1 /*
2 * Copyright (c) 1999, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/divnode.hpp"
35 #include "opto/inlinetypenode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/subtypenode.hpp"
44 #include "opto/superword.hpp"
45 #include "opto/vectornode.hpp"
46 #include "utilities/macros.hpp"
47
48 //=============================================================================
49 //------------------------------split_thru_phi---------------------------------
50 // Split Node 'n' through merge point if there is enough win.
51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
52 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
53 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
54 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
55 // so disable this for now
56 return nullptr;
57 }
58
59 // Splitting range check CastIIs through a loop induction Phi can
60 // cause new Phis to be created that are left unrelated to the loop
61 // induction Phi and prevent optimizations (vectorization)
62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
63 n->in(1) == region->as_CountedLoop()->phi()) {
64 return nullptr;
65 }
66
67 // Inline types should not be split through Phis because they cannot be merged
68 // through Phi nodes but each value input needs to be merged individually.
69 if (n->is_InlineType()) {
70 return nullptr;
71 }
72
73 if (cannot_split_division(n, region)) {
74 return nullptr;
75 }
76
77 int wins = 0;
78 assert(!n->is_CFG(), "");
79 assert(region->is_Region(), "");
80
81 const Type* type = n->bottom_type();
82 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
83 Node* phi;
84 if (t_oop != nullptr && t_oop->is_known_instance_field()) {
85 int iid = t_oop->instance_id();
86 int index = C->get_alias_index(t_oop);
87 int offset = t_oop->offset();
88 phi = new PhiNode(region, type, nullptr, iid, index, offset);
89 } else {
90 phi = PhiNode::make_blank(region, n);
91 }
92 uint old_unique = C->unique();
93 for (uint i = 1; i < region->req(); i++) {
94 Node* x;
95 Node* the_clone = nullptr;
96 if (region->in(i) == C->top()) {
97 x = C->top(); // Dead path? Use a dead data op
98 } else {
99 x = n->clone(); // Else clone up the data op
100 the_clone = x; // Remember for possible deletion.
101 // Alter data node to use pre-phi inputs
102 if (n->in(0) == region)
103 x->set_req( 0, region->in(i) );
104 for (uint j = 1; j < n->req(); j++) {
105 Node* in = n->in(j);
106 if (in->is_Phi() && in->in(0) == region)
107 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
108 }
109 }
110 // Check for a 'win' on some paths
111 const Type* t = x->Value(&_igvn);
112
113 bool singleton = t->singleton();
114
115 // A TOP singleton indicates that there are no possible values incoming
116 // along a particular edge. In most cases, this is OK, and the Phi will
117 // be eliminated later in an Ideal call. However, we can't allow this to
118 // happen if the singleton occurs on loop entry, as the elimination of
119 // the PhiNode may cause the resulting node to migrate back to a previous
120 // loop iteration.
121 if (singleton && t == Type::TOP) {
122 // Is_Loop() == false does not confirm the absence of a loop (e.g., an
123 // irreducible loop may not be indicated by an affirmative is_Loop());
124 // therefore, the only top we can split thru a phi is on a backedge of
125 // a loop.
126 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
127 }
128
129 if (singleton) {
130 wins++;
131 x = makecon(t);
132 } else {
133 // We now call Identity to try to simplify the cloned node.
134 // Note that some Identity methods call phase->type(this).
135 // Make sure that the type array is big enough for
136 // our new node, even though we may throw the node away.
137 // (Note: This tweaking with igvn only works because x is a new node.)
138 _igvn.set_type(x, t);
139 // If x is a TypeNode, capture any more-precise type permanently into Node
140 // otherwise it will be not updated during igvn->transform since
141 // igvn->type(x) is set to x->Value() already.
142 x->raise_bottom_type(t);
143 Node* y = x->Identity(&_igvn);
144 if (y != x) {
145 wins++;
146 x = y;
147 } else {
148 y = _igvn.hash_find(x);
149 if (y == nullptr) {
150 y = similar_subtype_check(x, region->in(i));
151 }
152 if (y) {
153 wins++;
154 x = y;
155 } else {
156 // Else x is a new node we are keeping
157 // We do not need register_new_node_with_optimizer
158 // because set_type has already been called.
159 _igvn._worklist.push(x);
160 }
161 }
162 }
163
164 phi->set_req( i, x );
165
166 if (the_clone == nullptr) {
167 continue;
168 }
169
170 if (the_clone != x) {
171 _igvn.remove_dead_node(the_clone);
172 } else if (region->is_Loop() && i == LoopNode::LoopBackControl &&
173 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) {
174 // it is not a win if 'x' moved from an outer to an inner loop
175 // this edge case can only happen for Load nodes
176 wins = 0;
177 break;
178 }
179 }
180 // Too few wins?
181 if (wins <= policy) {
182 _igvn.remove_dead_node(phi);
183 return nullptr;
184 }
185
186 // Record Phi
187 register_new_node( phi, region );
188
189 for (uint i2 = 1; i2 < phi->req(); i2++) {
190 Node *x = phi->in(i2);
191 // If we commoned up the cloned 'x' with another existing Node,
192 // the existing Node picks up a new use. We need to make the
193 // existing Node occur higher up so it dominates its uses.
194 Node *old_ctrl;
195 IdealLoopTree *old_loop;
196
197 if (x->is_Con()) {
198 assert(get_ctrl(x) == C->root(), "constant control is not root");
199 continue;
200 }
201 // The occasional new node
202 if (x->_idx >= old_unique) { // Found a new, unplaced node?
203 old_ctrl = nullptr;
204 old_loop = nullptr; // Not in any prior loop
205 } else {
206 old_ctrl = get_ctrl(x);
207 old_loop = get_loop(old_ctrl); // Get prior loop
208 }
209 // New late point must dominate new use
210 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
211 if (new_ctrl == old_ctrl) // Nothing is changed
212 continue;
213
214 IdealLoopTree *new_loop = get_loop(new_ctrl);
215
216 // Don't move x into a loop if its uses are
217 // outside of loop. Otherwise x will be cloned
218 // for each use outside of this loop.
219 IdealLoopTree *use_loop = get_loop(region);
220 if (!new_loop->is_member(use_loop) &&
221 (old_loop == nullptr || !new_loop->is_member(old_loop))) {
222 // Take early control, later control will be recalculated
223 // during next iteration of loop optimizations.
224 new_ctrl = get_early_ctrl(x);
225 new_loop = get_loop(new_ctrl);
226 }
227 // Set new location
228 set_ctrl(x, new_ctrl);
229 // If changing loop bodies, see if we need to collect into new body
230 if (old_loop != new_loop) {
231 if (old_loop && !old_loop->_child)
232 old_loop->_body.yank(x);
233 if (!new_loop->_child)
234 new_loop->_body.push(x); // Collect body info
235 }
236 }
237
238 return phi;
239 }
240
241 // Test whether node 'x' can move into an inner loop relative to node 'n'.
242 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop,
243 // BUT it can also return true and 'x' is in the outer loop
244 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) {
245 IdealLoopTree* n_loop_tree = get_loop(n_loop);
246 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x));
247 // x_loop_tree should be outer or same loop as n_loop_tree
248 return !x_loop_tree->is_member(n_loop_tree);
249 }
250
251 // Subtype checks that carry profile data don't common so look for a replacement by following edges
252 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) {
253 if (x->is_SubTypeCheck()) {
254 Node* in1 = x->in(1);
255 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) {
256 Node* u = in1->fast_out(i);
257 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) {
258 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
259 Node* bol = u->fast_out(j);
260 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) {
261 Node* iff = bol->fast_out(k);
262 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with
263 // unrelated profile
264 if (iff->is_If() && is_dominator(iff, r_in)) {
265 return u;
266 }
267 }
268 }
269 }
270 }
271 }
272 return nullptr;
273 }
274
275 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor
276 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In
277 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in
278 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could
279 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis).
280 // We also need to check other loop phis as they could have been created in the same split-if pass when applying
281 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi.
282 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const {
283 const Type* zero;
284 switch (n->Opcode()) {
285 case Op_DivI:
286 case Op_ModI:
287 case Op_UDivI:
288 case Op_UModI:
289 zero = TypeInt::ZERO;
290 break;
291 case Op_DivL:
292 case Op_ModL:
293 case Op_UDivL:
294 case Op_UModL:
295 zero = TypeLong::ZERO;
296 break;
297 default:
298 return false;
299 }
300
301 if (n->in(0) != nullptr) {
302 // Cannot split through phi if Div or Mod node has a control dependency to a zero check.
303 return true;
304 }
305
306 Node* divisor = n->in(2);
307 return is_divisor_loop_phi(divisor, region) &&
308 loop_phi_backedge_type_contains_zero(divisor, zero);
309 }
310
311 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) {
312 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop;
313 }
314
315 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const {
316 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP;
317 }
318
319 //------------------------------dominated_by------------------------------------
320 // Replace the dominated test with an obvious true or false. Place it on the
321 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the
322 // live path up to the dominating control.
323 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) {
324 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
325
326 // prevdom is the dominating projection of the dominating test.
327 assert(iff->Opcode() == Op_If ||
328 iff->Opcode() == Op_CountedLoopEnd ||
329 iff->Opcode() == Op_LongCountedLoopEnd ||
330 iff->Opcode() == Op_RangeCheck ||
331 iff->Opcode() == Op_ParsePredicate,
332 "Check this code when new subtype is added");
333
334 int pop = prevdom->Opcode();
335 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
336 if (flip) {
337 if (pop == Op_IfTrue)
338 pop = Op_IfFalse;
339 else
340 pop = Op_IfTrue;
341 }
342 // 'con' is set to true or false to kill the dominated test.
343 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
344 // Hack the dominated test
345 _igvn.replace_input_of(iff, 1, con);
346
347 // If I don't have a reachable TRUE and FALSE path following the IfNode then
348 // I can assume this path reaches an infinite loop. In this case it's not
349 // important to optimize the data Nodes - either the whole compilation will
350 // be tossed or this path (and all data Nodes) will go dead.
351 if (iff->outcnt() != 2) {
352 return;
353 }
354
355 // Make control-dependent data Nodes on the live path (path that will remain
356 // once the dominated IF is removed) become control-dependent on the
357 // dominating projection.
358 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue);
359
360 if (dp == nullptr) {
361 return;
362 }
363
364 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes);
365 }
366
367 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) {
368 IdealLoopTree* old_loop = get_loop(source);
369
370 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) {
371 Node* out = source->fast_out(i); // Control-dependent node
372 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
373 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) {
374 assert(out->in(0) == source, "must be control dependent on source");
375 _igvn.replace_input_of(out, 0, dominator);
376 if (pin_array_access_nodes) {
377 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range
378 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the
379 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate
380 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
381 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest
382 // dominating check.
383 Node* clone = out->pin_array_access_node();
384 if (clone != nullptr) {
385 clone = _igvn.register_new_node_with_optimizer(clone, out);
386 _igvn.replace_node(out, clone);
387 out = clone;
388 }
389 }
390 set_early_ctrl(out, false);
391 IdealLoopTree* new_loop = get_loop(get_ctrl(out));
392 if (old_loop != new_loop) {
393 if (!old_loop->_child) {
394 old_loop->_body.yank(out);
395 }
396 if (!new_loop->_child) {
397 new_loop->_body.push(out);
398 }
399 }
400 --i;
401 --imax;
402 }
403 }
404 }
405
406 //------------------------------has_local_phi_input----------------------------
407 // Return TRUE if 'n' has Phi inputs from its local block and no other
408 // block-local inputs (all non-local-phi inputs come from earlier blocks)
409 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
410 Node *n_ctrl = get_ctrl(n);
411 // See if some inputs come from a Phi in this block, or from before
412 // this block.
413 uint i;
414 for( i = 1; i < n->req(); i++ ) {
415 Node *phi = n->in(i);
416 if( phi->is_Phi() && phi->in(0) == n_ctrl )
417 break;
418 }
419 if( i >= n->req() )
420 return nullptr; // No Phi inputs; nowhere to clone thru
421
422 // Check for inputs created between 'n' and the Phi input. These
423 // must split as well; they have already been given the chance
424 // (courtesy of a post-order visit) and since they did not we must
425 // recover the 'cost' of splitting them by being very profitable
426 // when splitting 'n'. Since this is unlikely we simply give up.
427 for( i = 1; i < n->req(); i++ ) {
428 Node *m = n->in(i);
429 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
430 // We allow the special case of AddP's with no local inputs.
431 // This allows us to split-up address expressions.
432 if (m->is_AddP() &&
433 get_ctrl(m->in(AddPNode::Base)) != n_ctrl &&
434 get_ctrl(m->in(AddPNode::Address)) != n_ctrl &&
435 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) {
436 // Move the AddP up to the dominating point. That's fine because control of m's inputs
437 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl.
438 Node* c = find_non_split_ctrl(idom(n_ctrl));
439 if (c->is_OuterStripMinedLoop()) {
440 c->as_Loop()->verify_strip_mined(1);
441 c = c->in(LoopNode::EntryControl);
442 }
443 set_ctrl_and_loop(m, c);
444 continue;
445 }
446 return nullptr;
447 }
448 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
449 }
450
451 return n_ctrl;
452 }
453
454 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
455 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) {
456 assert(bt == T_INT || bt == T_LONG, "only for integers");
457 int n_op = n->Opcode();
458
459 if (n_op == Op_LShift(bt)) {
460 // Scale is loop invariant
461 Node* scale = n->in(2);
462 Node* scale_ctrl = get_ctrl(scale);
463 IdealLoopTree* scale_loop = get_loop(scale_ctrl);
464 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) {
465 return nullptr;
466 }
467 const TypeInt* scale_t = scale->bottom_type()->isa_int();
468 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) {
469 return nullptr; // Dont bother with byte/short masking
470 }
471 // Add must vary with loop (else shift would be loop-invariant)
472 Node* add = n->in(1);
473 Node* add_ctrl = get_ctrl(add);
474 IdealLoopTree* add_loop = get_loop(add_ctrl);
475 if (n_loop != add_loop) {
476 return nullptr; // happens w/ evil ZKM loops
477 }
478
479 // Convert I-V into I+ (0-V); same for V-I
480 if (add->Opcode() == Op_Sub(bt) &&
481 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) {
482 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, "");
483 Node* zero = integercon(0, bt);
484 Node* neg = SubNode::make(zero, add->in(2), bt);
485 register_new_node_with_ctrl_of(neg, add->in(2));
486 add = AddNode::make(add->in(1), neg, bt);
487 register_new_node(add, add_ctrl);
488 }
489 if (add->Opcode() != Op_Add(bt)) return nullptr;
490 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, "");
491 // See if one add input is loop invariant
492 Node* add_var = add->in(1);
493 Node* add_var_ctrl = get_ctrl(add_var);
494 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl);
495 Node* add_invar = add->in(2);
496 Node* add_invar_ctrl = get_ctrl(add_invar);
497 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl);
498 if (add_invar_loop == n_loop) {
499 // Swap to find the invariant part
500 add_invar = add_var;
501 add_invar_ctrl = add_var_ctrl;
502 add_invar_loop = add_var_loop;
503 add_var = add->in(2);
504 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant
505 return nullptr;
506 }
507 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) {
508 return nullptr; // No invariant part of the add?
509 }
510
511 // Yes! Reshape address expression!
512 Node* inv_scale = LShiftNode::make(add_invar, scale, bt);
513 Node* inv_scale_ctrl =
514 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
515 add_invar_ctrl : scale_ctrl;
516 register_new_node(inv_scale, inv_scale_ctrl);
517 Node* var_scale = LShiftNode::make(add_var, scale, bt);
518 register_new_node(var_scale, n_ctrl);
519 Node* var_add = AddNode::make(var_scale, inv_scale, bt);
520 register_new_node(var_add, n_ctrl);
521 _igvn.replace_node(n, var_add);
522 return var_add;
523 }
524 return nullptr;
525 }
526
527 //------------------------------remix_address_expressions----------------------
528 // Rework addressing expressions to get the most loop-invariant stuff
529 // moved out. We'd like to do all associative operators, but it's especially
530 // important (common) to do address expressions.
531 Node* PhaseIdealLoop::remix_address_expressions(Node* n) {
532 if (!has_ctrl(n)) return nullptr;
533 Node* n_ctrl = get_ctrl(n);
534 IdealLoopTree* n_loop = get_loop(n_ctrl);
535
536 // See if 'n' mixes loop-varying and loop-invariant inputs and
537 // itself is loop-varying.
538
539 // Only interested in binary ops (and AddP)
540 if (n->req() < 3 || n->req() > 4) return nullptr;
541
542 Node* n1_ctrl = get_ctrl(n->in( 1));
543 Node* n2_ctrl = get_ctrl(n->in( 2));
544 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
545 IdealLoopTree* n1_loop = get_loop(n1_ctrl);
546 IdealLoopTree* n2_loop = get_loop(n2_ctrl);
547 IdealLoopTree* n3_loop = get_loop(n3_ctrl);
548
549 // Does one of my inputs spin in a tighter loop than self?
550 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) ||
551 (n_loop->is_member(n2_loop) && n_loop != n2_loop) ||
552 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) {
553 return nullptr; // Leave well enough alone
554 }
555
556 // Is at least one of my inputs loop-invariant?
557 if (n1_loop == n_loop &&
558 n2_loop == n_loop &&
559 n3_loop == n_loop) {
560 return nullptr; // No loop-invariant inputs
561 }
562
563 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT);
564 if (res != nullptr) {
565 return res;
566 }
567 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG);
568 if (res != nullptr) {
569 return res;
570 }
571
572 int n_op = n->Opcode();
573 // Replace (I+V) with (V+I)
574 if (n_op == Op_AddI ||
575 n_op == Op_AddL ||
576 n_op == Op_AddF ||
577 n_op == Op_AddD ||
578 n_op == Op_MulI ||
579 n_op == Op_MulL ||
580 n_op == Op_MulF ||
581 n_op == Op_MulD) {
582 if (n2_loop == n_loop) {
583 assert(n1_loop != n_loop, "");
584 n->swap_edges(1, 2);
585 }
586 }
587
588 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
589 // but not if I2 is a constant. Skip for irreducible loops.
590 if (n_op == Op_AddP && n_loop->_head->is_Loop()) {
591 if (n2_loop == n_loop && n3_loop != n_loop) {
592 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) {
593 Node* n22_ctrl = get_ctrl(n->in(2)->in(2));
594 Node* n23_ctrl = get_ctrl(n->in(2)->in(3));
595 IdealLoopTree* n22loop = get_loop(n22_ctrl);
596 IdealLoopTree* n23_loop = get_loop(n23_ctrl);
597 if (n22loop != n_loop && n22loop->is_member(n_loop) &&
598 n23_loop == n_loop) {
599 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3));
600 // Stuff new AddP in the loop preheader
601 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
602 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3));
603 register_new_node(add2, n_ctrl);
604 _igvn.replace_node(n, add2);
605 return add2;
606 }
607 }
608 }
609
610 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
611 if (n2_loop != n_loop && n3_loop == n_loop) {
612 if (n->in(3)->Opcode() == Op_AddX) {
613 Node* V = n->in(3)->in(1);
614 Node* I = n->in(3)->in(2);
615 if (is_member(n_loop,get_ctrl(V))) {
616 } else {
617 Node *tmp = V; V = I; I = tmp;
618 }
619 if (!is_member(n_loop,get_ctrl(I))) {
620 Node* add1 = new AddPNode(n->in(1), n->in(2), I);
621 // Stuff new AddP in the loop preheader
622 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl));
623 Node* add2 = new AddPNode(n->in(1), add1, V);
624 register_new_node(add2, n_ctrl);
625 _igvn.replace_node(n, add2);
626 return add2;
627 }
628 }
629 }
630 }
631
632 return nullptr;
633 }
634
635 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
636 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
637 assert(n->Opcode() == Op_AddI, "sanity");
638 Node * nn = nullptr;
639 Node * in1 = n->in(1);
640 Node * in2 = n->in(2);
641 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
642 IdealLoopTree* loop_n = get_loop(get_ctrl(n));
643 if (loop_n->is_counted() &&
644 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) &&
645 Matcher::match_rule_supported(Op_MulAddVS2VI) &&
646 Matcher::match_rule_supported(Op_MulAddS2I)) {
647 Node* mul_in1 = in1->in(1);
648 Node* mul_in2 = in1->in(2);
649 Node* mul_in3 = in2->in(1);
650 Node* mul_in4 = in2->in(2);
651 if (mul_in1->Opcode() == Op_LoadS &&
652 mul_in2->Opcode() == Op_LoadS &&
653 mul_in3->Opcode() == Op_LoadS &&
654 mul_in4->Opcode() == Op_LoadS) {
655 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
656 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
657 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
658 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
659 IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
660 IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
661 // All nodes should be in the same counted loop.
662 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
663 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
664 Node* adr1 = mul_in1->in(MemNode::Address);
665 Node* adr2 = mul_in2->in(MemNode::Address);
666 Node* adr3 = mul_in3->in(MemNode::Address);
667 Node* adr4 = mul_in4->in(MemNode::Address);
668 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
669 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
670 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
671 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
672 register_new_node_with_ctrl_of(nn, n);
673 _igvn.replace_node(n, nn);
674 return nn;
675 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
676 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
677 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
678 register_new_node_with_ctrl_of(nn, n);
679 _igvn.replace_node(n, nn);
680 return nn;
681 }
682 }
683 }
684 }
685 }
686 }
687 return nn;
688 }
689
690 //------------------------------conditional_move-------------------------------
691 // Attempt to replace a Phi with a conditional move. We have some pretty
692 // strict profitability requirements. All Phis at the merge point must
693 // be converted, so we can remove the control flow. We need to limit the
694 // number of c-moves to a small handful. All code that was in the side-arms
695 // of the CFG diamond is now speculatively executed. This code has to be
696 // "cheap enough". We are pretty much limited to CFG diamonds that merge
697 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
698 Node *PhaseIdealLoop::conditional_move( Node *region ) {
699
700 assert(region->is_Region(), "sanity check");
701 if (region->req() != 3) return nullptr;
702
703 // Check for CFG diamond
704 Node *lp = region->in(1);
705 Node *rp = region->in(2);
706 if (!lp || !rp) return nullptr;
707 Node *lp_c = lp->in(0);
708 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr;
709 IfNode *iff = lp_c->as_If();
710
711 // Check for ops pinned in an arm of the diamond.
712 // Can't remove the control flow in this case
713 if (lp->outcnt() > 1) return nullptr;
714 if (rp->outcnt() > 1) return nullptr;
715
716 IdealLoopTree* r_loop = get_loop(region);
717 assert(r_loop == get_loop(iff), "sanity");
718 // Always convert to CMOVE if all results are used only outside this loop.
719 bool used_inside_loop = (r_loop == _ltree_root);
720
721 // Check profitability
722 int cost = 0;
723 int phis = 0;
724 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
725 Node *out = region->fast_out(i);
726 if (!out->is_Phi()) continue; // Ignore other control edges, etc
727 phis++;
728 PhiNode* phi = out->as_Phi();
729 BasicType bt = phi->type()->basic_type();
730 switch (bt) {
731 case T_DOUBLE:
732 case T_FLOAT:
733 if (C->use_cmove()) {
734 continue; //TODO: maybe we want to add some cost
735 }
736 cost += Matcher::float_cmove_cost(); // Could be very expensive
737 break;
738 case T_LONG: {
739 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
740 }
741 case T_INT: // These all CMOV fine
742 case T_ADDRESS: { // (RawPtr)
743 cost++;
744 break;
745 }
746 case T_NARROWOOP: // Fall through
747 case T_OBJECT: { // Base oops are OK, but not derived oops
748 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
749 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
750 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
751 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
752 // have a Phi for the base here that we convert to a CMOVE all is well
753 // and good. But if the base is dead, we'll not make a CMOVE. Later
754 // the allocator will have to produce a base by creating a CMOVE of the
755 // relevant bases. This puts the allocator in the business of
756 // manufacturing expensive instructions, generally a bad plan.
757 // Just Say No to Conditionally-Moved Derived Pointers.
758 if (tp && tp->offset() != 0)
759 return nullptr;
760 cost++;
761 break;
762 }
763 default:
764 return nullptr; // In particular, can't do memory or I/O
765 }
766 // Add in cost any speculative ops
767 for (uint j = 1; j < region->req(); j++) {
768 Node *proj = region->in(j);
769 Node *inp = phi->in(j);
770 if (inp->isa_InlineType()) {
771 // TODO 8302217 This prevents PhiNode::push_inline_types_through
772 return nullptr;
773 }
774 if (get_ctrl(inp) == proj) { // Found local op
775 cost++;
776 // Check for a chain of dependent ops; these will all become
777 // speculative in a CMOV.
778 for (uint k = 1; k < inp->req(); k++)
779 if (get_ctrl(inp->in(k)) == proj)
780 cost += ConditionalMoveLimit; // Too much speculative goo
781 }
782 }
783 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
784 // This will likely Split-If, a higher-payoff operation.
785 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
786 Node* use = phi->fast_out(k);
787 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
788 cost += ConditionalMoveLimit;
789 // Is there a use inside the loop?
790 // Note: check only basic types since CMoveP is pinned.
791 if (!used_inside_loop && is_java_primitive(bt)) {
792 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
793 if (r_loop == u_loop || r_loop->is_member(u_loop)) {
794 used_inside_loop = true;
795 }
796 }
797 }
798 }//for
799 Node* bol = iff->in(1);
800 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt");
801 if (bol->is_OpaqueTemplateAssertionPredicate()) {
802 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes.
803 return nullptr;
804 }
805 if (bol->is_OpaqueMultiversioning()) {
806 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared.");
807 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless,
808 // and will make the multiversion_if constant fold in the next IGVN round.
809 return nullptr;
810 }
811 if (!bol->is_Bool()) {
812 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]);
813 return nullptr;
814 }
815 int cmp_op = bol->in(1)->Opcode();
816 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
817 return nullptr;
818 }
819 // It is expensive to generate flags from a float compare.
820 // Avoid duplicated float compare.
821 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr;
822
823 float infrequent_prob = PROB_UNLIKELY_MAG(3);
824 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
825 if (used_inside_loop) {
826 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo
827
828 // BlockLayoutByFrequency optimization moves infrequent branch
829 // from hot path. No point in CMOV'ing in such case (110 is used
830 // instead of 100 to take into account not exactness of float value).
831 if (BlockLayoutByFrequency) {
832 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
833 }
834 }
835 // Check for highly predictable branch. No point in CMOV'ing if
836 // we are going to predict accurately all the time.
837 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
838 //keep going
839 } else if (iff->_prob < infrequent_prob ||
840 iff->_prob > (1.0f - infrequent_prob))
841 return nullptr;
842
843 // --------------
844 // Now replace all Phis with CMOV's
845 Node *cmov_ctrl = iff->in(0);
846 uint flip = (lp->Opcode() == Op_IfTrue);
847 Node_List wq;
848 while (1) {
849 PhiNode* phi = nullptr;
850 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
851 Node *out = region->fast_out(i);
852 if (out->is_Phi()) {
853 phi = out->as_Phi();
854 break;
855 }
856 }
857 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) {
858 break;
859 }
860 // Move speculative ops
861 wq.push(phi);
862 while (wq.size() > 0) {
863 Node *n = wq.pop();
864 for (uint j = 1; j < n->req(); j++) {
865 Node* m = n->in(j);
866 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) {
867 set_ctrl(m, cmov_ctrl);
868 wq.push(m);
869 }
870 }
871 }
872 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
873 register_new_node(cmov, cmov_ctrl);
874 _igvn.replace_node(phi, cmov);
875 #ifndef PRODUCT
876 if (TraceLoopOpts) {
877 tty->print("CMOV ");
878 r_loop->dump_head();
879 if (Verbose) {
880 bol->in(1)->dump(1);
881 cmov->dump(1);
882 }
883 }
884 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
885 #endif
886 }
887
888 // The useless CFG diamond will fold up later; see the optimization in
889 // RegionNode::Ideal.
890 _igvn._worklist.push(region);
891
892 return iff->in(1);
893 }
894
895 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
896 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
897 Node* u = m->fast_out(i);
898 if (u->is_CFG()) {
899 if (u->is_NeverBranch()) {
900 u = u->as_NeverBranch()->proj_out(0);
901 enqueue_cfg_uses(u, wq);
902 } else {
903 wq.push(u);
904 }
905 }
906 }
907 }
908
909 // Try moving a store out of a loop, right before the loop
910 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
911 // Store has to be first in the loop body
912 IdealLoopTree *n_loop = get_loop(n_ctrl);
913 if (n->is_Store() && n_loop != _ltree_root &&
914 n_loop->is_loop() && n_loop->_head->is_Loop() &&
915 n->in(0) != nullptr) {
916 Node* address = n->in(MemNode::Address);
917 Node* value = n->in(MemNode::ValueIn);
918 Node* mem = n->in(MemNode::Memory);
919 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
920 IdealLoopTree* value_loop = get_loop(get_ctrl(value));
921
922 // - address and value must be loop invariant
923 // - memory must be a memory Phi for the loop
924 // - Store must be the only store on this memory slice in the
925 // loop: if there's another store following this one then value
926 // written at iteration i by the second store could be overwritten
927 // at iteration i+n by the first store: it's not safe to move the
928 // first store out of the loop
929 // - nothing must observe the memory Phi: it guarantees no read
930 // before the store, we are also guaranteed the store post
931 // dominates the loop head (ignoring a possible early
932 // exit). Otherwise there would be extra Phi involved between the
933 // loop's Phi and the store.
934 // - there must be no early exit from the loop before the Store
935 // (such an exit most of the time would be an extra use of the
936 // memory Phi but sometimes is a bottom memory Phi that takes the
937 // store as input).
938
939 if (!n_loop->is_member(address_loop) &&
940 !n_loop->is_member(value_loop) &&
941 mem->is_Phi() && mem->in(0) == n_loop->_head &&
942 mem->outcnt() == 1 &&
943 mem->in(LoopNode::LoopBackControl) == n) {
944
945 assert(n_loop->_tail != nullptr, "need a tail");
946 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
947
948 // Verify that there's no early exit of the loop before the store.
949 bool ctrl_ok = false;
950 {
951 // Follow control from loop head until n, we exit the loop or
952 // we reach the tail
953 ResourceMark rm;
954 Unique_Node_List wq;
955 wq.push(n_loop->_head);
956
957 for (uint next = 0; next < wq.size(); ++next) {
958 Node *m = wq.at(next);
959 if (m == n->in(0)) {
960 ctrl_ok = true;
961 continue;
962 }
963 assert(!has_ctrl(m), "should be CFG");
964 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
965 ctrl_ok = false;
966 break;
967 }
968 enqueue_cfg_uses(m, wq);
969 if (wq.size() > 10) {
970 ctrl_ok = false;
971 break;
972 }
973 }
974 }
975 if (ctrl_ok) {
976 // move the Store
977 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
978 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
979 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
980 // Disconnect the phi now. An empty phi can confuse other
981 // optimizations in this pass of loop opts.
982 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
983 n_loop->_body.yank(mem);
984
985 set_ctrl_and_loop(n, n->in(0));
986
987 return n;
988 }
989 }
990 }
991 return nullptr;
992 }
993
994 // Try moving a store out of a loop, right after the loop
995 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
996 if (n->is_Store() && n->in(0) != nullptr) {
997 Node *n_ctrl = get_ctrl(n);
998 IdealLoopTree *n_loop = get_loop(n_ctrl);
999 // Store must be in a loop
1000 if (n_loop != _ltree_root && !n_loop->_irreducible) {
1001 Node* address = n->in(MemNode::Address);
1002 Node* value = n->in(MemNode::ValueIn);
1003 IdealLoopTree* address_loop = get_loop(get_ctrl(address));
1004 // address must be loop invariant
1005 if (!n_loop->is_member(address_loop)) {
1006 // Store must be last on this memory slice in the loop and
1007 // nothing in the loop must observe it
1008 Node* phi = nullptr;
1009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1010 Node* u = n->fast_out(i);
1011 if (has_ctrl(u)) { // control use?
1012 IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1013 if (!n_loop->is_member(u_loop)) {
1014 continue;
1015 }
1016 if (u->is_Phi() && u->in(0) == n_loop->_head) {
1017 assert(_igvn.type(u) == Type::MEMORY, "bad phi");
1018 // multiple phis on the same slice are possible
1019 if (phi != nullptr) {
1020 return;
1021 }
1022 phi = u;
1023 continue;
1024 }
1025 }
1026 return;
1027 }
1028 if (phi != nullptr) {
1029 // Nothing in the loop before the store (next iteration)
1030 // must observe the stored value
1031 bool mem_ok = true;
1032 {
1033 ResourceMark rm;
1034 Unique_Node_List wq;
1035 wq.push(phi);
1036 for (uint next = 0; next < wq.size() && mem_ok; ++next) {
1037 Node *m = wq.at(next);
1038 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
1039 Node* u = m->fast_out(i);
1040 if (u->is_Store() || u->is_Phi()) {
1041 if (u != n) {
1042 wq.push(u);
1043 mem_ok = (wq.size() <= 10);
1044 }
1045 } else {
1046 mem_ok = false;
1047 break;
1048 }
1049 }
1050 }
1051 }
1052 if (mem_ok) {
1053 // Move the store out of the loop if the LCA of all
1054 // users (except for the phi) is outside the loop.
1055 Node* hook = new Node(1);
1056 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
1057 _igvn.rehash_node_delayed(phi);
1058 int count = phi->replace_edge(n, hook, &_igvn);
1059 assert(count > 0, "inconsistent phi");
1060
1061 // Compute latest point this store can go
1062 Node* lca = get_late_ctrl(n, get_ctrl(n));
1063 if (lca->is_OuterStripMinedLoop()) {
1064 lca = lca->in(LoopNode::EntryControl);
1065 }
1066 if (n_loop->is_member(get_loop(lca))) {
1067 // LCA is in the loop - bail out
1068 _igvn.replace_node(hook, n);
1069 return;
1070 }
1071 #ifdef ASSERT
1072 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
1073 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
1074 n_loop->_head->as_Loop()->verify_strip_mined(1);
1075 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
1076 IdealLoopTree* outer_loop = get_loop(outer);
1077 assert(n_loop->_parent == outer_loop, "broken loop tree");
1078 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
1079 }
1080 #endif
1081 lca = place_outside_loop(lca, n_loop);
1082 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop");
1083 assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1084
1085 // Move store out of the loop
1086 _igvn.replace_node(hook, n->in(MemNode::Memory));
1087 _igvn.replace_input_of(n, 0, lca);
1088 set_ctrl_and_loop(n, lca);
1089
1090 // Disconnect the phi now. An empty phi can confuse other
1091 // optimizations in this pass of loop opts..
1092 if (phi->in(LoopNode::LoopBackControl) == phi) {
1093 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1094 n_loop->_body.yank(phi);
1095 }
1096 }
1097 }
1098 }
1099 }
1100 }
1101 }
1102
1103 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1104 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1105 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1106 // flat array checks out of loops, mainly to enable loop unswitching.
1107 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1108 // Skip checks for more than one array
1109 if (n->req() > 3) {
1110 return;
1111 }
1112 Node* mem = n->in(FlatArrayCheckNode::Memory);
1113 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1114 IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1115 IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1116
1117 // Check if array is loop invariant
1118 if (!check_loop->is_member(ary_loop)) {
1119 // Walk up memory graph from the check until we leave the loop
1120 VectorSet wq;
1121 wq.set(mem->_idx);
1122 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1123 if (mem->is_Phi()) {
1124 mem = mem->in(1);
1125 } else if (mem->is_MergeMem()) {
1126 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1127 } else if (mem->is_Proj()) {
1128 mem = mem->in(0);
1129 } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1130 mem = mem->in(TypeFunc::Memory);
1131 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1132 mem = mem->in(MemNode::Memory);
1133 } else {
1134 #ifdef ASSERT
1135 mem->dump();
1136 #endif
1137 ShouldNotReachHere();
1138 }
1139 if (wq.test_set(mem->_idx)) {
1140 return;
1141 }
1142 }
1143 // Replace memory input and re-compute ctrl to move the check out of the loop
1144 _igvn.replace_input_of(n, 1, mem);
1145 set_ctrl_and_loop(n, get_early_ctrl(n));
1146 Node* bol = n->unique_out();
1147 set_ctrl_and_loop(bol, get_early_ctrl(bol));
1148 }
1149 }
1150
1151 // Split some nodes that take a counted loop phi as input at a counted
1152 // loop can cause vectorization of some expressions to fail
1153 bool PhaseIdealLoop::split_thru_phi_could_prevent_vectorization(Node* n, Node* n_blk) {
1154 if (!n_blk->is_CountedLoop()) {
1155 return false;
1156 }
1157
1158 int opcode = n->Opcode();
1159
1160 if (opcode != Op_AndI &&
1161 opcode != Op_MulI &&
1162 opcode != Op_RotateRight &&
1163 opcode != Op_RShiftI) {
1164 return false;
1165 }
1166
1167 return n->in(1) == n_blk->as_BaseCountedLoop()->phi();
1168 }
1169
1170 //------------------------------split_if_with_blocks_pre-----------------------
1171 // Do the real work in a non-recursive function. Data nodes want to be
1172 // cloned in the pre-order so they can feed each other nicely.
1173 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1174 // Cloning these guys is unlikely to win
1175 int n_op = n->Opcode();
1176 if (n_op == Op_MergeMem) {
1177 return n;
1178 }
1179 if (n->is_Proj()) {
1180 return n;
1181 }
1182
1183 if (n->isa_FlatArrayCheck()) {
1184 move_flat_array_check_out_of_loop(n);
1185 return n;
1186 }
1187
1188 // Do not clone-up CmpFXXX variations, as these are always
1189 // followed by a CmpI
1190 if (n->is_Cmp()) {
1191 return n;
1192 }
1193 // Attempt to use a conditional move instead of a phi/branch
1194 if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1195 Node *cmov = conditional_move( n );
1196 if (cmov) {
1197 return cmov;
1198 }
1199 }
1200 if (n->is_CFG() || n->is_LoadStore()) {
1201 return n;
1202 }
1203 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1204 if (!C->major_progress()) { // If chance of no more loop opts...
1205 _igvn._worklist.push(n); // maybe we'll remove them
1206 }
1207 return n;
1208 }
1209
1210 if (n->is_Con()) {
1211 return n; // No cloning for Con nodes
1212 }
1213
1214 Node *n_ctrl = get_ctrl(n);
1215 if (!n_ctrl) {
1216 return n; // Dead node
1217 }
1218
1219 Node* res = try_move_store_before_loop(n, n_ctrl);
1220 if (res != nullptr) {
1221 return n;
1222 }
1223
1224 // Attempt to remix address expressions for loop invariants
1225 Node *m = remix_address_expressions( n );
1226 if( m ) return m;
1227
1228 if (n_op == Op_AddI) {
1229 Node *nn = convert_add_to_muladd( n );
1230 if ( nn ) return nn;
1231 }
1232
1233 if (n->is_ConstraintCast()) {
1234 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1235 // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1236 // Node control inputs don't necessarily agree with loop control info (due to
1237 // transformations happened in between), thus additional dominance check is needed
1238 // to keep loop info valid.
1239 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1240 _igvn.replace_node(n, dom_cast);
1241 return dom_cast;
1242 }
1243 }
1244
1245 // Determine if the Node has inputs from some local Phi.
1246 // Returns the block to clone thru.
1247 Node *n_blk = has_local_phi_input( n );
1248 if( !n_blk ) return n;
1249
1250 // Do not clone the trip counter through on a CountedLoop
1251 // (messes up the canonical shape).
1252 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) ||
1253 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) {
1254 return n;
1255 }
1256 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination
1257 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) &&
1258 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) {
1259 return n;
1260 }
1261
1262 if (split_thru_phi_could_prevent_vectorization(n, n_blk)) {
1263 return n;
1264 }
1265
1266 // Check for having no control input; not pinned. Allow
1267 // dominating control.
1268 if (n->in(0)) {
1269 Node *dom = idom(n_blk);
1270 if (dom_lca(n->in(0), dom) != n->in(0)) {
1271 return n;
1272 }
1273 }
1274 // Policy: when is it profitable. You must get more wins than
1275 // policy before it is considered profitable. Policy is usually 0,
1276 // so 1 win is considered profitable. Big merges will require big
1277 // cloning, so get a larger policy.
1278 int policy = n_blk->req() >> 2;
1279
1280 // If the loop is a candidate for range check elimination,
1281 // delay splitting through it's phi until a later loop optimization
1282 if (n_blk->is_BaseCountedLoop()) {
1283 IdealLoopTree *lp = get_loop(n_blk);
1284 if (lp && lp->_rce_candidate) {
1285 return n;
1286 }
1287 }
1288
1289 if (must_throttle_split_if()) return n;
1290
1291 // Split 'n' through the merge point if it is profitable
1292 Node *phi = split_thru_phi( n, n_blk, policy );
1293 if (!phi) return n;
1294
1295 // Found a Phi to split thru!
1296 // Replace 'n' with the new phi
1297 _igvn.replace_node( n, phi );
1298 // Moved a load around the loop, 'en-registering' something.
1299 if (n_blk->is_Loop() && n->is_Load() &&
1300 !phi->in(LoopNode::LoopBackControl)->is_Load())
1301 C->set_major_progress();
1302
1303 return phi;
1304 }
1305
1306 static bool merge_point_too_heavy(Compile* C, Node* region) {
1307 // Bail out if the region and its phis have too many users.
1308 int weight = 0;
1309 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1310 weight += region->fast_out(i)->outcnt();
1311 }
1312 int nodes_left = C->max_node_limit() - C->live_nodes();
1313 if (weight * 8 > nodes_left) {
1314 if (PrintOpto) {
1315 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1316 }
1317 return true;
1318 } else {
1319 return false;
1320 }
1321 }
1322
1323 static bool merge_point_safe(Node* region) {
1324 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1325 // having a PhiNode input. This sidesteps the dangerous case where the split
1326 // ConvI2LNode may become TOP if the input Value() does not
1327 // overlap the ConvI2L range, leaving a node which may not dominate its
1328 // uses.
1329 // A better fix for this problem can be found in the BugTraq entry, but
1330 // expediency for Mantis demands this hack.
1331 #ifdef _LP64
1332 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1333 Node* n = region->fast_out(i);
1334 if (n->is_Phi()) {
1335 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1336 Node* m = n->fast_out(j);
1337 if (m->Opcode() == Op_ConvI2L)
1338 return false;
1339 if (m->is_CastII()) {
1340 return false;
1341 }
1342 }
1343 }
1344 }
1345 #endif
1346 return true;
1347 }
1348
1349
1350 //------------------------------place_outside_loop---------------------------------
1351 // Place some computation outside of this loop on the path to the use passed as argument
1352 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const {
1353 Node* head = loop->_head;
1354 assert(!loop->is_member(get_loop(useblock)), "must be outside loop");
1355 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) {
1356 loop = loop->_parent;
1357 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop");
1358 }
1359
1360 // Pick control right outside the loop
1361 for (;;) {
1362 Node* dom = idom(useblock);
1363 if (loop->is_member(get_loop(dom))) {
1364 break;
1365 }
1366 useblock = dom;
1367 }
1368 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control");
1369 return useblock;
1370 }
1371
1372
1373 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1374 if (!n->is_If() || n->is_BaseCountedLoopEnd()) {
1375 return false;
1376 }
1377 if (!n->in(0)->is_Region()) {
1378 return false;
1379 }
1380
1381 Node* region = n->in(0);
1382 Node* dom = idom(region);
1383 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) {
1384 return false;
1385 }
1386 IfNode* dom_if = dom->as_If();
1387 Node* proj_true = dom_if->proj_out(1);
1388 Node* proj_false = dom_if->proj_out(0);
1389
1390 for (uint i = 1; i < region->req(); i++) {
1391 if (is_dominator(proj_true, region->in(i))) {
1392 continue;
1393 }
1394 if (is_dominator(proj_false, region->in(i))) {
1395 continue;
1396 }
1397 return false;
1398 }
1399
1400 return true;
1401 }
1402
1403
1404 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1405 if (must_throttle_split_if()) {
1406 return false;
1407 }
1408
1409 // Do not do 'split-if' if irreducible loops are present.
1410 if (_has_irreducible_loops) {
1411 return false;
1412 }
1413
1414 if (merge_point_too_heavy(C, n_ctrl)) {
1415 return false;
1416 }
1417
1418 // Do not do 'split-if' if some paths are dead. First do dead code
1419 // elimination and then see if its still profitable.
1420 for (uint i = 1; i < n_ctrl->req(); i++) {
1421 if (n_ctrl->in(i) == C->top()) {
1422 return false;
1423 }
1424 }
1425
1426 // If trying to do a 'Split-If' at the loop head, it is only
1427 // profitable if the cmp folds up on BOTH paths. Otherwise we
1428 // risk peeling a loop forever.
1429
1430 // CNC - Disabled for now. Requires careful handling of loop
1431 // body selection for the cloned code. Also, make sure we check
1432 // for any input path not being in the same loop as n_ctrl. For
1433 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1434 // because the alternative loop entry points won't be converted
1435 // into LoopNodes.
1436 IdealLoopTree *n_loop = get_loop(n_ctrl);
1437 for (uint j = 1; j < n_ctrl->req(); j++) {
1438 if (get_loop(n_ctrl->in(j)) != n_loop) {
1439 return false;
1440 }
1441 }
1442
1443 // Check for safety of the merge point.
1444 if (!merge_point_safe(n_ctrl)) {
1445 return false;
1446 }
1447
1448 return true;
1449 }
1450
1451 // Detect if the node is the inner strip-mined loop
1452 // Return: null if it's not the case, or the exit of outer strip-mined loop
1453 static Node* is_inner_of_stripmined_loop(const Node* out) {
1454 Node* out_le = nullptr;
1455
1456 if (out->is_CountedLoopEnd()) {
1457 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1458
1459 if (loop != nullptr && loop->is_strip_mined()) {
1460 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1461 }
1462 }
1463
1464 return out_le;
1465 }
1466
1467 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) {
1468 // If the CmpP is a subtype check for a value that has just been
1469 // loaded from an array, the subtype check guarantees the value
1470 // can't be stored in a flat array and the load of the value
1471 // happens with a flat array check then: push the type check
1472 // through the phi of the flat array check. This needs special
1473 // logic because the subtype check's input is not a phi but a
1474 // LoadKlass that must first be cloned through the phi.
1475 if (n->Opcode() != Op_CmpP) {
1476 return false;
1477 }
1478
1479 Node* klassptr = n->in(1);
1480 Node* klasscon = n->in(2);
1481
1482 if (klassptr->is_DecodeNarrowPtr()) {
1483 klassptr = klassptr->in(1);
1484 }
1485
1486 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1487 return false;
1488 }
1489
1490 if (!klasscon->is_Con()) {
1491 return false;
1492 }
1493
1494 Node* addr = klassptr->in(MemNode::Address);
1495
1496 if (!addr->is_AddP()) {
1497 return false;
1498 }
1499
1500 intptr_t offset;
1501 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1502
1503 if (obj == nullptr) {
1504 return false;
1505 }
1506
1507 assert(obj != nullptr && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?");
1508 if (obj->Opcode() == Op_CastPP) {
1509 obj = obj->in(1);
1510 }
1511
1512 if (!obj->is_Phi()) {
1513 return false;
1514 }
1515
1516 Node* region = obj->in(0);
1517
1518 Node* phi = PhiNode::make_blank(region, n->in(1));
1519 for (uint i = 1; i < region->req(); i++) {
1520 Node* in = obj->in(i);
1521 Node* ctrl = region->in(i);
1522 if (addr->in(AddPNode::Base) != obj) {
1523 Node* cast = addr->in(AddPNode::Base);
1524 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph");
1525 Node* cast_clone = cast->clone();
1526 cast_clone->set_req(0, ctrl);
1527 cast_clone->set_req(1, in);
1528 register_new_node(cast_clone, ctrl);
1529 const Type* tcast = cast_clone->Value(&_igvn);
1530 _igvn.set_type(cast_clone, tcast);
1531 cast_clone->as_Type()->set_type(tcast);
1532 in = cast_clone;
1533 }
1534 Node* addr_clone = addr->clone();
1535 addr_clone->set_req(AddPNode::Base, in);
1536 addr_clone->set_req(AddPNode::Address, in);
1537 register_new_node(addr_clone, ctrl);
1538 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1539 Node* klassptr_clone = klassptr->clone();
1540 klassptr_clone->set_req(2, addr_clone);
1541 register_new_node(klassptr_clone, ctrl);
1542 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1543 if (klassptr != n->in(1)) {
1544 Node* decode = n->in(1);
1545 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1546 Node* decode_clone = decode->clone();
1547 decode_clone->set_req(1, klassptr_clone);
1548 register_new_node(decode_clone, ctrl);
1549 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1550 klassptr_clone = decode_clone;
1551 }
1552 phi->set_req(i, klassptr_clone);
1553 }
1554 register_new_node(phi, region);
1555 Node* orig = n->in(1);
1556 _igvn.replace_input_of(n, 1, phi);
1557 split_if_with_blocks_post(n);
1558 if (n->outcnt() != 0) {
1559 _igvn.replace_input_of(n, 1, orig);
1560 _igvn.remove_dead_node(phi);
1561 }
1562 return true;
1563 }
1564
1565 //------------------------------split_if_with_blocks_post----------------------
1566 // Do the real work in a non-recursive function. CFG hackery wants to be
1567 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1568 // info.
1569 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1570
1571 if (flat_array_element_type_check(n)) {
1572 return;
1573 }
1574
1575 // Cloning Cmp through Phi's involves the split-if transform.
1576 // FastLock is not used by an If
1577 if (n->is_Cmp() && !n->is_FastLock()) {
1578 Node *n_ctrl = get_ctrl(n);
1579 // Determine if the Node has inputs from some local Phi.
1580 // Returns the block to clone thru.
1581 Node *n_blk = has_local_phi_input(n);
1582 if (n_blk != n_ctrl) {
1583 return;
1584 }
1585
1586 if (!can_split_if(n_ctrl)) {
1587 return;
1588 }
1589
1590 if (n->outcnt() != 1) {
1591 return; // Multiple bool's from 1 compare?
1592 }
1593 Node *bol = n->unique_out();
1594 assert(bol->is_Bool(), "expect a bool here");
1595 if (bol->outcnt() != 1) {
1596 return;// Multiple branches from 1 compare?
1597 }
1598 Node *iff = bol->unique_out();
1599
1600 // Check some safety conditions
1601 if (iff->is_If()) { // Classic split-if?
1602 if (iff->in(0) != n_ctrl) {
1603 return; // Compare must be in same blk as if
1604 }
1605 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1606 // Can't split CMove with different control.
1607 if (get_ctrl(iff) != n_ctrl) {
1608 return;
1609 }
1610 if (get_ctrl(iff->in(2)) == n_ctrl ||
1611 get_ctrl(iff->in(3)) == n_ctrl) {
1612 return; // Inputs not yet split-up
1613 }
1614 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1615 return; // Loop-invar test gates loop-varying CMOVE
1616 }
1617 } else {
1618 return; // some other kind of node, such as an Allocate
1619 }
1620
1621 // When is split-if profitable? Every 'win' on means some control flow
1622 // goes dead, so it's almost always a win.
1623 int policy = 0;
1624 // Split compare 'n' through the merge point if it is profitable
1625 Node *phi = split_thru_phi( n, n_ctrl, policy);
1626 if (!phi) {
1627 return;
1628 }
1629
1630 // Found a Phi to split thru!
1631 // Replace 'n' with the new phi
1632 _igvn.replace_node(n, phi);
1633
1634 // Now split the bool up thru the phi
1635 Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1636 guarantee(bolphi != nullptr, "null boolean phi node");
1637
1638 _igvn.replace_node(bol, bolphi);
1639 assert(iff->in(1) == bolphi, "");
1640
1641 if (bolphi->Value(&_igvn)->singleton()) {
1642 return;
1643 }
1644
1645 // Conditional-move? Must split up now
1646 if (!iff->is_If()) {
1647 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1648 _igvn.replace_node(iff, cmovphi);
1649 return;
1650 }
1651
1652 // Now split the IF
1653 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff);
1654 if (TraceLoopOpts) {
1655 tty->print_cr("Split-If");
1656 }
1657 do_split_if(iff);
1658 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff);
1659 return;
1660 }
1661
1662 // Two identical ifs back to back can be merged
1663 if (try_merge_identical_ifs(n)) {
1664 return;
1665 }
1666
1667 // Check for an IF ready to split; one that has its
1668 // condition codes input coming from a Phi at the block start.
1669 int n_op = n->Opcode();
1670
1671 // Check for an IF being dominated by another IF same test
1672 if (n_op == Op_If ||
1673 n_op == Op_RangeCheck) {
1674 Node *bol = n->in(1);
1675 uint max = bol->outcnt();
1676 // Check for same test used more than once?
1677 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) {
1678 // Search up IDOMs to see if this IF is dominated.
1679 Node* cmp = bol->in(1);
1680 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol);
1681
1682 // Now search up IDOMs till cutoff, looking for a dominating test
1683 Node *prevdom = n;
1684 Node *dom = idom(prevdom);
1685 while (dom != cutoff) {
1686 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom &&
1687 safe_for_if_replacement(dom)) {
1688 // It's invalid to move control dependent data nodes in the inner
1689 // strip-mined loop, because:
1690 // 1) break validation of LoopNode::verify_strip_mined()
1691 // 2) move code with side-effect in strip-mined loop
1692 // Move to the exit of outer strip-mined loop in that case.
1693 Node* out_le = is_inner_of_stripmined_loop(dom);
1694 if (out_le != nullptr) {
1695 prevdom = out_le;
1696 }
1697 // Replace the dominated test with an obvious true or false.
1698 // Place it on the IGVN worklist for later cleanup.
1699 C->set_major_progress();
1700 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if,
1701 // to prevent an array load from floating above its range check. There are three cases:
1702 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin
1703 // all its array accesses at that point.
1704 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array
1705 // accesses would start to float, since we don't pin at that point.
1706 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1707 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck &&
1708 prevdom->in(0)->Opcode() != Op_RangeCheck;
1709 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1710 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1711 return;
1712 }
1713 prevdom = dom;
1714 dom = idom(prevdom);
1715 }
1716 }
1717 }
1718
1719 try_sink_out_of_loop(n);
1720 if (C->failing()) {
1721 return;
1722 }
1723
1724 try_move_store_after_loop(n);
1725
1726 // Remove multiple allocations of the same inline type
1727 if (n->is_InlineType()) {
1728 n->as_InlineType()->remove_redundant_allocations(this);
1729 }
1730 }
1731
1732 // Transform:
1733 //
1734 // if (some_condition) {
1735 // // body 1
1736 // } else {
1737 // // body 2
1738 // }
1739 // if (some_condition) {
1740 // // body 3
1741 // } else {
1742 // // body 4
1743 // }
1744 //
1745 // into:
1746 //
1747 //
1748 // if (some_condition) {
1749 // // body 1
1750 // // body 3
1751 // } else {
1752 // // body 2
1753 // // body 4
1754 // }
1755 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) {
1756 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1757 Node *n_ctrl = n->in(0);
1758 IfNode* dom_if = idom(n_ctrl)->as_If();
1759 if (n->in(1) != dom_if->in(1)) {
1760 assert(n->in(1)->in(1)->is_SubTypeCheck() &&
1761 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr ||
1762 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached");
1763 _igvn.replace_input_of(n, 1, dom_if->in(1));
1764 }
1765 ProjNode* dom_proj_true = dom_if->proj_out(1);
1766 ProjNode* dom_proj_false = dom_if->proj_out(0);
1767
1768 // Now split the IF
1769 RegionNode* new_false_region;
1770 RegionNode* new_true_region;
1771 do_split_if(n, &new_false_region, &new_true_region);
1772 assert(new_false_region->req() == new_true_region->req(), "");
1773 #ifdef ASSERT
1774 for (uint i = 1; i < new_false_region->req(); ++i) {
1775 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if");
1776 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");
1777 }
1778 #endif
1779 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test");
1780
1781 // We now have:
1782 // if (some_condition) {
1783 // // body 1
1784 // if (some_condition) {
1785 // body3: // new_true_region
1786 // // body3
1787 // } else {
1788 // goto body4;
1789 // }
1790 // } else {
1791 // // body 2
1792 // if (some_condition) {
1793 // goto body3;
1794 // } else {
1795 // body4: // new_false_region
1796 // // body4;
1797 // }
1798 // }
1799 //
1800
1801 // clone pinned nodes thru the resulting regions
1802 push_pinned_nodes_thru_region(dom_if, new_true_region);
1803 push_pinned_nodes_thru_region(dom_if, new_false_region);
1804
1805 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent
1806 // 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
1807 // unrelated control dependency.
1808 for (uint i = 1; i < new_false_region->req(); i++) {
1809 if (is_dominator(dom_proj_true, new_false_region->in(i))) {
1810 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1811 } else {
1812 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if");
1813 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If());
1814 }
1815 }
1816 return true;
1817 }
1818 return false;
1819 }
1820
1821 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) {
1822 for (DUIterator i = region->outs(); region->has_out(i); i++) {
1823 Node* u = region->out(i);
1824 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) {
1825 continue;
1826 }
1827 assert(u->in(0) == region, "not a control dependent node?");
1828 uint j = 1;
1829 for (; j < u->req(); ++j) {
1830 Node* in = u->in(j);
1831 if (!is_dominator(ctrl_or_self(in), dom_if)) {
1832 break;
1833 }
1834 }
1835 if (j == u->req()) {
1836 Node *phi = PhiNode::make_blank(region, u);
1837 for (uint k = 1; k < region->req(); ++k) {
1838 Node* clone = u->clone();
1839 clone->set_req(0, region->in(k));
1840 register_new_node(clone, region->in(k));
1841 phi->init_req(k, clone);
1842 }
1843 register_new_node(phi, region);
1844 _igvn.replace_node(u, phi);
1845 --i;
1846 }
1847 }
1848 }
1849
1850 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const {
1851 if (!dom->is_CountedLoopEnd()) {
1852 return true;
1853 }
1854 CountedLoopEndNode* le = dom->as_CountedLoopEnd();
1855 CountedLoopNode* cl = le->loopnode();
1856 if (cl == nullptr) {
1857 return true;
1858 }
1859 if (!cl->is_main_loop()) {
1860 return true;
1861 }
1862 if (cl->is_canonical_loop_entry() == nullptr) {
1863 return true;
1864 }
1865 // Further unrolling is possible so loop exit condition might change
1866 return false;
1867 }
1868
1869 // See if a shared loop-varying computation has no loop-varying uses.
1870 // Happens if something is only used for JVM state in uncommon trap exits,
1871 // like various versions of induction variable+offset. Clone the
1872 // computation per usage to allow it to sink out of the loop.
1873 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) {
1874 if (has_ctrl(n) &&
1875 !n->is_Phi() &&
1876 !n->is_Bool() &&
1877 !n->is_Proj() &&
1878 !n->is_MergeMem() &&
1879 !n->is_CMove() &&
1880 !n->is_OpaqueNotNull() &&
1881 !n->is_OpaqueInitializedAssertionPredicate() &&
1882 !n->is_OpaqueTemplateAssertionPredicate() &&
1883 !n->is_Type()) {
1884 Node *n_ctrl = get_ctrl(n);
1885 IdealLoopTree *n_loop = get_loop(n_ctrl);
1886
1887 if (n->in(0) != nullptr) {
1888 IdealLoopTree* loop_ctrl = get_loop(n->in(0));
1889 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) {
1890 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example,
1891 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop).
1892 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not.
1893 Node* maybe_pinned_n = n;
1894 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl);
1895 if (n->depends_only_on_test()) {
1896 Node* pinned_clone = n->pin_array_access_node();
1897 if (pinned_clone != nullptr) {
1898 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1899 // range check for that access. If that condition is replaced by an identical dominating one, then an
1900 // unpinned load would risk floating above its range check.
1901 register_new_node(pinned_clone, n_ctrl);
1902 maybe_pinned_n = pinned_clone;
1903 _igvn.replace_node(n, pinned_clone);
1904 }
1905 }
1906 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl);
1907 }
1908 }
1909 if (n_loop != _ltree_root && n->outcnt() > 1) {
1910 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of
1911 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us.
1912 Node* early_ctrl = compute_early_ctrl(n, n_ctrl);
1913 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now
1914 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops!
1915 if (n->is_Store() || n->is_LoadStore()) {
1916 assert(false, "no node with a side effect");
1917 C->record_failure("no node with a side effect");
1918 return;
1919 }
1920 Node* outer_loop_clone = nullptr;
1921 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) {
1922 Node* u = n->last_out(j); // Clone private computation per use
1923 _igvn.rehash_node_delayed(u);
1924 Node* x = nullptr;
1925 if (n->depends_only_on_test()) {
1926 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a
1927 // range check for that access. If that condition is replaced by an identical dominating one, then an
1928 // unpinned load would risk floating above its range check.
1929 x = n->pin_array_access_node();
1930 }
1931 if (x == nullptr) {
1932 x = n->clone();
1933 }
1934 Node* x_ctrl = nullptr;
1935 if (u->is_Phi()) {
1936 // Replace all uses of normal nodes. Replace Phi uses
1937 // individually, so the separate Nodes can sink down
1938 // different paths.
1939 uint k = 1;
1940 while (u->in(k) != n) k++;
1941 u->set_req(k, x);
1942 // x goes next to Phi input path
1943 x_ctrl = u->in(0)->in(k);
1944 // Find control for 'x' next to use but not inside inner loops.
1945 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1946 --j;
1947 } else { // Normal use
1948 if (has_ctrl(u)) {
1949 x_ctrl = get_ctrl(u);
1950 } else {
1951 x_ctrl = u->in(0);
1952 }
1953 // Find control for 'x' next to use but not inside inner loops.
1954 x_ctrl = place_outside_loop(x_ctrl, n_loop);
1955 // Replace all uses
1956 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) {
1957 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary
1958 // anymore now that we're going to pin n as well
1959 _igvn.replace_node(u, x);
1960 --j;
1961 } else {
1962 int nb = u->replace_edge(n, x, &_igvn);
1963 j -= nb;
1964 }
1965 }
1966
1967 if (n->is_Load()) {
1968 // For loads, add a control edge to a CFG node outside of the loop
1969 // to force them to not combine and return back inside the loop
1970 // during GVN optimization (4641526).
1971 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked");
1972
1973 IdealLoopTree* x_loop = get_loop(x_ctrl);
1974 Node* x_head = x_loop->_head;
1975 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) {
1976 // Do not add duplicate LoadNodes to the outer strip mined loop
1977 if (outer_loop_clone != nullptr) {
1978 _igvn.replace_node(x, outer_loop_clone);
1979 continue;
1980 }
1981 outer_loop_clone = x;
1982 }
1983 x->set_req(0, x_ctrl);
1984 } else if (n->in(0) != nullptr){
1985 x->set_req(0, x_ctrl);
1986 }
1987 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1988 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop");
1989 register_new_node(x, x_ctrl);
1990
1991 // Chain of AddP nodes: (AddP base (AddP base (AddP base )))
1992 // All AddP nodes must keep the same base after sinking so:
1993 // 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,
1994 // their bases remain the same.
1995 // (see 2- below)
1996 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() ||
1997 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) ||
1998 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape");
1999 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() &&
2000 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) {
2001 assert(!x->is_Load(), "load should be pinned");
2002 // Use a cast node to pin clone out of loop
2003 Node* cast = nullptr;
2004 for (uint k = 0; k < x->req(); k++) {
2005 Node* in = x->in(k);
2006 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) {
2007 const Type* in_t = _igvn.type(in);
2008 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t,
2009 ConstraintCastNode::UnconditionalDependency, nullptr);
2010 }
2011 if (cast != nullptr) {
2012 Node* prev = _igvn.hash_find_insert(cast);
2013 if (prev != nullptr && get_ctrl(prev) == x_ctrl) {
2014 cast->destruct(&_igvn);
2015 cast = prev;
2016 } else {
2017 register_new_node(cast, x_ctrl);
2018 }
2019 x->replace_edge(in, cast);
2020 // Chain of AddP nodes:
2021 // 2- A CastPP of the base is only added now that all AddP nodes are sunk
2022 if (x->is_AddP() && k == AddPNode::Base) {
2023 update_addp_chain_base(x, n->in(AddPNode::Base), cast);
2024 }
2025 break;
2026 }
2027 }
2028 assert(cast != nullptr, "must have added a cast to pin the node");
2029 }
2030 }
2031 _igvn.remove_dead_node(n);
2032 }
2033 _dom_lca_tags_round = 0;
2034 }
2035 }
2036 }
2037
2038 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) {
2039 ResourceMark rm;
2040 Node_List wq;
2041 wq.push(x);
2042 while (wq.size() != 0) {
2043 Node* n = wq.pop();
2044 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2045 Node* u = n->fast_out(i);
2046 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) {
2047 _igvn.replace_input_of(u, AddPNode::Base, new_base);
2048 wq.push(u);
2049 }
2050 }
2051 }
2052 }
2053
2054 // Compute the early control of a node by following its inputs until we reach
2055 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes.
2056 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) {
2057 Node* early_ctrl = nullptr;
2058 ResourceMark rm;
2059 Unique_Node_List wq;
2060 wq.push(n);
2061 for (uint i = 0; i < wq.size(); i++) {
2062 Node* m = wq.at(i);
2063 Node* c = nullptr;
2064 if (m->is_CFG()) {
2065 c = m;
2066 } else if (m->pinned()) {
2067 c = m->in(0);
2068 } else {
2069 for (uint j = 0; j < m->req(); j++) {
2070 Node* in = m->in(j);
2071 if (in != nullptr) {
2072 wq.push(in);
2073 }
2074 }
2075 }
2076 if (c != nullptr) {
2077 assert(is_dominator(c, n_ctrl), "control input must dominate current control");
2078 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) {
2079 early_ctrl = c;
2080 }
2081 }
2082 }
2083 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control");
2084 return early_ctrl;
2085 }
2086
2087 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) {
2088 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2089 Node* u = n->fast_out(i);
2090 if (u->is_Opaque1()) {
2091 return false; // Found loop limit, bugfix for 4677003
2092 }
2093 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to
2094 // get_late_ctrl_with_anti_dep() use their own tag
2095 _dom_lca_tags_round++;
2096 assert(_dom_lca_tags_round != 0, "shouldn't wrap around");
2097
2098 if (u->is_Phi()) {
2099 for (uint j = 1; j < u->req(); ++j) {
2100 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) {
2101 return false;
2102 }
2103 }
2104 } else {
2105 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0);
2106 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) {
2107 return false;
2108 }
2109 }
2110 }
2111 return true;
2112 }
2113
2114 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) {
2115 if (n->is_Load()) {
2116 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl);
2117 }
2118 IdealLoopTree *u_loop = get_loop(ctrl);
2119 if (u_loop == n_loop) {
2120 return false; // Found loop-varying use
2121 }
2122 if (n_loop->is_member(u_loop)) {
2123 return false; // Found use in inner loop
2124 }
2125 // 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
2126 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit
2127 // test of the pre loop above the point in the graph where it's pinned.
2128 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) {
2129 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop();
2130 if (is_dominator(pre_loop->loopexit(), ctrl)) {
2131 return false;
2132 }
2133 }
2134 return true;
2135 }
2136
2137 //------------------------------split_if_with_blocks---------------------------
2138 // Check for aggressive application of 'split-if' optimization,
2139 // using basic block level info.
2140 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
2141 Node* root = C->root();
2142 visited.set(root->_idx); // first, mark root as visited
2143 // Do pre-visit work for root
2144 Node* n = split_if_with_blocks_pre(root);
2145 uint cnt = n->outcnt();
2146 uint i = 0;
2147
2148 while (true) {
2149 // Visit all children
2150 if (i < cnt) {
2151 Node* use = n->raw_out(i);
2152 ++i;
2153 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
2154 // Now do pre-visit work for this use
2155 use = split_if_with_blocks_pre(use);
2156 nstack.push(n, i); // Save parent and next use's index.
2157 n = use; // Process all children of current use.
2158 cnt = use->outcnt();
2159 i = 0;
2160 }
2161 }
2162 else {
2163 // All of n's children have been processed, complete post-processing.
2164 if (cnt != 0 && !n->is_Con()) {
2165 assert(has_node(n), "no dead nodes");
2166 split_if_with_blocks_post(n);
2167 if (C->failing()) {
2168 return;
2169 }
2170 }
2171 if (must_throttle_split_if()) {
2172 nstack.clear();
2173 }
2174 if (nstack.is_empty()) {
2175 // Finished all nodes on stack.
2176 break;
2177 }
2178 // Get saved parent node and next use's index. Visit the rest of uses.
2179 n = nstack.node();
2180 cnt = n->outcnt();
2181 i = nstack.index();
2182 nstack.pop();
2183 }
2184 }
2185 }
2186
2187
2188 //=============================================================================
2189 //
2190 // C L O N E A L O O P B O D Y
2191 //
2192
2193 //------------------------------clone_iff--------------------------------------
2194 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2195 // "Nearly" because all Nodes have been cloned from the original in the loop,
2196 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2197 // through the Phi recursively, and return a Bool.
2198 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) {
2199
2200 // Convert this Phi into a Phi merging Bools
2201 uint i;
2202 for (i = 1; i < phi->req(); i++) {
2203 Node* b = phi->in(i);
2204 if (b->is_Phi()) {
2205 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2206 } else {
2207 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2208 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2209 }
2210 }
2211 Node* n = phi->in(1);
2212 Node* sample_opaque = nullptr;
2213 Node *sample_bool = nullptr;
2214 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2215 sample_opaque = n;
2216 sample_bool = n->in(1);
2217 assert(sample_bool->is_Bool(), "wrong type");
2218 } else {
2219 sample_bool = n;
2220 }
2221 Node* sample_cmp = sample_bool->in(1);
2222 const Type* t = Type::TOP;
2223 const TypePtr* at = nullptr;
2224 if (sample_cmp->is_FlatArrayCheck()) {
2225 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2226 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2227 t = Type::MEMORY;
2228 at = TypeRawPtr::BOTTOM;
2229 }
2230
2231 // Make Phis to merge the Cmp's inputs.
2232 PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2233 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2234 for (i = 1; i < phi->req(); i++) {
2235 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2236 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2237 phi1->set_req(i, n1);
2238 phi2->set_req(i, n2);
2239 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2240 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2241 }
2242 // See if these Phis have been made before.
2243 // Register with optimizer
2244 Node *hit1 = _igvn.hash_find_insert(phi1);
2245 if (hit1) { // Hit, toss just made Phi
2246 _igvn.remove_dead_node(phi1); // Remove new phi
2247 assert(hit1->is_Phi(), "" );
2248 phi1 = (PhiNode*)hit1; // Use existing phi
2249 } else { // Miss
2250 _igvn.register_new_node_with_optimizer(phi1);
2251 }
2252 Node *hit2 = _igvn.hash_find_insert(phi2);
2253 if (hit2) { // Hit, toss just made Phi
2254 _igvn.remove_dead_node(phi2); // Remove new phi
2255 assert(hit2->is_Phi(), "" );
2256 phi2 = (PhiNode*)hit2; // Use existing phi
2257 } else { // Miss
2258 _igvn.register_new_node_with_optimizer(phi2);
2259 }
2260 // Register Phis with loop/block info
2261 set_ctrl(phi1, phi->in(0));
2262 set_ctrl(phi2, phi->in(0));
2263 // Make a new Cmp
2264 Node *cmp = sample_cmp->clone();
2265 cmp->set_req(1, phi1);
2266 cmp->set_req(2, phi2);
2267 _igvn.register_new_node_with_optimizer(cmp);
2268 set_ctrl(cmp, phi->in(0));
2269
2270 // Make a new Bool
2271 Node *b = sample_bool->clone();
2272 b->set_req(1,cmp);
2273 _igvn.register_new_node_with_optimizer(b);
2274 set_ctrl(b, phi->in(0));
2275
2276 if (sample_opaque != nullptr) {
2277 Node* opaque = sample_opaque->clone();
2278 opaque->set_req(1, b);
2279 _igvn.register_new_node_with_optimizer(opaque);
2280 set_ctrl(opaque, phi->in(0));
2281 return opaque;
2282 }
2283
2284 assert(b->is_Bool(), "");
2285 return b;
2286 }
2287
2288 //------------------------------clone_bool-------------------------------------
2289 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
2290 // "Nearly" because all Nodes have been cloned from the original in the loop,
2291 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
2292 // through the Phi recursively, and return a Bool.
2293 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) {
2294 uint i;
2295 // Convert this Phi into a Phi merging Bools
2296 for( i = 1; i < phi->req(); i++ ) {
2297 Node *b = phi->in(i);
2298 if( b->is_Phi() ) {
2299 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi()));
2300 } else {
2301 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
2302 }
2303 }
2304
2305 Node *sample_cmp = phi->in(1);
2306
2307 // Make Phis to merge the Cmp's inputs.
2308 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
2309 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
2310 for( uint j = 1; j < phi->req(); j++ ) {
2311 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
2312 Node *n1, *n2;
2313 if( cmp_top->is_Cmp() ) {
2314 n1 = cmp_top->in(1);
2315 n2 = cmp_top->in(2);
2316 } else {
2317 n1 = n2 = cmp_top;
2318 }
2319 phi1->set_req( j, n1 );
2320 phi2->set_req( j, n2 );
2321 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2322 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2323 }
2324
2325 // See if these Phis have been made before.
2326 // Register with optimizer
2327 Node *hit1 = _igvn.hash_find_insert(phi1);
2328 if( hit1 ) { // Hit, toss just made Phi
2329 _igvn.remove_dead_node(phi1); // Remove new phi
2330 assert( hit1->is_Phi(), "" );
2331 phi1 = (PhiNode*)hit1; // Use existing phi
2332 } else { // Miss
2333 _igvn.register_new_node_with_optimizer(phi1);
2334 }
2335 Node *hit2 = _igvn.hash_find_insert(phi2);
2336 if( hit2 ) { // Hit, toss just made Phi
2337 _igvn.remove_dead_node(phi2); // Remove new phi
2338 assert( hit2->is_Phi(), "" );
2339 phi2 = (PhiNode*)hit2; // Use existing phi
2340 } else { // Miss
2341 _igvn.register_new_node_with_optimizer(phi2);
2342 }
2343 // Register Phis with loop/block info
2344 set_ctrl(phi1, phi->in(0));
2345 set_ctrl(phi2, phi->in(0));
2346 // Make a new Cmp
2347 Node *cmp = sample_cmp->clone();
2348 cmp->set_req( 1, phi1 );
2349 cmp->set_req( 2, phi2 );
2350 _igvn.register_new_node_with_optimizer(cmp);
2351 set_ctrl(cmp, phi->in(0));
2352
2353 assert( cmp->is_Cmp(), "" );
2354 return (CmpNode*)cmp;
2355 }
2356
2357 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
2358 IdealLoopTree* loop, IdealLoopTree* outer_loop,
2359 Node_List*& split_if_set, Node_List*& split_bool_set,
2360 Node_List*& split_cex_set, Node_List& worklist,
2361 uint new_counter, CloneLoopMode mode) {
2362 Node* nnn = old_new[old->_idx];
2363 // Copy uses to a worklist, so I can munge the def-use info
2364 // with impunity.
2365 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2366 worklist.push(old->fast_out(j));
2367
2368 while( worklist.size() ) {
2369 Node *use = worklist.pop();
2370 if (!has_node(use)) continue; // Ignore dead nodes
2371 if (use->in(0) == C->top()) continue;
2372 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2373 // Check for data-use outside of loop - at least one of OLD or USE
2374 // must not be a CFG node.
2375 #ifdef ASSERT
2376 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) {
2377 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
2378 assert(mode != IgnoreStripMined, "incorrect cloning mode");
2379 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
2380 }
2381 #endif
2382 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
2383
2384 // If the Data use is an IF, that means we have an IF outside the
2385 // loop that is switching on a condition that is set inside the
2386 // loop. Happens if people set a loop-exit flag; then test the flag
2387 // in the loop to break the loop, then test is again outside the
2388 // loop to determine which way the loop exited.
2389 //
2390 // For several uses we need to make sure that there is no phi between,
2391 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here
2392 // to avoid such a phi in between.
2393 // For example, it is unexpected that there is a Phi between an
2394 // AllocateArray node and its ValidLengthTest input that could cause
2395 // split if to break.
2396 assert(!use->is_OpaqueTemplateAssertionPredicate(),
2397 "should not clone a Template Assertion Predicate which should be removed once it's useless");
2398 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() ||
2399 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) {
2400 // Since this code is highly unlikely, we lazily build the worklist
2401 // of such Nodes to go split.
2402 if (!split_if_set) {
2403 split_if_set = new Node_List();
2404 }
2405 split_if_set->push(use);
2406 }
2407 if (use->is_Bool()) {
2408 if (!split_bool_set) {
2409 split_bool_set = new Node_List();
2410 }
2411 split_bool_set->push(use);
2412 }
2413 if (use->Opcode() == Op_CreateEx) {
2414 if (!split_cex_set) {
2415 split_cex_set = new Node_List();
2416 }
2417 split_cex_set->push(use);
2418 }
2419
2420
2421 // Get "block" use is in
2422 uint idx = 0;
2423 while( use->in(idx) != old ) idx++;
2424 Node *prev = use->is_CFG() ? use : get_ctrl(use);
2425 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
2426 Node* cfg = (prev->_idx >= new_counter && prev->is_Region())
2427 ? prev->in(2)
2428 : idom(prev);
2429 if( use->is_Phi() ) // Phi use is in prior block
2430 cfg = prev->in(idx); // NOT in block of Phi itself
2431 if (cfg->is_top()) { // Use is dead?
2432 _igvn.replace_input_of(use, idx, C->top());
2433 continue;
2434 }
2435
2436 // If use is referenced through control edge... (idx == 0)
2437 if (mode == IgnoreStripMined && idx == 0) {
2438 LoopNode *head = loop->_head->as_Loop();
2439 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
2440 // That node is outside the inner loop, leave it outside the
2441 // outer loop as well to not confuse verification code.
2442 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
2443 _igvn.replace_input_of(use, 0, head->outer_loop_exit());
2444 continue;
2445 }
2446 }
2447
2448 while(!outer_loop->is_member(get_loop(cfg))) {
2449 prev = cfg;
2450 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg);
2451 }
2452 // If the use occurs after merging several exits from the loop, then
2453 // old value must have dominated all those exits. Since the same old
2454 // value was used on all those exits we did not need a Phi at this
2455 // merge point. NOW we do need a Phi here. Each loop exit value
2456 // is now merged with the peeled body exit; each exit gets its own
2457 // private Phi and those Phis need to be merged here.
2458 Node *phi;
2459 if( prev->is_Region() ) {
2460 if( idx == 0 ) { // Updating control edge?
2461 phi = prev; // Just use existing control
2462 } else { // Else need a new Phi
2463 phi = PhiNode::make( prev, old );
2464 // Now recursively fix up the new uses of old!
2465 for( uint i = 1; i < prev->req(); i++ ) {
2466 worklist.push(phi); // Onto worklist once for each 'old' input
2467 }
2468 }
2469 } else {
2470 // Get new RegionNode merging old and new loop exits
2471 prev = old_new[prev->_idx];
2472 assert( prev, "just made this in step 7" );
2473 if( idx == 0) { // Updating control edge?
2474 phi = prev; // Just use existing control
2475 } else { // Else need a new Phi
2476 // Make a new Phi merging data values properly
2477 phi = PhiNode::make( prev, old );
2478 phi->set_req( 1, nnn );
2479 }
2480 }
2481 // If inserting a new Phi, check for prior hits
2482 if( idx != 0 ) {
2483 Node *hit = _igvn.hash_find_insert(phi);
2484 if( hit == nullptr ) {
2485 _igvn.register_new_node_with_optimizer(phi); // Register new phi
2486 } else { // or
2487 // Remove the new phi from the graph and use the hit
2488 _igvn.remove_dead_node(phi);
2489 phi = hit; // Use existing phi
2490 }
2491 set_ctrl(phi, prev);
2492 }
2493 // Make 'use' use the Phi instead of the old loop body exit value
2494 assert(use->in(idx) == old, "old is still input of use");
2495 // We notify all uses of old, including use, and the indirect uses,
2496 // that may now be optimized because we have replaced old with phi.
2497 _igvn.add_users_to_worklist(old);
2498 if (idx == 0 &&
2499 use->depends_only_on_test()) {
2500 Node* pinned_clone = use->pin_array_access_node();
2501 if (pinned_clone != nullptr) {
2502 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
2503 // into the region is left, an array load could become dependent on a condition that's not a range check for
2504 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
2505 // floating above its range check.
2506 pinned_clone->set_req(0, phi);
2507 register_new_node_with_ctrl_of(pinned_clone, use);
2508 _igvn.replace_node(use, pinned_clone);
2509 continue;
2510 }
2511 }
2512 _igvn.replace_input_of(use, idx, phi);
2513 if( use->_idx >= new_counter ) { // If updating new phis
2514 // Not needed for correctness, but prevents a weak assert
2515 // in AddPNode from tripping (when we end up with different
2516 // base & derived Phis that will become the same after
2517 // IGVN does CSE).
2518 Node *hit = _igvn.hash_find_insert(use);
2519 if( hit ) // Go ahead and re-hash for hits.
2520 _igvn.replace_node( use, hit );
2521 }
2522 }
2523 }
2524 }
2525
2526 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
2527 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
2528 bool check_old_new) {
2529 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2530 Node* u = n->fast_out(j);
2531 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned");
2532 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) {
2533 Node* c = phase->get_ctrl(u);
2534 IdealLoopTree* u_loop = phase->get_loop(c);
2535 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only");
2536 if (!loop->is_member(u_loop)) {
2537 if (outer_loop->is_member(u_loop)) {
2538 wq.push(u);
2539 } else {
2540 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of
2541 // the outer loop too
2542 Node* u_c = u->in(0);
2543 if (u_c != nullptr) {
2544 IdealLoopTree* u_c_loop = phase->get_loop(u_c);
2545 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) {
2546 wq.push(u);
2547 }
2548 }
2549 }
2550 }
2551 }
2552 }
2553 }
2554
2555 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
2556 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
2557 Node_List& extra_data_nodes) {
2558 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2559 CountedLoopNode* cl = head->as_CountedLoop();
2560 Node* l = cl->outer_loop();
2561 Node* tail = cl->outer_loop_tail();
2562 IfNode* le = cl->outer_loop_end();
2563 Node* sfpt = cl->outer_safepoint();
2564 CountedLoopEndNode* cle = cl->loopexit();
2565 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
2566 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
2567 Node* cle_out = cle->proj_out(false);
2568
2569 Node* new_sfpt = nullptr;
2570 Node* new_cle_out = cle_out->clone();
2571 old_new.map(cle_out->_idx, new_cle_out);
2572 if (mode == CloneIncludesStripMined) {
2573 // clone outer loop body
2574 Node* new_l = l->clone();
2575 Node* new_tail = tail->clone();
2576 IfNode* new_le = le->clone()->as_If();
2577 new_sfpt = sfpt->clone();
2578
2579 set_loop(new_l, outer_loop->_parent);
2580 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
2581 set_loop(new_cle_out, outer_loop->_parent);
2582 set_idom(new_cle_out, new_cle, dd);
2583 set_loop(new_sfpt, outer_loop->_parent);
2584 set_idom(new_sfpt, new_cle_out, dd);
2585 set_loop(new_le, outer_loop->_parent);
2586 set_idom(new_le, new_sfpt, dd);
2587 set_loop(new_tail, outer_loop->_parent);
2588 set_idom(new_tail, new_le, dd);
2589 set_idom(new_cl, new_l, dd);
2590
2591 old_new.map(l->_idx, new_l);
2592 old_new.map(tail->_idx, new_tail);
2593 old_new.map(le->_idx, new_le);
2594 old_new.map(sfpt->_idx, new_sfpt);
2595
2596 new_l->set_req(LoopNode::LoopBackControl, new_tail);
2597 new_l->set_req(0, new_l);
2598 new_tail->set_req(0, new_le);
2599 new_le->set_req(0, new_sfpt);
2600 new_sfpt->set_req(0, new_cle_out);
2601 new_cle_out->set_req(0, new_cle);
2602 new_cl->set_req(LoopNode::EntryControl, new_l);
2603
2604 _igvn.register_new_node_with_optimizer(new_l);
2605 _igvn.register_new_node_with_optimizer(new_tail);
2606 _igvn.register_new_node_with_optimizer(new_le);
2607 } else {
2608 Node *newhead = old_new[loop->_head->_idx];
2609 newhead->as_Loop()->clear_strip_mined();
2610 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
2611 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2612 }
2613 // Look at data node that were assigned a control in the outer
2614 // loop: they are kept in the outer loop by the safepoint so start
2615 // from the safepoint node's inputs.
2616 IdealLoopTree* outer_loop = get_loop(l);
2617 Node_Stack stack(2);
2618 stack.push(sfpt, 1);
2619 uint new_counter = C->unique();
2620 while (stack.size() > 0) {
2621 Node* n = stack.node();
2622 uint i = stack.index();
2623 while (i < n->req() &&
2624 (n->in(i) == nullptr ||
2625 !has_ctrl(n->in(i)) ||
2626 get_loop(get_ctrl(n->in(i))) != outer_loop ||
2627 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
2628 i++;
2629 }
2630 if (i < n->req()) {
2631 stack.set_index(i+1);
2632 stack.push(n->in(i), 0);
2633 } else {
2634 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
2635 Node* m = n == sfpt ? new_sfpt : n->clone();
2636 if (m != nullptr) {
2637 for (uint i = 0; i < n->req(); i++) {
2638 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) {
2639 m->set_req(i, old_new[m->in(i)->_idx]);
2640 }
2641 }
2642 } else {
2643 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2644 }
2645 if (n != sfpt) {
2646 extra_data_nodes.push(n);
2647 _igvn.register_new_node_with_optimizer(m);
2648 assert(get_ctrl(n) == cle_out, "what other control?");
2649 set_ctrl(m, new_cle_out);
2650 old_new.map(n->_idx, m);
2651 }
2652 stack.pop();
2653 }
2654 }
2655 if (mode == CloneIncludesStripMined) {
2656 _igvn.register_new_node_with_optimizer(new_sfpt);
2657 _igvn.register_new_node_with_optimizer(new_cle_out);
2658 }
2659 // Some other transformation may have pessimistically assigned some
2660 // data nodes to the outer loop. Set their control so they are out
2661 // of the outer loop.
2662 ResourceMark rm;
2663 Unique_Node_List wq;
2664 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2665 Node* old = extra_data_nodes.at(i);
2666 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2667 }
2668
2669 for (uint i = 0; i < loop->_body.size(); i++) {
2670 Node* old = loop->_body.at(i);
2671 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true);
2672 }
2673
2674 Node* inner_out = sfpt->in(0);
2675 if (inner_out->outcnt() > 1) {
2676 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true);
2677 }
2678
2679 Node* new_ctrl = cl->outer_loop_exit();
2680 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2681 for (uint i = 0; i < wq.size(); i++) {
2682 Node* n = wq.at(i);
2683 set_ctrl(n, new_ctrl);
2684 if (n->in(0) != nullptr) {
2685 _igvn.replace_input_of(n, 0, new_ctrl);
2686 }
2687 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false);
2688 }
2689 } else {
2690 Node *newhead = old_new[loop->_head->_idx];
2691 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2692 }
2693 }
2694
2695 //------------------------------clone_loop-------------------------------------
2696 //
2697 // C L O N E A L O O P B O D Y
2698 //
2699 // This is the basic building block of the loop optimizations. It clones an
2700 // entire loop body. It makes an old_new loop body mapping; with this mapping
2701 // you can find the new-loop equivalent to an old-loop node. All new-loop
2702 // nodes are exactly equal to their old-loop counterparts, all edges are the
2703 // same. All exits from the old-loop now have a RegionNode that merges the
2704 // equivalent new-loop path. This is true even for the normal "loop-exit"
2705 // condition. All uses of loop-invariant old-loop values now come from (one
2706 // or more) Phis that merge their new-loop equivalents.
2707 //
2708 // This operation leaves the graph in an illegal state: there are two valid
2709 // control edges coming from the loop pre-header to both loop bodies. I'll
2710 // definitely have to hack the graph after running this transform.
2711 //
2712 // From this building block I will further edit edges to perform loop peeling
2713 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2714 //
2715 // Parameter side_by_size_idom:
2716 // When side_by_size_idom is null, the dominator tree is constructed for
2717 // the clone loop to dominate the original. Used in construction of
2718 // pre-main-post loop sequence.
2719 // When nonnull, the clone and original are side-by-side, both are
2720 // dominated by the side_by_side_idom node. Used in construction of
2721 // unswitched loops.
2722 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2723 CloneLoopMode mode, Node* side_by_side_idom) {
2724
2725 LoopNode* head = loop->_head->as_Loop();
2726 head->verify_strip_mined(1);
2727
2728 if (C->do_vector_loop() && PrintOpto) {
2729 const char* mname = C->method()->name()->as_quoted_ascii();
2730 if (mname != nullptr) {
2731 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2732 }
2733 }
2734
2735 CloneMap& cm = C->clone_map();
2736 if (C->do_vector_loop()) {
2737 cm.set_clone_idx(cm.max_gen()+1);
2738 #ifndef PRODUCT
2739 if (PrintOpto) {
2740 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2741 loop->dump_head();
2742 }
2743 #endif
2744 }
2745
2746 // Step 1: Clone the loop body. Make the old->new mapping.
2747 clone_loop_body(loop->_body, old_new, &cm);
2748
2749 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2750
2751 // Step 2: Fix the edges in the new body. If the old input is outside the
2752 // loop use it. If the old input is INside the loop, use the corresponding
2753 // new node instead.
2754 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false);
2755
2756 Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2757 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2758
2759 // Step 3: Now fix control uses. Loop varying control uses have already
2760 // been fixed up (as part of all input edges in Step 2). Loop invariant
2761 // control uses must be either an IfFalse or an IfTrue. Make a merge
2762 // point to merge the old and new IfFalse/IfTrue nodes; make the use
2763 // refer to this.
2764 Node_List worklist;
2765 uint new_counter = C->unique();
2766 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist);
2767
2768 // Step 4: If loop-invariant use is not control, it must be dominated by a
2769 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
2770 // there if needed. Make a Phi there merging old and new used values.
2771 Node_List *split_if_set = nullptr;
2772 Node_List *split_bool_set = nullptr;
2773 Node_List *split_cex_set = nullptr;
2774 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set);
2775
2776 for (uint i = 0; i < extra_data_nodes.size(); i++) {
2777 Node* old = extra_data_nodes.at(i);
2778 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2779 split_bool_set, split_cex_set, worklist, new_counter,
2780 mode);
2781 }
2782
2783 // Check for IFs that need splitting/cloning. Happens if an IF outside of
2784 // the loop uses a condition set in the loop. The original IF probably
2785 // takes control from one or more OLD Regions (which in turn get from NEW
2786 // Regions). In any case, there will be a set of Phis for each merge point
2787 // from the IF up to where the original BOOL def exists the loop.
2788 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
2789
2790 }
2791
2792 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) {
2793 if (split_if_set) {
2794 while (split_if_set->size()) {
2795 Node *iff = split_if_set->pop();
2796 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1;
2797 if (iff->in(input)->is_Phi()) {
2798 Node *b = clone_iff(iff->in(input)->as_Phi());
2799 _igvn.replace_input_of(iff, input, b);
2800 }
2801 }
2802 }
2803 if (split_bool_set) {
2804 while (split_bool_set->size()) {
2805 Node *b = split_bool_set->pop();
2806 Node *phi = b->in(1);
2807 assert(phi->is_Phi(), "");
2808 CmpNode *cmp = clone_bool((PhiNode*) phi);
2809 _igvn.replace_input_of(b, 1, cmp);
2810 }
2811 }
2812 if (split_cex_set) {
2813 while (split_cex_set->size()) {
2814 Node *b = split_cex_set->pop();
2815 assert(b->in(0)->is_Region(), "");
2816 assert(b->in(1)->is_Phi(), "");
2817 assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2818 split_up(b, b->in(0), nullptr);
2819 }
2820 }
2821 }
2822
2823 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop,
2824 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set,
2825 Node_List*& split_bool_set, Node_List*& split_cex_set) {
2826 for(uint i = 0; i < body.size(); i++ ) {
2827 Node* old = body.at(i);
2828 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2829 split_bool_set, split_cex_set, worklist, new_counter,
2830 mode);
2831 }
2832 }
2833
2834 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode,
2835 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) {
2836 LoopNode* head = loop->_head->as_Loop();
2837 for(uint i = 0; i < body.size(); i++ ) {
2838 Node* old = body.at(i);
2839 if( !old->is_CFG() ) continue;
2840
2841 // Copy uses to a worklist, so I can munge the def-use info
2842 // with impunity.
2843 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) {
2844 worklist.push(old->fast_out(j));
2845 }
2846
2847 while (worklist.size()) { // Visit all uses
2848 Node *use = worklist.pop();
2849 if (!has_node(use)) continue; // Ignore dead nodes
2850 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use );
2851 if (!loop->is_member(use_loop) && use->is_CFG()) {
2852 // Both OLD and USE are CFG nodes here.
2853 assert(use->is_Proj(), "" );
2854 Node* nnn = old_new[old->_idx];
2855
2856 Node* newuse = nullptr;
2857 if (head->is_strip_mined() && mode != IgnoreStripMined) {
2858 CountedLoopNode* cl = head->as_CountedLoop();
2859 CountedLoopEndNode* cle = cl->loopexit();
2860 Node* cle_out = cle->proj_out_or_null(false);
2861 if (use == cle_out) {
2862 IfNode* le = cl->outer_loop_end();
2863 use = le->proj_out(false);
2864 use_loop = get_loop(use);
2865 if (mode == CloneIncludesStripMined) {
2866 nnn = old_new[le->_idx];
2867 } else {
2868 newuse = old_new[cle_out->_idx];
2869 }
2870 }
2871 }
2872 if (newuse == nullptr) {
2873 newuse = use->clone();
2874 }
2875
2876 // Clone the loop exit control projection
2877 if (C->do_vector_loop() && cm != nullptr) {
2878 cm->verify_insert_and_clone(use, newuse, cm->clone_idx());
2879 }
2880 newuse->set_req(0,nnn);
2881 _igvn.register_new_node_with_optimizer(newuse);
2882 set_loop(newuse, use_loop);
2883 set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2884
2885 // We need a Region to merge the exit from the peeled body and the
2886 // exit from the old loop body.
2887 RegionNode *r = new RegionNode(3);
2888 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use));
2889 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" );
2890
2891 // The original user of 'use' uses 'r' instead.
2892 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2893 Node* useuse = use->last_out(l);
2894 _igvn.rehash_node_delayed(useuse);
2895 uint uses_found = 0;
2896 if (useuse->in(0) == use) {
2897 useuse->set_req(0, r);
2898 uses_found++;
2899 if (useuse->is_CFG()) {
2900 // This is not a dom_depth > dd_r because when new
2901 // control flow is constructed by a loop opt, a node and
2902 // its dominator can end up at the same dom_depth
2903 assert(dom_depth(useuse) >= dd_r, "");
2904 set_idom(useuse, r, dom_depth(useuse));
2905 }
2906 }
2907 for (uint k = 1; k < useuse->req(); k++) {
2908 if( useuse->in(k) == use ) {
2909 useuse->set_req(k, r);
2910 uses_found++;
2911 if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2912 // This is not a dom_depth > dd_r because when new
2913 // control flow is constructed by a loop opt, a node
2914 // and its dominator can end up at the same dom_depth
2915 assert(dom_depth(useuse) >= dd_r , "");
2916 set_idom(useuse, r, dom_depth(useuse));
2917 }
2918 }
2919 }
2920 l -= uses_found; // we deleted 1 or more copies of this edge
2921 }
2922
2923 assert(use->is_Proj(), "loop exit should be projection");
2924 // lazy_replace() below moves all nodes that are:
2925 // - control dependent on the loop exit or
2926 // - have control set to the loop exit
2927 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it
2928 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one
2929 // becomes dead and is "replaced" by the region.
2930 Node* use_clone = use->clone();
2931 register_control(use_clone, use_loop, idom(use), dom_depth(use));
2932 // Now finish up 'r'
2933 r->set_req(1, newuse);
2934 r->set_req(2, use_clone);
2935 _igvn.register_new_node_with_optimizer(r);
2936 set_loop(r, use_loop);
2937 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r);
2938 lazy_replace(use, r);
2939 // Map the (cloned) old use to the new merge point
2940 old_new.map(use_clone->_idx, r);
2941 } // End of if a loop-exit test
2942 }
2943 }
2944 }
2945
2946 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd,
2947 IdealLoopTree* parent, bool partial) {
2948 for(uint i = 0; i < body.size(); i++ ) {
2949 Node *old = body.at(i);
2950 Node *nnn = old_new[old->_idx];
2951 // Fix CFG/Loop controlling the new node
2952 if (has_ctrl(old)) {
2953 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2954 } else {
2955 set_loop(nnn, parent);
2956 if (old->outcnt() > 0) {
2957 Node* dom = idom(old);
2958 if (old_new[dom->_idx] != nullptr) {
2959 dom = old_new[dom->_idx];
2960 set_idom(nnn, dom, dd );
2961 }
2962 }
2963 }
2964 // Correct edges to the new node
2965 for (uint j = 0; j < nnn->req(); j++) {
2966 Node *n = nnn->in(j);
2967 if (n != nullptr) {
2968 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n);
2969 if (loop->is_member(old_in_loop)) {
2970 if (old_new[n->_idx] != nullptr) {
2971 nnn->set_req(j, old_new[n->_idx]);
2972 } else {
2973 assert(!body.contains(n), "");
2974 assert(partial, "node not cloned");
2975 }
2976 }
2977 }
2978 }
2979 _igvn.hash_find_insert(nnn);
2980 }
2981 }
2982
2983 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) {
2984 for (uint i = 0; i < body.size(); i++) {
2985 Node* old = body.at(i);
2986 Node* nnn = old->clone();
2987 old_new.map(old->_idx, nnn);
2988 if (C->do_vector_loop() && cm != nullptr) {
2989 cm->verify_insert_and_clone(old, nnn, cm->clone_idx());
2990 }
2991 _igvn.register_new_node_with_optimizer(nnn);
2992 }
2993 }
2994
2995
2996 //---------------------- stride_of_possible_iv -------------------------------------
2997 // Looks for an iff/bool/comp with one operand of the compare
2998 // being a cycle involving an add and a phi,
2999 // with an optional truncation (left-shift followed by a right-shift)
3000 // of the add. Returns zero if not an iv.
3001 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
3002 Node* trunc1 = nullptr;
3003 Node* trunc2 = nullptr;
3004 const TypeInteger* ttype = nullptr;
3005 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) {
3006 return 0;
3007 }
3008 BoolNode* bl = iff->in(1)->as_Bool();
3009 Node* cmp = bl->in(1);
3010 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
3011 return 0;
3012 }
3013 // Must have an invariant operand
3014 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
3015 return 0;
3016 }
3017 Node* add2 = nullptr;
3018 Node* cmp1 = cmp->in(1);
3019 if (cmp1->is_Phi()) {
3020 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
3021 Node* phi = cmp1;
3022 for (uint i = 1; i < phi->req(); i++) {
3023 Node* in = phi->in(i);
3024 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
3025 &trunc1, &trunc2, &ttype, T_INT);
3026 if (add && add->in(1) == phi) {
3027 add2 = add->in(2);
3028 break;
3029 }
3030 }
3031 } else {
3032 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
3033 Node* addtrunc = cmp1;
3034 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
3035 &trunc1, &trunc2, &ttype, T_INT);
3036 if (add && add->in(1)->is_Phi()) {
3037 Node* phi = add->in(1);
3038 for (uint i = 1; i < phi->req(); i++) {
3039 if (phi->in(i) == addtrunc) {
3040 add2 = add->in(2);
3041 break;
3042 }
3043 }
3044 }
3045 }
3046 if (add2 != nullptr) {
3047 const TypeInt* add2t = _igvn.type(add2)->is_int();
3048 if (add2t->is_con()) {
3049 return add2t->get_con();
3050 }
3051 }
3052 return 0;
3053 }
3054
3055
3056 //---------------------- stay_in_loop -------------------------------------
3057 // Return the (unique) control output node that's in the loop (if it exists.)
3058 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
3059 Node* unique = nullptr;
3060 if (!n) return nullptr;
3061 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3062 Node* use = n->fast_out(i);
3063 if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
3064 if (unique != nullptr) {
3065 return nullptr;
3066 }
3067 unique = use;
3068 }
3069 }
3070 return unique;
3071 }
3072
3073 //------------------------------ register_node -------------------------------------
3074 // Utility to register node "n" with PhaseIdealLoop
3075 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) {
3076 _igvn.register_new_node_with_optimizer(n);
3077 loop->_body.push(n);
3078 if (n->is_CFG()) {
3079 set_loop(n, loop);
3080 set_idom(n, pred, ddepth);
3081 } else {
3082 set_ctrl(n, pred);
3083 }
3084 }
3085
3086 //------------------------------ proj_clone -------------------------------------
3087 // Utility to create an if-projection
3088 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
3089 ProjNode* c = p->clone()->as_Proj();
3090 c->set_req(0, iff);
3091 return c;
3092 }
3093
3094 //------------------------------ short_circuit_if -------------------------------------
3095 // Force the iff control output to be the live_proj
3096 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
3097 guarantee(live_proj != nullptr, "null projection");
3098 int proj_con = live_proj->_con;
3099 assert(proj_con == 0 || proj_con == 1, "false or true projection");
3100 Node* con = intcon(proj_con);
3101 if (iff) {
3102 iff->set_req(1, con);
3103 }
3104 return con;
3105 }
3106
3107 //------------------------------ insert_if_before_proj -------------------------------------
3108 // Insert a new if before an if projection (* - new node)
3109 //
3110 // before
3111 // if(test)
3112 // / \
3113 // v v
3114 // other-proj proj (arg)
3115 //
3116 // after
3117 // if(test)
3118 // / \
3119 // / v
3120 // | * proj-clone
3121 // v |
3122 // other-proj v
3123 // * new_if(relop(cmp[IU](left,right)))
3124 // / \
3125 // v v
3126 // * new-proj proj
3127 // (returned)
3128 //
3129 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
3130 IfNode* iff = proj->in(0)->as_If();
3131 IdealLoopTree *loop = get_loop(proj);
3132 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3133 uint ddepth = dom_depth(proj);
3134
3135 _igvn.rehash_node_delayed(iff);
3136 _igvn.rehash_node_delayed(proj);
3137
3138 proj->set_req(0, nullptr); // temporary disconnect
3139 ProjNode* proj2 = proj_clone(proj, iff);
3140 register_node(proj2, loop, iff, ddepth);
3141
3142 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
3143 register_node(cmp, loop, proj2, ddepth);
3144
3145 BoolNode* bol = new BoolNode(cmp, relop);
3146 register_node(bol, loop, proj2, ddepth);
3147
3148 int opcode = iff->Opcode();
3149 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
3150 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol);
3151 register_node(new_if, loop, proj2, ddepth);
3152
3153 proj->set_req(0, new_if); // reattach
3154 set_idom(proj, new_if, ddepth);
3155
3156 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
3157 guarantee(new_exit != nullptr, "null exit node");
3158 register_node(new_exit, get_loop(other_proj), new_if, ddepth);
3159
3160 return new_exit;
3161 }
3162
3163 //------------------------------ insert_region_before_proj -------------------------------------
3164 // Insert a region before an if projection (* - new node)
3165 //
3166 // before
3167 // if(test)
3168 // / |
3169 // v |
3170 // proj v
3171 // other-proj
3172 //
3173 // after
3174 // if(test)
3175 // / |
3176 // v |
3177 // * proj-clone v
3178 // | other-proj
3179 // v
3180 // * new-region
3181 // |
3182 // v
3183 // * dum_if
3184 // / \
3185 // v \
3186 // * dum-proj v
3187 // proj
3188 //
3189 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
3190 IfNode* iff = proj->in(0)->as_If();
3191 IdealLoopTree *loop = get_loop(proj);
3192 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
3193 uint ddepth = dom_depth(proj);
3194
3195 _igvn.rehash_node_delayed(iff);
3196 _igvn.rehash_node_delayed(proj);
3197
3198 proj->set_req(0, nullptr); // temporary disconnect
3199 ProjNode* proj2 = proj_clone(proj, iff);
3200 register_node(proj2, loop, iff, ddepth);
3201
3202 RegionNode* reg = new RegionNode(2);
3203 reg->set_req(1, proj2);
3204 register_node(reg, loop, iff, ddepth);
3205
3206 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt);
3207 register_node(dum_if, loop, reg, ddepth);
3208
3209 proj->set_req(0, dum_if); // reattach
3210 set_idom(proj, dum_if, ddepth);
3211
3212 ProjNode* dum_proj = proj_clone(other_proj, dum_if);
3213 register_node(dum_proj, loop, dum_if, ddepth);
3214
3215 return reg;
3216 }
3217
3218 // Idea
3219 // ----
3220 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops
3221 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel
3222 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new
3223 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned
3224 // test alone instead:
3225 //
3226 // Before Partial Peeling:
3227 // Loop:
3228 // <peeled section>
3229 // Split off signed loop exit test
3230 // <-- CUT HERE -->
3231 // Unchanged unsigned loop exit test
3232 // <rest of unpeeled section>
3233 // goto Loop
3234 //
3235 // After Partial Peeling:
3236 // <cloned peeled section>
3237 // Cloned split off signed loop exit test
3238 // Loop:
3239 // Unchanged unsigned loop exit test
3240 // <rest of unpeeled section>
3241 // <peeled section>
3242 // Split off signed loop exit test
3243 // goto Loop
3244 //
3245 // Details
3246 // -------
3247 // Before:
3248 // if (i <u limit) Unsigned loop exit condition
3249 // / |
3250 // v v
3251 // exit-proj stay-in-loop-proj
3252 //
3253 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it
3254 // before the CmpU on the stay-in-loop path and keep both tests:
3255 //
3256 // if (i <u limit) Signed loop exit test
3257 // / |
3258 // / if (i <u limit) Unsigned loop exit test
3259 // / / |
3260 // v v v
3261 // exit-region stay-in-loop-proj
3262 //
3263 // Implementation
3264 // --------------
3265 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned
3266 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop
3267 // exit tests is preserved, and their loop nesting is correct.
3268 //
3269 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit
3270 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant
3271 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later:
3272 //
3273 // if (stay-in-loop-const) Killed original unsigned loop exit test
3274 // / |
3275 // / v
3276 // / if (i < limit) Split off signed loop exit test
3277 // / / |
3278 // / / v
3279 // / / if (i <u limit) Cloned unsigned loop exit test
3280 // / / / |
3281 // v v v |
3282 // exit-region |
3283 // | |
3284 // dummy-if |
3285 // / | |
3286 // dead | |
3287 // v v
3288 // exit-proj stay-in-loop-proj
3289 //
3290 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned
3291 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()).
3292 //
3293 // Requirements
3294 // ------------
3295 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly
3296 // the same as before with only a single unsigned test. This is only possible if certain requirements are met.
3297 // Otherwise, we need to bail out (see comments in the code below).
3298 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) {
3299 const bool Signed = true;
3300 const bool Unsigned = false;
3301
3302 BoolNode* bol = if_cmpu->in(1)->as_Bool();
3303 if (bol->_test._test != BoolTest::lt) {
3304 return nullptr;
3305 }
3306 CmpNode* cmpu = bol->in(1)->as_Cmp();
3307 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison");
3308
3309 int stride = stride_of_possible_iv(if_cmpu);
3310 if (stride == 0) {
3311 return nullptr;
3312 }
3313
3314 Node* lp_proj = stay_in_loop(if_cmpu, loop);
3315 guarantee(lp_proj != nullptr, "null loop node");
3316
3317 ProjNode* lp_continue = lp_proj->as_Proj();
3318 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
3319 if (!lp_exit->is_IfFalse()) {
3320 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit).
3321 // We therefore can't add a single exit condition.
3322 return nullptr;
3323 }
3324 // The unsigned loop exit condition is
3325 // !(i <u limit)
3326 // = i >=u limit
3327 //
3328 // First, we note that for any x for which
3329 // 0 <= x <= INT_MAX
3330 // we can convert x to an unsigned int and still get the same guarantee:
3331 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX
3332 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA)
3333 //
3334 // With that in mind, if
3335 // limit >= 0 (COND)
3336 // then the unsigned loop exit condition
3337 // i >=u limit (ULE)
3338 // is equivalent to
3339 // i < 0 || i >= limit (SLE-full)
3340 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned
3341 // (uint) i >=u MAX_INT >= limit >= 0
3342 // or otherwise
3343 // i >= limit >= 0
3344 // holds due to (LEMMA).
3345 //
3346 // For completeness, a counterexample with limit < 0:
3347 // Assume i = -3 and limit = -2:
3348 // i < 0
3349 // -2 < 0
3350 // is true and thus also "i < 0 || i >= limit". But
3351 // i >=u limit
3352 // -3 >=u -2
3353 // is false.
3354 Node* limit = cmpu->in(2);
3355 const TypeInt* type_limit = _igvn.type(limit)->is_int();
3356 if (type_limit->_lo < 0) {
3357 return nullptr;
3358 }
3359
3360 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride:
3361 // stride < 0:
3362 // i < 0 (SLE = SLE-negative)
3363 // stride > 0:
3364 // i >= limit (SLE = SLE-positive)
3365 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0):
3366 //
3367 // Loop:
3368 // <peeled section>
3369 // i >= limit (SLE-positive)
3370 // <-- CUT HERE -->
3371 // i >=u limit (ULE)
3372 // <rest of unpeeled section>
3373 // goto Loop
3374 //
3375 // We exit the loop if:
3376 // (SLE) is true OR (ULE) is true
3377 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly
3378 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure:
3379 // (SLE) IMPLIES (ULE)
3380 // This indeed holds when (COND) is given:
3381 // - stride > 0:
3382 // i >= limit // (SLE = SLE-positive)
3383 // i >= limit >= 0 // (COND)
3384 // i >=u limit >= 0 // (LEMMA)
3385 // which is the unsigned loop exit condition (ULE).
3386 // - stride < 0:
3387 // i < 0 // (SLE = SLE-negative)
3388 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned
3389 // MAX_INT >= limit >= 0 // (COND)
3390 // MAX_INT >=u limit >= 0 // (LEMMA)
3391 // and thus from (NEG) and (LEMMA):
3392 // i >=u limit
3393 // which is the unsigned loop exit condition (ULE).
3394 //
3395 //
3396 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0):
3397 // <cloned peeled section>
3398 // i >= limit (SLE-positive)
3399 // Loop:
3400 // i >=u limit (ULE)
3401 // <rest of unpeeled section>
3402 // <peeled section>
3403 // i >= limit (SLE-positive)
3404 // goto Loop
3405 Node* rhs_cmpi;
3406 if (stride > 0) {
3407 rhs_cmpi = limit; // For i >= limit
3408 } else {
3409 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0
3410 }
3411 // Create a new region on the exit path
3412 RegionNode* reg = insert_region_before_proj(lp_exit);
3413 guarantee(reg != nullptr, "null region node");
3414
3415 // Clone the if-cmpu-true-false using a signed compare
3416 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
3417 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue);
3418 reg->add_req(cmpi_exit);
3419
3420 // Clone the if-cmpu-true-false
3421 BoolTest::mask rel_u = bol->_test._test;
3422 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
3423 reg->add_req(cmpu_exit);
3424
3425 // Force original if to stay in loop.
3426 short_circuit_if(if_cmpu, lp_continue);
3427
3428 return cmpi_exit->in(0)->as_If();
3429 }
3430
3431 //------------------------------ remove_cmpi_loop_exit -------------------------------------
3432 // Remove a previously inserted signed compare loop exit.
3433 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
3434 Node* lp_proj = stay_in_loop(if_cmp, loop);
3435 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
3436 stay_in_loop(lp_proj, loop)->is_If() &&
3437 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
3438 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
3439 if_cmp->set_req(1, con);
3440 }
3441
3442 //------------------------------ scheduled_nodelist -------------------------------------
3443 // Create a post order schedule of nodes that are in the
3444 // "member" set. The list is returned in "sched".
3445 // The first node in "sched" is the loop head, followed by
3446 // nodes which have no inputs in the "member" set, and then
3447 // followed by the nodes that have an immediate input dependence
3448 // on a node in "sched".
3449 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
3450
3451 assert(member.test(loop->_head->_idx), "loop head must be in member set");
3452 VectorSet visited;
3453 Node_Stack nstack(loop->_body.size());
3454
3455 Node* n = loop->_head; // top of stack is cached in "n"
3456 uint idx = 0;
3457 visited.set(n->_idx);
3458
3459 // Initially push all with no inputs from within member set
3460 for(uint i = 0; i < loop->_body.size(); i++ ) {
3461 Node *elt = loop->_body.at(i);
3462 if (member.test(elt->_idx)) {
3463 bool found = false;
3464 for (uint j = 0; j < elt->req(); j++) {
3465 Node* def = elt->in(j);
3466 if (def && member.test(def->_idx) && def != elt) {
3467 found = true;
3468 break;
3469 }
3470 }
3471 if (!found && elt != loop->_head) {
3472 nstack.push(n, idx);
3473 n = elt;
3474 assert(!visited.test(n->_idx), "not seen yet");
3475 visited.set(n->_idx);
3476 }
3477 }
3478 }
3479
3480 // traverse out's that are in the member set
3481 while (true) {
3482 if (idx < n->outcnt()) {
3483 Node* use = n->raw_out(idx);
3484 idx++;
3485 if (!visited.test_set(use->_idx)) {
3486 if (member.test(use->_idx)) {
3487 nstack.push(n, idx);
3488 n = use;
3489 idx = 0;
3490 }
3491 }
3492 } else {
3493 // All outputs processed
3494 sched.push(n);
3495 if (nstack.is_empty()) break;
3496 n = nstack.node();
3497 idx = nstack.index();
3498 nstack.pop();
3499 }
3500 }
3501 }
3502
3503
3504 //------------------------------ has_use_in_set -------------------------------------
3505 // Has a use in the vector set
3506 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
3507 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3508 Node* use = n->fast_out(j);
3509 if (vset.test(use->_idx)) {
3510 return true;
3511 }
3512 }
3513 return false;
3514 }
3515
3516
3517 //------------------------------ has_use_internal_to_set -------------------------------------
3518 // Has use internal to the vector set (ie. not in a phi at the loop head)
3519 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
3520 Node* head = loop->_head;
3521 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3522 Node* use = n->fast_out(j);
3523 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
3524 return true;
3525 }
3526 }
3527 return false;
3528 }
3529
3530
3531 //------------------------------ clone_for_use_outside_loop -------------------------------------
3532 // clone "n" for uses that are outside of loop
3533 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
3534 int cloned = 0;
3535 assert(worklist.size() == 0, "should be empty");
3536 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3537 Node* use = n->fast_out(j);
3538 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
3539 worklist.push(use);
3540 }
3541 }
3542
3543 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor,
3544 "Too many clones required in clone_for_use_outside_loop in partial peeling")) {
3545 return -1;
3546 }
3547
3548 while( worklist.size() ) {
3549 Node *use = worklist.pop();
3550 if (!has_node(use) || use->in(0) == C->top()) continue;
3551 uint j;
3552 for (j = 0; j < use->req(); j++) {
3553 if (use->in(j) == n) break;
3554 }
3555 assert(j < use->req(), "must be there");
3556
3557 // clone "n" and insert it between the inputs of "n" and the use outside the loop
3558 Node* n_clone = n->clone();
3559 _igvn.replace_input_of(use, j, n_clone);
3560 cloned++;
3561 Node* use_c;
3562 if (!use->is_Phi()) {
3563 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
3564 } else {
3565 // Use in a phi is considered a use in the associated predecessor block
3566 use_c = use->in(0)->in(j);
3567 }
3568 set_ctrl(n_clone, use_c);
3569 assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
3570 get_loop(use_c)->_body.push(n_clone);
3571 _igvn.register_new_node_with_optimizer(n_clone);
3572 #ifndef PRODUCT
3573 if (TracePartialPeeling) {
3574 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
3575 }
3576 #endif
3577 }
3578 return cloned;
3579 }
3580
3581
3582 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
3583 // clone "n" for special uses that are in the not_peeled region.
3584 // If these def-uses occur in separate blocks, the code generator
3585 // marks the method as not compilable. For example, if a "BoolNode"
3586 // is in a different basic block than the "IfNode" that uses it, then
3587 // the compilation is aborted in the code generator.
3588 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
3589 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
3590 if (n->is_Phi() || n->is_Load()) {
3591 return;
3592 }
3593 assert(worklist.size() == 0, "should be empty");
3594 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3595 Node* use = n->fast_out(j);
3596 if ( not_peel.test(use->_idx) &&
3597 (use->is_If() || use->is_CMove() || use->is_Bool()) &&
3598 use->in(1) == n) {
3599 worklist.push(use);
3600 }
3601 }
3602 if (worklist.size() > 0) {
3603 // clone "n" and insert it between inputs of "n" and the use
3604 Node* n_clone = n->clone();
3605 loop->_body.push(n_clone);
3606 _igvn.register_new_node_with_optimizer(n_clone);
3607 set_ctrl(n_clone, get_ctrl(n));
3608 sink_list.push(n_clone);
3609 not_peel.set(n_clone->_idx);
3610 #ifndef PRODUCT
3611 if (TracePartialPeeling) {
3612 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
3613 }
3614 #endif
3615 while( worklist.size() ) {
3616 Node *use = worklist.pop();
3617 _igvn.rehash_node_delayed(use);
3618 for (uint j = 1; j < use->req(); j++) {
3619 if (use->in(j) == n) {
3620 use->set_req(j, n_clone);
3621 }
3622 }
3623 }
3624 }
3625 }
3626
3627
3628 //------------------------------ insert_phi_for_loop -------------------------------------
3629 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
3630 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
3631 Node *phi = PhiNode::make(lp, back_edge_val);
3632 phi->set_req(LoopNode::EntryControl, lp_entry_val);
3633 // Use existing phi if it already exists
3634 Node *hit = _igvn.hash_find_insert(phi);
3635 if( hit == nullptr ) {
3636 _igvn.register_new_node_with_optimizer(phi);
3637 set_ctrl(phi, lp);
3638 } else {
3639 // Remove the new phi from the graph and use the hit
3640 _igvn.remove_dead_node(phi);
3641 phi = hit;
3642 }
3643 _igvn.replace_input_of(use, idx, phi);
3644 }
3645
3646 #ifdef ASSERT
3647 //------------------------------ is_valid_loop_partition -------------------------------------
3648 // Validate the loop partition sets: peel and not_peel
3649 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
3650 VectorSet& not_peel ) {
3651 uint i;
3652 // Check that peel_list entries are in the peel set
3653 for (i = 0; i < peel_list.size(); i++) {
3654 if (!peel.test(peel_list.at(i)->_idx)) {
3655 return false;
3656 }
3657 }
3658 // Check at loop members are in one of peel set or not_peel set
3659 for (i = 0; i < loop->_body.size(); i++ ) {
3660 Node *def = loop->_body.at(i);
3661 uint di = def->_idx;
3662 // Check that peel set elements are in peel_list
3663 if (peel.test(di)) {
3664 if (not_peel.test(di)) {
3665 return false;
3666 }
3667 // Must be in peel_list also
3668 bool found = false;
3669 for (uint j = 0; j < peel_list.size(); j++) {
3670 if (peel_list.at(j)->_idx == di) {
3671 found = true;
3672 break;
3673 }
3674 }
3675 if (!found) {
3676 return false;
3677 }
3678 } else if (not_peel.test(di)) {
3679 if (peel.test(di)) {
3680 return false;
3681 }
3682 } else {
3683 return false;
3684 }
3685 }
3686 return true;
3687 }
3688
3689 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
3690 // Ensure a use outside of loop is of the right form
3691 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
3692 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3693 return (use->is_Phi() &&
3694 use_c->is_Region() && use_c->req() == 3 &&
3695 (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
3696 use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
3697 use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
3698 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
3699 }
3700
3701 //------------------------------ is_valid_clone_loop_form -------------------------------------
3702 // Ensure that all uses outside of loop are of the right form
3703 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
3704 uint orig_exit_idx, uint clone_exit_idx) {
3705 uint len = peel_list.size();
3706 for (uint i = 0; i < len; i++) {
3707 Node *def = peel_list.at(i);
3708
3709 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3710 Node *use = def->fast_out(j);
3711 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
3712 if (!loop->is_member(get_loop(use_c))) {
3713 // use is not in the loop, check for correct structure
3714 if (use->in(0) == def) {
3715 // Okay
3716 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
3717 return false;
3718 }
3719 }
3720 }
3721 }
3722 return true;
3723 }
3724 #endif
3725
3726 //------------------------------ partial_peel -------------------------------------
3727 // Partially peel (aka loop rotation) the top portion of a loop (called
3728 // the peel section below) by cloning it and placing one copy just before
3729 // the new loop head and the other copy at the bottom of the new loop.
3730 //
3731 // before after where it came from
3732 //
3733 // stmt1 stmt1
3734 // loop: stmt2 clone
3735 // stmt2 if condA goto exitA clone
3736 // if condA goto exitA new_loop: new
3737 // stmt3 stmt3 clone
3738 // if !condB goto loop if condB goto exitB clone
3739 // exitB: stmt2 orig
3740 // stmt4 if !condA goto new_loop orig
3741 // exitA: goto exitA
3742 // exitB:
3743 // stmt4
3744 // exitA:
3745 //
3746 // Step 1: find the cut point: an exit test on probable
3747 // induction variable.
3748 // Step 2: schedule (with cloning) operations in the peel
3749 // section that can be executed after the cut into
3750 // the section that is not peeled. This may need
3751 // to clone operations into exit blocks. For
3752 // instance, a reference to A[i] in the not-peel
3753 // section and a reference to B[i] in an exit block
3754 // may cause a left-shift of i by 2 to be placed
3755 // in the peel block. This step will clone the left
3756 // shift into the exit block and sink the left shift
3757 // from the peel to the not-peel section.
3758 // Step 3: clone the loop, retarget the control, and insert
3759 // phis for values that are live across the new loop
3760 // head. This is very dependent on the graph structure
3761 // from clone_loop. It creates region nodes for
3762 // exit control and associated phi nodes for values
3763 // flow out of the loop through that exit. The region
3764 // node is dominated by the clone's control projection.
3765 // So the clone's peel section is placed before the
3766 // new loop head, and the clone's not-peel section is
3767 // forms the top part of the new loop. The original
3768 // peel section forms the tail of the new loop.
3769 // Step 4: update the dominator tree and recompute the
3770 // dominator depth.
3771 //
3772 // orig
3773 //
3774 // stmt1
3775 // |
3776 // v
3777 // predicates
3778 // |
3779 // v
3780 // loop<----+
3781 // | |
3782 // stmt2 |
3783 // | |
3784 // v |
3785 // ifA |
3786 // / | |
3787 // v v |
3788 // false true ^ <-- last_peel
3789 // / | |
3790 // / ===|==cut |
3791 // / stmt3 | <-- first_not_peel
3792 // / | |
3793 // | v |
3794 // v ifB |
3795 // exitA: / \ |
3796 // / \ |
3797 // v v |
3798 // false true |
3799 // / \ |
3800 // / ----+
3801 // |
3802 // v
3803 // exitB:
3804 // stmt4
3805 //
3806 //
3807 // after clone loop
3808 //
3809 // stmt1
3810 // |
3811 // v
3812 // predicates
3813 // / \
3814 // clone / \ orig
3815 // / \
3816 // / \
3817 // v v
3818 // +---->loop loop<----+
3819 // | | | |
3820 // | stmt2 stmt2 |
3821 // | | | |
3822 // | v v |
3823 // | ifA ifA |
3824 // | | \ / | |
3825 // | v v v v |
3826 // ^ true false false true ^ <-- last_peel
3827 // | | ^ \ / | |
3828 // | cut==|== \ \ / ===|==cut |
3829 // | stmt3 \ \ / stmt3 | <-- first_not_peel
3830 // | | dom | | | |
3831 // | v \ 1v v2 v |
3832 // | ifB regionA ifB |
3833 // | / \ | / \ |
3834 // | / \ v / \ |
3835 // | v v exitA: v v |
3836 // | true false false true |
3837 // | / ^ \ / \ |
3838 // +---- \ \ / ----+
3839 // dom \ /
3840 // \ 1v v2
3841 // regionB
3842 // |
3843 // v
3844 // exitB:
3845 // stmt4
3846 //
3847 //
3848 // after partial peel
3849 //
3850 // stmt1
3851 // |
3852 // v
3853 // predicates
3854 // /
3855 // clone / orig
3856 // / TOP
3857 // / \
3858 // v v
3859 // TOP->loop loop----+
3860 // | | |
3861 // stmt2 stmt2 |
3862 // | | |
3863 // v v |
3864 // ifA ifA |
3865 // | \ / | |
3866 // v v v v |
3867 // true false false true | <-- last_peel
3868 // | ^ \ / +------|---+
3869 // +->newloop \ \ / === ==cut | |
3870 // | stmt3 \ \ / TOP | |
3871 // | | dom | | stmt3 | | <-- first_not_peel
3872 // | v \ 1v v2 v | |
3873 // | ifB regionA ifB ^ v
3874 // | / \ | / \ | |
3875 // | / \ v / \ | |
3876 // | v v exitA: v v | |
3877 // | true false false true | |
3878 // | / ^ \ / \ | |
3879 // | | \ \ / v | |
3880 // | | dom \ / TOP | |
3881 // | | \ 1v v2 | |
3882 // ^ v regionB | |
3883 // | | | | |
3884 // | | v ^ v
3885 // | | exitB: | |
3886 // | | stmt4 | |
3887 // | +------------>-----------------+ |
3888 // | |
3889 // +-----------------<---------------------+
3890 //
3891 //
3892 // final graph
3893 //
3894 // stmt1
3895 // |
3896 // v
3897 // predicates
3898 // |
3899 // v
3900 // stmt2 clone
3901 // |
3902 // v
3903 // ........> ifA clone
3904 // : / |
3905 // dom / |
3906 // : v v
3907 // : false true
3908 // : | |
3909 // : | v
3910 // : | newloop<-----+
3911 // : | | |
3912 // : | stmt3 clone |
3913 // : | | |
3914 // : | v |
3915 // : | ifB |
3916 // : | / \ |
3917 // : | v v |
3918 // : | false true |
3919 // : | | | |
3920 // : | v stmt2 |
3921 // : | exitB: | |
3922 // : | stmt4 v |
3923 // : | ifA orig |
3924 // : | / \ |
3925 // : | / \ |
3926 // : | v v |
3927 // : | false true |
3928 // : | / \ |
3929 // : v v -----+
3930 // RegionA
3931 // |
3932 // v
3933 // exitA
3934 //
3935 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3936
3937 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3938 if (!loop->_head->is_Loop()) {
3939 return false;
3940 }
3941 LoopNode *head = loop->_head->as_Loop();
3942
3943 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3944 return false;
3945 }
3946
3947 // Check for complex exit control
3948 for (uint ii = 0; ii < loop->_body.size(); ii++) {
3949 Node *n = loop->_body.at(ii);
3950 int opc = n->Opcode();
3951 if (n->is_Call() ||
3952 opc == Op_Catch ||
3953 opc == Op_CatchProj ||
3954 opc == Op_Jump ||
3955 opc == Op_JumpProj) {
3956 #ifndef PRODUCT
3957 if (TracePartialPeeling) {
3958 tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3959 }
3960 #endif
3961 return false;
3962 }
3963 }
3964
3965 int dd = dom_depth(head);
3966
3967 // Step 1: find cut point
3968
3969 // Walk up dominators to loop head looking for first loop exit
3970 // which is executed on every path thru loop.
3971 IfNode *peel_if = nullptr;
3972 IfNode *peel_if_cmpu = nullptr;
3973
3974 Node *iff = loop->tail();
3975 while (iff != head) {
3976 if (iff->is_If()) {
3977 Node *ctrl = get_ctrl(iff->in(1));
3978 if (ctrl->is_top()) return false; // Dead test on live IF.
3979 // If loop-varying exit-test, check for induction variable
3980 if (loop->is_member(get_loop(ctrl)) &&
3981 loop->is_loop_exit(iff) &&
3982 is_possible_iv_test(iff)) {
3983 Node* cmp = iff->in(1)->in(1);
3984 if (cmp->Opcode() == Op_CmpI) {
3985 peel_if = iff->as_If();
3986 } else {
3987 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3988 peel_if_cmpu = iff->as_If();
3989 }
3990 }
3991 }
3992 iff = idom(iff);
3993 }
3994
3995 // Prefer signed compare over unsigned compare.
3996 IfNode* new_peel_if = nullptr;
3997 if (peel_if == nullptr) {
3998 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) {
3999 return false; // No peel point found
4000 }
4001 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
4002 if (new_peel_if == nullptr) {
4003 return false; // No peel point found
4004 }
4005 peel_if = new_peel_if;
4006 }
4007 Node* last_peel = stay_in_loop(peel_if, loop);
4008 Node* first_not_peeled = stay_in_loop(last_peel, loop);
4009 if (first_not_peeled == nullptr || first_not_peeled == head) {
4010 return false;
4011 }
4012
4013 #ifndef PRODUCT
4014 if (TraceLoopOpts) {
4015 tty->print("PartialPeel ");
4016 loop->dump_head();
4017 }
4018
4019 if (TracePartialPeeling) {
4020 tty->print_cr("before partial peel one iteration");
4021 Node_List wl;
4022 Node* t = head->in(2);
4023 while (true) {
4024 wl.push(t);
4025 if (t == head) break;
4026 t = idom(t);
4027 }
4028 while (wl.size() > 0) {
4029 Node* tt = wl.pop();
4030 tt->dump();
4031 if (tt == last_peel) tty->print_cr("-- cut --");
4032 }
4033 }
4034 #endif
4035
4036 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head);
4037
4038 VectorSet peel;
4039 VectorSet not_peel;
4040 Node_List peel_list;
4041 Node_List worklist;
4042 Node_List sink_list;
4043
4044 uint estimate = loop->est_loop_clone_sz(1);
4045 if (exceeding_node_budget(estimate)) {
4046 return false;
4047 }
4048
4049 // Set of cfg nodes to peel are those that are executable from
4050 // the head through last_peel.
4051 assert(worklist.size() == 0, "should be empty");
4052 worklist.push(head);
4053 peel.set(head->_idx);
4054 while (worklist.size() > 0) {
4055 Node *n = worklist.pop();
4056 if (n != last_peel) {
4057 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
4058 Node* use = n->fast_out(j);
4059 if (use->is_CFG() &&
4060 loop->is_member(get_loop(use)) &&
4061 !peel.test_set(use->_idx)) {
4062 worklist.push(use);
4063 }
4064 }
4065 }
4066 }
4067
4068 // Set of non-cfg nodes to peel are those that are control
4069 // dependent on the cfg nodes.
4070 for (uint i = 0; i < loop->_body.size(); i++) {
4071 Node *n = loop->_body.at(i);
4072 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
4073 if (peel.test(n_c->_idx)) {
4074 peel.set(n->_idx);
4075 } else {
4076 not_peel.set(n->_idx);
4077 }
4078 }
4079
4080 // Step 2: move operations from the peeled section down into the
4081 // not-peeled section
4082
4083 // Get a post order schedule of nodes in the peel region
4084 // Result in right-most operand.
4085 scheduled_nodelist(loop, peel, peel_list);
4086
4087 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4088
4089 // For future check for too many new phis
4090 uint old_phi_cnt = 0;
4091 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
4092 Node* use = head->fast_out(j);
4093 if (use->is_Phi()) old_phi_cnt++;
4094 }
4095
4096 #ifndef PRODUCT
4097 if (TracePartialPeeling) {
4098 tty->print_cr("\npeeled list");
4099 }
4100 #endif
4101
4102 // Evacuate nodes in peel region into the not_peeled region if possible
4103 bool too_many_clones = false;
4104 uint new_phi_cnt = 0;
4105 uint cloned_for_outside_use = 0;
4106 for (uint i = 0; i < peel_list.size();) {
4107 Node* n = peel_list.at(i);
4108 #ifndef PRODUCT
4109 if (TracePartialPeeling) n->dump();
4110 #endif
4111 bool incr = true;
4112 if (!n->is_CFG()) {
4113 if (has_use_in_set(n, not_peel)) {
4114 // If not used internal to the peeled region,
4115 // move "n" from peeled to not_peeled region.
4116 if (!has_use_internal_to_set(n, peel, loop)) {
4117 // if not pinned and not a load (which maybe anti-dependent on a store)
4118 // and not a CMove (Matcher expects only bool->cmove).
4119 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) {
4120 int new_clones = clone_for_use_outside_loop(loop, n, worklist);
4121 if (C->failing()) return false;
4122 if (new_clones == -1) {
4123 too_many_clones = true;
4124 break;
4125 }
4126 cloned_for_outside_use += new_clones;
4127 sink_list.push(n);
4128 peel.remove(n->_idx);
4129 not_peel.set(n->_idx);
4130 peel_list.remove(i);
4131 incr = false;
4132 #ifndef PRODUCT
4133 if (TracePartialPeeling) {
4134 tty->print_cr("sink to not_peeled region: %d newbb: %d",
4135 n->_idx, get_ctrl(n)->_idx);
4136 }
4137 #endif
4138 }
4139 } else {
4140 // Otherwise check for special def-use cases that span
4141 // the peel/not_peel boundary such as bool->if
4142 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
4143 new_phi_cnt++;
4144 }
4145 }
4146 }
4147 if (incr) i++;
4148 }
4149
4150 estimate += cloned_for_outside_use + new_phi_cnt;
4151 bool exceed_node_budget = !may_require_nodes(estimate);
4152 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
4153
4154 if (too_many_clones || exceed_node_budget || exceed_phi_limit) {
4155 #ifndef PRODUCT
4156 if (TracePartialPeeling && exceed_phi_limit) {
4157 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
4158 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F');
4159 }
4160 #endif
4161 if (new_peel_if != nullptr) {
4162 remove_cmpi_loop_exit(new_peel_if, loop);
4163 }
4164 // Inhibit more partial peeling on this loop
4165 assert(!head->is_partial_peel_loop(), "not partial peeled");
4166 head->mark_partial_peel_failed();
4167 if (cloned_for_outside_use > 0) {
4168 // Terminate this round of loop opts because
4169 // the graph outside this loop was changed.
4170 C->set_major_progress();
4171 return true;
4172 }
4173 return false;
4174 }
4175
4176 // Step 3: clone loop, retarget control, and insert new phis
4177
4178 // Create new loop head for new phis and to hang
4179 // the nodes being moved (sinked) from the peel region.
4180 LoopNode* new_head = new LoopNode(last_peel, last_peel);
4181 new_head->set_unswitch_count(head->unswitch_count()); // Preserve
4182 _igvn.register_new_node_with_optimizer(new_head);
4183 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
4184 _igvn.replace_input_of(first_not_peeled, 0, new_head);
4185 set_loop(new_head, loop);
4186 loop->_body.push(new_head);
4187 not_peel.set(new_head->_idx);
4188 set_idom(new_head, last_peel, dom_depth(first_not_peeled));
4189 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
4190
4191 while (sink_list.size() > 0) {
4192 Node* n = sink_list.pop();
4193 set_ctrl(n, new_head);
4194 }
4195
4196 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
4197
4198 clone_loop(loop, old_new, dd, IgnoreStripMined);
4199
4200 const uint clone_exit_idx = 1;
4201 const uint orig_exit_idx = 2;
4202 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
4203
4204 Node* head_clone = old_new[head->_idx];
4205 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
4206 Node* orig_tail_clone = head_clone->in(2);
4207
4208 // Add phi if "def" node is in peel set and "use" is not
4209
4210 for (uint i = 0; i < peel_list.size(); i++) {
4211 Node *def = peel_list.at(i);
4212 if (!def->is_CFG()) {
4213 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
4214 Node *use = def->fast_out(j);
4215 if (has_node(use) && use->in(0) != C->top() &&
4216 (!peel.test(use->_idx) ||
4217 (use->is_Phi() && use->in(0) == head)) ) {
4218 worklist.push(use);
4219 }
4220 }
4221 while( worklist.size() ) {
4222 Node *use = worklist.pop();
4223 for (uint j = 1; j < use->req(); j++) {
4224 Node* n = use->in(j);
4225 if (n == def) {
4226
4227 // "def" is in peel set, "use" is not in peel set
4228 // or "use" is in the entry boundary (a phi) of the peel set
4229
4230 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
4231
4232 if ( loop->is_member(get_loop( use_c )) ) {
4233 // use is in loop
4234 if (old_new[use->_idx] != nullptr) { // null for dead code
4235 Node* use_clone = old_new[use->_idx];
4236 _igvn.replace_input_of(use, j, C->top());
4237 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
4238 }
4239 } else {
4240 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
4241 // use is not in the loop, check if the live range includes the cut
4242 Node* lp_if = use_c->in(orig_exit_idx)->in(0);
4243 if (not_peel.test(lp_if->_idx)) {
4244 assert(j == orig_exit_idx, "use from original loop");
4245 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
4246 }
4247 }
4248 }
4249 }
4250 }
4251 }
4252 }
4253
4254 // Step 3b: retarget control
4255
4256 // Redirect control to the new loop head if a cloned node in
4257 // the not_peeled region has control that points into the peeled region.
4258 // This necessary because the cloned peeled region will be outside
4259 // the loop.
4260 // from to
4261 // cloned-peeled <---+
4262 // new_head_clone: | <--+
4263 // cloned-not_peeled in(0) in(0)
4264 // orig-peeled
4265
4266 for (uint i = 0; i < loop->_body.size(); i++) {
4267 Node *n = loop->_body.at(i);
4268 if (!n->is_CFG() && n->in(0) != nullptr &&
4269 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
4270 Node* n_clone = old_new[n->_idx];
4271 if (n_clone->depends_only_on_test()) {
4272 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the
4273 // backedge is removed, an array load could become dependent on a condition that's not a range check for that
4274 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
4275 // floating above its range check.
4276 Node* pinned_clone = n_clone->pin_array_access_node();
4277 if (pinned_clone != nullptr) {
4278 register_new_node_with_ctrl_of(pinned_clone, n_clone);
4279 old_new.map(n->_idx, pinned_clone);
4280 _igvn.replace_node(n_clone, pinned_clone);
4281 n_clone = pinned_clone;
4282 }
4283 }
4284 _igvn.replace_input_of(n_clone, 0, new_head_clone);
4285 }
4286 }
4287
4288 // Backedge of the surviving new_head (the clone) is original last_peel
4289 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
4290
4291 // Cut first node in original not_peel set
4292 _igvn.rehash_node_delayed(new_head); // Multiple edge updates:
4293 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of
4294 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls
4295
4296 // Copy head_clone back-branch info to original head
4297 // and remove original head's loop entry and
4298 // clone head's back-branch
4299 _igvn.rehash_node_delayed(head); // Multiple edge updates
4300 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
4301 head->set_req(LoopNode::LoopBackControl, C->top());
4302 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
4303
4304 // Similarly modify the phis
4305 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
4306 Node* use = head->fast_out(k);
4307 if (use->is_Phi() && use->outcnt() > 0) {
4308 Node* use_clone = old_new[use->_idx];
4309 _igvn.rehash_node_delayed(use); // Multiple edge updates
4310 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
4311 use->set_req(LoopNode::LoopBackControl, C->top());
4312 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
4313 }
4314 }
4315
4316 // Step 4: update dominator tree and dominator depth
4317
4318 set_idom(head, orig_tail_clone, dd);
4319 recompute_dom_depth();
4320
4321 // Inhibit more partial peeling on this loop
4322 new_head_clone->set_partial_peel_loop();
4323 C->set_major_progress();
4324 loop->record_for_igvn();
4325
4326 #ifndef PRODUCT
4327 if (TracePartialPeeling) {
4328 tty->print_cr("\nafter partial peel one iteration");
4329 Node_List wl;
4330 Node* t = last_peel;
4331 while (true) {
4332 wl.push(t);
4333 if (t == head_clone) break;
4334 t = idom(t);
4335 }
4336 while (wl.size() > 0) {
4337 Node* tt = wl.pop();
4338 if (tt == head) tty->print_cr("orig head");
4339 else if (tt == new_head_clone) tty->print_cr("new head");
4340 else if (tt == head_clone) tty->print_cr("clone head");
4341 tt->dump();
4342 }
4343 }
4344 #endif
4345
4346 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone);
4347
4348 return true;
4349 }
4350
4351 // Transform:
4352 //
4353 // loop<-----------------+
4354 // | |
4355 // stmt1 stmt2 .. stmtn |
4356 // | | | |
4357 // \ | / |
4358 // v v v |
4359 // region |
4360 // | |
4361 // shared_stmt |
4362 // | |
4363 // v |
4364 // if |
4365 // / \ |
4366 // | -----------+
4367 // v
4368 //
4369 // into:
4370 //
4371 // loop<-------------------+
4372 // | |
4373 // v |
4374 // +->loop |
4375 // | | |
4376 // | stmt1 stmt2 .. stmtn |
4377 // | | | | |
4378 // | | \ / |
4379 // | | v v |
4380 // | | region1 |
4381 // | | | |
4382 // | shared_stmt shared_stmt |
4383 // | | | |
4384 // | v v |
4385 // | if if |
4386 // | /\ / \ |
4387 // +-- | | -------+
4388 // \ /
4389 // v v
4390 // region2
4391 //
4392 // (region2 is shown to merge mirrored projections of the loop exit
4393 // ifs to make the diagram clearer but they really merge the same
4394 // projection)
4395 //
4396 // Conditions for this transformation to trigger:
4397 // - the path through stmt1 is frequent enough
4398 // - the inner loop will be turned into a counted loop after transformation
4399 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) {
4400 if (!DuplicateBackedge) {
4401 return false;
4402 }
4403 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only");
4404 if (!loop->_head->is_Loop()) {
4405 return false;
4406 }
4407
4408 uint estimate = loop->est_loop_clone_sz(1);
4409 if (exceeding_node_budget(estimate)) {
4410 return false;
4411 }
4412
4413 LoopNode *head = loop->_head->as_Loop();
4414
4415 Node* region = nullptr;
4416 IfNode* exit_test = nullptr;
4417 uint inner;
4418 float f;
4419 if (StressDuplicateBackedge) {
4420 if (head->is_strip_mined()) {
4421 return false;
4422 }
4423 Node* c = head->in(LoopNode::LoopBackControl);
4424
4425 while (c != head) {
4426 if (c->is_Region()) {
4427 region = c;
4428 }
4429 c = idom(c);
4430 }
4431
4432 if (region == nullptr) {
4433 return false;
4434 }
4435
4436 inner = 1;
4437 } else {
4438 // Is the shape of the loop that of a counted loop...
4439 Node* back_control = loop_exit_control(head, loop);
4440 if (back_control == nullptr) {
4441 return false;
4442 }
4443
4444 BoolTest::mask bt = BoolTest::illegal;
4445 float cl_prob = 0;
4446 Node* incr = nullptr;
4447 Node* limit = nullptr;
4448 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
4449 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) {
4450 return false;
4451 }
4452
4453 // With an extra phi for the candidate iv?
4454 // Or the region node is the loop head
4455 if (!incr->is_Phi() || incr->in(0) == head) {
4456 return false;
4457 }
4458
4459 PathFrequency pf(head, this);
4460 region = incr->in(0);
4461
4462 // Go over all paths for the extra phi's region and see if that
4463 // path is frequent enough and would match the expected iv shape
4464 // if the extra phi is removed
4465 inner = 0;
4466 for (uint i = 1; i < incr->req(); ++i) {
4467 Node* in = incr->in(i);
4468 Node* trunc1 = nullptr;
4469 Node* trunc2 = nullptr;
4470 const TypeInteger* iv_trunc_t = nullptr;
4471 Node* orig_in = in;
4472 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) {
4473 continue;
4474 }
4475 assert(in->Opcode() == Op_AddI, "wrong increment code");
4476 Node* xphi = nullptr;
4477 Node* stride = loop_iv_stride(in, loop, xphi);
4478
4479 if (stride == nullptr) {
4480 continue;
4481 }
4482
4483 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop);
4484 if (phi == nullptr ||
4485 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
4486 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
4487 return false;
4488 }
4489
4490 f = pf.to(region->in(i));
4491 if (f > 0.5) {
4492 inner = i;
4493 break;
4494 }
4495 }
4496
4497 if (inner == 0) {
4498 return false;
4499 }
4500
4501 exit_test = back_control->in(0)->as_If();
4502 }
4503
4504 if (idom(region)->is_Catch()) {
4505 return false;
4506 }
4507
4508 // Collect all control nodes that need to be cloned (shared_stmt in the diagram)
4509 Unique_Node_List wq;
4510 wq.push(head->in(LoopNode::LoopBackControl));
4511 for (uint i = 0; i < wq.size(); i++) {
4512 Node* c = wq.at(i);
4513 assert(get_loop(c) == loop, "not in the right loop?");
4514 if (c->is_Region()) {
4515 if (c != region) {
4516 for (uint j = 1; j < c->req(); ++j) {
4517 wq.push(c->in(j));
4518 }
4519 }
4520 } else {
4521 wq.push(c->in(0));
4522 }
4523 assert(!is_strict_dominator(c, region), "shouldn't go above region");
4524 }
4525
4526 Node* region_dom = idom(region);
4527
4528 // Can't do the transformation if this would cause a membar pair to
4529 // be split
4530 for (uint i = 0; i < wq.size(); i++) {
4531 Node* c = wq.at(i);
4532 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) {
4533 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair");
4534 if (!wq.member(c->as_MemBar()->leading_membar())) {
4535 return false;
4536 }
4537 }
4538 }
4539
4540 // Collect data nodes that need to be clones as well
4541 int dd = dom_depth(head);
4542
4543 for (uint i = 0; i < loop->_body.size(); ++i) {
4544 Node* n = loop->_body.at(i);
4545 if (has_ctrl(n)) {
4546 Node* c = get_ctrl(n);
4547 if (wq.member(c)) {
4548 wq.push(n);
4549 }
4550 } else {
4551 set_idom(n, idom(n), dd);
4552 }
4553 }
4554
4555 // clone shared_stmt
4556 clone_loop_body(wq, old_new, nullptr);
4557
4558 Node* region_clone = old_new[region->_idx];
4559 region_clone->set_req(inner, C->top());
4560 set_idom(region, region->in(inner), dd);
4561
4562 // Prepare the outer loop
4563 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]);
4564 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl));
4565 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head);
4566 set_idom(head, outer_head, dd);
4567
4568 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true);
4569
4570 // Make one of the shared_stmt copies only reachable from stmt1, the
4571 // other only from stmt2..stmtn.
4572 Node* dom = nullptr;
4573 for (uint i = 1; i < region->req(); ++i) {
4574 if (i != inner) {
4575 _igvn.replace_input_of(region, i, C->top());
4576 }
4577 Node* in = region_clone->in(i);
4578 if (in->is_top()) {
4579 continue;
4580 }
4581 if (dom == nullptr) {
4582 dom = in;
4583 } else {
4584 dom = dom_lca(dom, in);
4585 }
4586 }
4587
4588 set_idom(region_clone, dom, dd);
4589
4590 // Set up the outer loop
4591 for (uint i = 0; i < head->outcnt(); i++) {
4592 Node* u = head->raw_out(i);
4593 if (u->is_Phi()) {
4594 Node* outer_phi = u->clone();
4595 outer_phi->set_req(0, outer_head);
4596 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx];
4597 if (backedge == nullptr) {
4598 backedge = u->in(LoopNode::LoopBackControl);
4599 }
4600 outer_phi->set_req(LoopNode::LoopBackControl, backedge);
4601 register_new_node(outer_phi, outer_head);
4602 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi);
4603 }
4604 }
4605
4606 // create control and data nodes for out of loop uses (including region2)
4607 Node_List worklist;
4608 uint new_counter = C->unique();
4609 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist);
4610
4611 Node_List *split_if_set = nullptr;
4612 Node_List *split_bool_set = nullptr;
4613 Node_List *split_cex_set = nullptr;
4614 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist,
4615 split_if_set, split_bool_set, split_cex_set);
4616
4617 finish_clone_loop(split_if_set, split_bool_set, split_cex_set);
4618
4619 if (exit_test != nullptr) {
4620 float cnt = exit_test->_fcnt;
4621 if (cnt != COUNT_UNKNOWN) {
4622 exit_test->_fcnt = cnt * f;
4623 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f);
4624 }
4625 }
4626
4627 C->set_major_progress();
4628
4629 return true;
4630 }
4631
4632 // AutoVectorize the loop: replace scalar ops with vector ops.
4633 PhaseIdealLoop::AutoVectorizeStatus
4634 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) {
4635 // Counted loop only
4636 if (!lpt->is_counted()) {
4637 return AutoVectorizeStatus::Impossible;
4638 }
4639
4640 // Main-loop only
4641 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4642 if (!cl->is_main_loop()) {
4643 return AutoVectorizeStatus::Impossible;
4644 }
4645
4646 VLoop vloop(lpt, false);
4647 if (!vloop.check_preconditions()) {
4648 return AutoVectorizeStatus::TriedAndFailed;
4649 }
4650
4651 // Ensure the shared data is cleared before each use
4652 vshared.clear();
4653
4654 const VLoopAnalyzer vloop_analyzer(vloop, vshared);
4655 if (!vloop_analyzer.success()) {
4656 return AutoVectorizeStatus::TriedAndFailed;
4657 }
4658
4659 SuperWord sw(vloop_analyzer);
4660 if (!sw.transform_loop()) {
4661 return AutoVectorizeStatus::TriedAndFailed;
4662 }
4663
4664 return AutoVectorizeStatus::Success;
4665 }
4666
4667 // Just before insert_pre_post_loops, we can multiversion the loop:
4668 //
4669 // multiversion_if
4670 // | |
4671 // fast_loop slow_loop
4672 //
4673 // In the fast_loop we can make speculative assumptions, and put the
4674 // conditions into the multiversion_if. If the conditions hold at runtime,
4675 // we enter the fast_loop, if the conditions fail, we take the slow_loop
4676 // instead which does not make any of the speculative assumptions.
4677 //
4678 // Note: we only multiversion the loop if the loop does not have any
4679 // auto vectorization check Predicate. If we have that predicate,
4680 // then we can simply add the speculative assumption checks to
4681 // that Predicate. This means we do not need to duplicate the
4682 // loop - we have a smaller graph and save compile time. Should
4683 // the conditions ever fail, then we deopt / trap at the Predicate
4684 // and recompile without that Predicate. At that point we will
4685 // multiversion the loop, so that we can still have speculative
4686 // runtime checks.
4687 //
4688 // We perform the multiversioning when the loop is still in its single
4689 // iteration form, even before we insert pre and post loops. This makes
4690 // the cloning much simpler. However, this means that both the fast
4691 // and the slow loop have to be optimized independently (adding pre
4692 // and post loops, unrolling the main loop, auto-vectorize etc.). And
4693 // we may end up not needing any speculative assumptions in the fast_loop
4694 // and then rejecting the slow_loop by constant folding the multiversion_if.
4695 //
4696 // Therefore, we "delay" the optimization of the slow_loop until we add
4697 // at least one speculative assumption for the fast_loop. If we never
4698 // add such a speculative runtime check, the OpaqueMultiversioningNode
4699 // of the multiversion_if constant folds to true after loop opts, and the
4700 // multiversion_if folds away the "delayed" slow_loop. If we add any
4701 // speculative assumption, then we notify the OpaqueMultiversioningNode
4702 // with "notify_slow_loop_that_it_can_resume_optimizations".
4703 //
4704 // Note: new runtime checks can be added to the multiversion_if with
4705 // PhaseIdealLoop::create_new_if_for_multiversion
4706 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) {
4707 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4708 LoopNode* outer_loop = cl->skip_strip_mined();
4709 Node* entry = outer_loop->in(LoopNode::EntryControl);
4710
4711 // Check we have multiversioning enabled, and are not already multiversioned.
4712 if (!LoopMultiversioning || cl->is_multiversion()) { return; }
4713
4714 // Check that we do not have a parse-predicate where we can add the runtime checks
4715 // during auto-vectorization.
4716 const Predicates predicates(entry);
4717 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block();
4718 if (predicate_block->has_parse_predicate()) { return; }
4719
4720 // Check node budget.
4721 uint estimate = lpt->est_loop_clone_sz(2);
4722 if (!may_require_nodes(estimate)) { return; }
4723
4724 do_multiversioning(lpt, old_new);
4725 }
4726
4727 // Returns true if the Reduction node is unordered.
4728 static bool is_unordered_reduction(Node* n) {
4729 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order();
4730 }
4731
4732 // Having ReductionNodes in the loop is expensive. They need to recursively
4733 // fold together the vector values, for every vectorized loop iteration. If
4734 // we encounter the following pattern, we can vector accumulate the values
4735 // inside the loop, and only have a single UnorderedReduction after the loop.
4736 //
4737 // Note: UnorderedReduction represents a ReductionNode which does not require
4738 // calculating in strict order.
4739 //
4740 // CountedLoop init
4741 // | |
4742 // +------+ | +-----------------------+
4743 // | | | |
4744 // PhiNode (s) |
4745 // | |
4746 // | Vector |
4747 // | | |
4748 // UnorderedReduction (first_ur) |
4749 // | |
4750 // ... Vector |
4751 // | | |
4752 // UnorderedReduction (last_ur) |
4753 // | |
4754 // +---------------------+
4755 //
4756 // We patch the graph to look like this:
4757 //
4758 // CountedLoop identity_vector
4759 // | |
4760 // +-------+ | +---------------+
4761 // | | | |
4762 // PhiNode (v) |
4763 // | |
4764 // | Vector |
4765 // | | |
4766 // VectorAccumulator |
4767 // | |
4768 // ... Vector |
4769 // | | |
4770 // init VectorAccumulator |
4771 // | | | |
4772 // UnorderedReduction +-----------+
4773 //
4774 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we
4775 // use vector_accumulators, which do the same reductions, but only element
4776 // wise. This is a single operation per vector_accumulator, rather than many
4777 // for a UnorderedReduction. We can then reduce the last vector_accumulator
4778 // after the loop, and also reduce the init value into it.
4779 //
4780 // We can not do this with all reductions. Some reductions do not allow the
4781 // reordering of operations (for example float addition/multiplication require
4782 // strict order).
4783 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) {
4784 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity");
4785
4786 // Find all Phi nodes with an unordered Reduction on backedge.
4787 CountedLoopNode* cl = loop->_head->as_CountedLoop();
4788 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) {
4789 Node* phi = cl->fast_out(j);
4790 // We have a phi with a single use, and an unordered Reduction on the backedge.
4791 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) {
4792 continue;
4793 }
4794
4795 ReductionNode* last_ur = phi->in(2)->as_Reduction();
4796 assert(!last_ur->requires_strict_order(), "must be");
4797
4798 // Determine types
4799 const TypeVect* vec_t = last_ur->vect_type();
4800 uint vector_length = vec_t->length();
4801 BasicType bt = vec_t->element_basic_type();
4802
4803 // Convert opcode from vector-reduction -> scalar -> normal-vector-op
4804 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt);
4805 const int vopc = VectorNode::opcode(sopc, bt);
4806 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) {
4807 DEBUG_ONLY( last_ur->dump(); )
4808 assert(false, "do not have normal vector op for this reduction");
4809 continue; // not implemented -> fails
4810 }
4811
4812 // Traverse up the chain of unordered Reductions, checking that it loops back to
4813 // the phi. Check that all unordered Reductions only have a single use, except for
4814 // the last (last_ur), which only has phi as a use in the loop, and all other uses
4815 // are outside the loop.
4816 ReductionNode* current = last_ur;
4817 ReductionNode* first_ur = nullptr;
4818 while (true) {
4819 assert(!current->requires_strict_order(), "sanity");
4820
4821 // Expect no ctrl and a vector_input from within the loop.
4822 Node* ctrl = current->in(0);
4823 Node* vector_input = current->in(2);
4824 if (ctrl != nullptr || get_ctrl(vector_input) != cl) {
4825 DEBUG_ONLY( current->dump(1); )
4826 assert(false, "reduction has ctrl or bad vector_input");
4827 break; // Chain traversal fails.
4828 }
4829
4830 assert(current->vect_type() != nullptr, "must have vector type");
4831 if (current->vect_type() != last_ur->vect_type()) {
4832 // Reductions do not have the same vector type (length and element type).
4833 break; // Chain traversal fails.
4834 }
4835
4836 // Expect single use of an unordered Reduction, except for last_ur.
4837 if (current == last_ur) {
4838 // Expect all uses to be outside the loop, except phi.
4839 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) {
4840 Node* use = current->fast_out(k);
4841 if (use != phi && ctrl_or_self(use) == cl) {
4842 DEBUG_ONLY( current->dump(-1); )
4843 assert(false, "reduction has use inside loop");
4844 // Should not be allowed by SuperWord::mark_reductions
4845 return; // bail out of optimization
4846 }
4847 }
4848 } else {
4849 if (current->outcnt() != 1) {
4850 break; // Chain traversal fails.
4851 }
4852 }
4853
4854 // Expect another unordered Reduction or phi as the scalar input.
4855 Node* scalar_input = current->in(1);
4856 if (is_unordered_reduction(scalar_input) &&
4857 scalar_input->Opcode() == current->Opcode()) {
4858 // Move up the unordered Reduction chain.
4859 current = scalar_input->as_Reduction();
4860 assert(!current->requires_strict_order(), "must be");
4861 } else if (scalar_input == phi) {
4862 // Chain terminates at phi.
4863 first_ur = current;
4864 current = nullptr;
4865 break; // Success.
4866 } else {
4867 // scalar_input is neither phi nor a matching reduction
4868 // Can for example be scalar reduction when we have
4869 // partial vectorization.
4870 break; // Chain traversal fails.
4871 }
4872 }
4873 if (current != nullptr) {
4874 // Chain traversal was not successful.
4875 continue;
4876 }
4877 assert(first_ur != nullptr, "must have successfully terminated chain traversal");
4878
4879 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt);
4880 set_root_as_ctrl(identity_scalar);
4881 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt);
4882 register_new_node(identity_vector, C->root());
4883 assert(vec_t == identity_vector->vect_type(), "matching vector type");
4884 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction");
4885
4886 // Turn the scalar phi into a vector phi.
4887 _igvn.rehash_node_delayed(phi);
4888 Node* init = phi->in(1); // Remember init before replacing it.
4889 phi->set_req_X(1, identity_vector, &_igvn);
4890 phi->as_Type()->set_type(vec_t);
4891 _igvn.set_type(phi, vec_t);
4892
4893 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators.
4894 current = first_ur;
4895 while (true) {
4896 // Create vector_accumulator to replace current.
4897 Node* last_vector_accumulator = current->in(1);
4898 Node* vector_input = current->in(2);
4899 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t);
4900 register_new_node(vector_accumulator, cl);
4901 _igvn.replace_node(current, vector_accumulator);
4902 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction");
4903 if (current == last_ur) {
4904 break;
4905 }
4906 current = vector_accumulator->unique_out()->as_Reduction();
4907 assert(!current->requires_strict_order(), "must be");
4908 }
4909
4910 // Create post-loop reduction.
4911 Node* last_accumulator = phi->in(2);
4912 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt);
4913
4914 // Take over uses of last_accumulator that are not in the loop.
4915 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) {
4916 Node* use = last_accumulator->out(i);
4917 if (use != phi && use != post_loop_reduction) {
4918 assert(ctrl_or_self(use) != cl, "use must be outside loop");
4919 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn);
4920 --i;
4921 }
4922 }
4923 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl));
4924 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction");
4925
4926 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction");
4927 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator");
4928 assert(phi->outcnt() == 1, "accumulator is the only use of phi");
4929 }
4930 }
4931
4932 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) {
4933 for (uint i = 0; i < _data_nodes.size(); i++) {
4934 clone(_data_nodes[i], new_ctrl);
4935 }
4936 }
4937
4938 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl.
4939 void DataNodeGraph::clone(Node* node, Node* new_ctrl) {
4940 Node* clone = node->clone();
4941 _phase->igvn().register_new_node_with_optimizer(clone);
4942 _orig_to_new.put(node, clone);
4943 _phase->set_ctrl(clone, new_ctrl);
4944 if (node->is_CastII()) {
4945 clone->set_req(0, new_ctrl);
4946 }
4947 }
4948
4949 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their
4950 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph.
4951 void DataNodeGraph::rewire_clones_to_cloned_inputs() {
4952 _orig_to_new.iterate_all([&](Node* node, Node* clone) {
4953 for (uint i = 1; i < node->req(); i++) {
4954 Node** cloned_input = _orig_to_new.get(node->in(i));
4955 if (cloned_input != nullptr) {
4956 // Input was also cloned -> rewire clone to the cloned input.
4957 _phase->igvn().replace_input_of(clone, i, *cloned_input);
4958 }
4959 }
4960 });
4961 }
4962
4963 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes.
4964 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes.
4965 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes(
4966 const TransformStrategyForOpaqueLoopNodes& transform_strategy,
4967 Node* new_ctrl) {
4968 for (uint i = 0; i < _data_nodes.size(); i++) {
4969 Node* data_node = _data_nodes[i];
4970 if (data_node->is_Opaque1()) {
4971 transform_opaque_node(transform_strategy, data_node);
4972 } else {
4973 clone(data_node, new_ctrl);
4974 }
4975 }
4976 }
4977
4978 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) {
4979 Node* transformed_node;
4980 if (node->is_OpaqueLoopInit()) {
4981 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit());
4982 } else {
4983 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode");
4984 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride());
4985 }
4986 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs().
4987 _orig_to_new.put(node, transformed_node);
4988 }