 clone |
3434 // : | | |
3435 // : | v |
3436 // : | ifB |
3437 // : | / \ |
3438 // : | v v |
3439 // : | false true |
3440 // : | | | |
3441 // : | v stmt2 |
3442 // : | exitB: | |
3443 // : | stmt4 v |
3444 // : | ifA orig |
3445 // : | / \ |
3446 // : | / \ |
3447 // : | v v |
3448 // : | false true |
3449 // : | / \ |
3450 // : v v -----+
3451 // RegionA
3452 // |
3453 // v
3454 // exitA
3455 //
3456 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3457
3458 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3459 if (!loop->_head->is_Loop()) {
3460 return false;
3461 }
3462 LoopNode *head = loop->_head->as_Loop();
3463
3464 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3465 return false;
3466 }
3467
3468 // Check for complex exit control
3469 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3470 Node *n = loop->_body.at(ii);
3471 int opc = n->Opcode();
3472 if (n->is_Call() ||
3473 opc == Op_Catch ||
3474 opc == Op_CatchProj ||
3475 opc == Op_Jump ||
3476 opc == Op_JumpProj) {
3477 #ifndef PRODUCT
3478 if (TracePartialPeeling) {
3479 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3480 }
3481 #endif
3482 return false;
3483 }
3484 }
3485
3486 int dd = dom_depth(head);
3487
3488 // Step 1: find cut point
3489
3490 // Walk up dominators to loop head looking for first loop exit
3491 // which is executed on every path thru loop.
3492 IfNode *peel_if = nullptr;
3493 IfNode *peel_if_cmpu = nullptr;
3494
3495 Node *iff = loop->tail();
3496 while (iff != head) {
3497 if (iff->is_If()) {
3498 Node *ctrl = get_ctrl(iff->in(1));
3499 if (ctrl->is_top()) return false; // Dead test on live IF.
3500 // If loop-varying exit-test, check for induction variable
3501 if (loop->is_member(get_loop(ctrl)) &&
3502 loop->is_loop_exit(iff) &&
3503 is_possible_iv_test(iff)) {
3504 Node* cmp = iff->in(1)->in(1);
3505 if (cmp->Opcode() == Op_CmpI) {
3506 peel_if = iff->as_If();
3507 } else {
3508 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3509 peel_if_cmpu = iff->as_If();
3510 }
3511 }
3512 }
3513 iff = idom(iff);
3514 }
3515
3516 // Prefer signed compare over unsigned compare.
3517 IfNode* new_peel_if = nullptr;
3518 if (peel_if == nullptr) {
3519 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3520 return false; // No peel point found
3521 }
3522 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3523 if (new_peel_if == nullptr) {
3524 return false; // No peel point found
3525 }
3526 peel_if = new_peel_if;
3527 }
3528 Node* last_peel = stay_in_loop(peel_if, loop);
3529 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3530 if (first_not_peeled == nullptr || first_not_peeled == head) {
3531 return false;
3532 }
3533
3534 #ifndef PRODUCT
3535 if (TraceLoopOpts) {
3536 tty->print("PartialPeel ");
3537 loop->dump_head();
3538 }
3539
3540 if (TracePartialPeeling) {
3541 tty->print_cr("before partial peel one iteration");
3542 Node_List wl;
3543 Node* t = head->in(2);
3544 while (true) {
3545 wl.push(t);
3546 if (t == head) break;
3547 t = idom(t);
3548 }
3549 while (wl.size() > 0) {
3550 Node* tt = wl.pop();
3551 tt->dump();
3552 if (tt == last_peel) tty->print_cr("-- cut --");
3553 }
3554 }
3555 #endif
3556 VectorSet peel;
3557 VectorSet not_peel;
3558 Node_List peel_list;
3559 Node_List worklist;
3560 Node_List sink_list;
3561
3562 uint estimate = loop->est_loop_clone_sz(1);
3563 if (exceeding_node_budget(estimate)) {
3564 return false;
3565 }
3566
3567 // Set of cfg nodes to peel are those that are executable from
3568 // the head through last_peel.
3569 assert(worklist.size() == 0, "should be empty");
3570 worklist.push(head);
3571 peel.set(head->_idx);
3572 while (worklist.size() > 0) {
3573 Node *n = worklist.pop();
3574 if (n != last_peel) {
3575 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3576 Node* use = n->fast_out(j);
3577 if (use->is_CFG() &&
3578 loop->is_member(get_loop(use)) &&
3579 !peel.test_set(use->_idx)) {
3580 worklist.push(use);
3581 }
3582 }
3583 }
3584 }
3585
3586 // Set of non-cfg nodes to peel are those that are control
3587 // dependent on the cfg nodes.
3588 for (uint i = 0; i < loop->_body.size(); i++) {
3589 Node *n = loop->_body.at(i);
3590 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3591 if (peel.test(n_c->_idx)) {
3592 peel.set(n->_idx);
3593 } else {
3594 not_peel.set(n->_idx);
3595 }
3596 }
3597
3598 // Step 2: move operations from the peeled section down into the
3599 // not-peeled section
3600
3601 // Get a post order schedule of nodes in the peel region
3602 // Result in right-most operand.
3603 scheduled_nodelist(loop, peel, peel_list);
3604
3605 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3606
3607 // For future check for too many new phis
3608 uint old_phi_cnt = 0;
3609 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3610 Node* use = head->fast_out(j);
3611 if (use->is_Phi()) old_phi_cnt++;
3612 }
3613
3614 #ifndef PRODUCT
3615 if (TracePartialPeeling) {
3616 tty->print_cr("\npeeled list");
3617 }
3618 #endif
3619
3620 // Evacuate nodes in peel region into the not_peeled region if possible
3621 bool too_many_clones = false;
3622 uint new_phi_cnt = 0;
3623 uint cloned_for_outside_use = 0;
3624 for (uint i = 0; i < peel_list.size();) {
3625 Node* n = peel_list.at(i);
3626 #ifndef PRODUCT
3627 if (TracePartialPeeling) n->dump();
3628 #endif
3629 bool incr = true;
3630 if (!n->is_CFG()) {
3631 if (has_use_in_set(n, not_peel)) {
3632 // If not used internal to the peeled region,
3633 // move "n" from peeled to not_peeled region.
3634 if (!has_use_internal_to_set(n, peel, loop)) {
3635 // if not pinned and not a load (which maybe anti-dependent on a store)
3636 // and not a CMove (Matcher expects only bool->cmove).
3637 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3638 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3639 if (new_clones == -1) {
3640 too_many_clones = true;
3641 break;
3642 }
3643 cloned_for_outside_use += new_clones;
3644 sink_list.push(n);
3645 peel.remove(n->_idx);
3646 not_peel.set(n->_idx);
3647 peel_list.remove(i);
3648 incr = false;
3649 #ifndef PRODUCT
3650 if (TracePartialPeeling) {
3651 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3652 n->_idx, get_ctrl(n)->_idx);
3653 }
3654 #endif
3655 }
3656 } else {
3657 // Otherwise check for special def-use cases that span
3658 // the peel/not_peel boundary such as bool->if
3659 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3660 new_phi_cnt++;
3661 }
3662 }
3663 }
3664 if (incr) i++;
3665 }
3666
3667 estimate += cloned_for_outside_use + new_phi_cnt;
3668 bool exceed_node_budget = !may_require_nodes(estimate);
3669 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3670
3671 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3672 #ifndef PRODUCT
3673 if (TracePartialPeeling && exceed_phi_limit) {
3674 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3675 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3676 }
3677 #endif
3678 if (new_peel_if != nullptr) {
3679 remove_cmpi_loop_exit(new_peel_if, loop);
3680 }
3681 // Inhibit more partial peeling on this loop
3682 assert(!head->is_partial_peel_loop(), "not partial peeled");
3683 head->mark_partial_peel_failed();
3684 if (cloned_for_outside_use > 0) {
3685 // Terminate this round of loop opts because
3686 // the graph outside this loop was changed.
3687 C->set_major_progress();
3688 return true;
3689 }
3690 return false;
3691 }
3692
3693 // Step 3: clone loop, retarget control, and insert new phis
3694
3695 // Create new loop head for new phis and to hang
3696 // the nodes being moved (sinked) from the peel region.
3697 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3698 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3699 _igvn.register_new_node_with_optimizer(new_head);
3700 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3701 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3702 set_loop(new_head, loop);
3703 loop->_body.push(new_head);
3704 not_peel.set(new_head->_idx);
3705 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3706 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3707
3708 while (sink_list.size() > 0) {
3709 Node* n = sink_list.pop();
3710 set_ctrl(n, new_head);
3711 }
3712
3713 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3714
3715 clone_loop(loop, old_new, dd, IgnoreStripMined);
3716
3717 const uint clone_exit_idx = 1;
3718 const uint orig_exit_idx = 2;
3719 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3720
3721 Node* head_clone = old_new[head->_idx];
3722 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
3723 Node* orig_tail_clone = head_clone->in(2);
3724
3725 // Add phi if "def" node is in peel set and "use" is not
3726
3727 for (uint i = 0; i < peel_list.size(); i++) {
3728 Node *def = peel_list.at(i);
3729 if (!def->is_CFG()) {
3730 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3731 Node *use = def->fast_out(j);
3732 if (has_node(use) && use->in(0) != C->top() &&
3733 (!peel.test(use->_idx) ||
3734 (use->is_Phi() && use->in(0) == head)) ) {
3735 worklist.push(use);
3736 }
3737 }
3738 while( worklist.size() ) {
3739 Node *use = worklist.pop();
3740 for (uint j = 1; j < use->req(); j++) {
3741 Node* n = use->in(j);
3742 if (n == def) {
3743
3744 // "def" is in peel set, "use" is not in peel set
3745 // or "use" is in the entry boundary (a phi) of the peel set
3746
3747 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3748
3749 if ( loop->is_member(get_loop( use_c )) ) {
3750 // use is in loop
3751 if (old_new[use->_idx] != nullptr) { // null for dead code
3752 Node* use_clone = old_new[use->_idx];
3753 _igvn.replace_input_of(use, j, C->top());
3754 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3755 }
3756 } else {
3757 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3758 // use is not in the loop, check if the live range includes the cut
3759 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3760 if (not_peel.test(lp_if->_idx)) {
3761 assert(j == orig_exit_idx, "use from original loop");
3762 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3763 }
3764 }
3765 }
3766 }
3767 }
3768 }
3769 }
3770
3771 // Step 3b: retarget control
3772
3773 // Redirect control to the new loop head if a cloned node in
3774 // the not_peeled region has control that points into the peeled region.
3775 // This necessary because the cloned peeled region will be outside
3776 // the loop.
3777 // from to
3778 // cloned-peeled <---+
3779 // new_head_clone: | <--+
3780 // cloned-not_peeled in(0) in(0)
3781 // orig-peeled
3782
3783 for (uint i = 0; i < loop->_body.size(); i++) {
3784 Node *n = loop->_body.at(i);
3785 if (!n->is_CFG() && n->in(0) != nullptr &&
3786 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3787 Node* n_clone = old_new[n->_idx];
3788 _igvn.replace_input_of(n_clone, 0, new_head_clone);
3789 }
3790 }
3791
3792 // Backedge of the surviving new_head (the clone) is original last_peel
3793 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3794
3795 // Cut first node in original not_peel set
3796 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
3797 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
3798 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
3799
3800 // Copy head_clone back-branch info to original head
3801 // and remove original head's loop entry and
3802 // clone head's back-branch
3803 _igvn.rehash_node_delayed(head); // Multiple edge updates
3804 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
3805 head->set_req(LoopNode::LoopBackControl, C->top());
3806 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3807
3808 // Similarly modify the phis
3809 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3810 Node* use = head->fast_out(k);
3811 if (use->is_Phi() && use->outcnt() > 0) {
3812 Node* use_clone = old_new[use->_idx];
3813 _igvn.rehash_node_delayed(use); // Multiple edge updates
3814 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
3815 use->set_req(LoopNode::LoopBackControl, C->top());
3816 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3817 }
3818 }
3819
3820 // Step 4: update dominator tree and dominator depth
3821
3822 set_idom(head, orig_tail_clone, dd);
3823 recompute_dom_depth();
3824
3825 // Inhibit more partial peeling on this loop
3826 new_head_clone->set_partial_peel_loop();
3827 C->set_major_progress();
3828 loop->record_for_igvn();
3829
3830 #ifndef PRODUCT
3831 if (TracePartialPeeling) {
3832 tty->print_cr("\nafter partial peel one iteration");
3833 Node_List wl;
3834 Node* t = last_peel;
3835 while (true) {
3836 wl.push(t);
3837 if (t == head_clone) break;
3838 t = idom(t);
3839 }
3840 while (wl.size() > 0) {
3841 Node* tt = wl.pop();
3842 if (tt == head) tty->print_cr("orig head");
3843 else if (tt == new_head_clone) tty->print_cr("new head");
3844 else if (tt == head_clone) tty->print_cr("clone head");
3845 tt->dump();
3846 }
3847 }
3848 #endif
3849 return true;
3850 }
3851
3852 // Transform:
3853 //
3854 // loop<-----------------+
3855 // | |
3856 // stmt1 stmt2 .. stmtn |
3857 // | | | |
3858 // \ | / |
3859 // v v v |
3860 // region |
3861 // | |
3862 // shared_stmt |
3863 // | |
3864 // v |
3865 // if |
3866 // / \ |
3867 // | -----------+
3868 // v
3869 //
3870 // into:
3871 //
3872 // loop<-------------------+
3873 // | |
3874 // v |
3875 // +->loop |
3876 // | | |
3877 // | stmt1 stmt2 .. stmtn |
3878 // | | | | |
3879 // | | \ / |
3880 // | | v v |
3881 // | | region1 |
3882 // | | | |
3883 // | shared_stmt shared_stmt |
3884 // | | | |
3885 // | v v |
3886 // | if if |
3887 // | /\ / \ |
3888 // +-- | | -------+
3889 // \ /
3890 // v v
3891 // region2
3892 //
3893 // (region2 is shown to merge mirrored projections of the loop exit
3894 // ifs to make the diagram clearer but they really merge the same
3895 // projection)
3896 //
3897 // Conditions for this transformation to trigger:
3898 // - the path through stmt1 is frequent enough
3899 // - the inner loop will be turned into a counted loop after transformation
3900 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
3901 if (!DuplicateBackedge) {
3902 return false;
3903 }
3904 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
3905 if (!loop->_head->is_Loop()) {
3906 return false;
3907 }
3908
3909 uint estimate = loop->est_loop_clone_sz(1);
3910 if (exceeding_node_budget(estimate)) {
3911 return false;
3912 }
3913
3914 LoopNode *head = loop->_head->as_Loop();
3915
3916 Node* region = nullptr;
3917 IfNode* exit_test = nullptr;
3918 uint inner;
3919 float f;
3920 if (StressDuplicateBackedge) {
3921 if (head->is_strip_mined()) {
3922 return false;
3923 }
3924 Node* c = head->in(LoopNode::LoopBackControl);
3925
3926 while (c != head) {
3927 if (c->is_Region()) {
3928 region = c;
3929 }
3930 c = idom(c);
3931 }
3932
3933 if (region == nullptr) {
3934 return false;
3935 }
3936
3937 inner = 1;
3938 } else {
3939 // Is the shape of the loop that of a counted loop...
3940 Node* back_control = loop_exit_control(head, loop);
3941 if (back_control == nullptr) {
3942 return false;
3943 }
3944
3945 BoolTest::mask bt = BoolTest::illegal;
3946 float cl_prob = 0;
3947 Node* incr = nullptr;
3948 Node* limit = nullptr;
3949 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
3950 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
3951 return false;
3952 }
3953
3954 // With an extra phi for the candidate iv?
3955 // Or the region node is the loop head
3956 if (!incr->is_Phi() || incr->in(0) == head) {
3957 return false;
3958 }
3959
3960 PathFrequency pf(head, this);
3961 region = incr->in(0);
3962
3963 // Go over all paths for the extra phi's region and see if that
3964 // path is frequent enough and would match the expected iv shape
3965 // if the extra phi is removed
3966 inner = 0;
3967 for (uint i = 1; i < incr->req(); ++i) {
3968 Node* in = incr->in(i);
3969 Node* trunc1 = nullptr;
3970 Node* trunc2 = nullptr;
3971 const TypeInteger* iv_trunc_t = nullptr;
3972 Node* orig_in = in;
3973 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
3974 continue;
3975 }
3976 assert(in->Opcode() == Op_AddI, "wrong increment code");
3977 Node* xphi = nullptr;
3978 Node* stride = loop_iv_stride(in, loop, xphi);
3979
3980 if (stride == nullptr) {
3981 continue;
3982 }
3983
3984 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
3985 if (phi == nullptr ||
3986 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
3987 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
3988 return false;
3989 }
3990
3991 f = pf.to(region->in(i));
3992 if (f > 0.5) {
3993 inner = i;
3994 break;
3995 }
3996 }
3997
3998 if (inner == 0) {
3999 return false;
4000 }
4001
4002 exit_test = back_control->in(0)->as_If();
4003 }
4004
4005 if (idom(region)->is_Catch()) {
4006 return false;
4007 }
4008
4009 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4010 Unique_Node_List wq;
4011 wq.push(head->in(LoopNode::LoopBackControl));
4012 for (uint i = 0; i < wq.size(); i++) {
4013 Node* c = wq.at(i);
4014 assert(get_loop(c) == loop, "not in the right loop?");
4015 if (c->is_Region()) {
4016 if (c != region) {
4017 for (uint j = 1; j < c->req(); ++j) {
4018 wq.push(c->in(j));
4019 }
4020 }
4021 } else {
4022 wq.push(c->in(0));
4023 }
4024 assert(!is_dominator(c, region) || c == region, "shouldn't go above region");
4025 }
4026
4027 Node* region_dom = idom(region);
4028
4029 // Can't do the transformation if this would cause a membar pair to
4030 // be split
4031 for (uint i = 0; i < wq.size(); i++) {
4032 Node* c = wq.at(i);
4033 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4034 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4035 if (!wq.member(c->as_MemBar()->leading_membar())) {
4036 return false;
4037 }
4038 }
4039 }
4040
4041 // Collect data nodes that need to be clones as well
4042 int dd = dom_depth(head);
4043
4044 for (uint i = 0; i < loop->_body.size(); ++i) {
4045 Node* n = loop->_body.at(i);
4046 if (has_ctrl(n)) {
4047 Node* c = get_ctrl(n);
4048 if (wq.member(c)) {
4049 wq.push(n);
4050 }
4051 } else {
4052 set_idom(n, idom(n), dd);
4053 }
4054 }
4055
4056 // clone shared_stmt
4057 clone_loop_body(wq, old_new, nullptr);
4058
4059 Node* region_clone = old_new[region->_idx];
4060 region_clone->set_req(inner, C->top());
4061 set_idom(region, region->in(inner), dd);
4062
4063 // Prepare the outer loop
4064 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4065 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4066 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4067 set_idom(head, outer_head, dd);
4068
4069 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4070
4071 // Make one of the shared_stmt copies only reachable from stmt1, the
4072 // other only from stmt2..stmtn.
4073 Node* dom = nullptr;
4074 for (uint i = 1; i < region->req(); ++i) {
4075 if (i != inner) {
4076 _igvn.replace_input_of(region, i, C->top());
4077 }
4078 Node* in = region_clone->in(i);
4079 if (in->is_top()) {
4080 continue;
4081 }
4082 if (dom == nullptr) {
4083 dom = in;
4084 } else {
4085 dom = dom_lca(dom, in);
4086 }
4087 }
4088
4089 set_idom(region_clone, dom, dd);
4090
4091 // Set up the outer loop
4092 for (uint i = 0; i < head->outcnt(); i++) {
4093 Node* u = head->raw_out(i);
4094 if (u->is_Phi()) {
4095 Node* outer_phi = u->clone();
4096 outer_phi->set_req(0, outer_head);
4097 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4098 if (backedge == nullptr) {
4099 backedge = u->in(LoopNode::LoopBackControl);
4100 }
4101 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4102 register_new_node(outer_phi, outer_head);
4103 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4104 }
4105 }
4106
4107 // create control and data nodes for out of loop uses (including region2)
4108 Node_List worklist;
4109 uint new_counter = C->unique();
4110 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4111
4112 Node_List *split_if_set = nullptr;
4113 Node_List *split_bool_set = nullptr;
4114 Node_List *split_cex_set = nullptr;
4115 fix_data_uses(wq, loop, ControlAroundStripMined, head->is_strip_mined() ? loop->_parent : loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
4116
4117 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4118
4119 if (exit_test != nullptr) {
4120 float cnt = exit_test->_fcnt;
4121 if (cnt != COUNT_UNKNOWN) {
4122 exit_test->_fcnt = cnt * f;
4123 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4124 }
4125 }
4126
4127 C->set_major_progress();
4128
4129 return true;
4130 }
4131
4132 // Having ReductionNodes in the loop is expensive. They need to recursively
4133 // fold together the vector values, for every vectorized loop iteration. If
4134 // we encounter the following pattern, we can vector accumulate the values
4135 // inside the loop, and only have a single UnorderedReduction after the loop.
4136 //
4137 // CountedLoop init
4138 // | |
4139 // +------+ | +-----------------------+
4140 // | | | |
4141 // PhiNode (s) |
4142 // | |
4143 // | Vector |
4144 // | | |
4145 // UnorderedReduction (first_ur) |
4146 // | |
4147 // ... Vector |
4148 // | | |
4149 // UnorderedReduction (last_ur) |
4150 // | |
4151 // +---------------------+
4152 //
4153 // We patch the graph to look like this:
4154 //
4155 // CountedLoop identity_vector
4156 // | |
4157 // +-------+ | +---------------+
4158 // | | | |
4159 // PhiNode (v) |
4160 // | |
4161 // | Vector |
4162 // | | |
4163 // VectorAccumulator |
4164 // | |
4165 // ... Vector |
4166 // | | |
4167 // init VectorAccumulator |
4168 // | | | |
4169 // UnorderedReduction +-----------+
4170 //
4171 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4172 // use vector_accumulators, which do the same reductions, but only element
4173 // wise. This is a single operation per vector_accumulator, rather than many
4174 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4175 // after the loop, and also reduce the init value into it.
4176 // We can not do this with all reductions. Some reductions do not allow the
4177 // reordering of operations (for example float addition).
4178 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4179 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4180
4181 // Find all Phi nodes with UnorderedReduction on backedge.
4182 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4183 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4184 Node* phi = cl->fast_out(j);
4185 // We have a phi with a single use, and a UnorderedReduction on the backedge.
4186 if (!phi->is_Phi() || phi->outcnt() != 1 || !phi->in(2)->is_UnorderedReduction()) {
4187 continue;
4188 }
4189
4190 UnorderedReductionNode* last_ur = phi->in(2)->as_UnorderedReduction();
4191
4192 // Determine types
4193 const TypeVect* vec_t = last_ur->vect_type();
4194 uint vector_length = vec_t->length();
4195 BasicType bt = vec_t->element_basic_type();
4196 const Type* bt_t = Type::get_const_basic_type(bt);
4197
4198 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4199 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4200 const int vopc = VectorNode::opcode(sopc, bt);
4201 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4202 DEBUG_ONLY( last_ur->dump(); )
4203 assert(false, "do not have normal vector op for this reduction");
4204 continue; // not implemented -> fails
4205 }
4206
4207 // Traverse up the chain of UnorderedReductions, checking that it loops back to
4208 // the phi. Check that all UnorderedReductions only have a single use, except for
4209 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4210 // are outside the loop.
4211 UnorderedReductionNode* current = last_ur;
4212 UnorderedReductionNode* first_ur = nullptr;
4213 while (true) {
4214 assert(current->is_UnorderedReduction(), "sanity");
4215
4216 // Expect no ctrl and a vector_input from within the loop.
4217 Node* ctrl = current->in(0);
4218 Node* vector_input = current->in(2);
4219 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4220 DEBUG_ONLY( current->dump(1); )
4221 assert(false, "reduction has ctrl or bad vector_input");
4222 break; // Chain traversal fails.
4223 }
4224
4225 // Expect single use of UnorderedReduction, except for last_ur.
4226 if (current == last_ur) {
4227 // Expect all uses to be outside the loop, except phi.
4228 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4229 Node* use = current->fast_out(k);
4230 if (use != phi && ctrl_or_self(use) == cl) {
4231 DEBUG_ONLY( current->dump(-1); )
4232 assert(false, "reduction has use inside loop");
4233 // Should not be allowed by SuperWord::mark_reductions
4234 return; // bail out of optimization
4235 }
4236 }
4237 } else {
4238 if (current->outcnt() != 1) {
4239 break; // Chain traversal fails.
4240 }
4241 }
4242
4243 // Expect another UnorderedReduction or phi as the scalar input.
4244 Node* scalar_input = current->in(1);
4245 if (scalar_input->is_UnorderedReduction() &&
4246 scalar_input->Opcode() == current->Opcode()) {
4247 // Move up the UnorderedReduction chain.
4248 current = scalar_input->as_UnorderedReduction();
4249 } else if (scalar_input == phi) {
4250 // Chain terminates at phi.
4251 first_ur = current;
4252 current = nullptr;
4253 break; // Success.
4254 } else {
4255 // scalar_input is neither phi nor a matching reduction
4256 // Can for example be scalar reduction when we have
4257 // partial vectorization.
4258 break; // Chain traversal fails.
4259 }
4260 }
4261 if (current != nullptr) {
4262 // Chain traversal was not successful.
4263 continue;
4264 }
4265 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4266
4267 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4268 set_ctrl(identity_scalar, C->root());
4269 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t);
4270 register_new_node(identity_vector, C->root());
4271 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4272 VectorNode::trace_new_vector(identity_vector, "UnorderedReduction");
4273
4274 // Turn the scalar phi into a vector phi.
4275 _igvn.rehash_node_delayed(phi);
4276 Node* init = phi->in(1); // Remember init before replacing it.
4277 phi->set_req_X(1, identity_vector, &_igvn);
4278 phi->as_Type()->set_type(vec_t);
4279 _igvn.set_type(phi, vec_t);
4280
4281 // Traverse down the chain of UnorderedReductions, and replace them with vector_accumulators.
4282 current = first_ur;
4283 while (true) {
4284 // Create vector_accumulator to replace current.
4285 Node* last_vector_accumulator = current->in(1);
4286 Node* vector_input = current->in(2);
4287 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4288 register_new_node(vector_accumulator, cl);
4289 _igvn.replace_node(current, vector_accumulator);
4290 VectorNode::trace_new_vector(vector_accumulator, "UnorderedReduction");
4291 if (current == last_ur) {
4292 break;
4293 }
4294 current = vector_accumulator->unique_out()->as_UnorderedReduction();
4295 }
4296
4297 // Create post-loop reduction.
4298 Node* last_accumulator = phi->in(2);
4299 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4300
4301 // Take over uses of last_accumulator that are not in the loop.
4302 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4303 Node* use = last_accumulator->out(i);
4304 if (use != phi && use != post_loop_reduction) {
4305 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4306 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4307 --i;
4308 }
4309 }
4310 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4311 VectorNode::trace_new_vector(post_loop_reduction, "UnorderedReduction");
4312
4313 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4314 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4315 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4316 }
4317 }