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