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/checkedCast.hpp"
45 #include "utilities/macros.hpp"
46
47 //=============================================================================
48 //------------------------------split_thru_phi---------------------------------
49 // Split Node 'n' through merge point if there is enough win.
50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
51 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
52 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
53 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
54 // so disable this for now
55 return nullptr;
56 }
57
58 // Splitting range check CastIIs through a loop induction Phi can
59 // cause new Phis to be created that are left unrelated to the loop
60 // induction Phi and prevent optimizations (vectorization)
61 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
62 n->in(1) == region->as_CountedLoop()->phi()) {
63 return nullptr;
64 }
65
66 if (cannot_split_division(n, region)) {
67 return nullptr;
68 }
69
70 SplitThruPhiWins wins(region);
71 assert(!n->is_CFG(), "");
72 assert(region->is_Region(), "");
73
74 const Type* type = n->bottom_type();
75 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
76 Node* phi;
77 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
78 int iid = t_oop->instance_id();
79 int index = C->get_alias_index(t_oop);
80 int offset = t_oop->offset();
81 phi = new PhiNode(region, type, nullptr, iid, index, offset);
82 } else {
83 phi = PhiNode::make_blank(region, n);
84 }
85 uint old_unique = C->unique();
86 for (uint i = 1; i < region->req(); i++) {
87 Node* x;
88 Node* the_clone = nullptr;
89 if (region->in(i) == C->top()) {
90 x = C->top(); // Dead path? Use a dead data op
91 } else {
92 x = n->clone(); // Else clone up the data op
93 the_clone = x; // Remember for possible deletion.
94 // Alter data node to use pre-phi inputs
95 if (n->in(0) == region)
96 x->set_req( 0, region->in(i) );
97 for (uint j = 1; j < n->req(); j++) {
98 Node* in = n->in(j);
99 if (in->is_Phi() && in->in(0) == region)
100 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
101 }
102 }
103 // Check for a 'win' on some paths
104 const Type* t = x->Value(&_igvn);
105
106 bool singleton = t->singleton();
107
108 // A TOP singleton indicates that there are no possible values incoming
109 // along a particular edge. In most cases, this is OK, and the Phi will
110 // be eliminated later in an Ideal call. However, we can't allow this to
111 // happen if the singleton occurs on loop entry, as the elimination of
112 // the PhiNode may cause the resulting node to migrate back to a previous
113 // loop iteration.
114 if (singleton && t == Type::TOP) {
115 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
116 // irreducible loop may not be indicated by an affirmative is_Loop());
117 // therefore, the only top we can split thru a phi is on a backedge of
118 // a loop.
119 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
120 }
121
122 if (singleton) {
123 wins.add_win(i);
124 x = makecon(t);
125 } else {
126 // We now call Identity to try to simplify the cloned node.
127 // Note that some Identity methods call phase->type(this).
128 // Make sure that the type array is big enough for
129 // our new node, even though we may throw the node away.
130 // (Note: This tweaking with igvn only works because x is a new node.)
131 _igvn.set_type(x, t);
132 // If x is a TypeNode, capture any more-precise type permanently into Node
133 // otherwise it will be not updated during igvn->transform since
134 // igvn->type(x) is set to x->Value() already.
135 x->raise_bottom_type(t);
136 Node* y = x->Identity(&_igvn);
137 if (y != x) {
138 wins.add_win(i);
139 x = y;
140 } else {
141 y = _igvn.hash_find(x);
142 if (y == nullptr) {
143 y = similar_subtype_check(x, region->in(i));
144 }
145 if (y) {
146 wins.add_win(i);
147 x = y;
148 } else {
149 // Else x is a new node we are keeping
150 // We do not need register_new_node_with_optimizer
151 // because set_type has already been called.
152 _igvn._worklist.push(x);
153 }
154 }
155 }
156
157 phi->set_req( i, x );
158
159 if (the_clone == nullptr) {
160 continue;
161 }
162
163 if (the_clone != x) {
164 _igvn.remove_dead_node(the_clone);
165 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
166 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
167 // it is not a win if 'x' moved from an outer to an inner loop
168 // this edge case can only happen for Load nodes
169 wins.reset();
170 break;
171 }
172 }
173 // Too few wins?
174 if (!wins.profitable(policy)) {
175 _igvn.remove_dead_node(phi);
176 return nullptr;
177 }
178
179 // Record Phi
180 register_new_node( phi, region );
181
182 for (uint i2 = 1; i2 < phi->req(); i2++) {
183 Node *x = phi->in(i2);
184 // If we commoned up the cloned 'x' with another existing Node,
185 // the existing Node picks up a new use. We need to make the
186 // existing Node occur higher up so it dominates its uses.
187 Node *old_ctrl;
188 IdealLoopTree *old_loop;
189
190 if (x->is_Con()) {
191 assert(get_ctrl(x) == C->root(), "constant control is not root");
192 continue;
193 }
194 // The occasional new node
195 if (x->_idx >= old_unique) { // Found a new, unplaced node?
196 old_ctrl = nullptr;
197 old_loop = nullptr; // Not in any prior loop
198 } else {
199 old_ctrl = get_ctrl(x);
200 old_loop = get_loop(old_ctrl); // Get prior loop
201 }
202 // New late point must dominate new use
203 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
204 if (new_ctrl == old_ctrl) // Nothing is changed
205 continue;
206
207 IdealLoopTree *new_loop = get_loop(new_ctrl);
208
209 // Don't move x into a loop if its uses are
210 // outside of loop. Otherwise x will be cloned
211 // for each use outside of this loop.
212 IdealLoopTree *use_loop = get_loop(region);
213 if (!new_loop->is_member(use_loop) &&
214 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
215 // Take early control, later control will be recalculated
216 // during next iteration of loop optimizations.
217 new_ctrl = get_early_ctrl(x);
218 new_loop = get_loop(new_ctrl);
219 }
220 // Set new location
221 set_ctrl(x, new_ctrl);
222 // If changing loop bodies, see if we need to collect into new body
223 if (old_loop != new_loop) {
224 if (old_loop && !old_loop->_child)
225 old_loop->_body.yank(x);
226 if (!new_loop->_child)
227 new_loop->_body.push(x); // Collect body info
228 }
229 }
230
231 #ifndef PRODUCT
232 if (TraceLoopOpts) {
233 tty->print_cr("Split %d %s through %d Phi in %d %s",
234 n->_idx, n->Name(), phi->_idx, region->_idx, region->Name());
235 }
236 #endif // !PRODUCT
237
238 return phi;
239 }
240
241 // Test whether node 'x' can move into an inner loop relative to node 'n'.
242 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
243 // BUT it can also return true and 'x' is in the outer loop
244 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
245 IdealLoopTree* n_loop_tree = get_loop(n_loop);
246 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
247 // x_loop_tree should be outer or same loop as n_loop_tree
248 return !x_loop_tree->is_member(n_loop_tree);
249 }
250
251 // Subtype checks that carry profile data don't common so look for a replacement by following edges
252 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
253 if (x->is_SubTypeCheck()) {
254 Node* in1 = x->in(1);
255 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
256 Node* u = in1->fast_out(i);
257 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
258 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
259 Node* bol = u->fast_out(j);
260 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
261 Node* iff = bol->fast_out(k);
262 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
263 // unrelated profile
264 if (iff->is_If() && is_dominator(iff, r_in)) {
265 return u;
266 }
267 }
268 }
269 }
270 }
271 }
272 return nullptr;
273 }
274
275 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
276 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
277 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
278 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
279 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
280 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
281 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
282 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
283 const Type* zero;
284 switch (n->Opcode()) {
285 case Op_DivI:
286 case Op_ModI:
287 case Op_UDivI:
288 case Op_UModI:
289 zero = TypeInt::ZERO;
290 break;
291 case Op_DivL:
292 case Op_ModL:
293 case Op_UDivL:
294 case Op_UModL:
295 zero = TypeLong::ZERO;
296 break;
297 default:
298 return false;
299 }
300
301 if (n->in(0) != nullptr) {
302 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
303 return true;
304 }
305
306 Node* divisor = n->in(2);
307 return is_divisor_loop_phi(divisor, region) &&
308 loop_phi_backedge_type_contains_zero(divisor, zero);
309 }
310
311 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
312 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
313 }
314
315 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
316 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
317 }
318
319 //------------------------------dominated_by------------------------------------
320 // Replace the dominated test with an obvious true or false. Place it on the
321 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
322 // live path up to the dominating control.
323 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
324 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
325
326 // prevdom is the dominating projection of the dominating test.
327 assert(iff->Opcode() == Op_If ||
328 iff->Opcode() == Op_CountedLoopEnd ||
329 iff->Opcode() == Op_LongCountedLoopEnd ||
330 iff->Opcode() == Op_RangeCheck ||
331 iff->Opcode() == Op_ParsePredicate,
332 "Check this code when new subtype is added");
333
334 int pop = prevdom->Opcode();
335 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
336 if (flip) {
337 if (pop == Op_IfTrue)
338 pop = Op_IfFalse;
339 else
340 pop = Op_IfTrue;
341 }
342 // 'con' is set to true or false to kill the dominated test.
343 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
344 // Hack the dominated test
345 _igvn.replace_input_of(iff, 1, con);
346
347 // If I don't have a reachable TRUE and FALSE path following the IfNode then
348 // I can assume this path reaches an infinite loop. In this case it's not
349 // important to optimize the data Nodes - either the whole compilation will
350 // be tossed or this path (and all data Nodes) will go dead.
351 if (iff->outcnt() != 2) {
352 return;
353 }
354
355 // Make control-dependent data Nodes on the live path (path that will remain
356 // once the dominated IF is removed) become control-dependent on the
357 // dominating projection.
358 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
359
360 if (dp == nullptr) {
361 return;
362 }
363
364 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
365 }
366
367 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
368 IdealLoopTree* old_loop = get_loop(source);
369
370 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
371 Node* out = source->fast_out(i); // Control-dependent node
372 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
373 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
374 assert(out->in(0) == source, "must be control dependent on source");
375 _igvn.replace_input_of(out, 0, dominator);
376 if (pin_array_access_nodes) {
377 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
378 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
379 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
380 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
381 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
382 // dominating check.
383 Node* clone = out->pin_array_access_node();
384 if (clone != nullptr) {
385 clone = _igvn.register_new_node_with_optimizer(clone, out);
386 _igvn.replace_node(out, clone);
387 out = clone;
388 }
389 }
390 set_early_ctrl(out, false);
391 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
392 if (old_loop != new_loop) {
393 if (!old_loop->_child) {
394 old_loop->_body.yank(out);
395 }
396 if (!new_loop->_child) {
397 new_loop->_body.push(out);
398 }
399 }
400 --i;
401 --imax;
402 }
403 }
404 }
405
406 //------------------------------has_local_phi_input----------------------------
407 // Return TRUE if 'n' has Phi inputs from its local block and no other
408 // block-local inputs (all non-local-phi inputs come from earlier blocks)
409 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
410 Node *n_ctrl = get_ctrl(n);
411 // See if some inputs come from a Phi in this block, or from before
412 // this block.
413 uint i;
414 for( i = 1; i < n->req(); i++ ) {
415 Node *phi = n->in(i);
416 if( phi->is_Phi() && phi->in(0) == n_ctrl )
417 break;
418 }
419 if( i >= n->req() )
420 return nullptr; // No Phi inputs; nowhere to clone thru
421
422 // Check for inputs created between 'n' and the Phi input. These
423 // must split as well; they have already been given the chance
424 // (courtesy of a post-order visit) and since they did not we must
425 // recover the 'cost' of splitting them by being very profitable
426 // when splitting 'n'. Since this is unlikely we simply give up.
427 for( i = 1; i < n->req(); i++ ) {
428 Node *m = n->in(i);
429 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
430 // We allow the special case of AddP's with no local inputs.
431 // This allows us to split-up address expressions.
432 if (m->is_AddP() &&
433 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
434 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
435 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
436 // Move the AddP up to the dominating point. That's fine because control of m's inputs
437 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
438 Node* c = find_non_split_ctrl(idom(n_ctrl));
439 if (c->is_OuterStripMinedLoop()) {
440 c->as_Loop()->verify_strip_mined(1);
441 c = c->in(LoopNode::EntryControl);
442 }
443 set_ctrl_and_loop(m, c);
444 continue;
445 }
446 return nullptr;
447 }
448 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
449 }
450
451 return n_ctrl;
452 }
453
454 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
455 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
456 assert(bt == T_INT || bt == T_LONG, "only for integers");
457 int n_op = n->Opcode();
458
459 if (n_op == Op_LShift(bt)) {
460 // Scale is loop invariant
461 Node* scale = n->in(2);
462 Node* scale_ctrl = get_ctrl(scale);
463 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
464 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
465 return nullptr;
466 }
467 const TypeInt* scale_t = scale->bottom_type()->isa_int();
468 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
469 return nullptr; // Dont bother with byte/short masking
470 }
471 // Add must vary with loop (else shift would be loop-invariant)
472 Node* add = n->in(1);
473 Node* add_ctrl = get_ctrl(add);
474 IdealLoopTree* add_loop = get_loop(add_ctrl);
475 if (n_loop != add_loop) {
476 return nullptr; // happens w/ evil ZKM loops
477 }
478
479 // Convert I-V into I+ (0-V); same for V-I
480 if (add->Opcode() == Op_Sub(bt) &&
481 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
482 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
483 Node* zero = integercon(0, bt);
484 Node* neg = SubNode::make(zero, add->in(2), bt);
485 register_new_node_with_ctrl_of(neg, add->in(2));
486 add = AddNode::make(add->in(1), neg, bt);
487 register_new_node(add, add_ctrl);
488 }
489 if (add->Opcode() != Op_Add(bt)) return nullptr;
490 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
491 // See if one add input is loop invariant
492 Node* add_var = add->in(1);
493 Node* add_var_ctrl = get_ctrl(add_var);
494 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
495 Node* add_invar = add->in(2);
496 Node* add_invar_ctrl = get_ctrl(add_invar);
497 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
498 if (add_invar_loop == n_loop) {
499 // Swap to find the invariant part
500 add_invar = add_var;
501 add_invar_ctrl = add_var_ctrl;
502 add_invar_loop = add_var_loop;
503 add_var = add->in(2);
504 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
505 return nullptr;
506 }
507 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
508 return nullptr; // No invariant part of the add?
509 }
510
511 // Yes! Reshape address expression!
512 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
513 Node* inv_scale_ctrl =
514 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
515 add_invar_ctrl : scale_ctrl;
516 register_new_node(inv_scale, inv_scale_ctrl);
517 Node* var_scale = LShiftNode::make(add_var, scale, bt);
518 register_new_node(var_scale, n_ctrl);
519 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
520 register_new_node(var_add, n_ctrl);
521 _igvn.replace_node(n, var_add);
522 return var_add;
523 }
524 return nullptr;
525 }
526
527 //------------------------------remix_address_expressions----------------------
528 // Rework addressing expressions to get the most loop-invariant stuff
529 // moved out. We'd like to do all associative operators, but it's especially
530 // important (common) to do address expressions.
531 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
532 if (!has_ctrl(n)) return nullptr;
533 Node* n_ctrl = get_ctrl(n);
534 IdealLoopTree* n_loop = get_loop(n_ctrl);
535
536 // See if 'n' mixes loop-varying and loop-invariant inputs and
537 // itself is loop-varying.
538
539 // Only interested in binary ops (and AddP)
540 if (n->req() < 3 || n->req() > 4) return nullptr;
541
542 Node* n1_ctrl = get_ctrl(n->in( 1));
543 Node* n2_ctrl = get_ctrl(n->in( 2));
544 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
545 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
546 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
547 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
548
549 // Does one of my inputs spin in a tighter loop than self?
550 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
551 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
552 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
553 return nullptr; // Leave well enough alone
554 }
555
556 // Is at least one of my inputs loop-invariant?
557 if (n1_loop == n_loop &&
558 n2_loop == n_loop &&
559 n3_loop == n_loop) {
560 return nullptr; // No loop-invariant inputs
561 }
562
563 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
564 if (res != nullptr) {
565 return res;
566 }
567 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
568 if (res != nullptr) {
569 return res;
570 }
571
572 int n_op = n->Opcode();
573 // Replace (I+V) with (V+I)
574 if (n_op == Op_AddI ||
575 n_op == Op_AddL ||
576 n_op == Op_AddF ||
577 n_op == Op_AddD ||
578 n_op == Op_MulI ||
579 n_op == Op_MulL ||
580 n_op == Op_MulF ||
581 n_op == Op_MulD) {
582 if (n2_loop == n_loop) {
583 assert(n1_loop != n_loop, "");
584 n->swap_edges(1, 2);
585 }
586 }
587
588 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
589 // but not if I2 is a constant. Skip for irreducible loops.
590 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
591 if (n2_loop == n_loop && n3_loop != n_loop) {
592 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
593 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
594 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
595 IdealLoopTree* n22loop = get_loop(n22_ctrl);
596 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
597 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
598 n23_loop == n_loop) {
599 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
600 // Stuff new AddP in the loop preheader
601 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
602 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
603 register_new_node(add2, n_ctrl);
604 _igvn.replace_node(n, add2);
605 return add2;
606 }
607 }
608 }
609
610 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
611 if (n2_loop != n_loop && n3_loop == n_loop) {
612 if (n->in(3)->Opcode() == Op_AddX) {
613 Node* V = n->in(3)->in(1);
614 Node* I = n->in(3)->in(2);
615 if (is_member(n_loop,get_ctrl(V))) {
616 } else {
617 Node *tmp = V; V = I; I = tmp;
618 }
619 if (!is_member(n_loop,get_ctrl(I))) {
620 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
621 // Stuff new AddP in the loop preheader
622 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
623 Node* add2 = new AddPNode(n->in(1), add1, V);
624 register_new_node(add2, n_ctrl);
625 _igvn.replace_node(n, add2);
626 return add2;
627 }
628 }
629 }
630 }
631
632 return nullptr;
633 }
634
635 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
636 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
637 assert(n->Opcode() == Op_AddI, "sanity");
638 Node * nn = nullptr;
639 Node * in1 = n->in(1);
640 Node * in2 = n->in(2);
641 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
642 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
643 if (loop_n->is_counted() &&
644 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
645 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
646 Matcher::match_rule_supported(Op_MulAddS2I)) {
647 Node* mul_in1 = in1->in(1);
648 Node* mul_in2 = in1->in(2);
649 Node* mul_in3 = in2->in(1);
650 Node* mul_in4 = in2->in(2);
651 if (mul_in1->Opcode() == Op_LoadS &&
652 mul_in2->Opcode() == Op_LoadS &&
653 mul_in3->Opcode() == Op_LoadS &&
654 mul_in4->Opcode() == Op_LoadS) {
655 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
656 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
657 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
658 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
659 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
660 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
661 // All nodes should be in the same counted loop.
662 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
663 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
664 Node* adr1 = mul_in1->in(MemNode::Address);
665 Node* adr2 = mul_in2->in(MemNode::Address);
666 Node* adr3 = mul_in3->in(MemNode::Address);
667 Node* adr4 = mul_in4->in(MemNode::Address);
668 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
669 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
670 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
671 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
672 register_new_node_with_ctrl_of(nn, n);
673 _igvn.replace_node(n, nn);
674 return nn;
675 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
676 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
677 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
678 register_new_node_with_ctrl_of(nn, n);
679 _igvn.replace_node(n, nn);
680 return nn;
681 }
682 }
683 }
684 }
685 }
686 }
687 return nn;
688 }
689
690 //------------------------------conditional_move-------------------------------
691 // Attempt to replace a Phi with a conditional move. We have some pretty
692 // strict profitability requirements. All Phis at the merge point must
693 // be converted, so we can remove the control flow. We need to limit the
694 // number of c-moves to a small handful. All code that was in the side-arms
695 // of the CFG diamond is now speculatively executed. This code has to be
696 // "cheap enough". We are pretty much limited to CFG diamonds that merge
697 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
698 Node *PhaseIdealLoop::conditional_move( Node *region ) {
699
700 assert(region->is_Region(), "sanity check");
701 if (region->req() != 3) return nullptr;
702
703 // Check for CFG diamond
704 Node *lp = region->in(1);
705 Node *rp = region->in(2);
706 if (!lp || !rp) return nullptr;
707 Node *lp_c = lp->in(0);
708 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
709 IfNode *iff = lp_c->as_If();
710
711 // Check for ops pinned in an arm of the diamond.
712 // Can't remove the control flow in this case
713 if (lp->outcnt() > 1) return nullptr;
714 if (rp->outcnt() > 1) return nullptr;
715
716 IdealLoopTree* r_loop = get_loop(region);
717 assert(r_loop == get_loop(iff), "sanity");
718 // Always convert to CMOVE if all results are used only outside this loop.
719 bool used_inside_loop = (r_loop == _ltree_root);
720
721 // Check profitability
722 int cost = 0;
723 int phis = 0;
724 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
725 Node *out = region->fast_out(i);
726 if (!out->is_Phi()) continue; // Ignore other control edges, etc
727 phis++;
728 PhiNode* phi = out->as_Phi();
729 BasicType bt = phi->type()->basic_type();
730 switch (bt) {
731 case T_DOUBLE:
732 case T_FLOAT:
733 if (C->use_cmove()) {
734 continue; //TODO: maybe we want to add some cost
735 }
736 cost += Matcher::float_cmove_cost(); // Could be very expensive
737 break;
738 case T_LONG: {
739 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
740 }
741 case T_INT: // These all CMOV fine
742 case T_ADDRESS: { // (RawPtr)
743 cost++;
744 break;
745 }
746 case T_NARROWOOP: // Fall through
747 case T_OBJECT: { // Base oops are OK, but not derived oops
748 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
749 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
750 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
751 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
752 // have a Phi for the base here that we convert to a CMOVE all is well
753 // and good. But if the base is dead, we'll not make a CMOVE. Later
754 // the allocator will have to produce a base by creating a CMOVE of the
755 // relevant bases. This puts the allocator in the business of
756 // manufacturing expensive instructions, generally a bad plan.
757 // Just Say No to Conditionally-Moved Derived Pointers.
758 if (tp && tp->offset() != 0)
759 return nullptr;
760 cost++;
761 break;
762 }
763 default:
764 return nullptr; // In particular, can't do memory or I/O
765 }
766 // Add in cost any speculative ops
767 for (uint j = 1; j < region->req(); j++) {
768 Node *proj = region->in(j);
769 Node *inp = phi->in(j);
770 if (get_ctrl(inp) == proj) { // Found local op
771 cost++;
772 // Check for a chain of dependent ops; these will all become
773 // speculative in a CMOV.
774 for (uint k = 1; k < inp->req(); k++)
775 if (get_ctrl(inp->in(k)) == proj)
776 cost += ConditionalMoveLimit; // Too much speculative goo
777 }
778 }
779 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
780 // This will likely Split-If, a higher-payoff operation.
781 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
782 Node* use = phi->fast_out(k);
783 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
784 cost += ConditionalMoveLimit;
785 // Is there a use inside the loop?
786 // Note: check only basic types since CMoveP is pinned.
787 if (!used_inside_loop && is_java_primitive(bt)) {
788 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
789 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
790 used_inside_loop = true;
791 }
792 }
793 }
794 }//for
795 Node* bol = iff->in(1);
796 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
797 if (bol->is_OpaqueTemplateAssertionPredicate()) {
798 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
799 return nullptr;
800 }
801 if (bol->is_OpaqueMultiversioning()) {
802 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
803 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
804 // and will make the multiversion_if constant fold in the next IGVN round.
805 return nullptr;
806 }
807 if (!bol->is_Bool()) {
808 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
809 return nullptr;
810 }
811 int cmp_op = bol->in(1)->Opcode();
812 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
813 return nullptr;
814 }
815 // It is expensive to generate flags from a float compare.
816 // Avoid duplicated float compare.
817 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
818
819 float infrequent_prob = PROB_UNLIKELY_MAG(3);
820 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
821 if (used_inside_loop) {
822 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
823
824 // BlockLayoutByFrequency optimization moves infrequent branch
825 // from hot path. No point in CMOV'ing in such case (110 is used
826 // instead of 100 to take into account not exactness of float value).
827 if (BlockLayoutByFrequency) {
828 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
829 }
830 }
831 // Check for highly predictable branch. No point in CMOV'ing if
832 // we are going to predict accurately all the time.
833 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
834 //keep going
835 } else if (iff->_prob < infrequent_prob ||
836 iff->_prob > (1.0f - infrequent_prob))
837 return nullptr;
838
839 // --------------
840 // Now replace all Phis with CMOV's
841 Node *cmov_ctrl = iff->in(0);
842 uint flip = (lp->Opcode() == Op_IfTrue);
843 Node_List wq;
844 while (1) {
845 PhiNode* phi = nullptr;
846 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
847 Node *out = region->fast_out(i);
848 if (out->is_Phi()) {
849 phi = out->as_Phi();
850 break;
851 }
852 }
853 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
854 break;
855 }
856 // Move speculative ops
857 wq.push(phi);
858 while (wq.size() > 0) {
859 Node *n = wq.pop();
860 for (uint j = 1; j < n->req(); j++) {
861 Node* m = n->in(j);
862 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
863 set_ctrl(m, cmov_ctrl);
864 wq.push(m);
865 }
866 }
867 }
868 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
869 register_new_node(cmov, cmov_ctrl);
870 _igvn.replace_node(phi, cmov);
871 #ifndef PRODUCT
872 if (TraceLoopOpts) {
873 tty->print("CMOV ");
874 r_loop->dump_head();
875 if (Verbose) {
876 bol->in(1)->dump(1);
877 cmov->dump(1);
878 }
879 }
880 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
881 #endif
882 }
883
884 // The useless CFG diamond will fold up later; see the optimization in
885 // RegionNode::Ideal.
886 _igvn._worklist.push(region);
887
888 return iff->in(1);
889 }
890
891 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
892 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
893 Node* u = m->fast_out(i);
894 if (u->is_CFG()) {
895 if (u->is_NeverBranch()) {
896 u = u->as_NeverBranch()->proj_out(0);
897 enqueue_cfg_uses(u, wq);
898 } else {
899 wq.push(u);
900 }
901 }
902 }
903 }
904
905 // Try moving a store out of a loop, right before the loop
906 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
907 // Store has to be first in the loop body
908 IdealLoopTree *n_loop = get_loop(n_ctrl);
909 if (n->is_Store() && n_loop != _ltree_root &&
910 n_loop->is_loop() && n_loop->_head->is_Loop() &&
911 n->in(0) != nullptr) {
912 Node* address = n->in(MemNode::Address);
913 Node* value = n->in(MemNode::ValueIn);
914 Node* mem = n->in(MemNode::Memory);
915 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
916 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
917
918 // - address and value must be loop invariant
919 // - memory must be a memory Phi for the loop
920 // - Store must be the only store on this memory slice in the
921 // loop: if there's another store following this one then value
922 // written at iteration i by the second store could be overwritten
923 // at iteration i+n by the first store: it's not safe to move the
924 // first store out of the loop
925 // - nothing must observe the memory Phi: it guarantees no read
926 // before the store, we are also guaranteed the store post
927 // dominates the loop head (ignoring a possible early
928 // exit). Otherwise there would be extra Phi involved between the
929 // loop's Phi and the store.
930 // - there must be no early exit from the loop before the Store
931 // (such an exit most of the time would be an extra use of the
932 // memory Phi but sometimes is a bottom memory Phi that takes the
933 // store as input).
934
935 if (!n_loop->is_member(address_loop) &&
936 !n_loop->is_member(value_loop) &&
937 mem->is_Phi() && mem->in(0) == n_loop->_head &&
938 mem->outcnt() == 1 &&
939 mem->in(LoopNode::LoopBackControl) == n) {
940
941 assert(n_loop->_tail != nullptr, "need a tail");
942 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
943
944 // Verify that there's no early exit of the loop before the store.
945 bool ctrl_ok = false;
946 {
947 // Follow control from loop head until n, we exit the loop or
948 // we reach the tail
949 ResourceMark rm;
950 Unique_Node_List wq;
951 wq.push(n_loop->_head);
952
953 for (uint next = 0; next < wq.size(); ++next) {
954 Node *m = wq.at(next);
955 if (m == n->in(0)) {
956 ctrl_ok = true;
957 continue;
958 }
959 assert(!has_ctrl(m), "should be CFG");
960 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
961 ctrl_ok = false;
962 break;
963 }
964 enqueue_cfg_uses(m, wq);
965 if (wq.size() > 10) {
966 ctrl_ok = false;
967 break;
968 }
969 }
970 }
971 if (ctrl_ok) {
972 // move the Store
973 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
974 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
975 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
976 // Disconnect the phi now. An empty phi can confuse other
977 // optimizations in this pass of loop opts.
978 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
979 n_loop->_body.yank(mem);
980
981 set_ctrl_and_loop(n, n->in(0));
982
983 return n;
984 }
985 }
986 }
987 return nullptr;
988 }
989
990 // Try moving a store out of a loop, right after the loop
991 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
992 if (n->is_Store() && n->in(0) != nullptr) {
993 Node *n_ctrl = get_ctrl(n);
994 IdealLoopTree *n_loop = get_loop(n_ctrl);
995 // Store must be in a loop
996 if (n_loop != _ltree_root && !n_loop->_irreducible) {
997 Node* address = n->in(MemNode::Address);
998 Node* value = n->in(MemNode::ValueIn);
999 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
1000 // address must be loop invariant
1001 if (!n_loop->is_member(address_loop)) {
1002 // Store must be last on this memory slice in the loop and
1003 // nothing in the loop must observe it
1004 Node* phi = nullptr;
1005 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1006 Node* u = n->fast_out(i);
1007 if (has_ctrl(u)) { // control use?
1008 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1009 if (!n_loop->is_member(u_loop)) {
1010 continue;
1011 }
1012 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1013 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1014 // multiple phis on the same slice are possible
1015 if (phi != nullptr) {
1016 return;
1017 }
1018 phi = u;
1019 continue;
1020 }
1021 }
1022 return;
1023 }
1024 if (phi != nullptr) {
1025 // Nothing in the loop before the store (next iteration)
1026 // must observe the stored value
1027 bool mem_ok = true;
1028 {
1029 ResourceMark rm;
1030 Unique_Node_List wq;
1031 wq.push(phi);
1032 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1033 Node *m = wq.at(next);
1034 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1035 Node* u = m->fast_out(i);
1036 if (u->is_Store() || u->is_Phi()) {
1037 if (u != n) {
1038 wq.push(u);
1039 mem_ok = (wq.size() <= 10);
1040 }
1041 } else {
1042 mem_ok = false;
1043 break;
1044 }
1045 }
1046 }
1047 }
1048 if (mem_ok) {
1049 // Move the store out of the loop if the LCA of all
1050 // users (except for the phi) is outside the loop.
1051 Node* hook = new Node(1);
1052 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1053 _igvn.rehash_node_delayed(phi);
1054 int count = phi->replace_edge(n, hook, &_igvn);
1055 assert(count > 0, "inconsistent phi");
1056
1057 // Compute latest point this store can go
1058 Node* lca = get_late_ctrl(n, get_ctrl(n));
1059 if (lca->is_OuterStripMinedLoop()) {
1060 lca = lca->in(LoopNode::EntryControl);
1061 }
1062 if (n_loop->is_member(get_loop(lca))) {
1063 // LCA is in the loop - bail out
1064 _igvn.replace_node(hook, n);
1065 return;
1066 }
1067 #ifdef ASSERT
1068 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1069 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1070 n_loop->_head->as_Loop()->verify_strip_mined(1);
1071 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1072 IdealLoopTree* outer_loop = get_loop(outer);
1073 assert(n_loop->_parent == outer_loop, "broken loop tree");
1074 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1075 }
1076 #endif
1077 lca = place_outside_loop(lca, n_loop);
1078 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1079 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");
1080
1081 // Move store out of the loop
1082 _igvn.replace_node(hook, n->in(MemNode::Memory));
1083 _igvn.replace_input_of(n, 0, lca);
1084 set_ctrl_and_loop(n, lca);
1085
1086 // Disconnect the phi now. An empty phi can confuse other
1087 // optimizations in this pass of loop opts..
1088 if (phi->in(LoopNode::LoopBackControl) == phi) {
1089 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1090 n_loop->_body.yank(phi);
1091 }
1092 }
1093 }
1094 }
1095 }
1096 }
1097 }
1098
1099 //------------------------------split_if_with_blocks_pre-----------------------
1100 // Do the real work in a non-recursive function. Data nodes want to be
1101 // cloned in the pre-order so they can feed each other nicely.
1102 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1103 // Cloning these guys is unlikely to win
1104 int n_op = n->Opcode();
1105 if (n_op == Op_MergeMem) {
1106 return n;
1107 }
1108 if (n->is_Proj()) {
1109 return n;
1110 }
1111 // Do not clone-up CmpFXXX variations, as these are always
1112 // followed by a CmpI
1113 if (n->is_Cmp()) {
1114 return n;
1115 }
1116 // Attempt to use a conditional move instead of a phi/branch
1117 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1118 Node *cmov = conditional_move( n );
1119 if (cmov) {
1120 return cmov;
1121 }
1122 }
1123 if (n->is_CFG() || n->is_LoadStore()) {
1124 return n;
1125 }
1126 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1127 if (!C->major_progress()) { // If chance of no more loop opts...
1128 _igvn._worklist.push(n); // maybe we'll remove them
1129 }
1130 return n;
1131 }
1132
1133 if (n->is_Con()) {
1134 return n; // No cloning for Con nodes
1135 }
1136
1137 Node *n_ctrl = get_ctrl(n);
1138 if (!n_ctrl) {
1139 return n; // Dead node
1140 }
1141
1142 Node* res = try_move_store_before_loop(n, n_ctrl);
1143 if (res != nullptr) {
1144 return n;
1145 }
1146
1147 // Attempt to remix address expressions for loop invariants
1148 Node *m = remix_address_expressions( n );
1149 if( m ) return m;
1150
1151 if (n_op == Op_AddI) {
1152 Node *nn = convert_add_to_muladd( n );
1153 if ( nn ) return nn;
1154 }
1155
1156 if (n->is_ConstraintCast()) {
1157 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1158 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1159 // Node control inputs don't necessarily agree with loop control info (due to
1160 // transformations happened in between), thus additional dominance check is needed
1161 // to keep loop info valid.
1162 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1163 _igvn.replace_node(n, dom_cast);
1164 return dom_cast;
1165 }
1166 }
1167
1168 // Determine if the Node has inputs from some local Phi.
1169 // Returns the block to clone thru.
1170 Node *n_blk = has_local_phi_input( n );
1171 if( !n_blk ) return n;
1172
1173 // Do not clone the trip counter through on a CountedLoop
1174 // (messes up the canonical shape).
1175 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1176 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1177 return n;
1178 }
1179 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1180 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1181 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1182 return n;
1183 }
1184
1185 // Check for having no control input; not pinned. Allow
1186 // dominating control.
1187 if (n->in(0)) {
1188 Node *dom = idom(n_blk);
1189 if (dom_lca(n->in(0), dom) != n->in(0)) {
1190 return n;
1191 }
1192 }
1193 // Policy: when is it profitable. You must get more wins than
1194 // policy before it is considered profitable. Policy is usually 0,
1195 // so 1 win is considered profitable. Big merges will require big
1196 // cloning, so get a larger policy.
1197 int policy = n_blk->req() >> 2;
1198
1199 // If the loop is a candidate for range check elimination,
1200 // delay splitting through it's phi until a later loop optimization
1201 if (n_blk->is_BaseCountedLoop()) {
1202 IdealLoopTree *lp = get_loop(n_blk);
1203 if (lp && lp->_rce_candidate) {
1204 return n;
1205 }
1206 }
1207
1208 if (must_throttle_split_if()) return n;
1209
1210 // Split 'n' through the merge point if it is profitable
1211 Node *phi = split_thru_phi( n, n_blk, policy );
1212 if (!phi) return n;
1213
1214 // Found a Phi to split thru!
1215 // Replace 'n' with the new phi
1216 _igvn.replace_node( n, phi );
1217 // Moved a load around the loop, 'en-registering' something.
1218 if (n_blk->is_Loop() && n->is_Load() &&
1219 !phi->in(LoopNode::LoopBackControl)->is_Load())
1220 C->set_major_progress();
1221
1222 return phi;
1223 }
1224
1225 static bool merge_point_too_heavy(Compile* C, Node* region) {
1226 // Bail out if the region and its phis have too many users.
1227 int weight = 0;
1228 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1229 weight += region->fast_out(i)->outcnt();
1230 }
1231 int nodes_left = C->max_node_limit() - C->live_nodes();
1232 if (weight * 8 > nodes_left) {
1233 if (PrintOpto) {
1234 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1235 }
1236 return true;
1237 } else {
1238 return false;
1239 }
1240 }
1241
1242 static bool merge_point_safe(Node* region) {
1243 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1244 // having a PhiNode input. This sidesteps the dangerous case where the split
1245 // ConvI2LNode may become TOP if the input Value() does not
1246 // overlap the ConvI2L range, leaving a node which may not dominate its
1247 // uses.
1248 // A better fix for this problem can be found in the BugTraq entry, but
1249 // expediency for Mantis demands this hack.
1250 #ifdef _LP64
1251 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1252 Node* n = region->fast_out(i);
1253 if (n->is_Phi()) {
1254 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1255 Node* m = n->fast_out(j);
1256 if (m->Opcode() == Op_ConvI2L)
1257 return false;
1258 if (m->is_CastII()) {
1259 return false;
1260 }
1261 }
1262 }
1263 }
1264 #endif
1265 return true;
1266 }
1267
1268
1269 //------------------------------place_outside_loop---------------------------------
1270 // Place some computation outside of this loop on the path to the use passed as argument
1271 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1272 Node* head = loop->_head;
1273 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1274 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1275 loop = loop->_parent;
1276 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1277 }
1278
1279 // Pick control right outside the loop
1280 for (;;) {
1281 Node* dom = idom(useblock);
1282 if (loop->is_member(get_loop(dom))) {
1283 break;
1284 }
1285 useblock = dom;
1286 }
1287 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1288 return useblock;
1289 }
1290
1291
1292 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1293 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1294 return false;
1295 }
1296 if (!n->in(0)->is_Region()) {
1297 return false;
1298 }
1299
1300 Node* region = n->in(0);
1301 Node* dom = idom(region);
1302 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1303 return false;
1304 }
1305 IfNode* dom_if = dom->as_If();
1306 Node* proj_true = dom_if->proj_out(1);
1307 Node* proj_false = dom_if->proj_out(0);
1308
1309 for (uint i = 1; i < region->req(); i++) {
1310 if (is_dominator(proj_true, region->in(i))) {
1311 continue;
1312 }
1313 if (is_dominator(proj_false, region->in(i))) {
1314 continue;
1315 }
1316 return false;
1317 }
1318
1319 return true;
1320 }
1321
1322
1323 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1324 if (must_throttle_split_if()) {
1325 return false;
1326 }
1327
1328 // Do not do 'split-if' if irreducible loops are present.
1329 if (_has_irreducible_loops) {
1330 return false;
1331 }
1332
1333 if (merge_point_too_heavy(C, n_ctrl)) {
1334 return false;
1335 }
1336
1337 // Do not do 'split-if' if some paths are dead. First do dead code
1338 // elimination and then see if its still profitable.
1339 for (uint i = 1; i < n_ctrl->req(); i++) {
1340 if (n_ctrl->in(i) == C->top()) {
1341 return false;
1342 }
1343 }
1344
1345 // If trying to do a 'Split-If' at the loop head, it is only
1346 // profitable if the cmp folds up on BOTH paths. Otherwise we
1347 // risk peeling a loop forever.
1348
1349 // CNC - Disabled for now. Requires careful handling of loop
1350 // body selection for the cloned code. Also, make sure we check
1351 // for any input path not being in the same loop as n_ctrl. For
1352 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1353 // because the alternative loop entry points won't be converted
1354 // into LoopNodes.
1355 IdealLoopTree *n_loop = get_loop(n_ctrl);
1356 for (uint j = 1; j < n_ctrl->req(); j++) {
1357 if (get_loop(n_ctrl->in(j)) != n_loop) {
1358 return false;
1359 }
1360 }
1361
1362 // Check for safety of the merge point.
1363 if (!merge_point_safe(n_ctrl)) {
1364 return false;
1365 }
1366
1367 return true;
1368 }
1369
1370 // Detect if the node is the inner strip-mined loop
1371 // Return: null if it's not the case, or the exit of outer strip-mined loop
1372 static Node* is_inner_of_stripmined_loop(const Node* out) {
1373 Node* out_le = nullptr;
1374
1375 if (out->is_CountedLoopEnd()) {
1376 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1377
1378 if (loop != nullptr && loop->is_strip_mined()) {
1379 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1380 }
1381 }
1382
1383 return out_le;
1384 }
1385
1386 //------------------------------split_if_with_blocks_post----------------------
1387 // Do the real work in a non-recursive function. CFG hackery wants to be
1388 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1389 // info.
1390 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1391
1392 // Cloning Cmp through Phi's involves the split-if transform.
1393 // FastLock is not used by an If
1394 if (n->is_Cmp() && !n->is_FastLock()) {
1395 Node *n_ctrl = get_ctrl(n);
1396 // Determine if the Node has inputs from some local Phi.
1397 // Returns the block to clone thru.
1398 Node *n_blk = has_local_phi_input(n);
1399 if (n_blk != n_ctrl) {
1400 return;
1401 }
1402
1403 if (!can_split_if(n_ctrl)) {
1404 return;
1405 }
1406
1407 if (n->outcnt() != 1) {
1408 return; // Multiple bool's from 1 compare?
1409 }
1410 Node *bol = n->unique_out();
1411 assert(bol->is_Bool(), "expect a bool here");
1412 if (bol->outcnt() != 1) {
1413 return;// Multiple branches from 1 compare?
1414 }
1415 Node *iff = bol->unique_out();
1416
1417 // Check some safety conditions
1418 if (iff->is_If()) { // Classic split-if?
1419 if (iff->in(0) != n_ctrl) {
1420 return; // Compare must be in same blk as if
1421 }
1422 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1423 // Can't split CMove with different control.
1424 if (get_ctrl(iff) != n_ctrl) {
1425 return;
1426 }
1427 if (get_ctrl(iff->in(2)) == n_ctrl ||
1428 get_ctrl(iff->in(3)) == n_ctrl) {
1429 return; // Inputs not yet split-up
1430 }
1431 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1432 return; // Loop-invar test gates loop-varying CMOVE
1433 }
1434 } else {
1435 return; // some other kind of node, such as an Allocate
1436 }
1437
1438 // When is split-if profitable? Every 'win' on means some control flow
1439 // goes dead, so it's almost always a win.
1440 int policy = 0;
1441 // Split compare 'n' through the merge point if it is profitable
1442 Node *phi = split_thru_phi( n, n_ctrl, policy);
1443 if (!phi) {
1444 return;
1445 }
1446
1447 // Found a Phi to split thru!
1448 // Replace 'n' with the new phi
1449 _igvn.replace_node(n, phi);
1450
1451 // Now split the bool up thru the phi
1452 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1453 guarantee(bolphi != nullptr, "null boolean phi node");
1454
1455 _igvn.replace_node(bol, bolphi);
1456 assert(iff->in(1) == bolphi, "");
1457
1458 if (bolphi->Value(&_igvn)->singleton()) {
1459 return;
1460 }
1461
1462 // Conditional-move? Must split up now
1463 if (!iff->is_If()) {
1464 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1465 _igvn.replace_node(iff, cmovphi);
1466 return;
1467 }
1468
1469 // Now split the IF
1470 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1471 if (TraceLoopOpts) {
1472 tty->print_cr("Split-If");
1473 }
1474 do_split_if(iff);
1475 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1476 return;
1477 }
1478
1479 // Two identical ifs back to back can be merged
1480 if (try_merge_identical_ifs(n)) {
1481 return;
1482 }
1483
1484 // Check for an IF ready to split; one that has its
1485 // condition codes input coming from a Phi at the block start.
1486 int n_op = n->Opcode();
1487
1488 // Check for an IF being dominated by another IF same test
1489 if (n_op == Op_If ||
1490 n_op == Op_RangeCheck) {
1491 Node *bol = n->in(1);
1492 uint max = bol->outcnt();
1493 // Check for same test used more than once?
1494 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1495 // Search up IDOMs to see if this IF is dominated.
1496 Node* cmp = bol->in(1);
1497 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1498
1499 // Now search up IDOMs till cutoff, looking for a dominating test
1500 Node *prevdom = n;
1501 Node *dom = idom(prevdom);
1502 while (dom != cutoff) {
1503 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1504 safe_for_if_replacement(dom)) {
1505 // It's invalid to move control dependent data nodes in the inner
1506 // strip-mined loop, because:
1507 // 1) break validation of LoopNode::verify_strip_mined()
1508 // 2) move code with side-effect in strip-mined loop
1509 // Move to the exit of outer strip-mined loop in that case.
1510 Node* out_le = is_inner_of_stripmined_loop(dom);
1511 if (out_le != nullptr) {
1512 prevdom = out_le;
1513 }
1514 // Replace the dominated test with an obvious true or false.
1515 // Place it on the IGVN worklist for later cleanup.
1516 C->set_major_progress();
1517 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1518 // to prevent an array load from floating above its range check. There are three cases:
1519 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1520 // all its array accesses at that point.
1521 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1522 // accesses would start to float, since we don't pin at that point.
1523 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1524 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1525 prevdom->in(0)->Opcode() != Op_RangeCheck;
1526 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1527 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1528 return;
1529 }
1530 prevdom = dom;
1531 dom = idom(prevdom);
1532 }
1533 }
1534 }
1535
1536 try_sink_out_of_loop(n);
1537 if (C->failing()) {
1538 return;
1539 }
1540
1541 try_move_store_after_loop(n);
1542 }
1543
1544 // Transform:
1545 //
1546 // if (some_condition) {
1547 // // body 1
1548 // } else {
1549 // // body 2
1550 // }
1551 // if (some_condition) {
1552 // // body 3
1553 // } else {
1554 // // body 4
1555 // }
1556 //
1557 // into:
1558 //
1559 //
1560 // if (some_condition) {
1561 // // body 1
1562 // // body 3
1563 // } else {
1564 // // body 2
1565 // // body 4
1566 // }
1567 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1568 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1569 Node *n_ctrl = n->in(0);
1570 IfNode* dom_if = idom(n_ctrl)->as_If();
1571 if (n->in(1) != dom_if->in(1)) {
1572 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1573 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1574 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1575 _igvn.replace_input_of(n, 1, dom_if->in(1));
1576 }
1577 ProjNode* dom_proj_true = dom_if->proj_out(1);
1578 ProjNode* dom_proj_false = dom_if->proj_out(0);
1579
1580 // Now split the IF
1581 RegionNode* new_false_region;
1582 RegionNode* new_true_region;
1583 do_split_if(n, &new_false_region, &new_true_region);
1584 assert(new_false_region->req() == new_true_region->req(), "");
1585 #ifdef ASSERT
1586 for (uint i = 1; i < new_false_region->req(); ++i) {
1587 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1588 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");
1589 }
1590 #endif
1591 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1592
1593 // We now have:
1594 // if (some_condition) {
1595 // // body 1
1596 // if (some_condition) {
1597 // body3: // new_true_region
1598 // // body3
1599 // } else {
1600 // goto body4;
1601 // }
1602 // } else {
1603 // // body 2
1604 // if (some_condition) {
1605 // goto body3;
1606 // } else {
1607 // body4: // new_false_region
1608 // // body4;
1609 // }
1610 // }
1611 //
1612
1613 // clone pinned nodes thru the resulting regions
1614 push_pinned_nodes_thru_region(dom_if, new_true_region);
1615 push_pinned_nodes_thru_region(dom_if, new_false_region);
1616
1617 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1618 // 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
1619 // unrelated control dependency.
1620 for (uint i = 1; i < new_false_region->req(); i++) {
1621 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1622 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1623 } else {
1624 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1625 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1626 }
1627 }
1628 return true;
1629 }
1630 return false;
1631 }
1632
1633 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1634 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1635 Node* u = region->out(i);
1636 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1637 continue;
1638 }
1639 assert(u->in(0) == region, "not a control dependent node?");
1640 uint j = 1;
1641 for (; j < u->req(); ++j) {
1642 Node* in = u->in(j);
1643 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1644 break;
1645 }
1646 }
1647 if (j == u->req()) {
1648 Node *phi = PhiNode::make_blank(region, u);
1649 for (uint k = 1; k < region->req(); ++k) {
1650 Node* clone = u->clone();
1651 clone->set_req(0, region->in(k));
1652 register_new_node(clone, region->in(k));
1653 phi->init_req(k, clone);
1654 }
1655 register_new_node(phi, region);
1656 _igvn.replace_node(u, phi);
1657 --i;
1658 }
1659 }
1660 }
1661
1662 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1663 if (!dom->is_CountedLoopEnd()) {
1664 return true;
1665 }
1666 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1667 CountedLoopNode* cl = le->loopnode();
1668 if (cl == nullptr) {
1669 return true;
1670 }
1671 if (!cl->is_main_loop()) {
1672 return true;
1673 }
1674 if (cl->is_canonical_loop_entry() == nullptr) {
1675 return true;
1676 }
1677 // Further unrolling is possible so loop exit condition might change
1678 return false;
1679 }
1680
1681 // See if a shared loop-varying computation has no loop-varying uses.
1682 // Happens if something is only used for JVM state in uncommon trap exits,
1683 // like various versions of induction variable+offset. Clone the
1684 // computation per usage to allow it to sink out of the loop.
1685 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1686 bool is_raw_to_oop_cast = n->is_ConstraintCast() &&
1687 n->in(1)->bottom_type()->isa_rawptr() &&
1688 !n->bottom_type()->isa_rawptr();
1689
1690 if (has_ctrl(n) &&
1691 !n->is_Phi() &&
1692 !n->is_Bool() &&
1693 !n->is_Proj() &&
1694 !n->is_MergeMem() &&
1695 !n->is_CMove() &&
1696 !n->is_OpaqueNotNull() &&
1697 !n->is_OpaqueInitializedAssertionPredicate() &&
1698 !n->is_OpaqueTemplateAssertionPredicate() &&
1699 !is_raw_to_oop_cast && // don't extend live ranges of raw oops
1700 (KillPathsReachableByDeadTypeNode || !n->is_Type())
1701 ) {
1702 Node *n_ctrl = get_ctrl(n);
1703 IdealLoopTree *n_loop = get_loop(n_ctrl);
1704
1705 if (n->in(0) != nullptr) {
1706 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1707 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1708 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1709 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1710 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1711 Node* maybe_pinned_n = n;
1712 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1713 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) {
1714 if (n->depends_only_on_test()) {
1715 Node* pinned_clone = n->pin_array_access_node();
1716 if (pinned_clone != nullptr) {
1717 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1718 // range check for that access. If that condition is replaced by an identical dominating one, then an
1719 // unpinned load would risk floating above its range check.
1720 register_new_node(pinned_clone, n_ctrl);
1721 maybe_pinned_n = pinned_clone;
1722 _igvn.replace_node(n, pinned_clone);
1723 }
1724 }
1725 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1726 }
1727 }
1728 }
1729 if (n_loop != _ltree_root && n->outcnt() > 1) {
1730 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1731 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1732 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1733 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1734 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1735 if (n->is_Store() || n->is_LoadStore()) {
1736 assert(false, "no node with a side effect");
1737 C->record_failure("no node with a side effect");
1738 return;
1739 }
1740 Node* outer_loop_clone = nullptr;
1741 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1742 Node* u = n->last_out(j); // Clone private computation per use
1743 _igvn.rehash_node_delayed(u);
1744 Node* x = nullptr;
1745 if (n->depends_only_on_test()) {
1746 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1747 // range check for that access. If that condition is replaced by an identical dominating one, then an
1748 // unpinned load would risk floating above its range check.
1749 x = n->pin_array_access_node();
1750 }
1751 if (x == nullptr) {
1752 x = n->clone();
1753 }
1754 Node* x_ctrl = nullptr;
1755 if (u->is_Phi()) {
1756 // Replace all uses of normal nodes. Replace Phi uses
1757 // individually, so the separate Nodes can sink down
1758 // different paths.
1759 uint k = 1;
1760 while (u->in(k) != n) k++;
1761 u->set_req(k, x);
1762 // x goes next to Phi input path
1763 x_ctrl = u->in(0)->in(k);
1764 // Find control for 'x' next to use but not inside inner loops.
1765 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1766 --j;
1767 } else { // Normal use
1768 if (has_ctrl(u)) {
1769 x_ctrl = get_ctrl(u);
1770 } else {
1771 x_ctrl = u->in(0);
1772 }
1773 // Find control for 'x' next to use but not inside inner loops.
1774 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1775 // Replace all uses
1776 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1777 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1778 // anymore now that we're going to pin n as well
1779 _igvn.replace_node(u, x);
1780 --j;
1781 } else {
1782 int nb = u->replace_edge(n, x, &_igvn);
1783 j -= nb;
1784 }
1785 }
1786
1787 if (n->is_Load()) {
1788 // For loads, add a control edge to a CFG node outside of the loop
1789 // to force them to not combine and return back inside the loop
1790 // during GVN optimization (4641526).
1791 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1792
1793 IdealLoopTree* x_loop = get_loop(x_ctrl);
1794 Node* x_head = x_loop->_head;
1795 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1796 // Do not add duplicate LoadNodes to the outer strip mined loop
1797 if (outer_loop_clone != nullptr) {
1798 _igvn.replace_node(x, outer_loop_clone);
1799 continue;
1800 }
1801 outer_loop_clone = x;
1802 }
1803 x->set_req(0, x_ctrl);
1804 } else if (n->in(0) != nullptr){
1805 x->set_req(0, x_ctrl);
1806 }
1807 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1808 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1809 register_new_node(x, x_ctrl);
1810
1811 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1812 // All AddP nodes must keep the same base after sinking so:
1813 // 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,
1814 // their bases remain the same.
1815 // (see 2- below)
1816 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1817 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1818 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1819 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
1820 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
1821 assert(!x->is_Load(), "load should be pinned");
1822 // Use a cast node to pin clone out of loop
1823 Node* cast = nullptr;
1824 for (uint k = 0; k < x->req(); k++) {
1825 Node* in = x->in(k);
1826 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
1827 const Type* in_t = _igvn.type(in);
1828 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
1829 ConstraintCastNode::UnconditionalDependency, nullptr);
1830 }
1831 if (cast != nullptr) {
1832 Node* prev = _igvn.hash_find_insert(cast);
1833 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
1834 cast->destruct(&_igvn);
1835 cast = prev;
1836 } else {
1837 register_new_node(cast, x_ctrl);
1838 }
1839 x->replace_edge(in, cast);
1840 // Chain of AddP nodes:
1841 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
1842 if (x->is_AddP() && k == AddPNode::Base) {
1843 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
1844 }
1845 break;
1846 }
1847 }
1848 assert(cast != nullptr, "must have added a cast to pin the node");
1849 }
1850 }
1851 _igvn.remove_dead_node(n);
1852 }
1853 _dom_lca_tags_round = 0;
1854 }
1855 }
1856 }
1857
1858 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
1859 ResourceMark rm;
1860 Node_List wq;
1861 wq.push(x);
1862 while (wq.size() != 0) {
1863 Node* n = wq.pop();
1864 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1865 Node* u = n->fast_out(i);
1866 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
1867 _igvn.replace_input_of(u, AddPNode::Base, new_base);
1868 wq.push(u);
1869 }
1870 }
1871 }
1872 }
1873
1874 // Compute the early control of a node by following its inputs until we reach
1875 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
1876 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
1877 Node* early_ctrl = nullptr;
1878 ResourceMark rm;
1879 Unique_Node_List wq;
1880 wq.push(n);
1881 for (uint i = 0; i < wq.size(); i++) {
1882 Node* m = wq.at(i);
1883 Node* c = nullptr;
1884 if (m->is_CFG()) {
1885 c = m;
1886 } else if (m->pinned()) {
1887 c = m->in(0);
1888 } else {
1889 for (uint j = 0; j < m->req(); j++) {
1890 Node* in = m->in(j);
1891 if (in != nullptr) {
1892 wq.push(in);
1893 }
1894 }
1895 }
1896 if (c != nullptr) {
1897 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
1898 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
1899 early_ctrl = c;
1900 }
1901 }
1902 }
1903 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
1904 return early_ctrl;
1905 }
1906
1907 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
1908 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1909 Node* u = n->fast_out(i);
1910 if (u->is_Opaque1()) {
1911 return false; // Found loop limit, bugfix for 4677003
1912 }
1913 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
1914 // get_late_ctrl_with_anti_dep() use their own tag
1915 _dom_lca_tags_round++;
1916 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
1917
1918 if (u->is_Phi()) {
1919 for (uint j = 1; j < u->req(); ++j) {
1920 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
1921 return false;
1922 }
1923 }
1924 } else {
1925 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
1926 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
1927 return false;
1928 }
1929 }
1930 }
1931 return true;
1932 }
1933
1934 // 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
1935 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
1936 // test of the pre loop above the point in the graph where it's pinned. This results in a broken graph. One way to avoid
1937 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code
1938 // is generated for the main loop.
1939 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) {
1940 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
1941 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
1942 if (is_dominator(pre_loop->loopexit(), ctrl)) {
1943 return true;
1944 }
1945 }
1946 return false;
1947 }
1948
1949 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
1950 if (n->is_Load()) {
1951 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
1952 }
1953 IdealLoopTree *u_loop = get_loop(ctrl);
1954 if (u_loop == n_loop) {
1955 return false; // Found loop-varying use
1956 }
1957 if (n_loop->is_member(u_loop)) {
1958 return false; // Found use in inner loop
1959 }
1960 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) {
1961 return false;
1962 }
1963 return true;
1964 }
1965
1966 //------------------------------split_if_with_blocks---------------------------
1967 // Check for aggressive application of 'split-if' optimization,
1968 // using basic block level info.
1969 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1970 Node* root = C->root();
1971 visited.set(root->_idx); // first, mark root as visited
1972 // Do pre-visit work for root
1973 Node* n = split_if_with_blocks_pre(root);
1974 uint cnt = n->outcnt();
1975 uint i = 0;
1976
1977 while (true) {
1978 // Visit all children
1979 if (i < cnt) {
1980 Node* use = n->raw_out(i);
1981 ++i;
1982 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1983 // Now do pre-visit work for this use
1984 use = split_if_with_blocks_pre(use);
1985 nstack.push(n, i); // Save parent and next use's index.
1986 n = use; // Process all children of current use.
1987 cnt = use->outcnt();
1988 i = 0;
1989 }
1990 }
1991 else {
1992 // All of n's children have been processed, complete post-processing.
1993 if (cnt != 0 && !n->is_Con()) {
1994 assert(has_node(n), "no dead nodes");
1995 split_if_with_blocks_post(n);
1996 if (C->failing()) {
1997 return;
1998 }
1999 }
2000 if (must_throttle_split_if()) {
2001 nstack.clear();
2002 }
2003 if (nstack.is_empty()) {
2004 // Finished all nodes on stack.
2005 break;
2006 }
2007 // Get saved parent node and next use's index. Visit the rest of uses.
2008 n = nstack.node();
2009 cnt = n->outcnt();
2010 i = nstack.index();
2011 nstack.pop();
2012 }
2013 }
2014 }
2015
2016
2017 //=============================================================================
2018 //
2019 // C L O N E A L O O P B O D Y
2020 //
2021
2022 //------------------------------clone_iff--------------------------------------
2023 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2024 // "Nearly" because all Nodes have been cloned from the original in the loop,
2025 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2026 // through the Phi recursively, and return a Bool.
2027 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2028
2029 // Convert this Phi into a Phi merging Bools
2030 uint i;
2031 for (i = 1; i < phi->req(); i++) {
2032 Node* b = phi->in(i);
2033 if (b->is_Phi()) {
2034 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2035 } else {
2036 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2037 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2038 }
2039 }
2040 Node* n = phi->in(1);
2041 Node* sample_opaque = nullptr;
2042 Node *sample_bool = nullptr;
2043 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2044 sample_opaque = n;
2045 sample_bool = n->in(1);
2046 assert(sample_bool->is_Bool(), "wrong type");
2047 } else {
2048 sample_bool = n;
2049 }
2050 Node *sample_cmp = sample_bool->in(1);
2051
2052 // Make Phis to merge the Cmp's inputs.
2053 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2054 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2055 for (i = 1; i < phi->req(); i++) {
2056 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2057 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2058 phi1->set_req(i, n1);
2059 phi2->set_req(i, n2);
2060 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2061 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2062 }
2063 // See if these Phis have been made before.
2064 // Register with optimizer
2065 Node *hit1 = _igvn.hash_find_insert(phi1);
2066 if (hit1) { // Hit, toss just made Phi
2067 _igvn.remove_dead_node(phi1); // Remove new phi
2068 assert(hit1->is_Phi(), "" );
2069 phi1 = (PhiNode*)hit1; // Use existing phi
2070 } else { // Miss
2071 _igvn.register_new_node_with_optimizer(phi1);
2072 }
2073 Node *hit2 = _igvn.hash_find_insert(phi2);
2074 if (hit2) { // Hit, toss just made Phi
2075 _igvn.remove_dead_node(phi2); // Remove new phi
2076 assert(hit2->is_Phi(), "" );
2077 phi2 = (PhiNode*)hit2; // Use existing phi
2078 } else { // Miss
2079 _igvn.register_new_node_with_optimizer(phi2);
2080 }
2081 // Register Phis with loop/block info
2082 set_ctrl(phi1, phi->in(0));
2083 set_ctrl(phi2, phi->in(0));
2084 // Make a new Cmp
2085 Node *cmp = sample_cmp->clone();
2086 cmp->set_req(1, phi1);
2087 cmp->set_req(2, phi2);
2088 _igvn.register_new_node_with_optimizer(cmp);
2089 set_ctrl(cmp, phi->in(0));
2090
2091 // Make a new Bool
2092 Node *b = sample_bool->clone();
2093 b->set_req(1,cmp);
2094 _igvn.register_new_node_with_optimizer(b);
2095 set_ctrl(b, phi->in(0));
2096
2097 if (sample_opaque != nullptr) {
2098 Node* opaque = sample_opaque->clone();
2099 opaque->set_req(1, b);
2100 _igvn.register_new_node_with_optimizer(opaque);
2101 set_ctrl(opaque, phi->in(0));
2102 return opaque;
2103 }
2104
2105 assert(b->is_Bool(), "");
2106 return b;
2107 }
2108
2109 //------------------------------clone_bool-------------------------------------
2110 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2111 // "Nearly" because all Nodes have been cloned from the original in the loop,
2112 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2113 // through the Phi recursively, and return a Bool.
2114 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2115 uint i;
2116 // Convert this Phi into a Phi merging Bools
2117 for( i = 1; i < phi->req(); i++ ) {
2118 Node *b = phi->in(i);
2119 if( b->is_Phi() ) {
2120 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2121 } else {
2122 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2123 }
2124 }
2125
2126 Node *sample_cmp = phi->in(1);
2127
2128 // Make Phis to merge the Cmp's inputs.
2129 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2130 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2131 for( uint j = 1; j < phi->req(); j++ ) {
2132 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2133 Node *n1, *n2;
2134 if( cmp_top->is_Cmp() ) {
2135 n1 = cmp_top->in(1);
2136 n2 = cmp_top->in(2);
2137 } else {
2138 n1 = n2 = cmp_top;
2139 }
2140 phi1->set_req( j, n1 );
2141 phi2->set_req( j, n2 );
2142 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2143 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2144 }
2145
2146 // See if these Phis have been made before.
2147 // Register with optimizer
2148 Node *hit1 = _igvn.hash_find_insert(phi1);
2149 if( hit1 ) { // Hit, toss just made Phi
2150 _igvn.remove_dead_node(phi1); // Remove new phi
2151 assert( hit1->is_Phi(), "" );
2152 phi1 = (PhiNode*)hit1; // Use existing phi
2153 } else { // Miss
2154 _igvn.register_new_node_with_optimizer(phi1);
2155 }
2156 Node *hit2 = _igvn.hash_find_insert(phi2);
2157 if( hit2 ) { // Hit, toss just made Phi
2158 _igvn.remove_dead_node(phi2); // Remove new phi
2159 assert( hit2->is_Phi(), "" );
2160 phi2 = (PhiNode*)hit2; // Use existing phi
2161 } else { // Miss
2162 _igvn.register_new_node_with_optimizer(phi2);
2163 }
2164 // Register Phis with loop/block info
2165 set_ctrl(phi1, phi->in(0));
2166 set_ctrl(phi2, phi->in(0));
2167 // Make a new Cmp
2168 Node *cmp = sample_cmp->clone();
2169 cmp->set_req( 1, phi1 );
2170 cmp->set_req( 2, phi2 );
2171 _igvn.register_new_node_with_optimizer(cmp);
2172 set_ctrl(cmp, phi->in(0));
2173
2174 assert( cmp->is_Cmp(), "" );
2175 return (CmpNode*)cmp;
2176 }
2177
2178 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2179 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2180 Node_List*& split_if_set, Node_List*& split_bool_set,
2181 Node_List*& split_cex_set, Node_List& worklist,
2182 uint new_counter, CloneLoopMode mode) {
2183 Node* nnn = old_new[old->_idx];
2184 // Copy uses to a worklist, so I can munge the def-use info
2185 // with impunity.
2186 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2187 worklist.push(old->fast_out(j));
2188
2189 while( worklist.size() ) {
2190 Node *use = worklist.pop();
2191 if (!has_node(use)) continue; // Ignore dead nodes
2192 if (use->in(0) == C->top()) continue;
2193 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2194 // Check for data-use outside of loop - at least one of OLD or USE
2195 // must not be a CFG node.
2196 #ifdef ASSERT
2197 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2198 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2199 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2200 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2201 }
2202 #endif
2203 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2204
2205 // If the Data use is an IF, that means we have an IF outside the
2206 // loop that is switching on a condition that is set inside the
2207 // loop. Happens if people set a loop-exit flag; then test the flag
2208 // in the loop to break the loop, then test is again outside the
2209 // loop to determine which way the loop exited.
2210 //
2211 // For several uses we need to make sure that there is no phi between,
2212 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2213 // to avoid such a phi in between.
2214 // For example, it is unexpected that there is a Phi between an
2215 // AllocateArray node and its ValidLengthTest input that could cause
2216 // split if to break.
2217 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2218 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2219 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2220 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2221 // Since this code is highly unlikely, we lazily build the worklist
2222 // of such Nodes to go split.
2223 if (!split_if_set) {
2224 split_if_set = new Node_List();
2225 }
2226 split_if_set->push(use);
2227 }
2228 if (use->is_Bool()) {
2229 if (!split_bool_set) {
2230 split_bool_set = new Node_List();
2231 }
2232 split_bool_set->push(use);
2233 }
2234 if (use->Opcode() == Op_CreateEx) {
2235 if (!split_cex_set) {
2236 split_cex_set = new Node_List();
2237 }
2238 split_cex_set->push(use);
2239 }
2240
2241
2242 // Get "block" use is in
2243 uint idx = 0;
2244 while( use->in(idx) != old ) idx++;
2245 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2246 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2247 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2248 ? prev->in(2)
2249 : idom(prev);
2250 if( use->is_Phi() ) // Phi use is in prior block
2251 cfg = prev->in(idx); // NOT in block of Phi itself
2252 if (cfg->is_top()) { // Use is dead?
2253 _igvn.replace_input_of(use, idx, C->top());
2254 continue;
2255 }
2256
2257 // If use is referenced through control edge... (idx == 0)
2258 if (mode == IgnoreStripMined && idx == 0) {
2259 LoopNode *head = loop->_head->as_Loop();
2260 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2261 // That node is outside the inner loop, leave it outside the
2262 // outer loop as well to not confuse verification code.
2263 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2264 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2265 continue;
2266 }
2267 }
2268
2269 while(!outer_loop->is_member(get_loop(cfg))) {
2270 prev = cfg;
2271 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2272 }
2273 // If the use occurs after merging several exits from the loop, then
2274 // old value must have dominated all those exits. Since the same old
2275 // value was used on all those exits we did not need a Phi at this
2276 // merge point. NOW we do need a Phi here. Each loop exit value
2277 // is now merged with the peeled body exit; each exit gets its own
2278 // private Phi and those Phis need to be merged here.
2279 Node *phi;
2280 if( prev->is_Region() ) {
2281 if( idx == 0 ) { // Updating control edge?
2282 phi = prev; // Just use existing control
2283 } else { // Else need a new Phi
2284 phi = PhiNode::make( prev, old );
2285 // Now recursively fix up the new uses of old!
2286 for( uint i = 1; i < prev->req(); i++ ) {
2287 worklist.push(phi); // Onto worklist once for each 'old' input
2288 }
2289 }
2290 } else {
2291 // Get new RegionNode merging old and new loop exits
2292 prev = old_new[prev->_idx];
2293 assert( prev, "just made this in step 7" );
2294 if( idx == 0) { // Updating control edge?
2295 phi = prev; // Just use existing control
2296 } else { // Else need a new Phi
2297 // Make a new Phi merging data values properly
2298 phi = PhiNode::make( prev, old );
2299 phi->set_req( 1, nnn );
2300 }
2301 }
2302 // If inserting a new Phi, check for prior hits
2303 if( idx != 0 ) {
2304 Node *hit = _igvn.hash_find_insert(phi);
2305 if( hit == nullptr ) {
2306 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2307 } else { // or
2308 // Remove the new phi from the graph and use the hit
2309 _igvn.remove_dead_node(phi);
2310 phi = hit; // Use existing phi
2311 }
2312 set_ctrl(phi, prev);
2313 }
2314 // Make 'use' use the Phi instead of the old loop body exit value
2315 assert(use->in(idx) == old, "old is still input of use");
2316 // We notify all uses of old, including use, and the indirect uses,
2317 // that may now be optimized because we have replaced old with phi.
2318 _igvn.add_users_to_worklist(old);
2319 if (idx == 0 &&
2320 use->depends_only_on_test()) {
2321 Node* pinned_clone = use->pin_array_access_node();
2322 if (pinned_clone != nullptr) {
2323 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2324 // into the region is left, an array load could become dependent on a condition that's not a range check for
2325 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2326 // floating above its range check.
2327 pinned_clone->set_req(0, phi);
2328 register_new_node_with_ctrl_of(pinned_clone, use);
2329 _igvn.replace_node(use, pinned_clone);
2330 continue;
2331 }
2332 }
2333 _igvn.replace_input_of(use, idx, phi);
2334 if( use->_idx >= new_counter ) { // If updating new phis
2335 // Not needed for correctness, but prevents a weak assert
2336 // in AddPNode from tripping (when we end up with different
2337 // base & derived Phis that will become the same after
2338 // IGVN does CSE).
2339 Node *hit = _igvn.hash_find_insert(use);
2340 if( hit ) // Go ahead and re-hash for hits.
2341 _igvn.replace_node( use, hit );
2342 }
2343 }
2344 }
2345 }
2346
2347 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2348 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2349 bool check_old_new) {
2350 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2351 Node* u = n->fast_out(j);
2352 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2353 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2354 Node* c = phase->get_ctrl(u);
2355 IdealLoopTree* u_loop = phase->get_loop(c);
2356 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2357 if (!loop->is_member(u_loop)) {
2358 if (outer_loop->is_member(u_loop)) {
2359 wq.push(u);
2360 } else {
2361 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2362 // the outer loop too
2363 Node* u_c = u->in(0);
2364 if (u_c != nullptr) {
2365 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2366 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2367 wq.push(u);
2368 }
2369 }
2370 }
2371 }
2372 }
2373 }
2374 }
2375
2376 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2377 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2378 Node_List& extra_data_nodes) {
2379 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2380 CountedLoopNode* cl = head->as_CountedLoop();
2381 Node* l = cl->outer_loop();
2382 Node* tail = cl->outer_loop_tail();
2383 IfNode* le = cl->outer_loop_end();
2384 Node* sfpt = cl->outer_safepoint();
2385 CountedLoopEndNode* cle = cl->loopexit();
2386 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2387 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2388 Node* cle_out = cle->proj_out(false);
2389
2390 Node* new_sfpt = nullptr;
2391 Node* new_cle_out = cle_out->clone();
2392 old_new.map(cle_out->_idx, new_cle_out);
2393 if (mode == CloneIncludesStripMined) {
2394 // clone outer loop body
2395 Node* new_l = l->clone();
2396 Node* new_tail = tail->clone();
2397 IfNode* new_le = le->clone()->as_If();
2398 new_sfpt = sfpt->clone();
2399
2400 set_loop(new_l, outer_loop->_parent);
2401 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2402 set_loop(new_cle_out, outer_loop->_parent);
2403 set_idom(new_cle_out, new_cle, dd);
2404 set_loop(new_sfpt, outer_loop->_parent);
2405 set_idom(new_sfpt, new_cle_out, dd);
2406 set_loop(new_le, outer_loop->_parent);
2407 set_idom(new_le, new_sfpt, dd);
2408 set_loop(new_tail, outer_loop->_parent);
2409 set_idom(new_tail, new_le, dd);
2410 set_idom(new_cl, new_l, dd);
2411
2412 old_new.map(l->_idx, new_l);
2413 old_new.map(tail->_idx, new_tail);
2414 old_new.map(le->_idx, new_le);
2415 old_new.map(sfpt->_idx, new_sfpt);
2416
2417 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2418 new_l->set_req(0, new_l);
2419 new_tail->set_req(0, new_le);
2420 new_le->set_req(0, new_sfpt);
2421 new_sfpt->set_req(0, new_cle_out);
2422 new_cle_out->set_req(0, new_cle);
2423 new_cl->set_req(LoopNode::EntryControl, new_l);
2424
2425 _igvn.register_new_node_with_optimizer(new_l);
2426 _igvn.register_new_node_with_optimizer(new_tail);
2427 _igvn.register_new_node_with_optimizer(new_le);
2428 } else {
2429 Node *newhead = old_new[loop->_head->_idx];
2430 newhead->as_Loop()->clear_strip_mined();
2431 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2432 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2433 }
2434 // Look at data node that were assigned a control in the outer
2435 // loop: they are kept in the outer loop by the safepoint so start
2436 // from the safepoint node's inputs.
2437 IdealLoopTree* outer_loop = get_loop(l);
2438 Node_Stack stack(2);
2439 stack.push(sfpt, 1);
2440 uint new_counter = C->unique();
2441 while (stack.size() > 0) {
2442 Node* n = stack.node();
2443 uint i = stack.index();
2444 while (i < n->req() &&
2445 (n->in(i) == nullptr ||
2446 !has_ctrl(n->in(i)) ||
2447 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2448 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2449 i++;
2450 }
2451 if (i < n->req()) {
2452 stack.set_index(i+1);
2453 stack.push(n->in(i), 0);
2454 } else {
2455 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2456 Node* m = n == sfpt ? new_sfpt : n->clone();
2457 if (m != nullptr) {
2458 for (uint i = 0; i < n->req(); i++) {
2459 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2460 m->set_req(i, old_new[m->in(i)->_idx]);
2461 }
2462 }
2463 } else {
2464 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2465 }
2466 if (n != sfpt) {
2467 extra_data_nodes.push(n);
2468 _igvn.register_new_node_with_optimizer(m);
2469 assert(get_ctrl(n) == cle_out, "what other control?");
2470 set_ctrl(m, new_cle_out);
2471 old_new.map(n->_idx, m);
2472 }
2473 stack.pop();
2474 }
2475 }
2476 if (mode == CloneIncludesStripMined) {
2477 _igvn.register_new_node_with_optimizer(new_sfpt);
2478 _igvn.register_new_node_with_optimizer(new_cle_out);
2479 }
2480 // Some other transformation may have pessimistically assigned some
2481 // data nodes to the outer loop. Set their control so they are out
2482 // of the outer loop.
2483 ResourceMark rm;
2484 Unique_Node_List wq;
2485 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2486 Node* old = extra_data_nodes.at(i);
2487 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2488 }
2489
2490 for (uint i = 0; i < loop->_body.size(); i++) {
2491 Node* old = loop->_body.at(i);
2492 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2493 }
2494
2495 Node* inner_out = sfpt->in(0);
2496 if (inner_out->outcnt() > 1) {
2497 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2498 }
2499
2500 Node* new_ctrl = cl->outer_loop_exit();
2501 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2502 for (uint i = 0; i < wq.size(); i++) {
2503 Node* n = wq.at(i);
2504 set_ctrl(n, new_ctrl);
2505 if (n->in(0) != nullptr) {
2506 _igvn.replace_input_of(n, 0, new_ctrl);
2507 }
2508 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2509 }
2510 } else {
2511 Node *newhead = old_new[loop->_head->_idx];
2512 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2513 }
2514 }
2515
2516 //------------------------------clone_loop-------------------------------------
2517 //
2518 // C L O N E A L O O P B O D Y
2519 //
2520 // This is the basic building block of the loop optimizations. It clones an
2521 // entire loop body. It makes an old_new loop body mapping; with this mapping
2522 // you can find the new-loop equivalent to an old-loop node. All new-loop
2523 // nodes are exactly equal to their old-loop counterparts, all edges are the
2524 // same. All exits from the old-loop now have a RegionNode that merges the
2525 // equivalent new-loop path. This is true even for the normal "loop-exit"
2526 // condition. All uses of loop-invariant old-loop values now come from (one
2527 // or more) Phis that merge their new-loop equivalents.
2528 //
2529 // This operation leaves the graph in an illegal state: there are two valid
2530 // control edges coming from the loop pre-header to both loop bodies. I'll
2531 // definitely have to hack the graph after running this transform.
2532 //
2533 // From this building block I will further edit edges to perform loop peeling
2534 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2535 //
2536 // Parameter side_by_size_idom:
2537 // When side_by_size_idom is null, the dominator tree is constructed for
2538 // the clone loop to dominate the original. Used in construction of
2539 // pre-main-post loop sequence.
2540 // When nonnull, the clone and original are side-by-side, both are
2541 // dominated by the side_by_side_idom node. Used in construction of
2542 // unswitched loops.
2543 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2544 CloneLoopMode mode, Node* side_by_side_idom) {
2545
2546 LoopNode* head = loop->_head->as_Loop();
2547 head->verify_strip_mined(1);
2548
2549 if (C->do_vector_loop() && PrintOpto) {
2550 const char* mname = C->method()->name()->as_quoted_ascii();
2551 if (mname != nullptr) {
2552 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2553 }
2554 }
2555
2556 CloneMap& cm = C->clone_map();
2557 if (C->do_vector_loop()) {
2558 cm.set_clone_idx(cm.max_gen()+1);
2559 #ifndef PRODUCT
2560 if (PrintOpto) {
2561 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2562 loop->dump_head();
2563 }
2564 #endif
2565 }
2566
2567 // Step 1: Clone the loop body. Make the old->new mapping.
2568 clone_loop_body(loop->_body, old_new, &cm);
2569
2570 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2571
2572 // Step 2: Fix the edges in the new body. If the old input is outside the
2573 // loop use it. If the old input is INside the loop, use the corresponding
2574 // new node instead.
2575 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2576
2577 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2578 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2579
2580 // Step 3: Now fix control uses. Loop varying control uses have already
2581 // been fixed up (as part of all input edges in Step 2). Loop invariant
2582 // control uses must be either an IfFalse or an IfTrue. Make a merge
2583 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2584 // refer to this.
2585 Node_List worklist;
2586 uint new_counter = C->unique();
2587 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2588
2589 // Step 4: If loop-invariant use is not control, it must be dominated by a
2590 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2591 // there if needed. Make a Phi there merging old and new used values.
2592 Node_List *split_if_set = nullptr;
2593 Node_List *split_bool_set = nullptr;
2594 Node_List *split_cex_set = nullptr;
2595 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2596
2597 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2598 Node* old = extra_data_nodes.at(i);
2599 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2600 split_bool_set, split_cex_set, worklist, new_counter,
2601 mode);
2602 }
2603
2604 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2605 // the loop uses a condition set in the loop. The original IF probably
2606 // takes control from one or more OLD Regions (which in turn get from NEW
2607 // Regions). In any case, there will be a set of Phis for each merge point
2608 // from the IF up to where the original BOOL def exists the loop.
2609 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2610
2611 }
2612
2613 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2614 if (split_if_set) {
2615 while (split_if_set->size()) {
2616 Node *iff = split_if_set->pop();
2617 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2618 if (iff->in(input)->is_Phi()) {
2619 Node *b = clone_iff(iff->in(input)->as_Phi());
2620 _igvn.replace_input_of(iff, input, b);
2621 }
2622 }
2623 }
2624 if (split_bool_set) {
2625 while (split_bool_set->size()) {
2626 Node *b = split_bool_set->pop();
2627 Node *phi = b->in(1);
2628 assert(phi->is_Phi(), "");
2629 CmpNode *cmp = clone_bool((PhiNode*) phi);
2630 _igvn.replace_input_of(b, 1, cmp);
2631 }
2632 }
2633 if (split_cex_set) {
2634 while (split_cex_set->size()) {
2635 Node *b = split_cex_set->pop();
2636 assert(b->in(0)->is_Region(), "");
2637 assert(b->in(1)->is_Phi(), "");
2638 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2639 split_up(b, b->in(0), nullptr);
2640 }
2641 }
2642 }
2643
2644 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2645 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2646 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2647 for(uint i = 0; i < body.size(); i++ ) {
2648 Node* old = body.at(i);
2649 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2650 split_bool_set, split_cex_set, worklist, new_counter,
2651 mode);
2652 }
2653 }
2654
2655 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2656 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2657 LoopNode* head = loop->_head->as_Loop();
2658 for(uint i = 0; i < body.size(); i++ ) {
2659 Node* old = body.at(i);
2660 if( !old->is_CFG() ) continue;
2661
2662 // Copy uses to a worklist, so I can munge the def-use info
2663 // with impunity.
2664 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2665 worklist.push(old->fast_out(j));
2666 }
2667
2668 while (worklist.size()) { // Visit all uses
2669 Node *use = worklist.pop();
2670 if (!has_node(use)) continue; // Ignore dead nodes
2671 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2672 if (!loop->is_member(use_loop) && use->is_CFG()) {
2673 // Both OLD and USE are CFG nodes here.
2674 assert(use->is_Proj(), "" );
2675 Node* nnn = old_new[old->_idx];
2676
2677 Node* newuse = nullptr;
2678 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2679 CountedLoopNode* cl = head->as_CountedLoop();
2680 CountedLoopEndNode* cle = cl->loopexit();
2681 Node* cle_out = cle->proj_out_or_null(false);
2682 if (use == cle_out) {
2683 IfNode* le = cl->outer_loop_end();
2684 use = le->proj_out(false);
2685 use_loop = get_loop(use);
2686 if (mode == CloneIncludesStripMined) {
2687 nnn = old_new[le->_idx];
2688 } else {
2689 newuse = old_new[cle_out->_idx];
2690 }
2691 }
2692 }
2693 if (newuse == nullptr) {
2694 newuse = use->clone();
2695 }
2696
2697 // Clone the loop exit control projection
2698 if (C->do_vector_loop() && cm != nullptr) {
2699 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2700 }
2701 newuse->set_req(0,nnn);
2702 _igvn.register_new_node_with_optimizer(newuse);
2703 set_loop(newuse, use_loop);
2704 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2705
2706 // We need a Region to merge the exit from the peeled body and the
2707 // exit from the old loop body.
2708 RegionNode *r = new RegionNode(3);
2709 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2710 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2711
2712 // The original user of 'use' uses 'r' instead.
2713 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2714 Node* useuse = use->last_out(l);
2715 _igvn.rehash_node_delayed(useuse);
2716 uint uses_found = 0;
2717 if (useuse->in(0) == use) {
2718 useuse->set_req(0, r);
2719 uses_found++;
2720 if (useuse->is_CFG()) {
2721 // This is not a dom_depth > dd_r because when new
2722 // control flow is constructed by a loop opt, a node and
2723 // its dominator can end up at the same dom_depth
2724 assert(dom_depth(useuse) >= dd_r, "");
2725 set_idom(useuse, r, dom_depth(useuse));
2726 }
2727 }
2728 for (uint k = 1; k < useuse->req(); k++) {
2729 if( useuse->in(k) == use ) {
2730 useuse->set_req(k, r);
2731 uses_found++;
2732 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2733 // This is not a dom_depth > dd_r because when new
2734 // control flow is constructed by a loop opt, a node
2735 // and its dominator can end up at the same dom_depth
2736 assert(dom_depth(useuse) >= dd_r , "");
2737 set_idom(useuse, r, dom_depth(useuse));
2738 }
2739 }
2740 }
2741 l -= uses_found; // we deleted 1 or more copies of this edge
2742 }
2743
2744 assert(use->is_Proj(), "loop exit should be projection");
2745 // lazy_replace() below moves all nodes that are:
2746 // - control dependent on the loop exit or
2747 // - have control set to the loop exit
2748 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2749 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2750 // becomes dead and is "replaced" by the region.
2751 Node* use_clone = use->clone();
2752 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2753 // Now finish up 'r'
2754 r->set_req(1, newuse);
2755 r->set_req(2, use_clone);
2756 _igvn.register_new_node_with_optimizer(r);
2757 set_loop(r, use_loop);
2758 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2759 lazy_replace(use, r);
2760 // Map the (cloned) old use to the new merge point
2761 old_new.map(use_clone->_idx, r);
2762 } // End of if a loop-exit test
2763 }
2764 }
2765 }
2766
2767 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2768 IdealLoopTree* parent, bool partial) {
2769 for(uint i = 0; i < body.size(); i++ ) {
2770 Node *old = body.at(i);
2771 Node *nnn = old_new[old->_idx];
2772 // Fix CFG/Loop controlling the new node
2773 if (has_ctrl(old)) {
2774 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2775 } else {
2776 set_loop(nnn, parent);
2777 if (old->outcnt() > 0) {
2778 Node* dom = idom(old);
2779 if (old_new[dom->_idx] != nullptr) {
2780 dom = old_new[dom->_idx];
2781 set_idom(nnn, dom, dd );
2782 }
2783 }
2784 }
2785 // Correct edges to the new node
2786 for (uint j = 0; j < nnn->req(); j++) {
2787 Node *n = nnn->in(j);
2788 if (n != nullptr) {
2789 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2790 if (loop->is_member(old_in_loop)) {
2791 if (old_new[n->_idx] != nullptr) {
2792 nnn->set_req(j, old_new[n->_idx]);
2793 } else {
2794 assert(!body.contains(n), "");
2795 assert(partial, "node not cloned");
2796 }
2797 }
2798 }
2799 }
2800 _igvn.hash_find_insert(nnn);
2801 }
2802 }
2803
2804 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2805 for (uint i = 0; i < body.size(); i++) {
2806 Node* old = body.at(i);
2807 Node* nnn = old->clone();
2808 old_new.map(old->_idx, nnn);
2809 if (C->do_vector_loop() && cm != nullptr) {
2810 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2811 }
2812 _igvn.register_new_node_with_optimizer(nnn);
2813 }
2814 }
2815
2816
2817 //---------------------- stride_of_possible_iv -------------------------------------
2818 // Looks for an iff/bool/comp with one operand of the compare
2819 // being a cycle involving an add and a phi,
2820 // with an optional truncation (left-shift followed by a right-shift)
2821 // of the add. Returns zero if not an iv.
2822 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2823 Node* trunc1 = nullptr;
2824 Node* trunc2 = nullptr;
2825 const TypeInteger* ttype = nullptr;
2826 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
2827 return 0;
2828 }
2829 BoolNode* bl = iff->in(1)->as_Bool();
2830 Node* cmp = bl->in(1);
2831 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2832 return 0;
2833 }
2834 // Must have an invariant operand
2835 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2836 return 0;
2837 }
2838 Node* add2 = nullptr;
2839 Node* cmp1 = cmp->in(1);
2840 if (cmp1->is_Phi()) {
2841 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2842 Node* phi = cmp1;
2843 for (uint i = 1; i < phi->req(); i++) {
2844 Node* in = phi->in(i);
2845 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2846 &trunc1, &trunc2, &ttype, T_INT);
2847 if (add && add->in(1) == phi) {
2848 add2 = add->in(2);
2849 break;
2850 }
2851 }
2852 } else {
2853 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2854 Node* addtrunc = cmp1;
2855 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2856 &trunc1, &trunc2, &ttype, T_INT);
2857 if (add && add->in(1)->is_Phi()) {
2858 Node* phi = add->in(1);
2859 for (uint i = 1; i < phi->req(); i++) {
2860 if (phi->in(i) == addtrunc) {
2861 add2 = add->in(2);
2862 break;
2863 }
2864 }
2865 }
2866 }
2867 if (add2 != nullptr) {
2868 const TypeInt* add2t = _igvn.type(add2)->is_int();
2869 if (add2t->is_con()) {
2870 return add2t->get_con();
2871 }
2872 }
2873 return 0;
2874 }
2875
2876
2877 //---------------------- stay_in_loop -------------------------------------
2878 // Return the (unique) control output node that's in the loop (if it exists.)
2879 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2880 Node* unique = nullptr;
2881 if (!n) return nullptr;
2882 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2883 Node* use = n->fast_out(i);
2884 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2885 if (unique != nullptr) {
2886 return nullptr;
2887 }
2888 unique = use;
2889 }
2890 }
2891 return unique;
2892 }
2893
2894 //------------------------------ register_node -------------------------------------
2895 // Utility to register node "n" with PhaseIdealLoop
2896 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
2897 _igvn.register_new_node_with_optimizer(n);
2898 loop->_body.push(n);
2899 if (n->is_CFG()) {
2900 set_loop(n, loop);
2901 set_idom(n, pred, ddepth);
2902 } else {
2903 set_ctrl(n, pred);
2904 }
2905 }
2906
2907 //------------------------------ proj_clone -------------------------------------
2908 // Utility to create an if-projection
2909 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2910 ProjNode* c = p->clone()->as_Proj();
2911 c->set_req(0, iff);
2912 return c;
2913 }
2914
2915 //------------------------------ short_circuit_if -------------------------------------
2916 // Force the iff control output to be the live_proj
2917 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2918 guarantee(live_proj != nullptr, "null projection");
2919 int proj_con = live_proj->_con;
2920 assert(proj_con == 0 || proj_con == 1, "false or true projection");
2921 Node* con = intcon(proj_con);
2922 if (iff) {
2923 iff->set_req(1, con);
2924 }
2925 return con;
2926 }
2927
2928 //------------------------------ insert_if_before_proj -------------------------------------
2929 // Insert a new if before an if projection (* - new node)
2930 //
2931 // before
2932 // if(test)
2933 // / \
2934 // v v
2935 // other-proj proj (arg)
2936 //
2937 // after
2938 // if(test)
2939 // / \
2940 // / v
2941 // | * proj-clone
2942 // v |
2943 // other-proj v
2944 // * new_if(relop(cmp[IU](left,right)))
2945 // / \
2946 // v v
2947 // * new-proj proj
2948 // (returned)
2949 //
2950 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2951 IfNode* iff = proj->in(0)->as_If();
2952 IdealLoopTree *loop = get_loop(proj);
2953 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2954 uint ddepth = dom_depth(proj);
2955
2956 _igvn.rehash_node_delayed(iff);
2957 _igvn.rehash_node_delayed(proj);
2958
2959 proj->set_req(0, nullptr); // temporary disconnect
2960 ProjNode* proj2 = proj_clone(proj, iff);
2961 register_node(proj2, loop, iff, ddepth);
2962
2963 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2964 register_node(cmp, loop, proj2, ddepth);
2965
2966 BoolNode* bol = new BoolNode(cmp, relop);
2967 register_node(bol, loop, proj2, ddepth);
2968
2969 int opcode = iff->Opcode();
2970 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2971 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
2972 register_node(new_if, loop, proj2, ddepth);
2973
2974 proj->set_req(0, new_if); // reattach
2975 set_idom(proj, new_if, ddepth);
2976
2977 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2978 guarantee(new_exit != nullptr, "null exit node");
2979 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2980
2981 return new_exit;
2982 }
2983
2984 //------------------------------ insert_region_before_proj -------------------------------------
2985 // Insert a region before an if projection (* - new node)
2986 //
2987 // before
2988 // if(test)
2989 // / |
2990 // v |
2991 // proj v
2992 // other-proj
2993 //
2994 // after
2995 // if(test)
2996 // / |
2997 // v |
2998 // * proj-clone v
2999 // | other-proj
3000 // v
3001 // * new-region
3002 // |
3003 // v
3004 // * dum_if
3005 // / \
3006 // v \
3007 // * dum-proj v
3008 // proj
3009 //
3010 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3011 IfNode* iff = proj->in(0)->as_If();
3012 IdealLoopTree *loop = get_loop(proj);
3013 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3014 uint ddepth = dom_depth(proj);
3015
3016 _igvn.rehash_node_delayed(iff);
3017 _igvn.rehash_node_delayed(proj);
3018
3019 proj->set_req(0, nullptr); // temporary disconnect
3020 ProjNode* proj2 = proj_clone(proj, iff);
3021 register_node(proj2, loop, iff, ddepth);
3022
3023 RegionNode* reg = new RegionNode(2);
3024 reg->set_req(1, proj2);
3025 register_node(reg, loop, iff, ddepth);
3026
3027 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3028 register_node(dum_if, loop, reg, ddepth);
3029
3030 proj->set_req(0, dum_if); // reattach
3031 set_idom(proj, dum_if, ddepth);
3032
3033 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3034 register_node(dum_proj, loop, dum_if, ddepth);
3035
3036 return reg;
3037 }
3038
3039 // Idea
3040 // ----
3041 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3042 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3043 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3044 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3045 // test alone instead:
3046 //
3047 // Before Partial Peeling:
3048 // Loop:
3049 // <peeled section>
3050 // Split off signed loop exit test
3051 // <-- CUT HERE -->
3052 // Unchanged unsigned loop exit test
3053 // <rest of unpeeled section>
3054 // goto Loop
3055 //
3056 // After Partial Peeling:
3057 // <cloned peeled section>
3058 // Cloned split off signed loop exit test
3059 // Loop:
3060 // Unchanged unsigned loop exit test
3061 // <rest of unpeeled section>
3062 // <peeled section>
3063 // Split off signed loop exit test
3064 // goto Loop
3065 //
3066 // Details
3067 // -------
3068 // Before:
3069 // if (i <u limit) Unsigned loop exit condition
3070 // / |
3071 // v v
3072 // exit-proj stay-in-loop-proj
3073 //
3074 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3075 // before the CmpU on the stay-in-loop path and keep both tests:
3076 //
3077 // if (i <u limit) Signed loop exit test
3078 // / |
3079 // / if (i <u limit) Unsigned loop exit test
3080 // / / |
3081 // v v v
3082 // exit-region stay-in-loop-proj
3083 //
3084 // Implementation
3085 // --------------
3086 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3087 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3088 // exit tests is preserved, and their loop nesting is correct.
3089 //
3090 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3091 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3092 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3093 //
3094 // if (stay-in-loop-const) Killed original unsigned loop exit test
3095 // / |
3096 // / v
3097 // / if (i < limit) Split off signed loop exit test
3098 // / / |
3099 // / / v
3100 // / / if (i <u limit) Cloned unsigned loop exit test
3101 // / / / |
3102 // v v v |
3103 // exit-region |
3104 // | |
3105 // dummy-if |
3106 // / | |
3107 // dead | |
3108 // v v
3109 // exit-proj stay-in-loop-proj
3110 //
3111 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3112 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3113 //
3114 // Requirements
3115 // ------------
3116 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3117 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3118 // Otherwise, we need to bail out (see comments in the code below).
3119 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3120 const bool Signed = true;
3121 const bool Unsigned = false;
3122
3123 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3124 if (bol->_test._test != BoolTest::lt) {
3125 return nullptr;
3126 }
3127 CmpNode* cmpu = bol->in(1)->as_Cmp();
3128 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3129
3130 int stride = stride_of_possible_iv(if_cmpu);
3131 if (stride == 0) {
3132 return nullptr;
3133 }
3134
3135 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3136 guarantee(lp_proj != nullptr, "null loop node");
3137
3138 ProjNode* lp_continue = lp_proj->as_Proj();
3139 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3140 if (!lp_exit->is_IfFalse()) {
3141 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3142 // We therefore can't add a single exit condition.
3143 return nullptr;
3144 }
3145 // The unsigned loop exit condition is
3146 // !(i <u limit)
3147 // = i >=u limit
3148 //
3149 // First, we note that for any x for which
3150 // 0 <= x <= INT_MAX
3151 // we can convert x to an unsigned int and still get the same guarantee:
3152 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3153 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3154 //
3155 // With that in mind, if
3156 // limit >= 0 (COND)
3157 // then the unsigned loop exit condition
3158 // i >=u limit (ULE)
3159 // is equivalent to
3160 // i < 0 || i >= limit (SLE-full)
3161 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3162 // (uint) i >=u MAX_INT >= limit >= 0
3163 // or otherwise
3164 // i >= limit >= 0
3165 // holds due to (LEMMA).
3166 //
3167 // For completeness, a counterexample with limit < 0:
3168 // Assume i = -3 and limit = -2:
3169 // i < 0
3170 // -2 < 0
3171 // is true and thus also "i < 0 || i >= limit". But
3172 // i >=u limit
3173 // -3 >=u -2
3174 // is false.
3175 Node* limit = cmpu->in(2);
3176 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3177 if (type_limit->_lo < 0) {
3178 return nullptr;
3179 }
3180
3181 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3182 // stride < 0:
3183 // i < 0 (SLE = SLE-negative)
3184 // stride > 0:
3185 // i >= limit (SLE = SLE-positive)
3186 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3187 //
3188 // Loop:
3189 // <peeled section>
3190 // i >= limit (SLE-positive)
3191 // <-- CUT HERE -->
3192 // i >=u limit (ULE)
3193 // <rest of unpeeled section>
3194 // goto Loop
3195 //
3196 // We exit the loop if:
3197 // (SLE) is true OR (ULE) is true
3198 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3199 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3200 // (SLE) IMPLIES (ULE)
3201 // This indeed holds when (COND) is given:
3202 // - stride > 0:
3203 // i >= limit // (SLE = SLE-positive)
3204 // i >= limit >= 0 // (COND)
3205 // i >=u limit >= 0 // (LEMMA)
3206 // which is the unsigned loop exit condition (ULE).
3207 // - stride < 0:
3208 // i < 0 // (SLE = SLE-negative)
3209 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3210 // MAX_INT >= limit >= 0 // (COND)
3211 // MAX_INT >=u limit >= 0 // (LEMMA)
3212 // and thus from (NEG) and (LEMMA):
3213 // i >=u limit
3214 // which is the unsigned loop exit condition (ULE).
3215 //
3216 //
3217 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3218 // <cloned peeled section>
3219 // i >= limit (SLE-positive)
3220 // Loop:
3221 // i >=u limit (ULE)
3222 // <rest of unpeeled section>
3223 // <peeled section>
3224 // i >= limit (SLE-positive)
3225 // goto Loop
3226 Node* rhs_cmpi;
3227 if (stride > 0) {
3228 rhs_cmpi = limit; // For i >= limit
3229 } else {
3230 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3231 }
3232 // Create a new region on the exit path
3233 RegionNode* reg = insert_region_before_proj(lp_exit);
3234 guarantee(reg != nullptr, "null region node");
3235
3236 // Clone the if-cmpu-true-false using a signed compare
3237 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3238 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3239 reg->add_req(cmpi_exit);
3240
3241 // Clone the if-cmpu-true-false
3242 BoolTest::mask rel_u = bol->_test._test;
3243 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3244 reg->add_req(cmpu_exit);
3245
3246 // Force original if to stay in loop.
3247 short_circuit_if(if_cmpu, lp_continue);
3248
3249 return cmpi_exit->in(0)->as_If();
3250 }
3251
3252 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3253 // Remove a previously inserted signed compare loop exit.
3254 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3255 Node* lp_proj = stay_in_loop(if_cmp, loop);
3256 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3257 stay_in_loop(lp_proj, loop)->is_If() &&
3258 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3259 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3260 if_cmp->set_req(1, con);
3261 }
3262
3263 //------------------------------ scheduled_nodelist -------------------------------------
3264 // Create a post order schedule of nodes that are in the
3265 // "member" set. The list is returned in "sched".
3266 // The first node in "sched" is the loop head, followed by
3267 // nodes which have no inputs in the "member" set, and then
3268 // followed by the nodes that have an immediate input dependence
3269 // on a node in "sched".
3270 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3271
3272 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3273 VectorSet visited;
3274 Node_Stack nstack(loop->_body.size());
3275
3276 Node* n = loop->_head; // top of stack is cached in "n"
3277 uint idx = 0;
3278 visited.set(n->_idx);
3279
3280 // Initially push all with no inputs from within member set
3281 for(uint i = 0; i < loop->_body.size(); i++ ) {
3282 Node *elt = loop->_body.at(i);
3283 if (member.test(elt->_idx)) {
3284 bool found = false;
3285 for (uint j = 0; j < elt->req(); j++) {
3286 Node* def = elt->in(j);
3287 if (def && member.test(def->_idx) && def != elt) {
3288 found = true;
3289 break;
3290 }
3291 }
3292 if (!found && elt != loop->_head) {
3293 nstack.push(n, idx);
3294 n = elt;
3295 assert(!visited.test(n->_idx), "not seen yet");
3296 visited.set(n->_idx);
3297 }
3298 }
3299 }
3300
3301 // traverse out's that are in the member set
3302 while (true) {
3303 if (idx < n->outcnt()) {
3304 Node* use = n->raw_out(idx);
3305 idx++;
3306 if (!visited.test_set(use->_idx)) {
3307 if (member.test(use->_idx)) {
3308 nstack.push(n, idx);
3309 n = use;
3310 idx = 0;
3311 }
3312 }
3313 } else {
3314 // All outputs processed
3315 sched.push(n);
3316 if (nstack.is_empty()) break;
3317 n = nstack.node();
3318 idx = nstack.index();
3319 nstack.pop();
3320 }
3321 }
3322 }
3323
3324
3325 //------------------------------ has_use_in_set -------------------------------------
3326 // Has a use in the vector set
3327 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3328 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3329 Node* use = n->fast_out(j);
3330 if (vset.test(use->_idx)) {
3331 return true;
3332 }
3333 }
3334 return false;
3335 }
3336
3337
3338 //------------------------------ has_use_internal_to_set -------------------------------------
3339 // Has use internal to the vector set (ie. not in a phi at the loop head)
3340 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3341 Node* head = loop->_head;
3342 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3343 Node* use = n->fast_out(j);
3344 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3345 return true;
3346 }
3347 }
3348 return false;
3349 }
3350
3351
3352 //------------------------------ clone_for_use_outside_loop -------------------------------------
3353 // clone "n" for uses that are outside of loop
3354 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3355 int cloned = 0;
3356 assert(worklist.size() == 0, "should be empty");
3357 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3358 Node* use = n->fast_out(j);
3359 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3360 worklist.push(use);
3361 }
3362 }
3363
3364 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3365 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3366 return -1;
3367 }
3368
3369 while( worklist.size() ) {
3370 Node *use = worklist.pop();
3371 if (!has_node(use) || use->in(0) == C->top()) continue;
3372 uint j;
3373 for (j = 0; j < use->req(); j++) {
3374 if (use->in(j) == n) break;
3375 }
3376 assert(j < use->req(), "must be there");
3377
3378 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3379 Node* n_clone = n->clone();
3380 _igvn.replace_input_of(use, j, n_clone);
3381 cloned++;
3382 Node* use_c;
3383 if (!use->is_Phi()) {
3384 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3385 } else {
3386 // Use in a phi is considered a use in the associated predecessor block
3387 use_c = use->in(0)->in(j);
3388 }
3389 set_ctrl(n_clone, use_c);
3390 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3391 get_loop(use_c)->_body.push(n_clone);
3392 _igvn.register_new_node_with_optimizer(n_clone);
3393 #ifndef PRODUCT
3394 if (TracePartialPeeling) {
3395 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3396 }
3397 #endif
3398 }
3399 return cloned;
3400 }
3401
3402
3403 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3404 // clone "n" for special uses that are in the not_peeled region.
3405 // If these def-uses occur in separate blocks, the code generator
3406 // marks the method as not compilable. For example, if a "BoolNode"
3407 // is in a different basic block than the "IfNode" that uses it, then
3408 // the compilation is aborted in the code generator.
3409 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3410 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3411 if (n->is_Phi() || n->is_Load()) {
3412 return;
3413 }
3414 assert(worklist.size() == 0, "should be empty");
3415 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3416 Node* use = n->fast_out(j);
3417 if ( not_peel.test(use->_idx) &&
3418 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) &&
3419 use->in(1) == n) {
3420 worklist.push(use);
3421 }
3422 }
3423 if (worklist.size() > 0) {
3424 // clone "n" and insert it between inputs of "n" and the use
3425 Node* n_clone = n->clone();
3426 loop->_body.push(n_clone);
3427 _igvn.register_new_node_with_optimizer(n_clone);
3428 set_ctrl(n_clone, get_ctrl(n));
3429 sink_list.push(n_clone);
3430 not_peel.set(n_clone->_idx);
3431 #ifndef PRODUCT
3432 if (TracePartialPeeling) {
3433 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3434 }
3435 #endif
3436 while( worklist.size() ) {
3437 Node *use = worklist.pop();
3438 _igvn.rehash_node_delayed(use);
3439 for (uint j = 1; j < use->req(); j++) {
3440 if (use->in(j) == n) {
3441 use->set_req(j, n_clone);
3442 }
3443 }
3444 }
3445 }
3446 }
3447
3448
3449 //------------------------------ insert_phi_for_loop -------------------------------------
3450 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3451 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3452 Node *phi = PhiNode::make(lp, back_edge_val);
3453 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3454 // Use existing phi if it already exists
3455 Node *hit = _igvn.hash_find_insert(phi);
3456 if( hit == nullptr ) {
3457 _igvn.register_new_node_with_optimizer(phi);
3458 set_ctrl(phi, lp);
3459 } else {
3460 // Remove the new phi from the graph and use the hit
3461 _igvn.remove_dead_node(phi);
3462 phi = hit;
3463 }
3464 _igvn.replace_input_of(use, idx, phi);
3465 }
3466
3467 #ifdef ASSERT
3468 //------------------------------ is_valid_loop_partition -------------------------------------
3469 // Validate the loop partition sets: peel and not_peel
3470 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3471 VectorSet& not_peel ) {
3472 uint i;
3473 // Check that peel_list entries are in the peel set
3474 for (i = 0; i < peel_list.size(); i++) {
3475 if (!peel.test(peel_list.at(i)->_idx)) {
3476 return false;
3477 }
3478 }
3479 // Check at loop members are in one of peel set or not_peel set
3480 for (i = 0; i < loop->_body.size(); i++ ) {
3481 Node *def = loop->_body.at(i);
3482 uint di = def->_idx;
3483 // Check that peel set elements are in peel_list
3484 if (peel.test(di)) {
3485 if (not_peel.test(di)) {
3486 return false;
3487 }
3488 // Must be in peel_list also
3489 bool found = false;
3490 for (uint j = 0; j < peel_list.size(); j++) {
3491 if (peel_list.at(j)->_idx == di) {
3492 found = true;
3493 break;
3494 }
3495 }
3496 if (!found) {
3497 return false;
3498 }
3499 } else if (not_peel.test(di)) {
3500 if (peel.test(di)) {
3501 return false;
3502 }
3503 } else {
3504 return false;
3505 }
3506 }
3507 return true;
3508 }
3509
3510 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3511 // Ensure a use outside of loop is of the right form
3512 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3513 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3514 return (use->is_Phi() &&
3515 use_c->is_Region() && use_c->req() == 3 &&
3516 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3517 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3518 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3519 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3520 }
3521
3522 //------------------------------ is_valid_clone_loop_form -------------------------------------
3523 // Ensure that all uses outside of loop are of the right form
3524 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3525 uint orig_exit_idx, uint clone_exit_idx) {
3526 uint len = peel_list.size();
3527 for (uint i = 0; i < len; i++) {
3528 Node *def = peel_list.at(i);
3529
3530 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3531 Node *use = def->fast_out(j);
3532 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3533 if (!loop->is_member(get_loop(use_c))) {
3534 // use is not in the loop, check for correct structure
3535 if (use->in(0) == def) {
3536 // Okay
3537 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3538 return false;
3539 }
3540 }
3541 }
3542 }
3543 return true;
3544 }
3545 #endif
3546
3547 //------------------------------ partial_peel -------------------------------------
3548 // Partially peel (aka loop rotation) the top portion of a loop (called
3549 // the peel section below) by cloning it and placing one copy just before
3550 // the new loop head and the other copy at the bottom of the new loop.
3551 //
3552 // before after where it came from
3553 //
3554 // stmt1 stmt1
3555 // loop: stmt2 clone
3556 // stmt2 if condA goto exitA clone
3557 // if condA goto exitA new_loop: new
3558 // stmt3 stmt3 clone
3559 // if !condB goto loop if condB goto exitB clone
3560 // exitB: stmt2 orig
3561 // stmt4 if !condA goto new_loop orig
3562 // exitA: goto exitA
3563 // exitB:
3564 // stmt4
3565 // exitA:
3566 //
3567 // Step 1: find the cut point: an exit test on probable
3568 // induction variable.
3569 // Step 2: schedule (with cloning) operations in the peel
3570 // section that can be executed after the cut into
3571 // the section that is not peeled. This may need
3572 // to clone operations into exit blocks. For
3573 // instance, a reference to A[i] in the not-peel
3574 // section and a reference to B[i] in an exit block
3575 // may cause a left-shift of i by 2 to be placed
3576 // in the peel block. This step will clone the left
3577 // shift into the exit block and sink the left shift
3578 // from the peel to the not-peel section.
3579 // Step 3: clone the loop, retarget the control, and insert
3580 // phis for values that are live across the new loop
3581 // head. This is very dependent on the graph structure
3582 // from clone_loop. It creates region nodes for
3583 // exit control and associated phi nodes for values
3584 // flow out of the loop through that exit. The region
3585 // node is dominated by the clone's control projection.
3586 // So the clone's peel section is placed before the
3587 // new loop head, and the clone's not-peel section is
3588 // forms the top part of the new loop. The original
3589 // peel section forms the tail of the new loop.
3590 // Step 4: update the dominator tree and recompute the
3591 // dominator depth.
3592 //
3593 // orig
3594 //
3595 // stmt1
3596 // |
3597 // v
3598 // predicates
3599 // |
3600 // v
3601 // loop<----+
3602 // | |
3603 // stmt2 |
3604 // | |
3605 // v |
3606 // ifA |
3607 // / | |
3608 // v v |
3609 // false true ^ <-- last_peel
3610 // / | |
3611 // / ===|==cut |
3612 // / stmt3 | <-- first_not_peel
3613 // / | |
3614 // | v |
3615 // v ifB |
3616 // exitA: / \ |
3617 // / \ |
3618 // v v |
3619 // false true |
3620 // / \ |
3621 // / ----+
3622 // |
3623 // v
3624 // exitB:
3625 // stmt4
3626 //
3627 //
3628 // after clone loop
3629 //
3630 // stmt1
3631 // |
3632 // v
3633 // predicates
3634 // / \
3635 // clone / \ orig
3636 // / \
3637 // / \
3638 // v v
3639 // +---->loop loop<----+
3640 // | | | |
3641 // | stmt2 stmt2 |
3642 // | | | |
3643 // | v v |
3644 // | ifA ifA |
3645 // | | \ / | |
3646 // | v v v v |
3647 // ^ true false false true ^ <-- last_peel
3648 // | | ^ \ / | |
3649 // | cut==|== \ \ / ===|==cut |
3650 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3651 // | | dom | | | |
3652 // | v \ 1v v2 v |
3653 // | ifB regionA ifB |
3654 // | / \ | / \ |
3655 // | / \ v / \ |
3656 // | v v exitA: v v |
3657 // | true false false true |
3658 // | / ^ \ / \ |
3659 // +---- \ \ / ----+
3660 // dom \ /
3661 // \ 1v v2
3662 // regionB
3663 // |
3664 // v
3665 // exitB:
3666 // stmt4
3667 //
3668 //
3669 // after partial peel
3670 //
3671 // stmt1
3672 // |
3673 // v
3674 // predicates
3675 // /
3676 // clone / orig
3677 // / TOP
3678 // / \
3679 // v v
3680 // TOP->loop loop----+
3681 // | | |
3682 // stmt2 stmt2 |
3683 // | | |
3684 // v v |
3685 // ifA ifA |
3686 // | \ / | |
3687 // v v v v |
3688 // true false false true | <-- last_peel
3689 // | ^ \ / +------|---+
3690 // +->newloop \ \ / === ==cut | |
3691 // | stmt3 \ \ / TOP | |
3692 // | | dom | | stmt3 | | <-- first_not_peel
3693 // | v \ 1v v2 v | |
3694 // | ifB regionA ifB ^ v
3695 // | / \ | / \ | |
3696 // | / \ v / \ | |
3697 // | v v exitA: v v | |
3698 // | true false false true | |
3699 // | / ^ \ / \ | |
3700 // | | \ \ / v | |
3701 // | | dom \ / TOP | |
3702 // | | \ 1v v2 | |
3703 // ^ v regionB | |
3704 // | | | | |
3705 // | | v ^ v
3706 // | | exitB: | |
3707 // | | stmt4 | |
3708 // | +------------>-----------------+ |
3709 // | |
3710 // +-----------------<---------------------+
3711 //
3712 //
3713 // final graph
3714 //
3715 // stmt1
3716 // |
3717 // v
3718 // predicates
3719 // |
3720 // v
3721 // stmt2 clone
3722 // |
3723 // v
3724 // ........> ifA clone
3725 // : / |
3726 // dom / |
3727 // : v v
3728 // : false true
3729 // : | |
3730 // : | v
3731 // : | newloop<-----+
3732 // : | | |
3733 // : | stmt3 clone |
3734 // : | | |
3735 // : | v |
3736 // : | ifB |
3737 // : | / \ |
3738 // : | v v |
3739 // : | false true |
3740 // : | | | |
3741 // : | v stmt2 |
3742 // : | exitB: | |
3743 // : | stmt4 v |
3744 // : | ifA orig |
3745 // : | / \ |
3746 // : | / \ |
3747 // : | v v |
3748 // : | false true |
3749 // : | / \ |
3750 // : v v -----+
3751 // RegionA
3752 // |
3753 // v
3754 // exitA
3755 //
3756 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3757
3758 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3759 if (!loop->_head->is_Loop()) {
3760 return false;
3761 }
3762 LoopNode *head = loop->_head->as_Loop();
3763
3764 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3765 return false;
3766 }
3767
3768 // Check for complex exit control
3769 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3770 Node *n = loop->_body.at(ii);
3771 int opc = n->Opcode();
3772 if (n->is_Call() ||
3773 opc == Op_Catch ||
3774 opc == Op_CatchProj ||
3775 opc == Op_Jump ||
3776 opc == Op_JumpProj) {
3777 #ifndef PRODUCT
3778 if (TracePartialPeeling) {
3779 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3780 }
3781 #endif
3782 return false;
3783 }
3784 }
3785
3786 int dd = dom_depth(head);
3787
3788 // Step 1: find cut point
3789
3790 // Walk up dominators to loop head looking for first loop exit
3791 // which is executed on every path thru loop.
3792 IfNode *peel_if = nullptr;
3793 IfNode *peel_if_cmpu = nullptr;
3794
3795 Node *iff = loop->tail();
3796 while (iff != head) {
3797 if (iff->is_If()) {
3798 Node *ctrl = get_ctrl(iff->in(1));
3799 if (ctrl->is_top()) return false; // Dead test on live IF.
3800 // If loop-varying exit-test, check for induction variable
3801 if (loop->is_member(get_loop(ctrl)) &&
3802 loop->is_loop_exit(iff) &&
3803 is_possible_iv_test(iff)) {
3804 Node* cmp = iff->in(1)->in(1);
3805 if (cmp->Opcode() == Op_CmpI) {
3806 peel_if = iff->as_If();
3807 } else {
3808 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3809 peel_if_cmpu = iff->as_If();
3810 }
3811 }
3812 }
3813 iff = idom(iff);
3814 }
3815
3816 // Prefer signed compare over unsigned compare.
3817 IfNode* new_peel_if = nullptr;
3818 if (peel_if == nullptr) {
3819 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3820 return false; // No peel point found
3821 }
3822 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3823 if (new_peel_if == nullptr) {
3824 return false; // No peel point found
3825 }
3826 peel_if = new_peel_if;
3827 }
3828 Node* last_peel = stay_in_loop(peel_if, loop);
3829 Node* first_not_peeled = stay_in_loop(last_peel, loop);
3830 if (first_not_peeled == nullptr || first_not_peeled == head) {
3831 return false;
3832 }
3833
3834 #ifndef PRODUCT
3835 if (TraceLoopOpts) {
3836 tty->print("PartialPeel ");
3837 loop->dump_head();
3838 }
3839
3840 if (TracePartialPeeling) {
3841 tty->print_cr("before partial peel one iteration");
3842 Node_List wl;
3843 Node* t = head->in(2);
3844 while (true) {
3845 wl.push(t);
3846 if (t == head) break;
3847 t = idom(t);
3848 }
3849 while (wl.size() > 0) {
3850 Node* tt = wl.pop();
3851 tt->dump();
3852 if (tt == last_peel) tty->print_cr("-- cut --");
3853 }
3854 }
3855 #endif
3856
3857 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
3858
3859 VectorSet peel;
3860 VectorSet not_peel;
3861 Node_List peel_list;
3862 Node_List worklist;
3863 Node_List sink_list;
3864
3865 uint estimate = loop->est_loop_clone_sz(1);
3866 if (exceeding_node_budget(estimate)) {
3867 return false;
3868 }
3869
3870 // Set of cfg nodes to peel are those that are executable from
3871 // the head through last_peel.
3872 assert(worklist.size() == 0, "should be empty");
3873 worklist.push(head);
3874 peel.set(head->_idx);
3875 while (worklist.size() > 0) {
3876 Node *n = worklist.pop();
3877 if (n != last_peel) {
3878 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3879 Node* use = n->fast_out(j);
3880 if (use->is_CFG() &&
3881 loop->is_member(get_loop(use)) &&
3882 !peel.test_set(use->_idx)) {
3883 worklist.push(use);
3884 }
3885 }
3886 }
3887 }
3888
3889 // Set of non-cfg nodes to peel are those that are control
3890 // dependent on the cfg nodes.
3891 for (uint i = 0; i < loop->_body.size(); i++) {
3892 Node *n = loop->_body.at(i);
3893 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3894 if (peel.test(n_c->_idx)) {
3895 peel.set(n->_idx);
3896 } else {
3897 not_peel.set(n->_idx);
3898 }
3899 }
3900
3901 // Step 2: move operations from the peeled section down into the
3902 // not-peeled section
3903
3904 // Get a post order schedule of nodes in the peel region
3905 // Result in right-most operand.
3906 scheduled_nodelist(loop, peel, peel_list);
3907
3908 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3909
3910 // For future check for too many new phis
3911 uint old_phi_cnt = 0;
3912 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3913 Node* use = head->fast_out(j);
3914 if (use->is_Phi()) old_phi_cnt++;
3915 }
3916
3917 #ifndef PRODUCT
3918 if (TracePartialPeeling) {
3919 tty->print_cr("\npeeled list");
3920 }
3921 #endif
3922
3923 // Evacuate nodes in peel region into the not_peeled region if possible
3924 bool too_many_clones = false;
3925 uint new_phi_cnt = 0;
3926 uint cloned_for_outside_use = 0;
3927 for (uint i = 0; i < peel_list.size();) {
3928 Node* n = peel_list.at(i);
3929 #ifndef PRODUCT
3930 if (TracePartialPeeling) n->dump();
3931 #endif
3932 bool incr = true;
3933 if (!n->is_CFG()) {
3934 if (has_use_in_set(n, not_peel)) {
3935 // If not used internal to the peeled region,
3936 // move "n" from peeled to not_peeled region.
3937 if (!has_use_internal_to_set(n, peel, loop)) {
3938 // if not pinned and not a load (which maybe anti-dependent on a store)
3939 // and not a CMove (Matcher expects only bool->cmove).
3940 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
3941 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
3942 if (C->failing()) return false;
3943 if (new_clones == -1) {
3944 too_many_clones = true;
3945 break;
3946 }
3947 cloned_for_outside_use += new_clones;
3948 sink_list.push(n);
3949 peel.remove(n->_idx);
3950 not_peel.set(n->_idx);
3951 peel_list.remove(i);
3952 incr = false;
3953 #ifndef PRODUCT
3954 if (TracePartialPeeling) {
3955 tty->print_cr("sink to not_peeled region: %d newbb: %d",
3956 n->_idx, get_ctrl(n)->_idx);
3957 }
3958 #endif
3959 }
3960 } else {
3961 // Otherwise check for special def-use cases that span
3962 // the peel/not_peel boundary such as bool->if
3963 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3964 new_phi_cnt++;
3965 }
3966 }
3967 }
3968 if (incr) i++;
3969 }
3970
3971 estimate += cloned_for_outside_use + new_phi_cnt;
3972 bool exceed_node_budget = !may_require_nodes(estimate);
3973 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3974
3975 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
3976 #ifndef PRODUCT
3977 if (TracePartialPeeling && exceed_phi_limit) {
3978 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
3979 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
3980 }
3981 #endif
3982 if (new_peel_if != nullptr) {
3983 remove_cmpi_loop_exit(new_peel_if, loop);
3984 }
3985 // Inhibit more partial peeling on this loop
3986 assert(!head->is_partial_peel_loop(), "not partial peeled");
3987 head->mark_partial_peel_failed();
3988 if (cloned_for_outside_use > 0) {
3989 // Terminate this round of loop opts because
3990 // the graph outside this loop was changed.
3991 C->set_major_progress();
3992 return true;
3993 }
3994 return false;
3995 }
3996
3997 // Step 3: clone loop, retarget control, and insert new phis
3998
3999 // Create new loop head for new phis and to hang
4000 // the nodes being moved (sinked) from the peel region.
4001 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4002 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4003 _igvn.register_new_node_with_optimizer(new_head);
4004 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4005 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4006 set_loop(new_head, loop);
4007 loop->_body.push(new_head);
4008 not_peel.set(new_head->_idx);
4009 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4010 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4011
4012 while (sink_list.size() > 0) {
4013 Node* n = sink_list.pop();
4014 set_ctrl(n, new_head);
4015 }
4016
4017 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4018
4019 clone_loop(loop, old_new, dd, IgnoreStripMined);
4020
4021 const uint clone_exit_idx = 1;
4022 const uint orig_exit_idx = 2;
4023 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4024
4025 Node* head_clone = old_new[head->_idx];
4026 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4027 Node* orig_tail_clone = head_clone->in(2);
4028
4029 // Add phi if "def" node is in peel set and "use" is not
4030
4031 for (uint i = 0; i < peel_list.size(); i++) {
4032 Node *def = peel_list.at(i);
4033 if (!def->is_CFG()) {
4034 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4035 Node *use = def->fast_out(j);
4036 if (has_node(use) && use->in(0) != C->top() &&
4037 (!peel.test(use->_idx) ||
4038 (use->is_Phi() && use->in(0) == head)) ) {
4039 worklist.push(use);
4040 }
4041 }
4042 while( worklist.size() ) {
4043 Node *use = worklist.pop();
4044 for (uint j = 1; j < use->req(); j++) {
4045 Node* n = use->in(j);
4046 if (n == def) {
4047
4048 // "def" is in peel set, "use" is not in peel set
4049 // or "use" is in the entry boundary (a phi) of the peel set
4050
4051 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4052
4053 if ( loop->is_member(get_loop( use_c )) ) {
4054 // use is in loop
4055 if (old_new[use->_idx] != nullptr) { // null for dead code
4056 Node* use_clone = old_new[use->_idx];
4057 _igvn.replace_input_of(use, j, C->top());
4058 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4059 }
4060 } else {
4061 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4062 // use is not in the loop, check if the live range includes the cut
4063 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4064 if (not_peel.test(lp_if->_idx)) {
4065 assert(j == orig_exit_idx, "use from original loop");
4066 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4067 }
4068 }
4069 }
4070 }
4071 }
4072 }
4073 }
4074
4075 // Step 3b: retarget control
4076
4077 // Redirect control to the new loop head if a cloned node in
4078 // the not_peeled region has control that points into the peeled region.
4079 // This necessary because the cloned peeled region will be outside
4080 // the loop.
4081 // from to
4082 // cloned-peeled <---+
4083 // new_head_clone: | <--+
4084 // cloned-not_peeled in(0) in(0)
4085 // orig-peeled
4086
4087 for (uint i = 0; i < loop->_body.size(); i++) {
4088 Node *n = loop->_body.at(i);
4089 if (!n->is_CFG() && n->in(0) != nullptr &&
4090 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4091 Node* n_clone = old_new[n->_idx];
4092 if (n_clone->depends_only_on_test()) {
4093 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4094 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4095 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4096 // floating above its range check.
4097 Node* pinned_clone = n_clone->pin_array_access_node();
4098 if (pinned_clone != nullptr) {
4099 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4100 old_new.map(n->_idx, pinned_clone);
4101 _igvn.replace_node(n_clone, pinned_clone);
4102 n_clone = pinned_clone;
4103 }
4104 }
4105 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4106 }
4107 }
4108
4109 // Backedge of the surviving new_head (the clone) is original last_peel
4110 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4111
4112 // Cut first node in original not_peel set
4113 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4114 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4115 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4116
4117 // Copy head_clone back-branch info to original head
4118 // and remove original head's loop entry and
4119 // clone head's back-branch
4120 _igvn.rehash_node_delayed(head); // Multiple edge updates
4121 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4122 head->set_req(LoopNode::LoopBackControl, C->top());
4123 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4124
4125 // Similarly modify the phis
4126 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4127 Node* use = head->fast_out(k);
4128 if (use->is_Phi() && use->outcnt() > 0) {
4129 Node* use_clone = old_new[use->_idx];
4130 _igvn.rehash_node_delayed(use); // Multiple edge updates
4131 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4132 use->set_req(LoopNode::LoopBackControl, C->top());
4133 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4134 }
4135 }
4136
4137 // Step 4: update dominator tree and dominator depth
4138
4139 set_idom(head, orig_tail_clone, dd);
4140 recompute_dom_depth();
4141
4142 // Inhibit more partial peeling on this loop
4143 new_head_clone->set_partial_peel_loop();
4144 C->set_major_progress();
4145 loop->record_for_igvn();
4146
4147 #ifndef PRODUCT
4148 if (TracePartialPeeling) {
4149 tty->print_cr("\nafter partial peel one iteration");
4150 Node_List wl;
4151 Node* t = last_peel;
4152 while (true) {
4153 wl.push(t);
4154 if (t == head_clone) break;
4155 t = idom(t);
4156 }
4157 while (wl.size() > 0) {
4158 Node* tt = wl.pop();
4159 if (tt == head) tty->print_cr("orig head");
4160 else if (tt == new_head_clone) tty->print_cr("new head");
4161 else if (tt == head_clone) tty->print_cr("clone head");
4162 tt->dump();
4163 }
4164 }
4165 #endif
4166
4167 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4168
4169 return true;
4170 }
4171
4172 // Transform:
4173 //
4174 // loop<-----------------+
4175 // | |
4176 // stmt1 stmt2 .. stmtn |
4177 // | | | |
4178 // \ | / |
4179 // v v v |
4180 // region |
4181 // | |
4182 // shared_stmt |
4183 // | |
4184 // v |
4185 // if |
4186 // / \ |
4187 // | -----------+
4188 // v
4189 //
4190 // into:
4191 //
4192 // loop<-------------------+
4193 // | |
4194 // v |
4195 // +->loop |
4196 // | | |
4197 // | stmt1 stmt2 .. stmtn |
4198 // | | | | |
4199 // | | \ / |
4200 // | | v v |
4201 // | | region1 |
4202 // | | | |
4203 // | shared_stmt shared_stmt |
4204 // | | | |
4205 // | v v |
4206 // | if if |
4207 // | /\ / \ |
4208 // +-- | | -------+
4209 // \ /
4210 // v v
4211 // region2
4212 //
4213 // (region2 is shown to merge mirrored projections of the loop exit
4214 // ifs to make the diagram clearer but they really merge the same
4215 // projection)
4216 //
4217 // Conditions for this transformation to trigger:
4218 // - the path through stmt1 is frequent enough
4219 // - the inner loop will be turned into a counted loop after transformation
4220 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4221 if (!DuplicateBackedge) {
4222 return false;
4223 }
4224 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4225 if (!loop->_head->is_Loop()) {
4226 return false;
4227 }
4228
4229 uint estimate = loop->est_loop_clone_sz(1);
4230 if (exceeding_node_budget(estimate)) {
4231 return false;
4232 }
4233
4234 LoopNode *head = loop->_head->as_Loop();
4235
4236 Node* region = nullptr;
4237 IfNode* exit_test = nullptr;
4238 uint inner;
4239 float f;
4240 if (StressDuplicateBackedge) {
4241 if (head->is_strip_mined()) {
4242 return false;
4243 }
4244 Node* c = head->in(LoopNode::LoopBackControl);
4245
4246 while (c != head) {
4247 if (c->is_Region()) {
4248 region = c;
4249 }
4250 c = idom(c);
4251 }
4252
4253 if (region == nullptr) {
4254 return false;
4255 }
4256
4257 inner = 1;
4258 } else {
4259 // Is the shape of the loop that of a counted loop...
4260 Node* back_control = loop_exit_control(head, loop);
4261 if (back_control == nullptr) {
4262 return false;
4263 }
4264
4265 BoolTest::mask bt = BoolTest::illegal;
4266 float cl_prob = 0;
4267 Node* incr = nullptr;
4268 Node* limit = nullptr;
4269 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4270 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4271 return false;
4272 }
4273
4274 // With an extra phi for the candidate iv?
4275 // Or the region node is the loop head
4276 if (!incr->is_Phi() || incr->in(0) == head) {
4277 return false;
4278 }
4279
4280 PathFrequency pf(head, this);
4281 region = incr->in(0);
4282
4283 // Go over all paths for the extra phi's region and see if that
4284 // path is frequent enough and would match the expected iv shape
4285 // if the extra phi is removed
4286 inner = 0;
4287 for (uint i = 1; i < incr->req(); ++i) {
4288 Node* in = incr->in(i);
4289 Node* trunc1 = nullptr;
4290 Node* trunc2 = nullptr;
4291 const TypeInteger* iv_trunc_t = nullptr;
4292 Node* orig_in = in;
4293 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4294 continue;
4295 }
4296 assert(in->Opcode() == Op_AddI, "wrong increment code");
4297 Node* xphi = nullptr;
4298 Node* stride = loop_iv_stride(in, xphi);
4299
4300 if (stride == nullptr) {
4301 continue;
4302 }
4303
4304 PhiNode* phi = loop_iv_phi(xphi, nullptr, head);
4305 if (phi == nullptr ||
4306 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4307 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4308 return false;
4309 }
4310
4311 f = pf.to(region->in(i));
4312 if (f > 0.5) {
4313 inner = i;
4314 break;
4315 }
4316 }
4317
4318 if (inner == 0) {
4319 return false;
4320 }
4321
4322 exit_test = back_control->in(0)->as_If();
4323 }
4324
4325 if (idom(region)->is_Catch()) {
4326 return false;
4327 }
4328
4329 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4330 Unique_Node_List wq;
4331 wq.push(head->in(LoopNode::LoopBackControl));
4332 for (uint i = 0; i < wq.size(); i++) {
4333 Node* c = wq.at(i);
4334 assert(get_loop(c) == loop, "not in the right loop?");
4335 if (c->is_Region()) {
4336 if (c != region) {
4337 for (uint j = 1; j < c->req(); ++j) {
4338 wq.push(c->in(j));
4339 }
4340 }
4341 } else {
4342 wq.push(c->in(0));
4343 }
4344 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4345 }
4346
4347 Node* region_dom = idom(region);
4348
4349 // Can't do the transformation if this would cause a membar pair to
4350 // be split
4351 for (uint i = 0; i < wq.size(); i++) {
4352 Node* c = wq.at(i);
4353 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4354 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4355 if (!wq.member(c->as_MemBar()->leading_membar())) {
4356 return false;
4357 }
4358 }
4359 }
4360 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head);
4361
4362 // Collect data nodes that need to be clones as well
4363 int dd = dom_depth(head);
4364
4365 for (uint i = 0; i < loop->_body.size(); ++i) {
4366 Node* n = loop->_body.at(i);
4367 if (has_ctrl(n)) {
4368 Node* c = get_ctrl(n);
4369 if (wq.member(c)) {
4370 wq.push(n);
4371 }
4372 } else {
4373 set_idom(n, idom(n), dd);
4374 }
4375 }
4376
4377 // clone shared_stmt
4378 clone_loop_body(wq, old_new, nullptr);
4379
4380 Node* region_clone = old_new[region->_idx];
4381 region_clone->set_req(inner, C->top());
4382 set_idom(region, region->in(inner), dd);
4383
4384 // Prepare the outer loop
4385 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4386 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4387 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4388 set_idom(head, outer_head, dd);
4389
4390 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4391
4392 // Make one of the shared_stmt copies only reachable from stmt1, the
4393 // other only from stmt2..stmtn.
4394 Node* dom = nullptr;
4395 for (uint i = 1; i < region->req(); ++i) {
4396 if (i != inner) {
4397 _igvn.replace_input_of(region, i, C->top());
4398 }
4399 Node* in = region_clone->in(i);
4400 if (in->is_top()) {
4401 continue;
4402 }
4403 if (dom == nullptr) {
4404 dom = in;
4405 } else {
4406 dom = dom_lca(dom, in);
4407 }
4408 }
4409
4410 set_idom(region_clone, dom, dd);
4411
4412 // Set up the outer loop
4413 for (uint i = 0; i < head->outcnt(); i++) {
4414 Node* u = head->raw_out(i);
4415 if (u->is_Phi()) {
4416 Node* outer_phi = u->clone();
4417 outer_phi->set_req(0, outer_head);
4418 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4419 if (backedge == nullptr) {
4420 backedge = u->in(LoopNode::LoopBackControl);
4421 }
4422 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4423 register_new_node(outer_phi, outer_head);
4424 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4425 }
4426 }
4427
4428 // create control and data nodes for out of loop uses (including region2)
4429 Node_List worklist;
4430 uint new_counter = C->unique();
4431 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4432
4433 Node_List *split_if_set = nullptr;
4434 Node_List *split_bool_set = nullptr;
4435 Node_List *split_cex_set = nullptr;
4436 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4437 split_if_set, split_bool_set, split_cex_set);
4438
4439 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4440
4441 if (exit_test != nullptr) {
4442 float cnt = exit_test->_fcnt;
4443 if (cnt != COUNT_UNKNOWN) {
4444 exit_test->_fcnt = cnt * f;
4445 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4446 }
4447 }
4448
4449 C->set_major_progress();
4450
4451 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head);
4452
4453 return true;
4454 }
4455
4456 // AutoVectorize the loop: replace scalar ops with vector ops.
4457 PhaseIdealLoop::AutoVectorizeStatus
4458 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4459 // Counted loop only
4460 if (!lpt->is_counted()) {
4461 return AutoVectorizeStatus::Impossible;
4462 }
4463
4464 // Main-loop only
4465 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4466 if (!cl->is_main_loop()) {
4467 return AutoVectorizeStatus::Impossible;
4468 }
4469
4470 VLoop vloop(lpt, false);
4471 if (!vloop.check_preconditions()) {
4472 return AutoVectorizeStatus::TriedAndFailed;
4473 }
4474
4475 // Ensure the shared data is cleared before each use
4476 vshared.clear();
4477
4478 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4479 if (!vloop_analyzer.success()) {
4480 return AutoVectorizeStatus::TriedAndFailed;
4481 }
4482
4483 SuperWord sw(vloop_analyzer);
4484 if (!sw.transform_loop()) {
4485 return AutoVectorizeStatus::TriedAndFailed;
4486 }
4487
4488 return AutoVectorizeStatus::Success;
4489 }
4490
4491 // Just before insert_pre_post_loops, we can multiversion the loop:
4492 //
4493 // multiversion_if
4494 // | |
4495 // fast_loop slow_loop
4496 //
4497 // In the fast_loop we can make speculative assumptions, and put the
4498 // conditions into the multiversion_if. If the conditions hold at runtime,
4499 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4500 // instead which does not make any of the speculative assumptions.
4501 //
4502 // Note: we only multiversion the loop if the loop does not have any
4503 // auto vectorization check Predicate. If we have that predicate,
4504 // then we can simply add the speculative assumption checks to
4505 // that Predicate. This means we do not need to duplicate the
4506 // loop - we have a smaller graph and save compile time. Should
4507 // the conditions ever fail, then we deopt / trap at the Predicate
4508 // and recompile without that Predicate. At that point we will
4509 // multiversion the loop, so that we can still have speculative
4510 // runtime checks.
4511 //
4512 // We perform the multiversioning when the loop is still in its single
4513 // iteration form, even before we insert pre and post loops. This makes
4514 // the cloning much simpler. However, this means that both the fast
4515 // and the slow loop have to be optimized independently (adding pre
4516 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4517 // we may end up not needing any speculative assumptions in the fast_loop
4518 // and then rejecting the slow_loop by constant folding the multiversion_if.
4519 //
4520 // Therefore, we "delay" the optimization of the slow_loop until we add
4521 // at least one speculative assumption for the fast_loop. If we never
4522 // add such a speculative runtime check, the OpaqueMultiversioningNode
4523 // of the multiversion_if constant folds to true after loop opts, and the
4524 // multiversion_if folds away the "delayed" slow_loop. If we add any
4525 // speculative assumption, then we notify the OpaqueMultiversioningNode
4526 // with "notify_slow_loop_that_it_can_resume_optimizations".
4527 //
4528 // Note: new runtime checks can be added to the multiversion_if with
4529 // PhaseIdealLoop::create_new_if_for_multiversion
4530 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4531 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4532 LoopNode* outer_loop = cl->skip_strip_mined();
4533 Node* entry = outer_loop->in(LoopNode::EntryControl);
4534
4535 // Check we have multiversioning enabled, and are not already multiversioned.
4536 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4537
4538 // Check that we do not have a parse-predicate where we can add the runtime checks
4539 // during auto-vectorization.
4540 const Predicates predicates(entry);
4541 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4542 if (predicate_block->has_parse_predicate()) { return; }
4543
4544 // Check node budget.
4545 uint estimate = lpt->est_loop_clone_sz(2);
4546 if (!may_require_nodes(estimate)) { return; }
4547
4548 do_multiversioning(lpt, old_new);
4549 }
4550
4551 // Returns true if the Reduction node is unordered.
4552 static bool is_unordered_reduction(Node* n) {
4553 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4554 }
4555
4556 // Having ReductionNodes in the loop is expensive. They need to recursively
4557 // fold together the vector values, for every vectorized loop iteration. If
4558 // we encounter the following pattern, we can vector accumulate the values
4559 // inside the loop, and only have a single UnorderedReduction after the loop.
4560 //
4561 // Note: UnorderedReduction represents a ReductionNode which does not require
4562 // calculating in strict order.
4563 //
4564 // CountedLoop init
4565 // | |
4566 // +------+ | +-----------------------+
4567 // | | | |
4568 // PhiNode (s) |
4569 // | |
4570 // | Vector |
4571 // | | |
4572 // UnorderedReduction (first_ur) |
4573 // | |
4574 // ... Vector |
4575 // | | |
4576 // UnorderedReduction (last_ur) |
4577 // | |
4578 // +---------------------+
4579 //
4580 // We patch the graph to look like this:
4581 //
4582 // CountedLoop identity_vector
4583 // | |
4584 // +-------+ | +---------------+
4585 // | | | |
4586 // PhiNode (v) |
4587 // | |
4588 // | Vector |
4589 // | | |
4590 // VectorAccumulator |
4591 // | |
4592 // ... Vector |
4593 // | | |
4594 // init VectorAccumulator |
4595 // | | | |
4596 // UnorderedReduction +-----------+
4597 //
4598 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4599 // use vector_accumulators, which do the same reductions, but only element
4600 // wise. This is a single operation per vector_accumulator, rather than many
4601 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4602 // after the loop, and also reduce the init value into it.
4603 //
4604 // We can not do this with all reductions. Some reductions do not allow the
4605 // reordering of operations (for example float addition/multiplication require
4606 // strict order).
4607 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4608 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4609
4610 // Find all Phi nodes with an unordered Reduction on backedge.
4611 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4612 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4613 Node* phi = cl->fast_out(j);
4614 // We have a phi with a single use, and an unordered Reduction on the backedge.
4615 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4616 continue;
4617 }
4618
4619 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4620 assert(!last_ur->requires_strict_order(), "must be");
4621
4622 // Determine types
4623 const TypeVect* vec_t = last_ur->vect_type();
4624 uint vector_length = vec_t->length();
4625 BasicType bt = vec_t->element_basic_type();
4626
4627 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4628 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4629 const int vopc = VectorNode::opcode(sopc, bt);
4630 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4631 DEBUG_ONLY( last_ur->dump(); )
4632 assert(false, "do not have normal vector op for this reduction");
4633 continue; // not implemented -> fails
4634 }
4635
4636 // Traverse up the chain of unordered Reductions, checking that it loops back to
4637 // the phi. Check that all unordered Reductions only have a single use, except for
4638 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4639 // are outside the loop.
4640 ReductionNode* current = last_ur;
4641 ReductionNode* first_ur = nullptr;
4642 while (true) {
4643 assert(!current->requires_strict_order(), "sanity");
4644
4645 // Expect no ctrl and a vector_input from within the loop.
4646 Node* ctrl = current->in(0);
4647 Node* vector_input = current->in(2);
4648 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4649 DEBUG_ONLY( current->dump(1); )
4650 assert(false, "reduction has ctrl or bad vector_input");
4651 break; // Chain traversal fails.
4652 }
4653
4654 assert(current->vect_type() != nullptr, "must have vector type");
4655 if (current->vect_type() != last_ur->vect_type()) {
4656 // Reductions do not have the same vector type (length and element type).
4657 break; // Chain traversal fails.
4658 }
4659
4660 // Expect single use of an unordered Reduction, except for last_ur.
4661 if (current == last_ur) {
4662 // Expect all uses to be outside the loop, except phi.
4663 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4664 Node* use = current->fast_out(k);
4665 if (use != phi && ctrl_or_self(use) == cl) {
4666 DEBUG_ONLY( current->dump(-1); )
4667 assert(false, "reduction has use inside loop");
4668 // Should not be allowed by SuperWord::mark_reductions
4669 return; // bail out of optimization
4670 }
4671 }
4672 } else {
4673 if (current->outcnt() != 1) {
4674 break; // Chain traversal fails.
4675 }
4676 }
4677
4678 // Expect another unordered Reduction or phi as the scalar input.
4679 Node* scalar_input = current->in(1);
4680 if (is_unordered_reduction(scalar_input) &&
4681 scalar_input->Opcode() == current->Opcode()) {
4682 // Move up the unordered Reduction chain.
4683 current = scalar_input->as_Reduction();
4684 assert(!current->requires_strict_order(), "must be");
4685 } else if (scalar_input == phi) {
4686 // Chain terminates at phi.
4687 first_ur = current;
4688 current = nullptr;
4689 break; // Success.
4690 } else {
4691 // scalar_input is neither phi nor a matching reduction
4692 // Can for example be scalar reduction when we have
4693 // partial vectorization.
4694 break; // Chain traversal fails.
4695 }
4696 }
4697 if (current != nullptr) {
4698 // Chain traversal was not successful.
4699 continue;
4700 }
4701 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4702
4703 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4704 set_root_as_ctrl(identity_scalar);
4705 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4706 register_new_node(identity_vector, C->root());
4707 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4708 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4709
4710 // Turn the scalar phi into a vector phi.
4711 _igvn.rehash_node_delayed(phi);
4712 Node* init = phi->in(1); // Remember init before replacing it.
4713 phi->set_req_X(1, identity_vector, &_igvn);
4714 phi->as_Type()->set_type(vec_t);
4715 _igvn.set_type(phi, vec_t);
4716
4717 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4718 current = first_ur;
4719 while (true) {
4720 // Create vector_accumulator to replace current.
4721 Node* last_vector_accumulator = current->in(1);
4722 Node* vector_input = current->in(2);
4723 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4724 register_new_node(vector_accumulator, cl);
4725 _igvn.replace_node(current, vector_accumulator);
4726 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4727 if (current == last_ur) {
4728 break;
4729 }
4730 current = vector_accumulator->unique_out()->as_Reduction();
4731 assert(!current->requires_strict_order(), "must be");
4732 }
4733
4734 // Create post-loop reduction.
4735 Node* last_accumulator = phi->in(2);
4736 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4737
4738 // Take over uses of last_accumulator that are not in the loop.
4739 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4740 Node* use = last_accumulator->out(i);
4741 if (use != phi && use != post_loop_reduction) {
4742 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4743 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4744 --i;
4745 }
4746 }
4747 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4748 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4749
4750 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4751 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4752 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4753 }
4754 }
4755
4756 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4757 for (uint i = 0; i < _data_nodes.size(); i++) {
4758 clone(_data_nodes[i], new_ctrl);
4759 }
4760 }
4761
4762 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4763 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4764 Node* clone = node->clone();
4765 _phase->igvn().register_new_node_with_optimizer(clone);
4766 _orig_to_new.put(node, clone);
4767 _phase->set_ctrl(clone, new_ctrl);
4768 if (node->is_CastII()) {
4769 clone->set_req(0, new_ctrl);
4770 }
4771 }
4772
4773 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4774 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4775 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4776 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4777 for (uint i = 1; i < node->req(); i++) {
4778 Node** cloned_input = _orig_to_new.get(node->in(i));
4779 if (cloned_input != nullptr) {
4780 // Input was also cloned -> rewire clone to the cloned input.
4781 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4782 }
4783 }
4784 });
4785 }
4786
4787 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4788 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4789 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4790 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4791 Node* new_ctrl) {
4792 for (uint i = 0; i < _data_nodes.size(); i++) {
4793 Node* data_node = _data_nodes[i];
4794 if (data_node->is_Opaque1()) {
4795 transform_opaque_node(transform_strategy, data_node);
4796 } else {
4797 clone(data_node, new_ctrl);
4798 }
4799 }
4800 }
4801
4802 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4803 Node* transformed_node;
4804 if (node->is_OpaqueLoopInit()) {
4805 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4806 } else {
4807 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4808 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4809 }
4810 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4811 _orig_to_new.put(node, transformed_node);
4812 }