1 /*
2 * Copyright (c) 1999, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/subtypenode.hpp"
42 #include "opto/superword.hpp"
43 #include "opto/vectornode.hpp"
44 #include "utilities/macros.hpp"
45
46 //=============================================================================
47 //------------------------------split_thru_phi---------------------------------
48 // Split Node 'n' through merge point if there is enough win.
49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
50 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
51 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
52 // ConvI2L/ConvL2I 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 = makecon(t);
124 } else {
125 // We now call Identity to try to simplify the cloned node.
126 // Note that some Identity methods call phase->type(this).
127 // Make sure that the type array is big enough for
128 // our new node, even though we may throw the node away.
129 // (Note: This tweaking with igvn only works because x is a new node.)
130 _igvn.set_type(x, t);
131 // If x is a TypeNode, capture any more-precise type permanently into Node
132 // otherwise it will be not updated during igvn->transform since
133 // igvn->type(x) is set to x->Value() already.
134 x->raise_bottom_type(t);
135 Node* y = x->Identity(&_igvn);
136 if (y != x) {
137 wins++;
138 x = y;
139 } else {
140 y = _igvn.hash_find(x);
141 if (y == nullptr) {
142 y = similar_subtype_check(x, region->in(i));
143 }
144 if (y) {
145 wins++;
146 x = y;
147 } else {
148 // Else x is a new node we are keeping
149 // We do not need register_new_node_with_optimizer
150 // because set_type has already been called.
151 _igvn._worklist.push(x);
152 }
153 }
154 }
155
156 phi->set_req( i, x );
157
158 if (the_clone == nullptr) {
159 continue;
160 }
161
162 if (the_clone != x) {
163 _igvn.remove_dead_node(the_clone);
164 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
165 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
166 // it is not a win if 'x' moved from an outer to an inner loop
167 // this edge case can only happen for Load nodes
168 wins = 0;
169 break;
170 }
171 }
172 // Too few wins?
173 if (wins <= policy) {
174 _igvn.remove_dead_node(phi);
175 return nullptr;
176 }
177
178 // Record Phi
179 register_new_node( phi, region );
180
181 for (uint i2 = 1; i2 < phi->req(); i2++) {
182 Node *x = phi->in(i2);
183 // If we commoned up the cloned 'x' with another existing Node,
184 // the existing Node picks up a new use. We need to make the
185 // existing Node occur higher up so it dominates its uses.
186 Node *old_ctrl;
187 IdealLoopTree *old_loop;
188
189 if (x->is_Con()) {
190 assert(get_ctrl(x) == C->root(), "constant control is not root");
191 continue;
192 }
193 // The occasional new node
194 if (x->_idx >= old_unique) { // Found a new, unplaced node?
195 old_ctrl = nullptr;
196 old_loop = nullptr; // Not in any prior loop
197 } else {
198 old_ctrl = get_ctrl(x);
199 old_loop = get_loop(old_ctrl); // Get prior loop
200 }
201 // New late point must dominate new use
202 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
203 if (new_ctrl == old_ctrl) // Nothing is changed
204 continue;
205
206 IdealLoopTree *new_loop = get_loop(new_ctrl);
207
208 // Don't move x into a loop if its uses are
209 // outside of loop. Otherwise x will be cloned
210 // for each use outside of this loop.
211 IdealLoopTree *use_loop = get_loop(region);
212 if (!new_loop->is_member(use_loop) &&
213 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
214 // Take early control, later control will be recalculated
215 // during next iteration of loop optimizations.
216 new_ctrl = get_early_ctrl(x);
217 new_loop = get_loop(new_ctrl);
218 }
219 // Set new location
220 set_ctrl(x, new_ctrl);
221 // If changing loop bodies, see if we need to collect into new body
222 if (old_loop != new_loop) {
223 if (old_loop && !old_loop->_child)
224 old_loop->_body.yank(x);
225 if (!new_loop->_child)
226 new_loop->_body.push(x); // Collect body info
227 }
228 }
229
230 return phi;
231 }
232
233 // Test whether node 'x' can move into an inner loop relative to node 'n'.
234 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
235 // BUT it can also return true and 'x' is in the outer loop
236 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
237 IdealLoopTree* n_loop_tree = get_loop(n_loop);
238 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
239 // x_loop_tree should be outer or same loop as n_loop_tree
240 return !x_loop_tree->is_member(n_loop_tree);
241 }
242
243 // Subtype checks that carry profile data don't common so look for a replacement by following edges
244 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
245 if (x->is_SubTypeCheck()) {
246 Node* in1 = x->in(1);
247 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
248 Node* u = in1->fast_out(i);
249 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
250 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
251 Node* bol = u->fast_out(j);
252 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
253 Node* iff = bol->fast_out(k);
254 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
255 // unrelated profile
256 if (iff->is_If() && is_dominator(iff, r_in)) {
257 return u;
258 }
259 }
260 }
261 }
262 }
263 }
264 return nullptr;
265 }
266
267 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
268 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
269 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
270 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
271 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
272 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
273 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
274 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
275 const Type* zero;
276 switch (n->Opcode()) {
277 case Op_DivI:
278 case Op_ModI:
279 case Op_UDivI:
280 case Op_UModI:
281 zero = TypeInt::ZERO;
282 break;
283 case Op_DivL:
284 case Op_ModL:
285 case Op_UDivL:
286 case Op_UModL:
287 zero = TypeLong::ZERO;
288 break;
289 default:
290 return false;
291 }
292
293 if (n->in(0) != nullptr) {
294 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
295 return true;
296 }
297
298 Node* divisor = n->in(2);
299 return is_divisor_loop_phi(divisor, region) &&
300 loop_phi_backedge_type_contains_zero(divisor, zero);
301 }
302
303 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
304 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
305 }
306
307 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
308 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
309 }
310
311 //------------------------------dominated_by------------------------------------
312 // Replace the dominated test with an obvious true or false. Place it on the
313 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
314 // live path up to the dominating control.
315 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
316 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
317
318 // prevdom is the dominating projection of the dominating test.
319 assert(iff->Opcode() == Op_If ||
320 iff->Opcode() == Op_CountedLoopEnd ||
321 iff->Opcode() == Op_LongCountedLoopEnd ||
322 iff->Opcode() == Op_RangeCheck ||
323 iff->Opcode() == Op_ParsePredicate,
324 "Check this code when new subtype is added");
325
326 int pop = prevdom->Opcode();
327 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
328 if (flip) {
329 if (pop == Op_IfTrue)
330 pop = Op_IfFalse;
331 else
332 pop = Op_IfTrue;
333 }
334 // 'con' is set to true or false to kill the dominated test.
335 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
336 // Hack the dominated test
337 _igvn.replace_input_of(iff, 1, con);
338
339 // If I don't have a reachable TRUE and FALSE path following the IfNode then
340 // I can assume this path reaches an infinite loop. In this case it's not
341 // important to optimize the data Nodes - either the whole compilation will
342 // be tossed or this path (and all data Nodes) will go dead.
343 if (iff->outcnt() != 2) {
344 return;
345 }
346
347 // Make control-dependent data Nodes on the live path (path that will remain
348 // once the dominated IF is removed) become control-dependent on the
349 // dominating projection.
350 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
351
352 if (dp == nullptr) {
353 return;
354 }
355
356 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
357 }
358
359 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
360 IdealLoopTree* old_loop = get_loop(source);
361
362 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
363 Node* out = source->fast_out(i); // Control-dependent node
364 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
365 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
366 assert(out->in(0) == source, "must be control dependent on source");
367 _igvn.replace_input_of(out, 0, dominator);
368 if (pin_array_access_nodes) {
369 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
370 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
371 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
372 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
373 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
374 // dominating check.
375 Node* clone = out->pin_array_access_node();
376 if (clone != nullptr) {
377 clone = _igvn.register_new_node_with_optimizer(clone, out);
378 _igvn.replace_node(out, clone);
379 out = clone;
380 }
381 }
382 set_early_ctrl(out, false);
383 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
384 if (old_loop != new_loop) {
385 if (!old_loop->_child) {
386 old_loop->_body.yank(out);
387 }
388 if (!new_loop->_child) {
389 new_loop->_body.push(out);
390 }
391 }
392 --i;
393 --imax;
394 }
395 }
396 }
397
398 //------------------------------has_local_phi_input----------------------------
399 // Return TRUE if 'n' has Phi inputs from its local block and no other
400 // block-local inputs (all non-local-phi inputs come from earlier blocks)
401 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
402 Node *n_ctrl = get_ctrl(n);
403 // See if some inputs come from a Phi in this block, or from before
404 // this block.
405 uint i;
406 for( i = 1; i < n->req(); i++ ) {
407 Node *phi = n->in(i);
408 if( phi->is_Phi() && phi->in(0) == n_ctrl )
409 break;
410 }
411 if( i >= n->req() )
412 return nullptr; // No Phi inputs; nowhere to clone thru
413
414 // Check for inputs created between 'n' and the Phi input. These
415 // must split as well; they have already been given the chance
416 // (courtesy of a post-order visit) and since they did not we must
417 // recover the 'cost' of splitting them by being very profitable
418 // when splitting 'n'. Since this is unlikely we simply give up.
419 for( i = 1; i < n->req(); i++ ) {
420 Node *m = n->in(i);
421 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
422 // We allow the special case of AddP's with no local inputs.
423 // This allows us to split-up address expressions.
424 if (m->is_AddP() &&
425 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
426 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
427 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
428 // Move the AddP up to the dominating point. That's fine because control of m's inputs
429 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
430 Node* c = find_non_split_ctrl(idom(n_ctrl));
431 if (c->is_OuterStripMinedLoop()) {
432 c->as_Loop()->verify_strip_mined(1);
433 c = c->in(LoopNode::EntryControl);
434 }
435 set_ctrl_and_loop(m, c);
436 continue;
437 }
438 return nullptr;
439 }
440 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
441 }
442
443 return n_ctrl;
444 }
445
446 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
447 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
448 assert(bt == T_INT || bt == T_LONG, "only for integers");
449 int n_op = n->Opcode();
450
451 if (n_op == Op_LShift(bt)) {
452 // Scale is loop invariant
453 Node* scale = n->in(2);
454 Node* scale_ctrl = get_ctrl(scale);
455 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
456 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
457 return nullptr;
458 }
459 const TypeInt* scale_t = scale->bottom_type()->isa_int();
460 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
461 return nullptr; // Dont bother with byte/short masking
462 }
463 // Add must vary with loop (else shift would be loop-invariant)
464 Node* add = n->in(1);
465 Node* add_ctrl = get_ctrl(add);
466 IdealLoopTree* add_loop = get_loop(add_ctrl);
467 if (n_loop != add_loop) {
468 return nullptr; // happens w/ evil ZKM loops
469 }
470
471 // Convert I-V into I+ (0-V); same for V-I
472 if (add->Opcode() == Op_Sub(bt) &&
473 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
474 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
475 Node* zero = integercon(0, bt);
476 Node* neg = SubNode::make(zero, add->in(2), bt);
477 register_new_node_with_ctrl_of(neg, add->in(2));
478 add = AddNode::make(add->in(1), neg, bt);
479 register_new_node(add, add_ctrl);
480 }
481 if (add->Opcode() != Op_Add(bt)) return nullptr;
482 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
483 // See if one add input is loop invariant
484 Node* add_var = add->in(1);
485 Node* add_var_ctrl = get_ctrl(add_var);
486 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
487 Node* add_invar = add->in(2);
488 Node* add_invar_ctrl = get_ctrl(add_invar);
489 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
490 if (add_invar_loop == n_loop) {
491 // Swap to find the invariant part
492 add_invar = add_var;
493 add_invar_ctrl = add_var_ctrl;
494 add_invar_loop = add_var_loop;
495 add_var = add->in(2);
496 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
497 return nullptr;
498 }
499 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
500 return nullptr; // No invariant part of the add?
501 }
502
503 // Yes! Reshape address expression!
504 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
505 Node* inv_scale_ctrl =
506 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
507 add_invar_ctrl : scale_ctrl;
508 register_new_node(inv_scale, inv_scale_ctrl);
509 Node* var_scale = LShiftNode::make(add_var, scale, bt);
510 register_new_node(var_scale, n_ctrl);
511 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
512 register_new_node(var_add, n_ctrl);
513 _igvn.replace_node(n, var_add);
514 return var_add;
515 }
516 return nullptr;
517 }
518
519 //------------------------------remix_address_expressions----------------------
520 // Rework addressing expressions to get the most loop-invariant stuff
521 // moved out. We'd like to do all associative operators, but it's especially
522 // important (common) to do address expressions.
523 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
524 if (!has_ctrl(n)) return nullptr;
525 Node* n_ctrl = get_ctrl(n);
526 IdealLoopTree* n_loop = get_loop(n_ctrl);
527
528 // See if 'n' mixes loop-varying and loop-invariant inputs and
529 // itself is loop-varying.
530
531 // Only interested in binary ops (and AddP)
532 if (n->req() < 3 || n->req() > 4) return nullptr;
533
534 Node* n1_ctrl = get_ctrl(n->in( 1));
535 Node* n2_ctrl = get_ctrl(n->in( 2));
536 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
537 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
538 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
539 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
540
541 // Does one of my inputs spin in a tighter loop than self?
542 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
543 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
544 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
545 return nullptr; // Leave well enough alone
546 }
547
548 // Is at least one of my inputs loop-invariant?
549 if (n1_loop == n_loop &&
550 n2_loop == n_loop &&
551 n3_loop == n_loop) {
552 return nullptr; // No loop-invariant inputs
553 }
554
555 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
556 if (res != nullptr) {
557 return res;
558 }
559 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
560 if (res != nullptr) {
561 return res;
562 }
563
564 int n_op = n->Opcode();
565 // Replace (I+V) with (V+I)
566 if (n_op == Op_AddI ||
567 n_op == Op_AddL ||
568 n_op == Op_AddF ||
569 n_op == Op_AddD ||
570 n_op == Op_MulI ||
571 n_op == Op_MulL ||
572 n_op == Op_MulF ||
573 n_op == Op_MulD) {
574 if (n2_loop == n_loop) {
575 assert(n1_loop != n_loop, "");
576 n->swap_edges(1, 2);
577 }
578 }
579
580 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
581 // but not if I2 is a constant. Skip for irreducible loops.
582 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
583 if (n2_loop == n_loop && n3_loop != n_loop) {
584 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
585 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
586 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
587 IdealLoopTree* n22loop = get_loop(n22_ctrl);
588 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
589 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
590 n23_loop == n_loop) {
591 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
592 // Stuff new AddP in the loop preheader
593 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
594 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
595 register_new_node(add2, n_ctrl);
596 _igvn.replace_node(n, add2);
597 return add2;
598 }
599 }
600 }
601
602 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
603 if (n2_loop != n_loop && n3_loop == n_loop) {
604 if (n->in(3)->Opcode() == Op_AddX) {
605 Node* V = n->in(3)->in(1);
606 Node* I = n->in(3)->in(2);
607 if (is_member(n_loop,get_ctrl(V))) {
608 } else {
609 Node *tmp = V; V = I; I = tmp;
610 }
611 if (!is_member(n_loop,get_ctrl(I))) {
612 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
613 // Stuff new AddP in the loop preheader
614 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
615 Node* add2 = new AddPNode(n->in(1), add1, V);
616 register_new_node(add2, n_ctrl);
617 _igvn.replace_node(n, add2);
618 return add2;
619 }
620 }
621 }
622 }
623
624 return nullptr;
625 }
626
627 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
628 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
629 assert(n->Opcode() == Op_AddI, "sanity");
630 Node * nn = nullptr;
631 Node * in1 = n->in(1);
632 Node * in2 = n->in(2);
633 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
634 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
635 if (loop_n->is_counted() &&
636 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
637 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
638 Matcher::match_rule_supported(Op_MulAddS2I)) {
639 Node* mul_in1 = in1->in(1);
640 Node* mul_in2 = in1->in(2);
641 Node* mul_in3 = in2->in(1);
642 Node* mul_in4 = in2->in(2);
643 if (mul_in1->Opcode() == Op_LoadS &&
644 mul_in2->Opcode() == Op_LoadS &&
645 mul_in3->Opcode() == Op_LoadS &&
646 mul_in4->Opcode() == Op_LoadS) {
647 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
648 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
649 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
650 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
651 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
652 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
653 // All nodes should be in the same counted loop.
654 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
655 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
656 Node* adr1 = mul_in1->in(MemNode::Address);
657 Node* adr2 = mul_in2->in(MemNode::Address);
658 Node* adr3 = mul_in3->in(MemNode::Address);
659 Node* adr4 = mul_in4->in(MemNode::Address);
660 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
661 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
662 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
663 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
664 register_new_node_with_ctrl_of(nn, n);
665 _igvn.replace_node(n, nn);
666 return nn;
667 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
668 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
669 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
670 register_new_node_with_ctrl_of(nn, n);
671 _igvn.replace_node(n, nn);
672 return nn;
673 }
674 }
675 }
676 }
677 }
678 }
679 return nn;
680 }
681
682 //------------------------------conditional_move-------------------------------
683 // Attempt to replace a Phi with a conditional move. We have some pretty
684 // strict profitability requirements. All Phis at the merge point must
685 // be converted, so we can remove the control flow. We need to limit the
686 // number of c-moves to a small handful. All code that was in the side-arms
687 // of the CFG diamond is now speculatively executed. This code has to be
688 // "cheap enough". We are pretty much limited to CFG diamonds that merge
689 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
690 Node *PhaseIdealLoop::conditional_move( Node *region ) {
691
692 assert(region->is_Region(), "sanity check");
693 if (region->req() != 3) return nullptr;
694
695 // Check for CFG diamond
696 Node *lp = region->in(1);
697 Node *rp = region->in(2);
698 if (!lp || !rp) return nullptr;
699 Node *lp_c = lp->in(0);
700 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
701 IfNode *iff = lp_c->as_If();
702
703 // Check for ops pinned in an arm of the diamond.
704 // Can't remove the control flow in this case
705 if (lp->outcnt() > 1) return nullptr;
706 if (rp->outcnt() > 1) return nullptr;
707
708 IdealLoopTree* r_loop = get_loop(region);
709 assert(r_loop == get_loop(iff), "sanity");
710 // Always convert to CMOVE if all results are used only outside this loop.
711 bool used_inside_loop = (r_loop == _ltree_root);
712
713 // Check profitability
714 int cost = 0;
715 int phis = 0;
716 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
717 Node *out = region->fast_out(i);
718 if (!out->is_Phi()) continue; // Ignore other control edges, etc
719 phis++;
720 PhiNode* phi = out->as_Phi();
721 BasicType bt = phi->type()->basic_type();
722 switch (bt) {
723 case T_DOUBLE:
724 case T_FLOAT:
725 if (C->use_cmove()) {
726 continue; //TODO: maybe we want to add some cost
727 }
728 cost += Matcher::float_cmove_cost(); // Could be very expensive
729 break;
730 case T_LONG: {
731 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
732 }
733 case T_INT: // These all CMOV fine
734 case T_ADDRESS: { // (RawPtr)
735 cost++;
736 break;
737 }
738 case T_NARROWOOP: // Fall through
739 case T_OBJECT: { // Base oops are OK, but not derived oops
740 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
741 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
742 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
743 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
744 // have a Phi for the base here that we convert to a CMOVE all is well
745 // and good. But if the base is dead, we'll not make a CMOVE. Later
746 // the allocator will have to produce a base by creating a CMOVE of the
747 // relevant bases. This puts the allocator in the business of
748 // manufacturing expensive instructions, generally a bad plan.
749 // Just Say No to Conditionally-Moved Derived Pointers.
750 if (tp && tp->offset() != 0)
751 return nullptr;
752 cost++;
753 break;
754 }
755 default:
756 return nullptr; // In particular, can't do memory or I/O
757 }
758 // Add in cost any speculative ops
759 for (uint j = 1; j < region->req(); j++) {
760 Node *proj = region->in(j);
761 Node *inp = phi->in(j);
762 if (get_ctrl(inp) == proj) { // Found local op
763 cost++;
764 // Check for a chain of dependent ops; these will all become
765 // speculative in a CMOV.
766 for (uint k = 1; k < inp->req(); k++)
767 if (get_ctrl(inp->in(k)) == proj)
768 cost += ConditionalMoveLimit; // Too much speculative goo
769 }
770 }
771 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
772 // This will likely Split-If, a higher-payoff operation.
773 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
774 Node* use = phi->fast_out(k);
775 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
776 cost += ConditionalMoveLimit;
777 // Is there a use inside the loop?
778 // Note: check only basic types since CMoveP is pinned.
779 if (!used_inside_loop && is_java_primitive(bt)) {
780 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
781 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
782 used_inside_loop = true;
783 }
784 }
785 }
786 }//for
787 Node* bol = iff->in(1);
788 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
789 if (bol->is_OpaqueTemplateAssertionPredicate()) {
790 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
791 return nullptr;
792 }
793 if (bol->is_OpaqueMultiversioning()) {
794 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
795 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
796 // and will make the multiversion_if constant fold in the next IGVN round.
797 return nullptr;
798 }
799 if (!bol->is_Bool()) {
800 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
801 return nullptr;
802 }
803 int cmp_op = bol->in(1)->Opcode();
804 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
805 return nullptr;
806 }
807 // It is expensive to generate flags from a float compare.
808 // Avoid duplicated float compare.
809 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
810
811 float infrequent_prob = PROB_UNLIKELY_MAG(3);
812 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
813 if (used_inside_loop) {
814 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
815
816 // BlockLayoutByFrequency optimization moves infrequent branch
817 // from hot path. No point in CMOV'ing in such case (110 is used
818 // instead of 100 to take into account not exactness of float value).
819 if (BlockLayoutByFrequency) {
820 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
821 }
822 }
823 // Check for highly predictable branch. No point in CMOV'ing if
824 // we are going to predict accurately all the time.
825 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
826 //keep going
827 } else if (iff->_prob < infrequent_prob ||
828 iff->_prob > (1.0f - infrequent_prob))
829 return nullptr;
830
831 // --------------
832 // Now replace all Phis with CMOV's
833 Node *cmov_ctrl = iff->in(0);
834 uint flip = (lp->Opcode() == Op_IfTrue);
835 Node_List wq;
836 while (1) {
837 PhiNode* phi = nullptr;
838 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
839 Node *out = region->fast_out(i);
840 if (out->is_Phi()) {
841 phi = out->as_Phi();
842 break;
843 }
844 }
845 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
846 break;
847 }
848 // Move speculative ops
849 wq.push(phi);
850 while (wq.size() > 0) {
851 Node *n = wq.pop();
852 for (uint j = 1; j < n->req(); j++) {
853 Node* m = n->in(j);
854 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
855 set_ctrl(m, cmov_ctrl);
856 wq.push(m);
857 }
858 }
859 }
860 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
861 register_new_node(cmov, cmov_ctrl);
862 _igvn.replace_node(phi, cmov);
863 #ifndef PRODUCT
864 if (TraceLoopOpts) {
865 tty->print("CMOV ");
866 r_loop->dump_head();
867 if (Verbose) {
868 bol->in(1)->dump(1);
869 cmov->dump(1);
870 }
871 }
872 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
873 #endif
874 }
875
876 // The useless CFG diamond will fold up later; see the optimization in
877 // RegionNode::Ideal.
878 _igvn._worklist.push(region);
879
880 return iff->in(1);
881 }
882
883 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
884 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
885 Node* u = m->fast_out(i);
886 if (u->is_CFG()) {
887 if (u->is_NeverBranch()) {
888 u = u->as_NeverBranch()->proj_out(0);
889 enqueue_cfg_uses(u, wq);
890 } else {
891 wq.push(u);
892 }
893 }
894 }
895 }
896
897 // Try moving a store out of a loop, right before the loop
898 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
899 // Store has to be first in the loop body
900 IdealLoopTree *n_loop = get_loop(n_ctrl);
901 if (n->is_Store() && n_loop != _ltree_root &&
902 n_loop->is_loop() && n_loop->_head->is_Loop() &&
903 n->in(0) != nullptr) {
904 Node* address = n->in(MemNode::Address);
905 Node* value = n->in(MemNode::ValueIn);
906 Node* mem = n->in(MemNode::Memory);
907 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
908 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
909
910 // - address and value must be loop invariant
911 // - memory must be a memory Phi for the loop
912 // - Store must be the only store on this memory slice in the
913 // loop: if there's another store following this one then value
914 // written at iteration i by the second store could be overwritten
915 // at iteration i+n by the first store: it's not safe to move the
916 // first store out of the loop
917 // - nothing must observe the memory Phi: it guarantees no read
918 // before the store, we are also guaranteed the store post
919 // dominates the loop head (ignoring a possible early
920 // exit). Otherwise there would be extra Phi involved between the
921 // loop's Phi and the store.
922 // - there must be no early exit from the loop before the Store
923 // (such an exit most of the time would be an extra use of the
924 // memory Phi but sometimes is a bottom memory Phi that takes the
925 // store as input).
926
927 if (!n_loop->is_member(address_loop) &&
928 !n_loop->is_member(value_loop) &&
929 mem->is_Phi() && mem->in(0) == n_loop->_head &&
930 mem->outcnt() == 1 &&
931 mem->in(LoopNode::LoopBackControl) == n) {
932
933 assert(n_loop->_tail != nullptr, "need a tail");
934 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
935
936 // Verify that there's no early exit of the loop before the store.
937 bool ctrl_ok = false;
938 {
939 // Follow control from loop head until n, we exit the loop or
940 // we reach the tail
941 ResourceMark rm;
942 Unique_Node_List wq;
943 wq.push(n_loop->_head);
944
945 for (uint next = 0; next < wq.size(); ++next) {
946 Node *m = wq.at(next);
947 if (m == n->in(0)) {
948 ctrl_ok = true;
949 continue;
950 }
951 assert(!has_ctrl(m), "should be CFG");
952 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
953 ctrl_ok = false;
954 break;
955 }
956 enqueue_cfg_uses(m, wq);
957 if (wq.size() > 10) {
958 ctrl_ok = false;
959 break;
960 }
961 }
962 }
963 if (ctrl_ok) {
964 // move the Store
965 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
966 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
967 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
968 // Disconnect the phi now. An empty phi can confuse other
969 // optimizations in this pass of loop opts.
970 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
971 n_loop->_body.yank(mem);
972
973 set_ctrl_and_loop(n, n->in(0));
974
975 return n;
976 }
977 }
978 }
979 return nullptr;
980 }
981
982 // Try moving a store out of a loop, right after the loop
983 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
984 if (n->is_Store() && n->in(0) != nullptr) {
985 Node *n_ctrl = get_ctrl(n);
986 IdealLoopTree *n_loop = get_loop(n_ctrl);
987 // Store must be in a loop
988 if (n_loop != _ltree_root && !n_loop->_irreducible) {
989 Node* address = n->in(MemNode::Address);
990 Node* value = n->in(MemNode::ValueIn);
991 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
992 // address must be loop invariant
993 if (!n_loop->is_member(address_loop)) {
994 // Store must be last on this memory slice in the loop and
995 // nothing in the loop must observe it
996 Node* phi = nullptr;
997 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
998 Node* u = n->fast_out(i);
999 if (has_ctrl(u)) { // control use?
1000 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1001 if (!n_loop->is_member(u_loop)) {
1002 continue;
1003 }
1004 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1005 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1006 // multiple phis on the same slice are possible
1007 if (phi != nullptr) {
1008 return;
1009 }
1010 phi = u;
1011 continue;
1012 }
1013 }
1014 return;
1015 }
1016 if (phi != nullptr) {
1017 // Nothing in the loop before the store (next iteration)
1018 // must observe the stored value
1019 bool mem_ok = true;
1020 {
1021 ResourceMark rm;
1022 Unique_Node_List wq;
1023 wq.push(phi);
1024 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1025 Node *m = wq.at(next);
1026 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1027 Node* u = m->fast_out(i);
1028 if (u->is_Store() || u->is_Phi()) {
1029 if (u != n) {
1030 wq.push(u);
1031 mem_ok = (wq.size() <= 10);
1032 }
1033 } else {
1034 mem_ok = false;
1035 break;
1036 }
1037 }
1038 }
1039 }
1040 if (mem_ok) {
1041 // Move the store out of the loop if the LCA of all
1042 // users (except for the phi) is outside the loop.
1043 Node* hook = new Node(1);
1044 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1045 _igvn.rehash_node_delayed(phi);
1046 int count = phi->replace_edge(n, hook, &_igvn);
1047 assert(count > 0, "inconsistent phi");
1048
1049 // Compute latest point this store can go
1050 Node* lca = get_late_ctrl(n, get_ctrl(n));
1051 if (lca->is_OuterStripMinedLoop()) {
1052 lca = lca->in(LoopNode::EntryControl);
1053 }
1054 if (n_loop->is_member(get_loop(lca))) {
1055 // LCA is in the loop - bail out
1056 _igvn.replace_node(hook, n);
1057 return;
1058 }
1059 #ifdef ASSERT
1060 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1061 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1062 n_loop->_head->as_Loop()->verify_strip_mined(1);
1063 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1064 IdealLoopTree* outer_loop = get_loop(outer);
1065 assert(n_loop->_parent == outer_loop, "broken loop tree");
1066 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1067 }
1068 #endif
1069 lca = place_outside_loop(lca, n_loop);
1070 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1071 assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1072
1073 // Move store out of the loop
1074 _igvn.replace_node(hook, n->in(MemNode::Memory));
1075 _igvn.replace_input_of(n, 0, lca);
1076 set_ctrl_and_loop(n, lca);
1077
1078 // Disconnect the phi now. An empty phi can confuse other
1079 // optimizations in this pass of loop opts..
1080 if (phi->in(LoopNode::LoopBackControl) == phi) {
1081 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1082 n_loop->_body.yank(phi);
1083 }
1084 }
1085 }
1086 }
1087 }
1088 }
1089 }
1090
1091 //------------------------------split_if_with_blocks_pre-----------------------
1092 // Do the real work in a non-recursive function. Data nodes want to be
1093 // cloned in the pre-order so they can feed each other nicely.
1094 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1095 // Cloning these guys is unlikely to win
1096 int n_op = n->Opcode();
1097 if (n_op == Op_MergeMem) {
1098 return n;
1099 }
1100 if (n->is_Proj()) {
1101 return n;
1102 }
1103 // Do not clone-up CmpFXXX variations, as these are always
1104 // followed by a CmpI
1105 if (n->is_Cmp()) {
1106 return n;
1107 }
1108 // Attempt to use a conditional move instead of a phi/branch
1109 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1110 Node *cmov = conditional_move( n );
1111 if (cmov) {
1112 return cmov;
1113 }
1114 }
1115 if (n->is_CFG() || n->is_LoadStore()) {
1116 return n;
1117 }
1118 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1119 if (!C->major_progress()) { // If chance of no more loop opts...
1120 _igvn._worklist.push(n); // maybe we'll remove them
1121 }
1122 return n;
1123 }
1124
1125 if (n->is_Con()) {
1126 return n; // No cloning for Con nodes
1127 }
1128
1129 Node *n_ctrl = get_ctrl(n);
1130 if (!n_ctrl) {
1131 return n; // Dead node
1132 }
1133
1134 Node* res = try_move_store_before_loop(n, n_ctrl);
1135 if (res != nullptr) {
1136 return n;
1137 }
1138
1139 // Attempt to remix address expressions for loop invariants
1140 Node *m = remix_address_expressions( n );
1141 if( m ) return m;
1142
1143 if (n_op == Op_AddI) {
1144 Node *nn = convert_add_to_muladd( n );
1145 if ( nn ) return nn;
1146 }
1147
1148 if (n->is_ConstraintCast()) {
1149 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1150 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1151 // Node control inputs don't necessarily agree with loop control info (due to
1152 // transformations happened in between), thus additional dominance check is needed
1153 // to keep loop info valid.
1154 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1155 _igvn.replace_node(n, dom_cast);
1156 return dom_cast;
1157 }
1158 }
1159
1160 // Determine if the Node has inputs from some local Phi.
1161 // Returns the block to clone thru.
1162 Node *n_blk = has_local_phi_input( n );
1163 if( !n_blk ) return n;
1164
1165 // Do not clone the trip counter through on a CountedLoop
1166 // (messes up the canonical shape).
1167 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1168 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1169 return n;
1170 }
1171 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1172 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1173 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1174 return n;
1175 }
1176
1177 // Check for having no control input; not pinned. Allow
1178 // dominating control.
1179 if (n->in(0)) {
1180 Node *dom = idom(n_blk);
1181 if (dom_lca(n->in(0), dom) != n->in(0)) {
1182 return n;
1183 }
1184 }
1185 // Policy: when is it profitable. You must get more wins than
1186 // policy before it is considered profitable. Policy is usually 0,
1187 // so 1 win is considered profitable. Big merges will require big
1188 // cloning, so get a larger policy.
1189 int policy = n_blk->req() >> 2;
1190
1191 // If the loop is a candidate for range check elimination,
1192 // delay splitting through it's phi until a later loop optimization
1193 if (n_blk->is_BaseCountedLoop()) {
1194 IdealLoopTree *lp = get_loop(n_blk);
1195 if (lp && lp->_rce_candidate) {
1196 return n;
1197 }
1198 }
1199
1200 if (must_throttle_split_if()) return n;
1201
1202 // Split 'n' through the merge point if it is profitable
1203 Node *phi = split_thru_phi( n, n_blk, policy );
1204 if (!phi) return n;
1205
1206 // Found a Phi to split thru!
1207 // Replace 'n' with the new phi
1208 _igvn.replace_node( n, phi );
1209 // Moved a load around the loop, 'en-registering' something.
1210 if (n_blk->is_Loop() && n->is_Load() &&
1211 !phi->in(LoopNode::LoopBackControl)->is_Load())
1212 C->set_major_progress();
1213
1214 return phi;
1215 }
1216
1217 static bool merge_point_too_heavy(Compile* C, Node* region) {
1218 // Bail out if the region and its phis have too many users.
1219 int weight = 0;
1220 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1221 weight += region->fast_out(i)->outcnt();
1222 }
1223 int nodes_left = C->max_node_limit() - C->live_nodes();
1224 if (weight * 8 > nodes_left) {
1225 if (PrintOpto) {
1226 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1227 }
1228 return true;
1229 } else {
1230 return false;
1231 }
1232 }
1233
1234 static bool merge_point_safe(Node* region) {
1235 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1236 // having a PhiNode input. This sidesteps the dangerous case where the split
1237 // ConvI2LNode may become TOP if the input Value() does not
1238 // overlap the ConvI2L range, leaving a node which may not dominate its
1239 // uses.
1240 // A better fix for this problem can be found in the BugTraq entry, but
1241 // expediency for Mantis demands this hack.
1242 #ifdef _LP64
1243 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1244 Node* n = region->fast_out(i);
1245 if (n->is_Phi()) {
1246 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1247 Node* m = n->fast_out(j);
1248 if (m->Opcode() == Op_ConvI2L)
1249 return false;
1250 if (m->is_CastII()) {
1251 return false;
1252 }
1253 }
1254 }
1255 }
1256 #endif
1257 return true;
1258 }
1259
1260
1261 //------------------------------place_outside_loop---------------------------------
1262 // Place some computation outside of this loop on the path to the use passed as argument
1263 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1264 Node* head = loop->_head;
1265 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1266 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1267 loop = loop->_parent;
1268 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1269 }
1270
1271 // Pick control right outside the loop
1272 for (;;) {
1273 Node* dom = idom(useblock);
1274 if (loop->is_member(get_loop(dom))) {
1275 break;
1276 }
1277 useblock = dom;
1278 }
1279 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1280 return useblock;
1281 }
1282
1283
1284 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1285 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1286 return false;
1287 }
1288 if (!n->in(0)->is_Region()) {
1289 return false;
1290 }
1291
1292 Node* region = n->in(0);
1293 Node* dom = idom(region);
1294 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1295 return false;
1296 }
1297 IfNode* dom_if = dom->as_If();
1298 Node* proj_true = dom_if->proj_out(1);
1299 Node* proj_false = dom_if->proj_out(0);
1300
1301 for (uint i = 1; i < region->req(); i++) {
1302 if (is_dominator(proj_true, region->in(i))) {
1303 continue;
1304 }
1305 if (is_dominator(proj_false, region->in(i))) {
1306 continue;
1307 }
1308 return false;
1309 }
1310
1311 return true;
1312 }
1313
1314
1315 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1316 if (must_throttle_split_if()) {
1317 return false;
1318 }
1319
1320 // Do not do 'split-if' if irreducible loops are present.
1321 if (_has_irreducible_loops) {
1322 return false;
1323 }
1324
1325 if (merge_point_too_heavy(C, n_ctrl)) {
1326 return false;
1327 }
1328
1329 // Do not do 'split-if' if some paths are dead. First do dead code
1330 // elimination and then see if its still profitable.
1331 for (uint i = 1; i < n_ctrl->req(); i++) {
1332 if (n_ctrl->in(i) == C->top()) {
1333 return false;
1334 }
1335 }
1336
1337 // If trying to do a 'Split-If' at the loop head, it is only
1338 // profitable if the cmp folds up on BOTH paths. Otherwise we
1339 // risk peeling a loop forever.
1340
1341 // CNC - Disabled for now. Requires careful handling of loop
1342 // body selection for the cloned code. Also, make sure we check
1343 // for any input path not being in the same loop as n_ctrl. For
1344 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1345 // because the alternative loop entry points won't be converted
1346 // into LoopNodes.
1347 IdealLoopTree *n_loop = get_loop(n_ctrl);
1348 for (uint j = 1; j < n_ctrl->req(); j++) {
1349 if (get_loop(n_ctrl->in(j)) != n_loop) {
1350 return false;
1351 }
1352 }
1353
1354 // Check for safety of the merge point.
1355 if (!merge_point_safe(n_ctrl)) {
1356 return false;
1357 }
1358
1359 return true;
1360 }
1361
1362 // Detect if the node is the inner strip-mined loop
1363 // Return: null if it's not the case, or the exit of outer strip-mined loop
1364 static Node* is_inner_of_stripmined_loop(const Node* out) {
1365 Node* out_le = nullptr;
1366
1367 if (out->is_CountedLoopEnd()) {
1368 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1369
1370 if (loop != nullptr && loop->is_strip_mined()) {
1371 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1372 }
1373 }
1374
1375 return out_le;
1376 }
1377
1378 //------------------------------split_if_with_blocks_post----------------------
1379 // Do the real work in a non-recursive function. CFG hackery wants to be
1380 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1381 // info.
1382 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1383
1384 // Cloning Cmp through Phi's involves the split-if transform.
1385 // FastLock is not used by an If
1386 if (n->is_Cmp() && !n->is_FastLock()) {
1387 Node *n_ctrl = get_ctrl(n);
1388 // Determine if the Node has inputs from some local Phi.
1389 // Returns the block to clone thru.
1390 Node *n_blk = has_local_phi_input(n);
1391 if (n_blk != n_ctrl) {
1392 return;
1393 }
1394
1395 if (!can_split_if(n_ctrl)) {
1396 return;
1397 }
1398
1399 if (n->outcnt() != 1) {
1400 return; // Multiple bool's from 1 compare?
1401 }
1402 Node *bol = n->unique_out();
1403 assert(bol->is_Bool(), "expect a bool here");
1404 if (bol->outcnt() != 1) {
1405 return;// Multiple branches from 1 compare?
1406 }
1407 Node *iff = bol->unique_out();
1408
1409 // Check some safety conditions
1410 if (iff->is_If()) { // Classic split-if?
1411 if (iff->in(0) != n_ctrl) {
1412 return; // Compare must be in same blk as if
1413 }
1414 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1415 // Can't split CMove with different control.
1416 if (get_ctrl(iff) != n_ctrl) {
1417 return;
1418 }
1419 if (get_ctrl(iff->in(2)) == n_ctrl ||
1420 get_ctrl(iff->in(3)) == n_ctrl) {
1421 return; // Inputs not yet split-up
1422 }
1423 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1424 return; // Loop-invar test gates loop-varying CMOVE
1425 }
1426 } else {
1427 return; // some other kind of node, such as an Allocate
1428 }
1429
1430 // When is split-if profitable? Every 'win' on means some control flow
1431 // goes dead, so it's almost always a win.
1432 int policy = 0;
1433 // Split compare 'n' through the merge point if it is profitable
1434 Node *phi = split_thru_phi( n, n_ctrl, policy);
1435 if (!phi) {
1436 return;
1437 }
1438
1439 // Found a Phi to split thru!
1440 // Replace 'n' with the new phi
1441 _igvn.replace_node(n, phi);
1442
1443 // Now split the bool up thru the phi
1444 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1445 guarantee(bolphi != nullptr, "null boolean phi node");
1446
1447 _igvn.replace_node(bol, bolphi);
1448 assert(iff->in(1) == bolphi, "");
1449
1450 if (bolphi->Value(&_igvn)->singleton()) {
1451 return;
1452 }
1453
1454 // Conditional-move? Must split up now
1455 if (!iff->is_If()) {
1456 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1457 _igvn.replace_node(iff, cmovphi);
1458 return;
1459 }
1460
1461 // Now split the IF
1462 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1463 if (TraceLoopOpts) {
1464 tty->print_cr("Split-If");
1465 }
1466 do_split_if(iff);
1467 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1468 return;
1469 }
1470
1471 // Two identical ifs back to back can be merged
1472 if (try_merge_identical_ifs(n)) {
1473 return;
1474 }
1475
1476 // Check for an IF ready to split; one that has its
1477 // condition codes input coming from a Phi at the block start.
1478 int n_op = n->Opcode();
1479
1480 // Check for an IF being dominated by another IF same test
1481 if (n_op == Op_If ||
1482 n_op == Op_RangeCheck) {
1483 Node *bol = n->in(1);
1484 uint max = bol->outcnt();
1485 // Check for same test used more than once?
1486 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1487 // Search up IDOMs to see if this IF is dominated.
1488 Node* cmp = bol->in(1);
1489 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1490
1491 // Now search up IDOMs till cutoff, looking for a dominating test
1492 Node *prevdom = n;
1493 Node *dom = idom(prevdom);
1494 while (dom != cutoff) {
1495 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1496 safe_for_if_replacement(dom)) {
1497 // It's invalid to move control dependent data nodes in the inner
1498 // strip-mined loop, because:
1499 // 1) break validation of LoopNode::verify_strip_mined()
1500 // 2) move code with side-effect in strip-mined loop
1501 // Move to the exit of outer strip-mined loop in that case.
1502 Node* out_le = is_inner_of_stripmined_loop(dom);
1503 if (out_le != nullptr) {
1504 prevdom = out_le;
1505 }
1506 // Replace the dominated test with an obvious true or false.
1507 // Place it on the IGVN worklist for later cleanup.
1508 C->set_major_progress();
1509 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1510 // to prevent an array load from floating above its range check. There are three cases:
1511 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1512 // all its array accesses at that point.
1513 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1514 // accesses would start to float, since we don't pin at that point.
1515 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1516 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1517 prevdom->in(0)->Opcode() != Op_RangeCheck;
1518 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1519 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1520 return;
1521 }
1522 prevdom = dom;
1523 dom = idom(prevdom);
1524 }
1525 }
1526 }
1527
1528 try_sink_out_of_loop(n);
1529 if (C->failing()) {
1530 return;
1531 }
1532
1533 try_move_store_after_loop(n);
1534 }
1535
1536 // Transform:
1537 //
1538 // if (some_condition) {
1539 // // body 1
1540 // } else {
1541 // // body 2
1542 // }
1543 // if (some_condition) {
1544 // // body 3
1545 // } else {
1546 // // body 4
1547 // }
1548 //
1549 // into:
1550 //
1551 //
1552 // if (some_condition) {
1553 // // body 1
1554 // // body 3
1555 // } else {
1556 // // body 2
1557 // // body 4
1558 // }
1559 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1560 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1561 Node *n_ctrl = n->in(0);
1562 IfNode* dom_if = idom(n_ctrl)->as_If();
1563 if (n->in(1) != dom_if->in(1)) {
1564 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1565 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1566 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1567 _igvn.replace_input_of(n, 1, dom_if->in(1));
1568 }
1569 ProjNode* dom_proj_true = dom_if->proj_out(1);
1570 ProjNode* dom_proj_false = dom_if->proj_out(0);
1571
1572 // Now split the IF
1573 RegionNode* new_false_region;
1574 RegionNode* new_true_region;
1575 do_split_if(n, &new_false_region, &new_true_region);
1576 assert(new_false_region->req() == new_true_region->req(), "");
1577 #ifdef ASSERT
1578 for (uint i = 1; i < new_false_region->req(); ++i) {
1579 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1580 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");
1581 }
1582 #endif
1583 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1584
1585 // We now have:
1586 // if (some_condition) {
1587 // // body 1
1588 // if (some_condition) {
1589 // body3: // new_true_region
1590 // // body3
1591 // } else {
1592 // goto body4;
1593 // }
1594 // } else {
1595 // // body 2
1596 // if (some_condition) {
1597 // goto body3;
1598 // } else {
1599 // body4: // new_false_region
1600 // // body4;
1601 // }
1602 // }
1603 //
1604
1605 // clone pinned nodes thru the resulting regions
1606 push_pinned_nodes_thru_region(dom_if, new_true_region);
1607 push_pinned_nodes_thru_region(dom_if, new_false_region);
1608
1609 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1610 // 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
1611 // unrelated control dependency.
1612 for (uint i = 1; i < new_false_region->req(); i++) {
1613 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1614 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1615 } else {
1616 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1617 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1618 }
1619 }
1620 return true;
1621 }
1622 return false;
1623 }
1624
1625 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1626 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1627 Node* u = region->out(i);
1628 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1629 continue;
1630 }
1631 assert(u->in(0) == region, "not a control dependent node?");
1632 uint j = 1;
1633 for (; j < u->req(); ++j) {
1634 Node* in = u->in(j);
1635 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1636 break;
1637 }
1638 }
1639 if (j == u->req()) {
1640 Node *phi = PhiNode::make_blank(region, u);
1641 for (uint k = 1; k < region->req(); ++k) {
1642 Node* clone = u->clone();
1643 clone->set_req(0, region->in(k));
1644 register_new_node(clone, region->in(k));
1645 phi->init_req(k, clone);
1646 }
1647 register_new_node(phi, region);
1648 _igvn.replace_node(u, phi);
1649 --i;
1650 }
1651 }
1652 }
1653
1654 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1655 if (!dom->is_CountedLoopEnd()) {
1656 return true;
1657 }
1658 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1659 CountedLoopNode* cl = le->loopnode();
1660 if (cl == nullptr) {
1661 return true;
1662 }
1663 if (!cl->is_main_loop()) {
1664 return true;
1665 }
1666 if (cl->is_canonical_loop_entry() == nullptr) {
1667 return true;
1668 }
1669 // Further unrolling is possible so loop exit condition might change
1670 return false;
1671 }
1672
1673 // See if a shared loop-varying computation has no loop-varying uses.
1674 // Happens if something is only used for JVM state in uncommon trap exits,
1675 // like various versions of induction variable+offset. Clone the
1676 // computation per usage to allow it to sink out of the loop.
1677 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1678 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1679 n->in(1)->bottom_type()->isa_rawptr() &&
1680 !n->bottom_type()->isa_rawptr();
1681
1682 if (has_ctrl(n) &&
1683 !n->is_Phi() &&
1684 !n->is_Bool() &&
1685 !n->is_Proj() &&
1686 !n->is_MergeMem() &&
1687 !n->is_CMove() &&
1688 !n->is_OpaqueNotNull() &&
1689 !n->is_OpaqueInitializedAssertionPredicate() &&
1690 !n->is_OpaqueTemplateAssertionPredicate() &&
1691 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1692 (KillPathsReachableByDeadTypeNode || !n->is_Type())
1693 ) {
1694 Node *n_ctrl = get_ctrl(n);
1695 IdealLoopTree *n_loop = get_loop(n_ctrl);
1696
1697 if (n->in(0) != nullptr) {
1698 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1699 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1700 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1701 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1702 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1703 Node* maybe_pinned_n = n;
1704 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1705 if (n->depends_only_on_test()) {
1706 Node* pinned_clone = n->pin_array_access_node();
1707 if (pinned_clone != nullptr) {
1708 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1709 // range check for that access. If that condition is replaced by an identical dominating one, then an
1710 // unpinned load would risk floating above its range check.
1711 register_new_node(pinned_clone, n_ctrl);
1712 maybe_pinned_n = pinned_clone;
1713 _igvn.replace_node(n, pinned_clone);
1714 }
1715 }
1716 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1717 }
1718 }
1719 if (n_loop != _ltree_root && n->outcnt() > 1) {
1720 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1721 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1722 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1723 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1724 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1725 if (n->is_Store() || n->is_LoadStore()) {
1726 assert(false, "no node with a side effect");
1727 C->record_failure("no node with a side effect");
1728 return;
1729 }
1730 Node* outer_loop_clone = nullptr;
1731 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1732 Node* u = n->last_out(j); // Clone private computation per use
1733 _igvn.rehash_node_delayed(u);
1734 Node* x = nullptr;
1735 if (n->depends_only_on_test()) {
1736 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1737 // range check for that access. If that condition is replaced by an identical dominating one, then an
1738 // unpinned load would risk floating above its range check.
1739 x = n->pin_array_access_node();
1740 }
1741 if (x == nullptr) {
1742 x = n->clone();
1743 }
1744 Node* x_ctrl = nullptr;
1745 if (u->is_Phi()) {
1746 // Replace all uses of normal nodes. Replace Phi uses
1747 // individually, so the separate Nodes can sink down
1748 // different paths.
1749 uint k = 1;
1750 while (u->in(k) != n) k++;
1751 u->set_req(k, x);
1752 // x goes next to Phi input path
1753 x_ctrl = u->in(0)->in(k);
1754 // Find control for 'x' next to use but not inside inner loops.
1755 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1756 --j;
1757 } else { // Normal use
1758 if (has_ctrl(u)) {
1759 x_ctrl = get_ctrl(u);
1760 } else {
1761 x_ctrl = u->in(0);
1762 }
1763 // Find control for 'x' next to use but not inside inner loops.
1764 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1765 // Replace all uses
1766 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1767 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1768 // anymore now that we're going to pin n as well
1769 _igvn.replace_node(u, x);
1770 --j;
1771 } else {
1772 int nb = u->replace_edge(n, x, &_igvn);
1773 j -= nb;
1774 }
1775 }
1776
1777 if (n->is_Load()) {
1778 // For loads, add a control edge to a CFG node outside of the loop
1779 // to force them to not combine and return back inside the loop
1780 // during GVN optimization (4641526).
1781 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1782
1783 IdealLoopTree* x_loop = get_loop(x_ctrl);
1784 Node* x_head = x_loop->_head;
1785 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1786 // Do not add duplicate LoadNodes to the outer strip mined loop
1787 if (outer_loop_clone != nullptr) {
1788 _igvn.replace_node(x, outer_loop_clone);
1789 continue;
1790 }
1791 outer_loop_clone = x;
1792 }
1793 x->set_req(0, x_ctrl);
1794 } else if (n->in(0) != nullptr){
1795 x->set_req(0, x_ctrl);
1796 }
1797 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1798 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1799 register_new_node(x, x_ctrl);
1800
1801 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1802 // All AddP nodes must keep the same base after sinking so:
1803 // 1- We don't add a CastPP here until the last one of the chain is sunk: if part of the chain is not sunk,
1804 // their bases remain the same.
1805 // (see 2- below)
1806 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1807 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1808 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1809 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1810 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1811 assert(!x->is_Load(), "load should be pinned");
1812 // Use a cast node to pin clone out of loop
1813 Node* cast = nullptr;
1814 for (uint k = 0; k < x->req(); k++) {
1815 Node* in = x->in(k);
1816 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1817 const Type* in_t = _igvn.type(in);
1818 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1819 ConstraintCastNode::UnconditionalDependency, nullptr);
1820 }
1821 if (cast != nullptr) {
1822 Node* prev = _igvn.hash_find_insert(cast);
1823 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1824 cast->destruct(&_igvn);
1825 cast = prev;
1826 } else {
1827 register_new_node(cast, x_ctrl);
1828 }
1829 x->replace_edge(in, cast);
1830 // Chain of AddP nodes:
1831 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1832 if (x->is_AddP() && k == AddPNode::Base) {
1833 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1834 }
1835 break;
1836 }
1837 }
1838 assert(cast != nullptr, "must have added a cast to pin the node");
1839 }
1840 }
1841 _igvn.remove_dead_node(n);
1842 }
1843 _dom_lca_tags_round = 0;
1844 }
1845 }
1846 }
1847
1848 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1849 ResourceMark rm;
1850 Node_List wq;
1851 wq.push(x);
1852 while (wq.size() != 0) {
1853 Node* n = wq.pop();
1854 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1855 Node* u = n->fast_out(i);
1856 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1857 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1858 wq.push(u);
1859 }
1860 }
1861 }
1862 }
1863
1864 // Compute the early control of a node by following its inputs until we reach
1865 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1866 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1867 Node* early_ctrl = nullptr;
1868 ResourceMark rm;
1869 Unique_Node_List wq;
1870 wq.push(n);
1871 for (uint i = 0; i < wq.size(); i++) {
1872 Node* m = wq.at(i);
1873 Node* c = nullptr;
1874 if (m->is_CFG()) {
1875 c = m;
1876 } else if (m->pinned()) {
1877 c = m->in(0);
1878 } else {
1879 for (uint j = 0; j < m->req(); j++) {
1880 Node* in = m->in(j);
1881 if (in != nullptr) {
1882 wq.push(in);
1883 }
1884 }
1885 }
1886 if (c != nullptr) {
1887 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1888 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1889 early_ctrl = c;
1890 }
1891 }
1892 }
1893 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1894 return early_ctrl;
1895 }
1896
1897 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1898 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1899 Node* u = n->fast_out(i);
1900 if (u->is_Opaque1()) {
1901 return false; // Found loop limit, bugfix for 4677003
1902 }
1903 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1904 // get_late_ctrl_with_anti_dep() use their own tag
1905 _dom_lca_tags_round++;
1906 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1907
1908 if (u->is_Phi()) {
1909 for (uint j = 1; j < u->req(); ++j) {
1910 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1911 return false;
1912 }
1913 }
1914 } else {
1915 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1916 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1917 return false;
1918 }
1919 }
1920 }
1921 return true;
1922 }
1923
1924 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1925 if (n->is_Load()) {
1926 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1927 }
1928 IdealLoopTree *u_loop = get_loop(ctrl);
1929 if (u_loop == n_loop) {
1930 return false; // Found loop-varying use
1931 }
1932 if (n_loop->is_member(u_loop)) {
1933 return false; // Found use in inner loop
1934 }
1935 // Sinking a node from a pre loop to its main loop pins the node between the pre and main loops. If that node is input
1936 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1937 // test of the pre loop above the point in the graph where it's pinned.
1938 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1939 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1940 if (is_dominator(pre_loop->loopexit(), ctrl)) {
1941 return false;
1942 }
1943 }
1944 return true;
1945 }
1946
1947 //------------------------------split_if_with_blocks---------------------------
1948 // Check for aggressive application of 'split-if' optimization,
1949 // using basic block level info.
1950 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1951 Node* root = C->root();
1952 visited.set(root->_idx); // first, mark root as visited
1953 // Do pre-visit work for root
1954 Node* n = split_if_with_blocks_pre(root);
1955 uint cnt = n->outcnt();
1956 uint i = 0;
1957
1958 while (true) {
1959 // Visit all children
1960 if (i < cnt) {
1961 Node* use = n->raw_out(i);
1962 ++i;
1963 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1964 // Now do pre-visit work for this use
1965 use = split_if_with_blocks_pre(use);
1966 nstack.push(n, i); // Save parent and next use's index.
1967 n = use; // Process all children of current use.
1968 cnt = use->outcnt();
1969 i = 0;
1970 }
1971 }
1972 else {
1973 // All of n's children have been processed, complete post-processing.
1974 if (cnt != 0 && !n->is_Con()) {
1975 assert(has_node(n), "no dead nodes");
1976 split_if_with_blocks_post(n);
1977 if (C->failing()) {
1978 return;
1979 }
1980 }
1981 if (must_throttle_split_if()) {
1982 nstack.clear();
1983 }
1984 if (nstack.is_empty()) {
1985 // Finished all nodes on stack.
1986 break;
1987 }
1988 // Get saved parent node and next use's index. Visit the rest of uses.
1989 n = nstack.node();
1990 cnt = n->outcnt();
1991 i = nstack.index();
1992 nstack.pop();
1993 }
1994 }
1995 }
1996
1997
1998 //=============================================================================
1999 //
2000 // C L O N E A L O O P B O D Y
2001 //
2002
2003 //------------------------------clone_iff--------------------------------------
2004 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2005 // "Nearly" because all Nodes have been cloned from the original in the loop,
2006 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2007 // through the Phi recursively, and return a Bool.
2008 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2009
2010 // Convert this Phi into a Phi merging Bools
2011 uint i;
2012 for (i = 1; i < phi->req(); i++) {
2013 Node* b = phi->in(i);
2014 if (b->is_Phi()) {
2015 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2016 } else {
2017 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2018 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2019 }
2020 }
2021 Node* n = phi->in(1);
2022 Node* sample_opaque = nullptr;
2023 Node *sample_bool = nullptr;
2024 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2025 sample_opaque = n;
2026 sample_bool = n->in(1);
2027 assert(sample_bool->is_Bool(), "wrong type");
2028 } else {
2029 sample_bool = n;
2030 }
2031 Node *sample_cmp = sample_bool->in(1);
2032
2033 // Make Phis to merge the Cmp's inputs.
2034 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2035 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2036 for (i = 1; i < phi->req(); i++) {
2037 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2038 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2039 phi1->set_req(i, n1);
2040 phi2->set_req(i, n2);
2041 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2042 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2043 }
2044 // See if these Phis have been made before.
2045 // Register with optimizer
2046 Node *hit1 = _igvn.hash_find_insert(phi1);
2047 if (hit1) { // Hit, toss just made Phi
2048 _igvn.remove_dead_node(phi1); // Remove new phi
2049 assert(hit1->is_Phi(), "" );
2050 phi1 = (PhiNode*)hit1; // Use existing phi
2051 } else { // Miss
2052 _igvn.register_new_node_with_optimizer(phi1);
2053 }
2054 Node *hit2 = _igvn.hash_find_insert(phi2);
2055 if (hit2) { // Hit, toss just made Phi
2056 _igvn.remove_dead_node(phi2); // Remove new phi
2057 assert(hit2->is_Phi(), "" );
2058 phi2 = (PhiNode*)hit2; // Use existing phi
2059 } else { // Miss
2060 _igvn.register_new_node_with_optimizer(phi2);
2061 }
2062 // Register Phis with loop/block info
2063 set_ctrl(phi1, phi->in(0));
2064 set_ctrl(phi2, phi->in(0));
2065 // Make a new Cmp
2066 Node *cmp = sample_cmp->clone();
2067 cmp->set_req(1, phi1);
2068 cmp->set_req(2, phi2);
2069 _igvn.register_new_node_with_optimizer(cmp);
2070 set_ctrl(cmp, phi->in(0));
2071
2072 // Make a new Bool
2073 Node *b = sample_bool->clone();
2074 b->set_req(1,cmp);
2075 _igvn.register_new_node_with_optimizer(b);
2076 set_ctrl(b, phi->in(0));
2077
2078 if (sample_opaque != nullptr) {
2079 Node* opaque = sample_opaque->clone();
2080 opaque->set_req(1, b);
2081 _igvn.register_new_node_with_optimizer(opaque);
2082 set_ctrl(opaque, phi->in(0));
2083 return opaque;
2084 }
2085
2086 assert(b->is_Bool(), "");
2087 return b;
2088 }
2089
2090 //------------------------------clone_bool-------------------------------------
2091 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2092 // "Nearly" because all Nodes have been cloned from the original in the loop,
2093 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2094 // through the Phi recursively, and return a Bool.
2095 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2096 uint i;
2097 // Convert this Phi into a Phi merging Bools
2098 for( i = 1; i < phi->req(); i++ ) {
2099 Node *b = phi->in(i);
2100 if( b->is_Phi() ) {
2101 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2102 } else {
2103 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2104 }
2105 }
2106
2107 Node *sample_cmp = phi->in(1);
2108
2109 // Make Phis to merge the Cmp's inputs.
2110 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2111 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2112 for( uint j = 1; j < phi->req(); j++ ) {
2113 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2114 Node *n1, *n2;
2115 if( cmp_top->is_Cmp() ) {
2116 n1 = cmp_top->in(1);
2117 n2 = cmp_top->in(2);
2118 } else {
2119 n1 = n2 = cmp_top;
2120 }
2121 phi1->set_req( j, n1 );
2122 phi2->set_req( j, n2 );
2123 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2124 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2125 }
2126
2127 // See if these Phis have been made before.
2128 // Register with optimizer
2129 Node *hit1 = _igvn.hash_find_insert(phi1);
2130 if( hit1 ) { // Hit, toss just made Phi
2131 _igvn.remove_dead_node(phi1); // Remove new phi
2132 assert( hit1->is_Phi(), "" );
2133 phi1 = (PhiNode*)hit1; // Use existing phi
2134 } else { // Miss
2135 _igvn.register_new_node_with_optimizer(phi1);
2136 }
2137 Node *hit2 = _igvn.hash_find_insert(phi2);
2138 if( hit2 ) { // Hit, toss just made Phi
2139 _igvn.remove_dead_node(phi2); // Remove new phi
2140 assert( hit2->is_Phi(), "" );
2141 phi2 = (PhiNode*)hit2; // Use existing phi
2142 } else { // Miss
2143 _igvn.register_new_node_with_optimizer(phi2);
2144 }
2145 // Register Phis with loop/block info
2146 set_ctrl(phi1, phi->in(0));
2147 set_ctrl(phi2, phi->in(0));
2148 // Make a new Cmp
2149 Node *cmp = sample_cmp->clone();
2150 cmp->set_req( 1, phi1 );
2151 cmp->set_req( 2, phi2 );
2152 _igvn.register_new_node_with_optimizer(cmp);
2153 set_ctrl(cmp, phi->in(0));
2154
2155 assert( cmp->is_Cmp(), "" );
2156 return (CmpNode*)cmp;
2157 }
2158
2159 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2160 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2161 Node_List*& split_if_set, Node_List*& split_bool_set,
2162 Node_List*& split_cex_set, Node_List& worklist,
2163 uint new_counter, CloneLoopMode mode) {
2164 Node* nnn = old_new[old->_idx];
2165 // Copy uses to a worklist, so I can munge the def-use info
2166 // with impunity.
2167 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2168 worklist.push(old->fast_out(j));
2169
2170 while( worklist.size() ) {
2171 Node *use = worklist.pop();
2172 if (!has_node(use)) continue; // Ignore dead nodes
2173 if (use->in(0) == C->top()) continue;
2174 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2175 // Check for data-use outside of loop - at least one of OLD or USE
2176 // must not be a CFG node.
2177 #ifdef ASSERT
2178 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2179 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2180 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2181 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2182 }
2183 #endif
2184 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2185
2186 // If the Data use is an IF, that means we have an IF outside the
2187 // loop that is switching on a condition that is set inside the
2188 // loop. Happens if people set a loop-exit flag; then test the flag
2189 // in the loop to break the loop, then test is again outside the
2190 // loop to determine which way the loop exited.
2191 //
2192 // For several uses we need to make sure that there is no phi between,
2193 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2194 // to avoid such a phi in between.
2195 // For example, it is unexpected that there is a Phi between an
2196 // AllocateArray node and its ValidLengthTest input that could cause
2197 // split if to break.
2198 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2199 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2200 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2201 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2202 // Since this code is highly unlikely, we lazily build the worklist
2203 // of such Nodes to go split.
2204 if (!split_if_set) {
2205 split_if_set = new Node_List();
2206 }
2207 split_if_set->push(use);
2208 }
2209 if (use->is_Bool()) {
2210 if (!split_bool_set) {
2211 split_bool_set = new Node_List();
2212 }
2213 split_bool_set->push(use);
2214 }
2215 if (use->Opcode() == Op_CreateEx) {
2216 if (!split_cex_set) {
2217 split_cex_set = new Node_List();
2218 }
2219 split_cex_set->push(use);
2220 }
2221
2222
2223 // Get "block" use is in
2224 uint idx = 0;
2225 while( use->in(idx) != old ) idx++;
2226 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2227 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2228 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2229 ? prev->in(2)
2230 : idom(prev);
2231 if( use->is_Phi() ) // Phi use is in prior block
2232 cfg = prev->in(idx); // NOT in block of Phi itself
2233 if (cfg->is_top()) { // Use is dead?
2234 _igvn.replace_input_of(use, idx, C->top());
2235 continue;
2236 }
2237
2238 // If use is referenced through control edge... (idx == 0)
2239 if (mode == IgnoreStripMined && idx == 0) {
2240 LoopNode *head = loop->_head->as_Loop();
2241 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2242 // That node is outside the inner loop, leave it outside the
2243 // outer loop as well to not confuse verification code.
2244 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2245 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2246 continue;
2247 }
2248 }
2249
2250 while(!outer_loop->is_member(get_loop(cfg))) {
2251 prev = cfg;
2252 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2253 }
2254 // If the use occurs after merging several exits from the loop, then
2255 // old value must have dominated all those exits. Since the same old
2256 // value was used on all those exits we did not need a Phi at this
2257 // merge point. NOW we do need a Phi here. Each loop exit value
2258 // is now merged with the peeled body exit; each exit gets its own
2259 // private Phi and those Phis need to be merged here.
2260 Node *phi;
2261 if( prev->is_Region() ) {
2262 if( idx == 0 ) { // Updating control edge?
2263 phi = prev; // Just use existing control
2264 } else { // Else need a new Phi
2265 phi = PhiNode::make( prev, old );
2266 // Now recursively fix up the new uses of old!
2267 for( uint i = 1; i < prev->req(); i++ ) {
2268 worklist.push(phi); // Onto worklist once for each 'old' input
2269 }
2270 }
2271 } else {
2272 // Get new RegionNode merging old and new loop exits
2273 prev = old_new[prev->_idx];
2274 assert( prev, "just made this in step 7" );
2275 if( idx == 0) { // Updating control edge?
2276 phi = prev; // Just use existing control
2277 } else { // Else need a new Phi
2278 // Make a new Phi merging data values properly
2279 phi = PhiNode::make( prev, old );
2280 phi->set_req( 1, nnn );
2281 }
2282 }
2283 // If inserting a new Phi, check for prior hits
2284 if( idx != 0 ) {
2285 Node *hit = _igvn.hash_find_insert(phi);
2286 if( hit == nullptr ) {
2287 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2288 } else { // or
2289 // Remove the new phi from the graph and use the hit
2290 _igvn.remove_dead_node(phi);
2291 phi = hit; // Use existing phi
2292 }
2293 set_ctrl(phi, prev);
2294 }
2295 // Make 'use' use the Phi instead of the old loop body exit value
2296 assert(use->in(idx) == old, "old is still input of use");
2297 // We notify all uses of old, including use, and the indirect uses,
2298 // that may now be optimized because we have replaced old with phi.
2299 _igvn.add_users_to_worklist(old);
2300 if (idx == 0 &&
2301 use->depends_only_on_test()) {
2302 Node* pinned_clone = use->pin_array_access_node();
2303 if (pinned_clone != nullptr) {
2304 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2305 // into the region is left, an array load could become dependent on a condition that's not a range check for
2306 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2307 // floating above its range check.
2308 pinned_clone->set_req(0, phi);
2309 register_new_node_with_ctrl_of(pinned_clone, use);
2310 _igvn.replace_node(use, pinned_clone);
2311 continue;
2312 }
2313 }
2314 _igvn.replace_input_of(use, idx, phi);
2315 if( use->_idx >= new_counter ) { // If updating new phis
2316 // Not needed for correctness, but prevents a weak assert
2317 // in AddPNode from tripping (when we end up with different
2318 // base & derived Phis that will become the same after
2319 // IGVN does CSE).
2320 Node *hit = _igvn.hash_find_insert(use);
2321 if( hit ) // Go ahead and re-hash for hits.
2322 _igvn.replace_node( use, hit );
2323 }
2324 }
2325 }
2326 }
2327
2328 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2329 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2330 bool check_old_new) {
2331 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2332 Node* u = n->fast_out(j);
2333 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2334 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2335 Node* c = phase->get_ctrl(u);
2336 IdealLoopTree* u_loop = phase->get_loop(c);
2337 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2338 if (!loop->is_member(u_loop)) {
2339 if (outer_loop->is_member(u_loop)) {
2340 wq.push(u);
2341 } else {
2342 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2343 // the outer loop too
2344 Node* u_c = u->in(0);
2345 if (u_c != nullptr) {
2346 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2347 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2348 wq.push(u);
2349 }
2350 }
2351 }
2352 }
2353 }
2354 }
2355 }
2356
2357 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2358 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2359 Node_List& extra_data_nodes) {
2360 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2361 CountedLoopNode* cl = head->as_CountedLoop();
2362 Node* l = cl->outer_loop();
2363 Node* tail = cl->outer_loop_tail();
2364 IfNode* le = cl->outer_loop_end();
2365 Node* sfpt = cl->outer_safepoint();
2366 CountedLoopEndNode* cle = cl->loopexit();
2367 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2368 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2369 Node* cle_out = cle->proj_out(false);
2370
2371 Node* new_sfpt = nullptr;
2372 Node* new_cle_out = cle_out->clone();
2373 old_new.map(cle_out->_idx, new_cle_out);
2374 if (mode == CloneIncludesStripMined) {
2375 // clone outer loop body
2376 Node* new_l = l->clone();
2377 Node* new_tail = tail->clone();
2378 IfNode* new_le = le->clone()->as_If();
2379 new_sfpt = sfpt->clone();
2380
2381 set_loop(new_l, outer_loop->_parent);
2382 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2383 set_loop(new_cle_out, outer_loop->_parent);
2384 set_idom(new_cle_out, new_cle, dd);
2385 set_loop(new_sfpt, outer_loop->_parent);
2386 set_idom(new_sfpt, new_cle_out, dd);
2387 set_loop(new_le, outer_loop->_parent);
2388 set_idom(new_le, new_sfpt, dd);
2389 set_loop(new_tail, outer_loop->_parent);
2390 set_idom(new_tail, new_le, dd);
2391 set_idom(new_cl, new_l, dd);
2392
2393 old_new.map(l->_idx, new_l);
2394 old_new.map(tail->_idx, new_tail);
2395 old_new.map(le->_idx, new_le);
2396 old_new.map(sfpt->_idx, new_sfpt);
2397
2398 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2399 new_l->set_req(0, new_l);
2400 new_tail->set_req(0, new_le);
2401 new_le->set_req(0, new_sfpt);
2402 new_sfpt->set_req(0, new_cle_out);
2403 new_cle_out->set_req(0, new_cle);
2404 new_cl->set_req(LoopNode::EntryControl, new_l);
2405
2406 _igvn.register_new_node_with_optimizer(new_l);
2407 _igvn.register_new_node_with_optimizer(new_tail);
2408 _igvn.register_new_node_with_optimizer(new_le);
2409 } else {
2410 Node *newhead = old_new[loop->_head->_idx];
2411 newhead->as_Loop()->clear_strip_mined();
2412 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2413 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2414 }
2415 // Look at data node that were assigned a control in the outer
2416 // loop: they are kept in the outer loop by the safepoint so start
2417 // from the safepoint node's inputs.
2418 IdealLoopTree* outer_loop = get_loop(l);
2419 Node_Stack stack(2);
2420 stack.push(sfpt, 1);
2421 uint new_counter = C->unique();
2422 while (stack.size() > 0) {
2423 Node* n = stack.node();
2424 uint i = stack.index();
2425 while (i < n->req() &&
2426 (n->in(i) == nullptr ||
2427 !has_ctrl(n->in(i)) ||
2428 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2429 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2430 i++;
2431 }
2432 if (i < n->req()) {
2433 stack.set_index(i+1);
2434 stack.push(n->in(i), 0);
2435 } else {
2436 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2437 Node* m = n == sfpt ? new_sfpt : n->clone();
2438 if (m != nullptr) {
2439 for (uint i = 0; i < n->req(); i++) {
2440 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2441 m->set_req(i, old_new[m->in(i)->_idx]);
2442 }
2443 }
2444 } else {
2445 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2446 }
2447 if (n != sfpt) {
2448 extra_data_nodes.push(n);
2449 _igvn.register_new_node_with_optimizer(m);
2450 assert(get_ctrl(n) == cle_out, "what other control?");
2451 set_ctrl(m, new_cle_out);
2452 old_new.map(n->_idx, m);
2453 }
2454 stack.pop();
2455 }
2456 }
2457 if (mode == CloneIncludesStripMined) {
2458 _igvn.register_new_node_with_optimizer(new_sfpt);
2459 _igvn.register_new_node_with_optimizer(new_cle_out);
2460 }
2461 // Some other transformation may have pessimistically assigned some
2462 // data nodes to the outer loop. Set their control so they are out
2463 // of the outer loop.
2464 ResourceMark rm;
2465 Unique_Node_List wq;
2466 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2467 Node* old = extra_data_nodes.at(i);
2468 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2469 }
2470
2471 for (uint i = 0; i < loop->_body.size(); i++) {
2472 Node* old = loop->_body.at(i);
2473 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2474 }
2475
2476 Node* inner_out = sfpt->in(0);
2477 if (inner_out->outcnt() > 1) {
2478 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2479 }
2480
2481 Node* new_ctrl = cl->outer_loop_exit();
2482 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2483 for (uint i = 0; i < wq.size(); i++) {
2484 Node* n = wq.at(i);
2485 set_ctrl(n, new_ctrl);
2486 if (n->in(0) != nullptr) {
2487 _igvn.replace_input_of(n, 0, new_ctrl);
2488 }
2489 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2490 }
2491 } else {
2492 Node *newhead = old_new[loop->_head->_idx];
2493 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2494 }
2495 }
2496
2497 //------------------------------clone_loop-------------------------------------
2498 //
2499 // C L O N E A L O O P B O D Y
2500 //
2501 // This is the basic building block of the loop optimizations. It clones an
2502 // entire loop body. It makes an old_new loop body mapping; with this mapping
2503 // you can find the new-loop equivalent to an old-loop node. All new-loop
2504 // nodes are exactly equal to their old-loop counterparts, all edges are the
2505 // same. All exits from the old-loop now have a RegionNode that merges the
2506 // equivalent new-loop path. This is true even for the normal "loop-exit"
2507 // condition. All uses of loop-invariant old-loop values now come from (one
2508 // or more) Phis that merge their new-loop equivalents.
2509 //
2510 // This operation leaves the graph in an illegal state: there are two valid
2511 // control edges coming from the loop pre-header to both loop bodies. I'll
2512 // definitely have to hack the graph after running this transform.
2513 //
2514 // From this building block I will further edit edges to perform loop peeling
2515 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2516 //
2517 // Parameter side_by_size_idom:
2518 // When side_by_size_idom is null, the dominator tree is constructed for
2519 // the clone loop to dominate the original. Used in construction of
2520 // pre-main-post loop sequence.
2521 // When nonnull, the clone and original are side-by-side, both are
2522 // dominated by the side_by_side_idom node. Used in construction of
2523 // unswitched loops.
2524 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2525 CloneLoopMode mode, Node* side_by_side_idom) {
2526
2527 LoopNode* head = loop->_head->as_Loop();
2528 head->verify_strip_mined(1);
2529
2530 if (C->do_vector_loop() && PrintOpto) {
2531 const char* mname = C->method()->name()->as_quoted_ascii();
2532 if (mname != nullptr) {
2533 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2534 }
2535 }
2536
2537 CloneMap& cm = C->clone_map();
2538 if (C->do_vector_loop()) {
2539 cm.set_clone_idx(cm.max_gen()+1);
2540 #ifndef PRODUCT
2541 if (PrintOpto) {
2542 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2543 loop->dump_head();
2544 }
2545 #endif
2546 }
2547
2548 // Step 1: Clone the loop body. Make the old->new mapping.
2549 clone_loop_body(loop->_body, old_new, &cm);
2550
2551 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2552
2553 // Step 2: Fix the edges in the new body. If the old input is outside the
2554 // loop use it. If the old input is INside the loop, use the corresponding
2555 // new node instead.
2556 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2557
2558 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2559 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2560
2561 // Step 3: Now fix control uses. Loop varying control uses have already
2562 // been fixed up (as part of all input edges in Step 2). Loop invariant
2563 // control uses must be either an IfFalse or an IfTrue. Make a merge
2564 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2565 // refer to this.
2566 Node_List worklist;
2567 uint new_counter = C->unique();
2568 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2569
2570 // Step 4: If loop-invariant use is not control, it must be dominated by a
2571 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2572 // there if needed. Make a Phi there merging old and new used values.
2573 Node_List *split_if_set = nullptr;
2574 Node_List *split_bool_set = nullptr;
2575 Node_List *split_cex_set = nullptr;
2576 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2577
2578 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2579 Node* old = extra_data_nodes.at(i);
2580 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2581 split_bool_set, split_cex_set, worklist, new_counter,
2582 mode);
2583 }
2584
2585 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2586 // the loop uses a condition set in the loop. The original IF probably
2587 // takes control from one or more OLD Regions (which in turn get from NEW
2588 // Regions). In any case, there will be a set of Phis for each merge point
2589 // from the IF up to where the original BOOL def exists the loop.
2590 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2591
2592 }
2593
2594 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2595 if (split_if_set) {
2596 while (split_if_set->size()) {
2597 Node *iff = split_if_set->pop();
2598 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2599 if (iff->in(input)->is_Phi()) {
2600 Node *b = clone_iff(iff->in(input)->as_Phi());
2601 _igvn.replace_input_of(iff, input, b);
2602 }
2603 }
2604 }
2605 if (split_bool_set) {
2606 while (split_bool_set->size()) {
2607 Node *b = split_bool_set->pop();
2608 Node *phi = b->in(1);
2609 assert(phi->is_Phi(), "");
2610 CmpNode *cmp = clone_bool((PhiNode*) phi);
2611 _igvn.replace_input_of(b, 1, cmp);
2612 }
2613 }
2614 if (split_cex_set) {
2615 while (split_cex_set->size()) {
2616 Node *b = split_cex_set->pop();
2617 assert(b->in(0)->is_Region(), "");
2618 assert(b->in(1)->is_Phi(), "");
2619 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2620 split_up(b, b->in(0), nullptr);
2621 }
2622 }
2623 }
2624
2625 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2626 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2627 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2628 for(uint i = 0; i < body.size(); i++ ) {
2629 Node* old = body.at(i);
2630 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2631 split_bool_set, split_cex_set, worklist, new_counter,
2632 mode);
2633 }
2634 }
2635
2636 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2637 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2638 LoopNode* head = loop->_head->as_Loop();
2639 for(uint i = 0; i < body.size(); i++ ) {
2640 Node* old = body.at(i);
2641 if( !old->is_CFG() ) continue;
2642
2643 // Copy uses to a worklist, so I can munge the def-use info
2644 // with impunity.
2645 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2646 worklist.push(old->fast_out(j));
2647 }
2648
2649 while (worklist.size()) { // Visit all uses
2650 Node *use = worklist.pop();
2651 if (!has_node(use)) continue; // Ignore dead nodes
2652 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2653 if (!loop->is_member(use_loop) && use->is_CFG()) {
2654 // Both OLD and USE are CFG nodes here.
2655 assert(use->is_Proj(), "" );
2656 Node* nnn = old_new[old->_idx];
2657
2658 Node* newuse = nullptr;
2659 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2660 CountedLoopNode* cl = head->as_CountedLoop();
2661 CountedLoopEndNode* cle = cl->loopexit();
2662 Node* cle_out = cle->proj_out_or_null(false);
2663 if (use == cle_out) {
2664 IfNode* le = cl->outer_loop_end();
2665 use = le->proj_out(false);
2666 use_loop = get_loop(use);
2667 if (mode == CloneIncludesStripMined) {
2668 nnn = old_new[le->_idx];
2669 } else {
2670 newuse = old_new[cle_out->_idx];
2671 }
2672 }
2673 }
2674 if (newuse == nullptr) {
2675 newuse = use->clone();
2676 }
2677
2678 // Clone the loop exit control projection
2679 if (C->do_vector_loop() && cm != nullptr) {
2680 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2681 }
2682 newuse->set_req(0,nnn);
2683 _igvn.register_new_node_with_optimizer(newuse);
2684 set_loop(newuse, use_loop);
2685 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2686
2687 // We need a Region to merge the exit from the peeled body and the
2688 // exit from the old loop body.
2689 RegionNode *r = new RegionNode(3);
2690 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2691 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2692
2693 // The original user of 'use' uses 'r' instead.
2694 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2695 Node* useuse = use->last_out(l);
2696 _igvn.rehash_node_delayed(useuse);
2697 uint uses_found = 0;
2698 if (useuse->in(0) == use) {
2699 useuse->set_req(0, r);
2700 uses_found++;
2701 if (useuse->is_CFG()) {
2702 // This is not a dom_depth > dd_r because when new
2703 // control flow is constructed by a loop opt, a node and
2704 // its dominator can end up at the same dom_depth
2705 assert(dom_depth(useuse) >= dd_r, "");
2706 set_idom(useuse, r, dom_depth(useuse));
2707 }
2708 }
2709 for (uint k = 1; k < useuse->req(); k++) {
2710 if( useuse->in(k) == use ) {
2711 useuse->set_req(k, r);
2712 uses_found++;
2713 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2714 // This is not a dom_depth > dd_r because when new
2715 // control flow is constructed by a loop opt, a node
2716 // and its dominator can end up at the same dom_depth
2717 assert(dom_depth(useuse) >= dd_r , "");
2718 set_idom(useuse, r, dom_depth(useuse));
2719 }
2720 }
2721 }
2722 l -= uses_found; // we deleted 1 or more copies of this edge
2723 }
2724
2725 assert(use->is_Proj(), "loop exit should be projection");
2726 // lazy_replace() below moves all nodes that are:
2727 // - control dependent on the loop exit or
2728 // - have control set to the loop exit
2729 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2730 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2731 // becomes dead and is "replaced" by the region.
2732 Node* use_clone = use->clone();
2733 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2734 // Now finish up 'r'
2735 r->set_req(1, newuse);
2736 r->set_req(2, use_clone);
2737 _igvn.register_new_node_with_optimizer(r);
2738 set_loop(r, use_loop);
2739 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2740 lazy_replace(use, r);
2741 // Map the (cloned) old use to the new merge point
2742 old_new.map(use_clone->_idx, r);
2743 } // End of if a loop-exit test
2744 }
2745 }
2746 }
2747
2748 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2749 IdealLoopTree* parent, bool partial) {
2750 for(uint i = 0; i < body.size(); i++ ) {
2751 Node *old = body.at(i);
2752 Node *nnn = old_new[old->_idx];
2753 // Fix CFG/Loop controlling the new node
2754 if (has_ctrl(old)) {
2755 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2756 } else {
2757 set_loop(nnn, parent);
2758 if (old->outcnt() > 0) {
2759 Node* dom = idom(old);
2760 if (old_new[dom->_idx] != nullptr) {
2761 dom = old_new[dom->_idx];
2762 set_idom(nnn, dom, dd );
2763 }
2764 }
2765 }
2766 // Correct edges to the new node
2767 for (uint j = 0; j < nnn->req(); j++) {
2768 Node *n = nnn->in(j);
2769 if (n != nullptr) {
2770 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2771 if (loop->is_member(old_in_loop)) {
2772 if (old_new[n->_idx] != nullptr) {
2773 nnn->set_req(j, old_new[n->_idx]);
2774 } else {
2775 assert(!body.contains(n), "");
2776 assert(partial, "node not cloned");
2777 }
2778 }
2779 }
2780 }
2781 _igvn.hash_find_insert(nnn);
2782 }
2783 }
2784
2785 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2786 for (uint i = 0; i < body.size(); i++) {
2787 Node* old = body.at(i);
2788 Node* nnn = old->clone();
2789 old_new.map(old->_idx, nnn);
2790 if (C->do_vector_loop() && cm != nullptr) {
2791 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2792 }
2793 _igvn.register_new_node_with_optimizer(nnn);
2794 }
2795 }
2796
2797
2798 //---------------------- stride_of_possible_iv -------------------------------------
2799 // Looks for an iff/bool/comp with one operand of the compare
2800 // being a cycle involving an add and a phi,
2801 // with an optional truncation (left-shift followed by a right-shift)
2802 // of the add. Returns zero if not an iv.
2803 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2804 Node* trunc1 = nullptr;
2805 Node* trunc2 = nullptr;
2806 const TypeInteger* ttype = nullptr;
2807 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2808 return 0;
2809 }
2810 BoolNode* bl = iff->in(1)->as_Bool();
2811 Node* cmp = bl->in(1);
2812 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2813 return 0;
2814 }
2815 // Must have an invariant operand
2816 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2817 return 0;
2818 }
2819 Node* add2 = nullptr;
2820 Node* cmp1 = cmp->in(1);
2821 if (cmp1->is_Phi()) {
2822 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2823 Node* phi = cmp1;
2824 for (uint i = 1; i < phi->req(); i++) {
2825 Node* in = phi->in(i);
2826 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2827 &trunc1, &trunc2, &ttype, T_INT);
2828 if (add && add->in(1) == phi) {
2829 add2 = add->in(2);
2830 break;
2831 }
2832 }
2833 } else {
2834 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2835 Node* addtrunc = cmp1;
2836 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2837 &trunc1, &trunc2, &ttype, T_INT);
2838 if (add && add->in(1)->is_Phi()) {
2839 Node* phi = add->in(1);
2840 for (uint i = 1; i < phi->req(); i++) {
2841 if (phi->in(i) == addtrunc) {
2842 add2 = add->in(2);
2843 break;
2844 }
2845 }
2846 }
2847 }
2848 if (add2 != nullptr) {
2849 const TypeInt* add2t = _igvn.type(add2)->is_int();
2850 if (add2t->is_con()) {
2851 return add2t->get_con();
2852 }
2853 }
2854 return 0;
2855 }
2856
2857
2858 //---------------------- stay_in_loop -------------------------------------
2859 // Return the (unique) control output node that's in the loop (if it exists.)
2860 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2861 Node* unique = nullptr;
2862 if (!n) return nullptr;
2863 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2864 Node* use = n->fast_out(i);
2865 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2866 if (unique != nullptr) {
2867 return nullptr;
2868 }
2869 unique = use;
2870 }
2871 }
2872 return unique;
2873 }
2874
2875 //------------------------------ register_node -------------------------------------
2876 // Utility to register node "n" with PhaseIdealLoop
2877 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2878 _igvn.register_new_node_with_optimizer(n);
2879 loop->_body.push(n);
2880 if (n->is_CFG()) {
2881 set_loop(n, loop);
2882 set_idom(n, pred, ddepth);
2883 } else {
2884 set_ctrl(n, pred);
2885 }
2886 }
2887
2888 //------------------------------ proj_clone -------------------------------------
2889 // Utility to create an if-projection
2890 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2891 ProjNode* c = p->clone()->as_Proj();
2892 c->set_req(0, iff);
2893 return c;
2894 }
2895
2896 //------------------------------ short_circuit_if -------------------------------------
2897 // Force the iff control output to be the live_proj
2898 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2899 guarantee(live_proj != nullptr, "null projection");
2900 int proj_con = live_proj->_con;
2901 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2902 Node* con = intcon(proj_con);
2903 if (iff) {
2904 iff->set_req(1, con);
2905 }
2906 return con;
2907 }
2908
2909 //------------------------------ insert_if_before_proj -------------------------------------
2910 // Insert a new if before an if projection (* - new node)
2911 //
2912 // before
2913 // if(test)
2914 // / \
2915 // v v
2916 // other-proj proj (arg)
2917 //
2918 // after
2919 // if(test)
2920 // / \
2921 // / v
2922 // | * proj-clone
2923 // v |
2924 // other-proj v
2925 // * new_if(relop(cmp[IU](left,right)))
2926 // / \
2927 // v v
2928 // * new-proj proj
2929 // (returned)
2930 //
2931 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2932 IfNode* iff = proj->in(0)->as_If();
2933 IdealLoopTree *loop = get_loop(proj);
2934 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2935 uint ddepth = dom_depth(proj);
2936
2937 _igvn.rehash_node_delayed(iff);
2938 _igvn.rehash_node_delayed(proj);
2939
2940 proj->set_req(0, nullptr); // temporary disconnect
2941 ProjNode* proj2 = proj_clone(proj, iff);
2942 register_node(proj2, loop, iff, ddepth);
2943
2944 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2945 register_node(cmp, loop, proj2, ddepth);
2946
2947 BoolNode* bol = new BoolNode(cmp, relop);
2948 register_node(bol, loop, proj2, ddepth);
2949
2950 int opcode = iff->Opcode();
2951 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2952 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2953 register_node(new_if, loop, proj2, ddepth);
2954
2955 proj->set_req(0, new_if); // reattach
2956 set_idom(proj, new_if, ddepth);
2957
2958 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2959 guarantee(new_exit != nullptr, "null exit node");
2960 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2961
2962 return new_exit;
2963 }
2964
2965 //------------------------------ insert_region_before_proj -------------------------------------
2966 // Insert a region before an if projection (* - new node)
2967 //
2968 // before
2969 // if(test)
2970 // / |
2971 // v |
2972 // proj v
2973 // other-proj
2974 //
2975 // after
2976 // if(test)
2977 // / |
2978 // v |
2979 // * proj-clone v
2980 // | other-proj
2981 // v
2982 // * new-region
2983 // |
2984 // v
2985 // * dum_if
2986 // / \
2987 // v \
2988 // * dum-proj v
2989 // proj
2990 //
2991 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2992 IfNode* iff = proj->in(0)->as_If();
2993 IdealLoopTree *loop = get_loop(proj);
2994 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2995 uint ddepth = dom_depth(proj);
2996
2997 _igvn.rehash_node_delayed(iff);
2998 _igvn.rehash_node_delayed(proj);
2999
3000 proj->set_req(0, nullptr); // temporary disconnect
3001 ProjNode* proj2 = proj_clone(proj, iff);
3002 register_node(proj2, loop, iff, ddepth);
3003
3004 RegionNode* reg = new RegionNode(2);
3005 reg->set_req(1, proj2);
3006 register_node(reg, loop, iff, ddepth);
3007
3008 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3009 register_node(dum_if, loop, reg, ddepth);
3010
3011 proj->set_req(0, dum_if); // reattach
3012 set_idom(proj, dum_if, ddepth);
3013
3014 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3015 register_node(dum_proj, loop, dum_if, ddepth);
3016
3017 return reg;
3018 }
3019
3020 // Idea
3021 // ----
3022 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3023 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3024 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3025 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3026 // test alone instead:
3027 //
3028 // Before Partial Peeling:
3029 // Loop:
3030 // <peeled section>
3031 // Split off signed loop exit test
3032 // <-- CUT HERE -->
3033 // Unchanged unsigned loop exit test
3034 // <rest of unpeeled section>
3035 // goto Loop
3036 //
3037 // After Partial Peeling:
3038 // <cloned peeled section>
3039 // Cloned split off signed loop exit test
3040 // Loop:
3041 // Unchanged unsigned loop exit test
3042 // <rest of unpeeled section>
3043 // <peeled section>
3044 // Split off signed loop exit test
3045 // goto Loop
3046 //
3047 // Details
3048 // -------
3049 // Before:
3050 // if (i <u limit) Unsigned loop exit condition
3051 // / |
3052 // v v
3053 // exit-proj stay-in-loop-proj
3054 //
3055 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3056 // before the CmpU on the stay-in-loop path and keep both tests:
3057 //
3058 // if (i <u limit) Signed loop exit test
3059 // / |
3060 // / if (i <u limit) Unsigned loop exit test
3061 // / / |
3062 // v v v
3063 // exit-region stay-in-loop-proj
3064 //
3065 // Implementation
3066 // --------------
3067 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3068 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3069 // exit tests is preserved, and their loop nesting is correct.
3070 //
3071 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3072 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3073 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3074 //
3075 // if (stay-in-loop-const) Killed original unsigned loop exit test
3076 // / |
3077 // / v
3078 // / if (i < limit) Split off signed loop exit test
3079 // / / |
3080 // / / v
3081 // / / if (i <u limit) Cloned unsigned loop exit test
3082 // / / / |
3083 // v v v |
3084 // exit-region |
3085 // | |
3086 // dummy-if |
3087 // / | |
3088 // dead | |
3089 // v v
3090 // exit-proj stay-in-loop-proj
3091 //
3092 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3093 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3094 //
3095 // Requirements
3096 // ------------
3097 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3098 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3099 // Otherwise, we need to bail out (see comments in the code below).
3100 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3101 const bool Signed = true;
3102 const bool Unsigned = false;
3103
3104 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3105 if (bol->_test._test != BoolTest::lt) {
3106 return nullptr;
3107 }
3108 CmpNode* cmpu = bol->in(1)->as_Cmp();
3109 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3110
3111 int stride = stride_of_possible_iv(if_cmpu);
3112 if (stride == 0) {
3113 return nullptr;
3114 }
3115
3116 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3117 guarantee(lp_proj != nullptr, "null loop node");
3118
3119 ProjNode* lp_continue = lp_proj->as_Proj();
3120 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3121 if (!lp_exit->is_IfFalse()) {
3122 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3123 // We therefore can't add a single exit condition.
3124 return nullptr;
3125 }
3126 // The unsigned loop exit condition is
3127 // !(i <u limit)
3128 // = i >=u limit
3129 //
3130 // First, we note that for any x for which
3131 // 0 <= x <= INT_MAX
3132 // we can convert x to an unsigned int and still get the same guarantee:
3133 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3134 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3135 //
3136 // With that in mind, if
3137 // limit >= 0 (COND)
3138 // then the unsigned loop exit condition
3139 // i >=u limit (ULE)
3140 // is equivalent to
3141 // i < 0 || i >= limit (SLE-full)
3142 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3143 // (uint) i >=u MAX_INT >= limit >= 0
3144 // or otherwise
3145 // i >= limit >= 0
3146 // holds due to (LEMMA).
3147 //
3148 // For completeness, a counterexample with limit < 0:
3149 // Assume i = -3 and limit = -2:
3150 // i < 0
3151 // -2 < 0
3152 // is true and thus also "i < 0 || i >= limit". But
3153 // i >=u limit
3154 // -3 >=u -2
3155 // is false.
3156 Node* limit = cmpu->in(2);
3157 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3158 if (type_limit->_lo < 0) {
3159 return nullptr;
3160 }
3161
3162 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3163 // stride < 0:
3164 // i < 0 (SLE = SLE-negative)
3165 // stride > 0:
3166 // i >= limit (SLE = SLE-positive)
3167 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3168 //
3169 // Loop:
3170 // <peeled section>
3171 // i >= limit (SLE-positive)
3172 // <-- CUT HERE -->
3173 // i >=u limit (ULE)
3174 // <rest of unpeeled section>
3175 // goto Loop
3176 //
3177 // We exit the loop if:
3178 // (SLE) is true OR (ULE) is true
3179 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3180 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3181 // (SLE) IMPLIES (ULE)
3182 // This indeed holds when (COND) is given:
3183 // - stride > 0:
3184 // i >= limit // (SLE = SLE-positive)
3185 // i >= limit >= 0 // (COND)
3186 // i >=u limit >= 0 // (LEMMA)
3187 // which is the unsigned loop exit condition (ULE).
3188 // - stride < 0:
3189 // i < 0 // (SLE = SLE-negative)
3190 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3191 // MAX_INT >= limit >= 0 // (COND)
3192 // MAX_INT >=u limit >= 0 // (LEMMA)
3193 // and thus from (NEG) and (LEMMA):
3194 // i >=u limit
3195 // which is the unsigned loop exit condition (ULE).
3196 //
3197 //
3198 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3199 // <cloned peeled section>
3200 // i >= limit (SLE-positive)
3201 // Loop:
3202 // i >=u limit (ULE)
3203 // <rest of unpeeled section>
3204 // <peeled section>
3205 // i >= limit (SLE-positive)
3206 // goto Loop
3207 Node* rhs_cmpi;
3208 if (stride > 0) {
3209 rhs_cmpi = limit; // For i >= limit
3210 } else {
3211 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3212 }
3213 // Create a new region on the exit path
3214 RegionNode* reg = insert_region_before_proj(lp_exit);
3215 guarantee(reg != nullptr, "null region node");
3216
3217 // Clone the if-cmpu-true-false using a signed compare
3218 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3219 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3220 reg->add_req(cmpi_exit);
3221
3222 // Clone the if-cmpu-true-false
3223 BoolTest::mask rel_u = bol->_test._test;
3224 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3225 reg->add_req(cmpu_exit);
3226
3227 // Force original if to stay in loop.
3228 short_circuit_if(if_cmpu, lp_continue);
3229
3230 return cmpi_exit->in(0)->as_If();
3231 }
3232
3233 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3234 // Remove a previously inserted signed compare loop exit.
3235 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3236 Node* lp_proj = stay_in_loop(if_cmp, loop);
3237 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3238 stay_in_loop(lp_proj, loop)->is_If() &&
3239 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3240 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3241 if_cmp->set_req(1, con);
3242 }
3243
3244 //------------------------------ scheduled_nodelist -------------------------------------
3245 // Create a post order schedule of nodes that are in the
3246 // "member" set. The list is returned in "sched".
3247 // The first node in "sched" is the loop head, followed by
3248 // nodes which have no inputs in the "member" set, and then
3249 // followed by the nodes that have an immediate input dependence
3250 // on a node in "sched".
3251 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3252
3253 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3254 VectorSet visited;
3255 Node_Stack nstack(loop->_body.size());
3256
3257 Node* n = loop->_head; // top of stack is cached in "n"
3258 uint idx = 0;
3259 visited.set(n->_idx);
3260
3261 // Initially push all with no inputs from within member set
3262 for(uint i = 0; i < loop->_body.size(); i++ ) {
3263 Node *elt = loop->_body.at(i);
3264 if (member.test(elt->_idx)) {
3265 bool found = false;
3266 for (uint j = 0; j < elt->req(); j++) {
3267 Node* def = elt->in(j);
3268 if (def && member.test(def->_idx) && def != elt) {
3269 found = true;
3270 break;
3271 }
3272 }
3273 if (!found && elt != loop->_head) {
3274 nstack.push(n, idx);
3275 n = elt;
3276 assert(!visited.test(n->_idx), "not seen yet");
3277 visited.set(n->_idx);
3278 }
3279 }
3280 }
3281
3282 // traverse out's that are in the member set
3283 while (true) {
3284 if (idx < n->outcnt()) {
3285 Node* use = n->raw_out(idx);
3286 idx++;
3287 if (!visited.test_set(use->_idx)) {
3288 if (member.test(use->_idx)) {
3289 nstack.push(n, idx);
3290 n = use;
3291 idx = 0;
3292 }
3293 }
3294 } else {
3295 // All outputs processed
3296 sched.push(n);
3297 if (nstack.is_empty()) break;
3298 n = nstack.node();
3299 idx = nstack.index();
3300 nstack.pop();
3301 }
3302 }
3303 }
3304
3305
3306 //------------------------------ has_use_in_set -------------------------------------
3307 // Has a use in the vector set
3308 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3309 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3310 Node* use = n->fast_out(j);
3311 if (vset.test(use->_idx)) {
3312 return true;
3313 }
3314 }
3315 return false;
3316 }
3317
3318
3319 //------------------------------ has_use_internal_to_set -------------------------------------
3320 // Has use internal to the vector set (ie. not in a phi at the loop head)
3321 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3322 Node* head = loop->_head;
3323 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3324 Node* use = n->fast_out(j);
3325 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3326 return true;
3327 }
3328 }
3329 return false;
3330 }
3331
3332
3333 //------------------------------ clone_for_use_outside_loop -------------------------------------
3334 // clone "n" for uses that are outside of loop
3335 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3336 int cloned = 0;
3337 assert(worklist.size() == 0, "should be empty");
3338 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3339 Node* use = n->fast_out(j);
3340 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3341 worklist.push(use);
3342 }
3343 }
3344
3345 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3346 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3347 return -1;
3348 }
3349
3350 while( worklist.size() ) {
3351 Node *use = worklist.pop();
3352 if (!has_node(use) || use->in(0) == C->top()) continue;
3353 uint j;
3354 for (j = 0; j < use->req(); j++) {
3355 if (use->in(j) == n) break;
3356 }
3357 assert(j < use->req(), "must be there");
3358
3359 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3360 Node* n_clone = n->clone();
3361 _igvn.replace_input_of(use, j, n_clone);
3362 cloned++;
3363 Node* use_c;
3364 if (!use->is_Phi()) {
3365 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3366 } else {
3367 // Use in a phi is considered a use in the associated predecessor block
3368 use_c = use->in(0)->in(j);
3369 }
3370 set_ctrl(n_clone, use_c);
3371 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3372 get_loop(use_c)->_body.push(n_clone);
3373 _igvn.register_new_node_with_optimizer(n_clone);
3374 #ifndef PRODUCT
3375 if (TracePartialPeeling) {
3376 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3377 }
3378 #endif
3379 }
3380 return cloned;
3381 }
3382
3383
3384 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3385 // clone "n" for special uses that are in the not_peeled region.
3386 // If these def-uses occur in separate blocks, the code generator
3387 // marks the method as not compilable. For example, if a "BoolNode"
3388 // is in a different basic block than the "IfNode" that uses it, then
3389 // the compilation is aborted in the code generator.
3390 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3391 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3392 if (n->is_Phi() || n->is_Load()) {
3393 return;
3394 }
3395 assert(worklist.size() == 0, "should be empty");
3396 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3397 Node* use = n->fast_out(j);
3398 if ( not_peel.test(use->_idx) &&
3399 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3400 use->in(1) == n) {
3401 worklist.push(use);
3402 }
3403 }
3404 if (worklist.size() > 0) {
3405 // clone "n" and insert it between inputs of "n" and the use
3406 Node* n_clone = n->clone();
3407 loop->_body.push(n_clone);
3408 _igvn.register_new_node_with_optimizer(n_clone);
3409 set_ctrl(n_clone, get_ctrl(n));
3410 sink_list.push(n_clone);
3411 not_peel.set(n_clone->_idx);
3412 #ifndef PRODUCT
3413 if (TracePartialPeeling) {
3414 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3415 }
3416 #endif
3417 while( worklist.size() ) {
3418 Node *use = worklist.pop();
3419 _igvn.rehash_node_delayed(use);
3420 for (uint j = 1; j < use->req(); j++) {
3421 if (use->in(j) == n) {
3422 use->set_req(j, n_clone);
3423 }
3424 }
3425 }
3426 }
3427 }
3428
3429
3430 //------------------------------ insert_phi_for_loop -------------------------------------
3431 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3432 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3433 Node *phi = PhiNode::make(lp, back_edge_val);
3434 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3435 // Use existing phi if it already exists
3436 Node *hit = _igvn.hash_find_insert(phi);
3437 if( hit == nullptr ) {
3438 _igvn.register_new_node_with_optimizer(phi);
3439 set_ctrl(phi, lp);
3440 } else {
3441 // Remove the new phi from the graph and use the hit
3442 _igvn.remove_dead_node(phi);
3443 phi = hit;
3444 }
3445 _igvn.replace_input_of(use, idx, phi);
3446 }
3447
3448 #ifdef ASSERT
3449 //------------------------------ is_valid_loop_partition -------------------------------------
3450 // Validate the loop partition sets: peel and not_peel
3451 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3452 VectorSet& not_peel ) {
3453 uint i;
3454 // Check that peel_list entries are in the peel set
3455 for (i = 0; i < peel_list.size(); i++) {
3456 if (!peel.test(peel_list.at(i)->_idx)) {
3457 return false;
3458 }
3459 }
3460 // Check at loop members are in one of peel set or not_peel set
3461 for (i = 0; i < loop->_body.size(); i++ ) {
3462 Node *def = loop->_body.at(i);
3463 uint di = def->_idx;
3464 // Check that peel set elements are in peel_list
3465 if (peel.test(di)) {
3466 if (not_peel.test(di)) {
3467 return false;
3468 }
3469 // Must be in peel_list also
3470 bool found = false;
3471 for (uint j = 0; j < peel_list.size(); j++) {
3472 if (peel_list.at(j)->_idx == di) {
3473 found = true;
3474 break;
3475 }
3476 }
3477 if (!found) {
3478 return false;
3479 }
3480 } else if (not_peel.test(di)) {
3481 if (peel.test(di)) {
3482 return false;
3483 }
3484 } else {
3485 return false;
3486 }
3487 }
3488 return true;
3489 }
3490
3491 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3492 // Ensure a use outside of loop is of the right form
3493 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3494 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3495 return (use->is_Phi() &&
3496 use_c->is_Region() && use_c->req() == 3 &&
3497 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3498 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3499 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3500 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3501 }
3502
3503 //------------------------------ is_valid_clone_loop_form -------------------------------------
3504 // Ensure that all uses outside of loop are of the right form
3505 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3506 uint orig_exit_idx, uint clone_exit_idx) {
3507 uint len = peel_list.size();
3508 for (uint i = 0; i < len; i++) {
3509 Node *def = peel_list.at(i);
3510
3511 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3512 Node *use = def->fast_out(j);
3513 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3514 if (!loop->is_member(get_loop(use_c))) {
3515 // use is not in the loop, check for correct structure
3516 if (use->in(0) == def) {
3517 // Okay
3518 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3519 return false;
3520 }
3521 }
3522 }
3523 }
3524 return true;
3525 }
3526 #endif
3527
3528 //------------------------------ partial_peel -------------------------------------
3529 // Partially peel (aka loop rotation) the top portion of a loop (called
3530 // the peel section below) by cloning it and placing one copy just before
3531 // the new loop head and the other copy at the bottom of the new loop.
3532 //
3533 // before after where it came from
3534 //
3535 // stmt1 stmt1
3536 // loop: stmt2 clone
3537 // stmt2 if condA goto exitA clone
3538 // if condA goto exitA new_loop: new
3539 // stmt3 stmt3 clone
3540 // if !condB goto loop if condB goto exitB clone
3541 // exitB: stmt2 orig
3542 // stmt4 if !condA goto new_loop orig
3543 // exitA: goto exitA
3544 // exitB:
3545 // stmt4
3546 // exitA:
3547 //
3548 // Step 1: find the cut point: an exit test on probable
3549 // induction variable.
3550 // Step 2: schedule (with cloning) operations in the peel
3551 // section that can be executed after the cut into
3552 // the section that is not peeled. This may need
3553 // to clone operations into exit blocks. For
3554 // instance, a reference to A[i] in the not-peel
3555 // section and a reference to B[i] in an exit block
3556 // may cause a left-shift of i by 2 to be placed
3557 // in the peel block. This step will clone the left
3558 // shift into the exit block and sink the left shift
3559 // from the peel to the not-peel section.
3560 // Step 3: clone the loop, retarget the control, and insert
3561 // phis for values that are live across the new loop
3562 // head. This is very dependent on the graph structure
3563 // from clone_loop. It creates region nodes for
3564 // exit control and associated phi nodes for values
3565 // flow out of the loop through that exit. The region
3566 // node is dominated by the clone's control projection.
3567 // So the clone's peel section is placed before the
3568 // new loop head, and the clone's not-peel section is
3569 // forms the top part of the new loop. The original
3570 // peel section forms the tail of the new loop.
3571 // Step 4: update the dominator tree and recompute the
3572 // dominator depth.
3573 //
3574 // orig
3575 //
3576 // stmt1
3577 // |
3578 // v
3579 // predicates
3580 // |
3581 // v
3582 // loop<----+
3583 // | |
3584 // stmt2 |
3585 // | |
3586 // v |
3587 // ifA |
3588 // / | |
3589 // v v |
3590 // false true ^ <-- last_peel
3591 // / | |
3592 // / ===|==cut |
3593 // / stmt3 | <-- first_not_peel
3594 // / | |
3595 // | v |
3596 // v ifB |
3597 // exitA: / \ |
3598 // / \ |
3599 // v v |
3600 // false true |
3601 // / \ |
3602 // / ----+
3603 // |
3604 // v
3605 // exitB:
3606 // stmt4
3607 //
3608 //
3609 // after clone loop
3610 //
3611 // stmt1
3612 // |
3613 // v
3614 // predicates
3615 // / \
3616 // clone / \ orig
3617 // / \
3618 // / \
3619 // v v
3620 // +---->loop loop<----+
3621 // | | | |
3622 // | stmt2 stmt2 |
3623 // | | | |
3624 // | v v |
3625 // | ifA ifA |
3626 // | | \ / | |
3627 // | v v v v |
3628 // ^ true false false true ^ <-- last_peel
3629 // | | ^ \ / | |
3630 // | cut==|== \ \ / ===|==cut |
3631 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3632 // | | dom | | | |
3633 // | v \ 1v v2 v |
3634 // | ifB regionA ifB |
3635 // | / \ | / \ |
3636 // | / \ v / \ |
3637 // | v v exitA: v v |
3638 // | true false false true |
3639 // | / ^ \ / \ |
3640 // +---- \ \ / ----+
3641 // dom \ /
3642 // \ 1v v2
3643 // regionB
3644 // |
3645 // v
3646 // exitB:
3647 // stmt4
3648 //
3649 //
3650 // after partial peel
3651 //
3652 // stmt1
3653 // |
3654 // v
3655 // predicates
3656 // /
3657 // clone / orig
3658 // / TOP
3659 // / \
3660 // v v
3661 // TOP->loop loop----+
3662 // | | |
3663 // stmt2 stmt2 |
3664 // | | |
3665 // v v |
3666 // ifA ifA |
3667 // | \ / | |
3668 // v v v v |
3669 // true false false true | <-- last_peel
3670 // | ^ \ / +------|---+
3671 // +->newloop \ \ / === ==cut | |
3672 // | stmt3 \ \ / TOP | |
3673 // | | dom | | stmt3 | | <-- first_not_peel
3674 // | v \ 1v v2 v | |
3675 // | ifB regionA ifB ^ v
3676 // | / \ | / \ | |
3677 // | / \ v / \ | |
3678 // | v v exitA: v v | |
3679 // | true false false true | |
3680 // | / ^ \ / \ | |
3681 // | | \ \ / v | |
3682 // | | dom \ / TOP | |
3683 // | | \ 1v v2 | |
3684 // ^ v regionB | |
3685 // | | | | |
3686 // | | v ^ v
3687 // | | exitB: | |
3688 // | | stmt4 | |
3689 // | +------------>-----------------+ |
3690 // | |
3691 // +-----------------<---------------------+
3692 //
3693 //
3694 // final graph
3695 //
3696 // stmt1
3697 // |
3698 // v
3699 // predicates
3700 // |
3701 // v
3702 // stmt2 clone
3703 // |
3704 // v
3705 // ........> ifA clone
3706 // : / |
3707 // dom / |
3708 // : v v
3709 // : false true
3710 // : | |
3711 // : | v
3712 // : | newloop<-----+
3713 // : | | |
3714 // : | stmt3 clone |
3715 // : | | |
3716 // : | v |
3717 // : | ifB |
3718 // : | / \ |
3719 // : | v v |
3720 // : | false true |
3721 // : | | | |
3722 // : | v stmt2 |
3723 // : | exitB: | |
3724 // : | stmt4 v |
3725 // : | ifA orig |
3726 // : | / \ |
3727 // : | / \ |
3728 // : | v v |
3729 // : | false true |
3730 // : | / \ |
3731 // : v v -----+
3732 // RegionA
3733 // |
3734 // v
3735 // exitA
3736 //
3737 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3738
3739 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3740 if (!loop->_head->is_Loop()) {
3741 return false;
3742 }
3743 LoopNode *head = loop->_head->as_Loop();
3744
3745 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3746 return false;
3747 }
3748
3749 // Check for complex exit control
3750 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3751 Node *n = loop->_body.at(ii);
3752 int opc = n->Opcode();
3753 if (n->is_Call() ||
3754 opc == Op_Catch ||
3755 opc == Op_CatchProj ||
3756 opc == Op_Jump ||
3757 opc == Op_JumpProj) {
3758 #ifndef PRODUCT
3759 if (TracePartialPeeling) {
3760 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3761 }
3762 #endif
3763 return false;
3764 }
3765 }
3766
3767 int dd = dom_depth(head);
3768
3769 // Step 1: find cut point
3770
3771 // Walk up dominators to loop head looking for first loop exit
3772 // which is executed on every path thru loop.
3773 IfNode *peel_if = nullptr;
3774 IfNode *peel_if_cmpu = nullptr;
3775
3776 Node *iff = loop->tail();
3777 while (iff != head) {
3778 if (iff->is_If()) {
3779 Node *ctrl = get_ctrl(iff->in(1));
3780 if (ctrl->is_top()) return false; // Dead test on live IF.
3781 // If loop-varying exit-test, check for induction variable
3782 if (loop->is_member(get_loop(ctrl)) &&
3783 loop->is_loop_exit(iff) &&
3784 is_possible_iv_test(iff)) {
3785 Node* cmp = iff->in(1)->in(1);
3786 if (cmp->Opcode() == Op_CmpI) {
3787 peel_if = iff->as_If();
3788 } else {
3789 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3790 peel_if_cmpu = iff->as_If();
3791 }
3792 }
3793 }
3794 iff = idom(iff);
3795 }
3796
3797 // Prefer signed compare over unsigned compare.
3798 IfNode* new_peel_if = nullptr;
3799 if (peel_if == nullptr) {
3800 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3801 return false; // No peel point found
3802 }
3803 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3804 if (new_peel_if == nullptr) {
3805 return false; // No peel point found
3806 }
3807 peel_if = new_peel_if;
3808 }
3809 Node* last_peel = stay_in_loop(peel_if, loop);
3810 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3811 if (first_not_peeled == nullptr || first_not_peeled == head) {
3812 return false;
3813 }
3814
3815 #ifndef PRODUCT
3816 if (TraceLoopOpts) {
3817 tty->print("PartialPeel ");
3818 loop->dump_head();
3819 }
3820
3821 if (TracePartialPeeling) {
3822 tty->print_cr("before partial peel one iteration");
3823 Node_List wl;
3824 Node* t = head->in(2);
3825 while (true) {
3826 wl.push(t);
3827 if (t == head) break;
3828 t = idom(t);
3829 }
3830 while (wl.size() > 0) {
3831 Node* tt = wl.pop();
3832 tt->dump();
3833 if (tt == last_peel) tty->print_cr("-- cut --");
3834 }
3835 }
3836 #endif
3837
3838 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3839
3840 VectorSet peel;
3841 VectorSet not_peel;
3842 Node_List peel_list;
3843 Node_List worklist;
3844 Node_List sink_list;
3845
3846 uint estimate = loop->est_loop_clone_sz(1);
3847 if (exceeding_node_budget(estimate)) {
3848 return false;
3849 }
3850
3851 // Set of cfg nodes to peel are those that are executable from
3852 // the head through last_peel.
3853 assert(worklist.size() == 0, "should be empty");
3854 worklist.push(head);
3855 peel.set(head->_idx);
3856 while (worklist.size() > 0) {
3857 Node *n = worklist.pop();
3858 if (n != last_peel) {
3859 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3860 Node* use = n->fast_out(j);
3861 if (use->is_CFG() &&
3862 loop->is_member(get_loop(use)) &&
3863 !peel.test_set(use->_idx)) {
3864 worklist.push(use);
3865 }
3866 }
3867 }
3868 }
3869
3870 // Set of non-cfg nodes to peel are those that are control
3871 // dependent on the cfg nodes.
3872 for (uint i = 0; i < loop->_body.size(); i++) {
3873 Node *n = loop->_body.at(i);
3874 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3875 if (peel.test(n_c->_idx)) {
3876 peel.set(n->_idx);
3877 } else {
3878 not_peel.set(n->_idx);
3879 }
3880 }
3881
3882 // Step 2: move operations from the peeled section down into the
3883 // not-peeled section
3884
3885 // Get a post order schedule of nodes in the peel region
3886 // Result in right-most operand.
3887 scheduled_nodelist(loop, peel, peel_list);
3888
3889 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3890
3891 // For future check for too many new phis
3892 uint old_phi_cnt = 0;
3893 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3894 Node* use = head->fast_out(j);
3895 if (use->is_Phi()) old_phi_cnt++;
3896 }
3897
3898 #ifndef PRODUCT
3899 if (TracePartialPeeling) {
3900 tty->print_cr("\npeeled list");
3901 }
3902 #endif
3903
3904 // Evacuate nodes in peel region into the not_peeled region if possible
3905 bool too_many_clones = false;
3906 uint new_phi_cnt = 0;
3907 uint cloned_for_outside_use = 0;
3908 for (uint i = 0; i < peel_list.size();) {
3909 Node* n = peel_list.at(i);
3910 #ifndef PRODUCT
3911 if (TracePartialPeeling) n->dump();
3912 #endif
3913 bool incr = true;
3914 if (!n->is_CFG()) {
3915 if (has_use_in_set(n, not_peel)) {
3916 // If not used internal to the peeled region,
3917 // move "n" from peeled to not_peeled region.
3918 if (!has_use_internal_to_set(n, peel, loop)) {
3919 // if not pinned and not a load (which maybe anti-dependent on a store)
3920 // and not a CMove (Matcher expects only bool->cmove).
3921 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3922 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3923 if (C->failing()) return false;
3924 if (new_clones == -1) {
3925 too_many_clones = true;
3926 break;
3927 }
3928 cloned_for_outside_use += new_clones;
3929 sink_list.push(n);
3930 peel.remove(n->_idx);
3931 not_peel.set(n->_idx);
3932 peel_list.remove(i);
3933 incr = false;
3934 #ifndef PRODUCT
3935 if (TracePartialPeeling) {
3936 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3937 n->_idx, get_ctrl(n)->_idx);
3938 }
3939 #endif
3940 }
3941 } else {
3942 // Otherwise check for special def-use cases that span
3943 // the peel/not_peel boundary such as bool->if
3944 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3945 new_phi_cnt++;
3946 }
3947 }
3948 }
3949 if (incr) i++;
3950 }
3951
3952 estimate += cloned_for_outside_use + new_phi_cnt;
3953 bool exceed_node_budget = !may_require_nodes(estimate);
3954 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3955
3956 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3957 #ifndef PRODUCT
3958 if (TracePartialPeeling && exceed_phi_limit) {
3959 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3960 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3961 }
3962 #endif
3963 if (new_peel_if != nullptr) {
3964 remove_cmpi_loop_exit(new_peel_if, loop);
3965 }
3966 // Inhibit more partial peeling on this loop
3967 assert(!head->is_partial_peel_loop(), "not partial peeled");
3968 head->mark_partial_peel_failed();
3969 if (cloned_for_outside_use > 0) {
3970 // Terminate this round of loop opts because
3971 // the graph outside this loop was changed.
3972 C->set_major_progress();
3973 return true;
3974 }
3975 return false;
3976 }
3977
3978 // Step 3: clone loop, retarget control, and insert new phis
3979
3980 // Create new loop head for new phis and to hang
3981 // the nodes being moved (sinked) from the peel region.
3982 LoopNode* new_head = new LoopNode(last_peel, last_peel);
3983 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3984 _igvn.register_new_node_with_optimizer(new_head);
3985 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3986 _igvn.replace_input_of(first_not_peeled, 0, new_head);
3987 set_loop(new_head, loop);
3988 loop->_body.push(new_head);
3989 not_peel.set(new_head->_idx);
3990 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3991 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3992
3993 while (sink_list.size() > 0) {
3994 Node* n = sink_list.pop();
3995 set_ctrl(n, new_head);
3996 }
3997
3998 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3999
4000 clone_loop(loop, old_new, dd, IgnoreStripMined);
4001
4002 const uint clone_exit_idx = 1;
4003 const uint orig_exit_idx = 2;
4004 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4005
4006 Node* head_clone = old_new[head->_idx];
4007 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4008 Node* orig_tail_clone = head_clone->in(2);
4009
4010 // Add phi if "def" node is in peel set and "use" is not
4011
4012 for (uint i = 0; i < peel_list.size(); i++) {
4013 Node *def = peel_list.at(i);
4014 if (!def->is_CFG()) {
4015 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4016 Node *use = def->fast_out(j);
4017 if (has_node(use) && use->in(0) != C->top() &&
4018 (!peel.test(use->_idx) ||
4019 (use->is_Phi() && use->in(0) == head)) ) {
4020 worklist.push(use);
4021 }
4022 }
4023 while( worklist.size() ) {
4024 Node *use = worklist.pop();
4025 for (uint j = 1; j < use->req(); j++) {
4026 Node* n = use->in(j);
4027 if (n == def) {
4028
4029 // "def" is in peel set, "use" is not in peel set
4030 // or "use" is in the entry boundary (a phi) of the peel set
4031
4032 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4033
4034 if ( loop->is_member(get_loop( use_c )) ) {
4035 // use is in loop
4036 if (old_new[use->_idx] != nullptr) { // null for dead code
4037 Node* use_clone = old_new[use->_idx];
4038 _igvn.replace_input_of(use, j, C->top());
4039 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4040 }
4041 } else {
4042 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4043 // use is not in the loop, check if the live range includes the cut
4044 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4045 if (not_peel.test(lp_if->_idx)) {
4046 assert(j == orig_exit_idx, "use from original loop");
4047 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4048 }
4049 }
4050 }
4051 }
4052 }
4053 }
4054 }
4055
4056 // Step 3b: retarget control
4057
4058 // Redirect control to the new loop head if a cloned node in
4059 // the not_peeled region has control that points into the peeled region.
4060 // This necessary because the cloned peeled region will be outside
4061 // the loop.
4062 // from to
4063 // cloned-peeled <---+
4064 // new_head_clone: | <--+
4065 // cloned-not_peeled in(0) in(0)
4066 // orig-peeled
4067
4068 for (uint i = 0; i < loop->_body.size(); i++) {
4069 Node *n = loop->_body.at(i);
4070 if (!n->is_CFG() && n->in(0) != nullptr &&
4071 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4072 Node* n_clone = old_new[n->_idx];
4073 if (n_clone->depends_only_on_test()) {
4074 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4075 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4076 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4077 // floating above its range check.
4078 Node* pinned_clone = n_clone->pin_array_access_node();
4079 if (pinned_clone != nullptr) {
4080 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4081 old_new.map(n->_idx, pinned_clone);
4082 _igvn.replace_node(n_clone, pinned_clone);
4083 n_clone = pinned_clone;
4084 }
4085 }
4086 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4087 }
4088 }
4089
4090 // Backedge of the surviving new_head (the clone) is original last_peel
4091 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4092
4093 // Cut first node in original not_peel set
4094 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4095 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4096 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4097
4098 // Copy head_clone back-branch info to original head
4099 // and remove original head's loop entry and
4100 // clone head's back-branch
4101 _igvn.rehash_node_delayed(head); // Multiple edge updates
4102 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4103 head->set_req(LoopNode::LoopBackControl, C->top());
4104 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4105
4106 // Similarly modify the phis
4107 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4108 Node* use = head->fast_out(k);
4109 if (use->is_Phi() && use->outcnt() > 0) {
4110 Node* use_clone = old_new[use->_idx];
4111 _igvn.rehash_node_delayed(use); // Multiple edge updates
4112 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4113 use->set_req(LoopNode::LoopBackControl, C->top());
4114 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4115 }
4116 }
4117
4118 // Step 4: update dominator tree and dominator depth
4119
4120 set_idom(head, orig_tail_clone, dd);
4121 recompute_dom_depth();
4122
4123 // Inhibit more partial peeling on this loop
4124 new_head_clone->set_partial_peel_loop();
4125 C->set_major_progress();
4126 loop->record_for_igvn();
4127
4128 #ifndef PRODUCT
4129 if (TracePartialPeeling) {
4130 tty->print_cr("\nafter partial peel one iteration");
4131 Node_List wl;
4132 Node* t = last_peel;
4133 while (true) {
4134 wl.push(t);
4135 if (t == head_clone) break;
4136 t = idom(t);
4137 }
4138 while (wl.size() > 0) {
4139 Node* tt = wl.pop();
4140 if (tt == head) tty->print_cr("orig head");
4141 else if (tt == new_head_clone) tty->print_cr("new head");
4142 else if (tt == head_clone) tty->print_cr("clone head");
4143 tt->dump();
4144 }
4145 }
4146 #endif
4147
4148 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4149
4150 return true;
4151 }
4152
4153 // Transform:
4154 //
4155 // loop<-----------------+
4156 // | |
4157 // stmt1 stmt2 .. stmtn |
4158 // | | | |
4159 // \ | / |
4160 // v v v |
4161 // region |
4162 // | |
4163 // shared_stmt |
4164 // | |
4165 // v |
4166 // if |
4167 // / \ |
4168 // | -----------+
4169 // v
4170 //
4171 // into:
4172 //
4173 // loop<-------------------+
4174 // | |
4175 // v |
4176 // +->loop |
4177 // | | |
4178 // | stmt1 stmt2 .. stmtn |
4179 // | | | | |
4180 // | | \ / |
4181 // | | v v |
4182 // | | region1 |
4183 // | | | |
4184 // | shared_stmt shared_stmt |
4185 // | | | |
4186 // | v v |
4187 // | if if |
4188 // | /\ / \ |
4189 // +-- | | -------+
4190 // \ /
4191 // v v
4192 // region2
4193 //
4194 // (region2 is shown to merge mirrored projections of the loop exit
4195 // ifs to make the diagram clearer but they really merge the same
4196 // projection)
4197 //
4198 // Conditions for this transformation to trigger:
4199 // - the path through stmt1 is frequent enough
4200 // - the inner loop will be turned into a counted loop after transformation
4201 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4202 if (!DuplicateBackedge) {
4203 return false;
4204 }
4205 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4206 if (!loop->_head->is_Loop()) {
4207 return false;
4208 }
4209
4210 uint estimate = loop->est_loop_clone_sz(1);
4211 if (exceeding_node_budget(estimate)) {
4212 return false;
4213 }
4214
4215 LoopNode *head = loop->_head->as_Loop();
4216
4217 Node* region = nullptr;
4218 IfNode* exit_test = nullptr;
4219 uint inner;
4220 float f;
4221 if (StressDuplicateBackedge) {
4222 if (head->is_strip_mined()) {
4223 return false;
4224 }
4225 Node* c = head->in(LoopNode::LoopBackControl);
4226
4227 while (c != head) {
4228 if (c->is_Region()) {
4229 region = c;
4230 }
4231 c = idom(c);
4232 }
4233
4234 if (region == nullptr) {
4235 return false;
4236 }
4237
4238 inner = 1;
4239 } else {
4240 // Is the shape of the loop that of a counted loop...
4241 Node* back_control = loop_exit_control(head, loop);
4242 if (back_control == nullptr) {
4243 return false;
4244 }
4245
4246 BoolTest::mask bt = BoolTest::illegal;
4247 float cl_prob = 0;
4248 Node* incr = nullptr;
4249 Node* limit = nullptr;
4250 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4251 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4252 return false;
4253 }
4254
4255 // With an extra phi for the candidate iv?
4256 // Or the region node is the loop head
4257 if (!incr->is_Phi() || incr->in(0) == head) {
4258 return false;
4259 }
4260
4261 PathFrequency pf(head, this);
4262 region = incr->in(0);
4263
4264 // Go over all paths for the extra phi's region and see if that
4265 // path is frequent enough and would match the expected iv shape
4266 // if the extra phi is removed
4267 inner = 0;
4268 for (uint i = 1; i < incr->req(); ++i) {
4269 Node* in = incr->in(i);
4270 Node* trunc1 = nullptr;
4271 Node* trunc2 = nullptr;
4272 const TypeInteger* iv_trunc_t = nullptr;
4273 Node* orig_in = in;
4274 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4275 continue;
4276 }
4277 assert(in->Opcode() == Op_AddI, "wrong increment code");
4278 Node* xphi = nullptr;
4279 Node* stride = loop_iv_stride(in, xphi);
4280
4281 if (stride == nullptr) {
4282 continue;
4283 }
4284
4285 PhiNode* phi = loop_iv_phi(xphi, nullptr, head);
4286 if (phi == nullptr ||
4287 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4288 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4289 return false;
4290 }
4291
4292 f = pf.to(region->in(i));
4293 if (f > 0.5) {
4294 inner = i;
4295 break;
4296 }
4297 }
4298
4299 if (inner == 0) {
4300 return false;
4301 }
4302
4303 exit_test = back_control->in(0)->as_If();
4304 }
4305
4306 if (idom(region)->is_Catch()) {
4307 return false;
4308 }
4309
4310 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4311 Unique_Node_List wq;
4312 wq.push(head->in(LoopNode::LoopBackControl));
4313 for (uint i = 0; i < wq.size(); i++) {
4314 Node* c = wq.at(i);
4315 assert(get_loop(c) == loop, "not in the right loop?");
4316 if (c->is_Region()) {
4317 if (c != region) {
4318 for (uint j = 1; j < c->req(); ++j) {
4319 wq.push(c->in(j));
4320 }
4321 }
4322 } else {
4323 wq.push(c->in(0));
4324 }
4325 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4326 }
4327
4328 Node* region_dom = idom(region);
4329
4330 // Can't do the transformation if this would cause a membar pair to
4331 // be split
4332 for (uint i = 0; i < wq.size(); i++) {
4333 Node* c = wq.at(i);
4334 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4335 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4336 if (!wq.member(c->as_MemBar()->leading_membar())) {
4337 return false;
4338 }
4339 }
4340 }
4341
4342 // Collect data nodes that need to be clones as well
4343 int dd = dom_depth(head);
4344
4345 for (uint i = 0; i < loop->_body.size(); ++i) {
4346 Node* n = loop->_body.at(i);
4347 if (has_ctrl(n)) {
4348 Node* c = get_ctrl(n);
4349 if (wq.member(c)) {
4350 wq.push(n);
4351 }
4352 } else {
4353 set_idom(n, idom(n), dd);
4354 }
4355 }
4356
4357 // clone shared_stmt
4358 clone_loop_body(wq, old_new, nullptr);
4359
4360 Node* region_clone = old_new[region->_idx];
4361 region_clone->set_req(inner, C->top());
4362 set_idom(region, region->in(inner), dd);
4363
4364 // Prepare the outer loop
4365 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4366 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4367 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4368 set_idom(head, outer_head, dd);
4369
4370 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4371
4372 // Make one of the shared_stmt copies only reachable from stmt1, the
4373 // other only from stmt2..stmtn.
4374 Node* dom = nullptr;
4375 for (uint i = 1; i < region->req(); ++i) {
4376 if (i != inner) {
4377 _igvn.replace_input_of(region, i, C->top());
4378 }
4379 Node* in = region_clone->in(i);
4380 if (in->is_top()) {
4381 continue;
4382 }
4383 if (dom == nullptr) {
4384 dom = in;
4385 } else {
4386 dom = dom_lca(dom, in);
4387 }
4388 }
4389
4390 set_idom(region_clone, dom, dd);
4391
4392 // Set up the outer loop
4393 for (uint i = 0; i < head->outcnt(); i++) {
4394 Node* u = head->raw_out(i);
4395 if (u->is_Phi()) {
4396 Node* outer_phi = u->clone();
4397 outer_phi->set_req(0, outer_head);
4398 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4399 if (backedge == nullptr) {
4400 backedge = u->in(LoopNode::LoopBackControl);
4401 }
4402 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4403 register_new_node(outer_phi, outer_head);
4404 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4405 }
4406 }
4407
4408 // create control and data nodes for out of loop uses (including region2)
4409 Node_List worklist;
4410 uint new_counter = C->unique();
4411 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4412
4413 Node_List *split_if_set = nullptr;
4414 Node_List *split_bool_set = nullptr;
4415 Node_List *split_cex_set = nullptr;
4416 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4417 split_if_set, split_bool_set, split_cex_set);
4418
4419 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4420
4421 if (exit_test != nullptr) {
4422 float cnt = exit_test->_fcnt;
4423 if (cnt != COUNT_UNKNOWN) {
4424 exit_test->_fcnt = cnt * f;
4425 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4426 }
4427 }
4428
4429 C->set_major_progress();
4430
4431 return true;
4432 }
4433
4434 // AutoVectorize the loop: replace scalar ops with vector ops.
4435 PhaseIdealLoop::AutoVectorizeStatus
4436 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4437 // Counted loop only
4438 if (!lpt->is_counted()) {
4439 return AutoVectorizeStatus::Impossible;
4440 }
4441
4442 // Main-loop only
4443 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4444 if (!cl->is_main_loop()) {
4445 return AutoVectorizeStatus::Impossible;
4446 }
4447
4448 VLoop vloop(lpt, false);
4449 if (!vloop.check_preconditions()) {
4450 return AutoVectorizeStatus::TriedAndFailed;
4451 }
4452
4453 // Ensure the shared data is cleared before each use
4454 vshared.clear();
4455
4456 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4457 if (!vloop_analyzer.success()) {
4458 return AutoVectorizeStatus::TriedAndFailed;
4459 }
4460
4461 SuperWord sw(vloop_analyzer);
4462 if (!sw.transform_loop()) {
4463 return AutoVectorizeStatus::TriedAndFailed;
4464 }
4465
4466 return AutoVectorizeStatus::Success;
4467 }
4468
4469 // Just before insert_pre_post_loops, we can multiversion the loop:
4470 //
4471 // multiversion_if
4472 // | |
4473 // fast_loop slow_loop
4474 //
4475 // In the fast_loop we can make speculative assumptions, and put the
4476 // conditions into the multiversion_if. If the conditions hold at runtime,
4477 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4478 // instead which does not make any of the speculative assumptions.
4479 //
4480 // Note: we only multiversion the loop if the loop does not have any
4481 // auto vectorization check Predicate. If we have that predicate,
4482 // then we can simply add the speculative assumption checks to
4483 // that Predicate. This means we do not need to duplicate the
4484 // loop - we have a smaller graph and save compile time. Should
4485 // the conditions ever fail, then we deopt / trap at the Predicate
4486 // and recompile without that Predicate. At that point we will
4487 // multiversion the loop, so that we can still have speculative
4488 // runtime checks.
4489 //
4490 // We perform the multiversioning when the loop is still in its single
4491 // iteration form, even before we insert pre and post loops. This makes
4492 // the cloning much simpler. However, this means that both the fast
4493 // and the slow loop have to be optimized independently (adding pre
4494 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4495 // we may end up not needing any speculative assumptions in the fast_loop
4496 // and then rejecting the slow_loop by constant folding the multiversion_if.
4497 //
4498 // Therefore, we "delay" the optimization of the slow_loop until we add
4499 // at least one speculative assumption for the fast_loop. If we never
4500 // add such a speculative runtime check, the OpaqueMultiversioningNode
4501 // of the multiversion_if constant folds to true after loop opts, and the
4502 // multiversion_if folds away the "delayed" slow_loop. If we add any
4503 // speculative assumption, then we notify the OpaqueMultiversioningNode
4504 // with "notify_slow_loop_that_it_can_resume_optimizations".
4505 //
4506 // Note: new runtime checks can be added to the multiversion_if with
4507 // PhaseIdealLoop::create_new_if_for_multiversion
4508 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4509 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4510 LoopNode* outer_loop = cl->skip_strip_mined();
4511 Node* entry = outer_loop->in(LoopNode::EntryControl);
4512
4513 // Check we have multiversioning enabled, and are not already multiversioned.
4514 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4515
4516 // Check that we do not have a parse-predicate where we can add the runtime checks
4517 // during auto-vectorization.
4518 const Predicates predicates(entry);
4519 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4520 if (predicate_block->has_parse_predicate()) { return; }
4521
4522 // Check node budget.
4523 uint estimate = lpt->est_loop_clone_sz(2);
4524 if (!may_require_nodes(estimate)) { return; }
4525
4526 do_multiversioning(lpt, old_new);
4527 }
4528
4529 // Returns true if the Reduction node is unordered.
4530 static bool is_unordered_reduction(Node* n) {
4531 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4532 }
4533
4534 // Having ReductionNodes in the loop is expensive. They need to recursively
4535 // fold together the vector values, for every vectorized loop iteration. If
4536 // we encounter the following pattern, we can vector accumulate the values
4537 // inside the loop, and only have a single UnorderedReduction after the loop.
4538 //
4539 // Note: UnorderedReduction represents a ReductionNode which does not require
4540 // calculating in strict order.
4541 //
4542 // CountedLoop init
4543 // | |
4544 // +------+ | +-----------------------+
4545 // | | | |
4546 // PhiNode (s) |
4547 // | |
4548 // | Vector |
4549 // | | |
4550 // UnorderedReduction (first_ur) |
4551 // | |
4552 // ... Vector |
4553 // | | |
4554 // UnorderedReduction (last_ur) |
4555 // | |
4556 // +---------------------+
4557 //
4558 // We patch the graph to look like this:
4559 //
4560 // CountedLoop identity_vector
4561 // | |
4562 // +-------+ | +---------------+
4563 // | | | |
4564 // PhiNode (v) |
4565 // | |
4566 // | Vector |
4567 // | | |
4568 // VectorAccumulator |
4569 // | |
4570 // ... Vector |
4571 // | | |
4572 // init VectorAccumulator |
4573 // | | | |
4574 // UnorderedReduction +-----------+
4575 //
4576 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4577 // use vector_accumulators, which do the same reductions, but only element
4578 // wise. This is a single operation per vector_accumulator, rather than many
4579 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4580 // after the loop, and also reduce the init value into it.
4581 //
4582 // We can not do this with all reductions. Some reductions do not allow the
4583 // reordering of operations (for example float addition/multiplication require
4584 // strict order).
4585 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4586 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4587
4588 // Find all Phi nodes with an unordered Reduction on backedge.
4589 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4590 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4591 Node* phi = cl->fast_out(j);
4592 // We have a phi with a single use, and an unordered Reduction on the backedge.
4593 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4594 continue;
4595 }
4596
4597 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4598 assert(!last_ur->requires_strict_order(), "must be");
4599
4600 // Determine types
4601 const TypeVect* vec_t = last_ur->vect_type();
4602 uint vector_length = vec_t->length();
4603 BasicType bt = vec_t->element_basic_type();
4604
4605 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4606 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4607 const int vopc = VectorNode::opcode(sopc, bt);
4608 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4609 DEBUG_ONLY( last_ur->dump(); )
4610 assert(false, "do not have normal vector op for this reduction");
4611 continue; // not implemented -> fails
4612 }
4613
4614 // Traverse up the chain of unordered Reductions, checking that it loops back to
4615 // the phi. Check that all unordered Reductions only have a single use, except for
4616 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4617 // are outside the loop.
4618 ReductionNode* current = last_ur;
4619 ReductionNode* first_ur = nullptr;
4620 while (true) {
4621 assert(!current->requires_strict_order(), "sanity");
4622
4623 // Expect no ctrl and a vector_input from within the loop.
4624 Node* ctrl = current->in(0);
4625 Node* vector_input = current->in(2);
4626 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4627 DEBUG_ONLY( current->dump(1); )
4628 assert(false, "reduction has ctrl or bad vector_input");
4629 break; // Chain traversal fails.
4630 }
4631
4632 assert(current->vect_type() != nullptr, "must have vector type");
4633 if (current->vect_type() != last_ur->vect_type()) {
4634 // Reductions do not have the same vector type (length and element type).
4635 break; // Chain traversal fails.
4636 }
4637
4638 // Expect single use of an unordered Reduction, except for last_ur.
4639 if (current == last_ur) {
4640 // Expect all uses to be outside the loop, except phi.
4641 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4642 Node* use = current->fast_out(k);
4643 if (use != phi && ctrl_or_self(use) == cl) {
4644 DEBUG_ONLY( current->dump(-1); )
4645 assert(false, "reduction has use inside loop");
4646 // Should not be allowed by SuperWord::mark_reductions
4647 return; // bail out of optimization
4648 }
4649 }
4650 } else {
4651 if (current->outcnt() != 1) {
4652 break; // Chain traversal fails.
4653 }
4654 }
4655
4656 // Expect another unordered Reduction or phi as the scalar input.
4657 Node* scalar_input = current->in(1);
4658 if (is_unordered_reduction(scalar_input) &&
4659 scalar_input->Opcode() == current->Opcode()) {
4660 // Move up the unordered Reduction chain.
4661 current = scalar_input->as_Reduction();
4662 assert(!current->requires_strict_order(), "must be");
4663 } else if (scalar_input == phi) {
4664 // Chain terminates at phi.
4665 first_ur = current;
4666 current = nullptr;
4667 break; // Success.
4668 } else {
4669 // scalar_input is neither phi nor a matching reduction
4670 // Can for example be scalar reduction when we have
4671 // partial vectorization.
4672 break; // Chain traversal fails.
4673 }
4674 }
4675 if (current != nullptr) {
4676 // Chain traversal was not successful.
4677 continue;
4678 }
4679 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4680
4681 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4682 set_root_as_ctrl(identity_scalar);
4683 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4684 register_new_node(identity_vector, C->root());
4685 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4686 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4687
4688 // Turn the scalar phi into a vector phi.
4689 _igvn.rehash_node_delayed(phi);
4690 Node* init = phi->in(1); // Remember init before replacing it.
4691 phi->set_req_X(1, identity_vector, &_igvn);
4692 phi->as_Type()->set_type(vec_t);
4693 _igvn.set_type(phi, vec_t);
4694
4695 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4696 current = first_ur;
4697 while (true) {
4698 // Create vector_accumulator to replace current.
4699 Node* last_vector_accumulator = current->in(1);
4700 Node* vector_input = current->in(2);
4701 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4702 register_new_node(vector_accumulator, cl);
4703 _igvn.replace_node(current, vector_accumulator);
4704 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4705 if (current == last_ur) {
4706 break;
4707 }
4708 current = vector_accumulator->unique_out()->as_Reduction();
4709 assert(!current->requires_strict_order(), "must be");
4710 }
4711
4712 // Create post-loop reduction.
4713 Node* last_accumulator = phi->in(2);
4714 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4715
4716 // Take over uses of last_accumulator that are not in the loop.
4717 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4718 Node* use = last_accumulator->out(i);
4719 if (use != phi && use != post_loop_reduction) {
4720 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4721 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4722 --i;
4723 }
4724 }
4725 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4726 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4727
4728 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4729 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4730 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4731 }
4732 }
4733
4734 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4735 for (uint i = 0; i < _data_nodes.size(); i++) {
4736 clone(_data_nodes[i], new_ctrl);
4737 }
4738 }
4739
4740 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4741 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4742 Node* clone = node->clone();
4743 _phase->igvn().register_new_node_with_optimizer(clone);
4744 _orig_to_new.put(node, clone);
4745 _phase->set_ctrl(clone, new_ctrl);
4746 if (node->is_CastII()) {
4747 clone->set_req(0, new_ctrl);
4748 }
4749 }
4750
4751 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4752 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4753 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4754 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4755 for (uint i = 1; i < node->req(); i++) {
4756 Node** cloned_input = _orig_to_new.get(node->in(i));
4757 if (cloned_input != nullptr) {
4758 // Input was also cloned -> rewire clone to the cloned input.
4759 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4760 }
4761 }
4762 });
4763 }
4764
4765 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4766 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4767 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4768 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4769 Node* new_ctrl) {
4770 for (uint i = 0; i < _data_nodes.size(); i++) {
4771 Node* data_node = _data_nodes[i];
4772 if (data_node->is_Opaque1()) {
4773 transform_opaque_node(transform_strategy, data_node);
4774 } else {
4775 clone(data_node, new_ctrl);
4776 }
4777 }
4778 }
4779
4780 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4781 Node* transformed_node;
4782 if (node->is_OpaqueLoopInit()) {
4783 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4784 } else {
4785 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4786 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4787 }
4788 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4789 _orig_to_new.put(node, transformed_node);
4790 }