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