1 /*
2 * Copyright (c) 1998, 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 "ci/ciMethodData.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/connode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/mulnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/opcodes.hpp"
45 #include "opto/predicates.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/vectorization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/checkedCast.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //=============================================================================
54 //--------------------------is_cloop_ind_var-----------------------------------
55 // Determine if a node is a counted loop induction variable.
56 // NOTE: The method is declared in "node.hpp".
57 bool Node::is_cloop_ind_var() const {
58 return (is_Phi() &&
59 as_Phi()->region()->is_CountedLoop() &&
60 as_Phi()->region()->as_CountedLoop()->phi() == this);
61 }
62
63 //=============================================================================
64 //------------------------------dump_spec--------------------------------------
65 // Dump special per-node info
66 #ifndef PRODUCT
67 void LoopNode::dump_spec(outputStream *st) const {
68 RegionNode::dump_spec(st);
69 if (is_inner_loop()) st->print( "inner " );
70 if (is_partial_peel_loop()) st->print( "partial_peel " );
71 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
72 }
73 #endif
74
75 //------------------------------is_valid_counted_loop-------------------------
76 bool LoopNode::is_valid_counted_loop(BasicType bt) const {
77 if (is_BaseCountedLoop() && as_BaseCountedLoop()->bt() == bt) {
78 BaseCountedLoopNode* l = as_BaseCountedLoop();
79 BaseCountedLoopEndNode* le = l->loopexit_or_null();
80 if (le != nullptr &&
81 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
82 Node* phi = l->phi();
83 Node* exit = le->proj_out_or_null(0 /* false */);
84 if (exit != nullptr && exit->Opcode() == Op_IfFalse &&
85 phi != nullptr && phi->is_Phi() &&
86 phi->in(LoopNode::LoopBackControl) == l->incr() &&
87 le->loopnode() == l && le->stride_is_con()) {
88 return true;
89 }
90 }
91 }
92 return false;
93 }
94
95 //------------------------------get_early_ctrl---------------------------------
96 // Compute earliest legal control
97 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
98 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
99 uint i;
100 Node *early;
101 if (n->in(0) && !n->is_expensive()) {
102 early = n->in(0);
103 if (!early->is_CFG()) // Might be a non-CFG multi-def
104 early = get_ctrl(early); // So treat input as a straight data input
105 i = 1;
106 } else {
107 early = get_ctrl(n->in(1));
108 i = 2;
109 }
110 uint e_d = dom_depth(early);
111 assert( early, "" );
112 for (; i < n->req(); i++) {
113 Node *cin = get_ctrl(n->in(i));
114 assert( cin, "" );
115 // Keep deepest dominator depth
116 uint c_d = dom_depth(cin);
117 if (c_d > e_d) { // Deeper guy?
118 early = cin; // Keep deepest found so far
119 e_d = c_d;
120 } else if (c_d == e_d && // Same depth?
121 early != cin) { // If not equal, must use slower algorithm
122 // If same depth but not equal, one _must_ dominate the other
123 // and we want the deeper (i.e., dominated) guy.
124 Node *n1 = early;
125 Node *n2 = cin;
126 while (1) {
127 n1 = idom(n1); // Walk up until break cycle
128 n2 = idom(n2);
129 if (n1 == cin || // Walked early up to cin
130 dom_depth(n2) < c_d)
131 break; // early is deeper; keep him
132 if (n2 == early || // Walked cin up to early
133 dom_depth(n1) < c_d) {
134 early = cin; // cin is deeper; keep him
135 break;
136 }
137 }
138 e_d = dom_depth(early); // Reset depth register cache
139 }
140 }
141
142 // Return earliest legal location
143 assert(early == find_non_split_ctrl(early), "unexpected early control");
144
145 if (n->is_expensive() && !_verify_only && !_verify_me) {
146 assert(n->in(0), "should have control input");
147 early = get_early_ctrl_for_expensive(n, early);
148 }
149
150 return early;
151 }
152
153 //------------------------------get_early_ctrl_for_expensive---------------------------------
154 // Move node up the dominator tree as high as legal while still beneficial
155 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
156 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
157 assert(OptimizeExpensiveOps, "optimization off?");
158
159 Node* ctl = n->in(0);
160 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
161 uint min_dom_depth = dom_depth(earliest);
162 #ifdef ASSERT
163 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
164 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
165 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
166 }
167 #endif
168 if (dom_depth(ctl) < min_dom_depth) {
169 return earliest;
170 }
171
172 while (true) {
173 Node* next = ctl;
174 // Moving the node out of a loop on the projection of an If
175 // confuses Loop Predication. So, once we hit a loop in an If branch
176 // that doesn't branch to an UNC, we stop. The code that process
177 // expensive nodes will notice the loop and skip over it to try to
178 // move the node further up.
179 if (ctl->is_CountedLoop() && ctl->in(1) != nullptr && ctl->in(1)->in(0) != nullptr && ctl->in(1)->in(0)->is_If()) {
180 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern()) {
181 break;
182 }
183 next = idom(ctl->in(1)->in(0));
184 } else if (ctl->is_Proj()) {
185 // We only move it up along a projection if the projection is
186 // the single control projection for its parent: same code path,
187 // if it's a If with UNC or fallthrough of a call.
188 Node* parent_ctl = ctl->in(0);
189 if (parent_ctl == nullptr) {
190 break;
191 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != nullptr) {
192 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
193 } else if (parent_ctl->is_If()) {
194 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern()) {
195 break;
196 }
197 assert(idom(ctl) == parent_ctl, "strange");
198 next = idom(parent_ctl);
199 } else if (ctl->is_CatchProj()) {
200 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
201 break;
202 }
203 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
204 next = parent_ctl->in(0)->in(0)->in(0);
205 } else {
206 // Check if parent control has a single projection (this
207 // control is the only possible successor of the parent
208 // control). If so, we can try to move the node above the
209 // parent control.
210 int nb_ctl_proj = 0;
211 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
212 Node *p = parent_ctl->fast_out(i);
213 if (p->is_Proj() && p->is_CFG()) {
214 nb_ctl_proj++;
215 if (nb_ctl_proj > 1) {
216 break;
217 }
218 }
219 }
220
221 if (nb_ctl_proj > 1) {
222 break;
223 }
224 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
225 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
226 assert(idom(ctl) == parent_ctl, "strange");
227 next = idom(parent_ctl);
228 }
229 } else {
230 next = idom(ctl);
231 }
232 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
233 break;
234 }
235 ctl = next;
236 }
237
238 if (ctl != n->in(0)) {
239 _igvn.replace_input_of(n, 0, ctl);
240 _igvn.hash_insert(n);
241 }
242
243 return ctl;
244 }
245
246
247 //------------------------------set_early_ctrl---------------------------------
248 // Set earliest legal control
249 void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {
250 Node *early = get_early_ctrl(n);
251
252 // Record earliest legal location
253 set_ctrl(n, early);
254 IdealLoopTree *loop = get_loop(early);
255 if (update_body && loop->_child == nullptr) {
256 loop->_body.push(n);
257 }
258 }
259
260 //------------------------------set_subtree_ctrl-------------------------------
261 // set missing _ctrl entries on new nodes
262 void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {
263 // Already set? Get out.
264 if (_loop_or_ctrl[n->_idx]) return;
265 // Recursively set _loop_or_ctrl array to indicate where the Node goes
266 uint i;
267 for (i = 0; i < n->req(); ++i) {
268 Node *m = n->in(i);
269 if (m && m != C->root()) {
270 set_subtree_ctrl(m, update_body);
271 }
272 }
273
274 // Fixup self
275 set_early_ctrl(n, update_body);
276 }
277
278 IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {
279 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
280 IdealLoopTree* parent = loop->_parent;
281 IdealLoopTree* sibling = parent->_child;
282 if (sibling == loop) {
283 parent->_child = outer_ilt;
284 } else {
285 while (sibling->_next != loop) {
286 sibling = sibling->_next;
287 }
288 sibling->_next = outer_ilt;
289 }
290 outer_ilt->_next = loop->_next;
291 outer_ilt->_parent = parent;
292 outer_ilt->_child = loop;
293 outer_ilt->_nest = loop->_nest;
294 loop->_parent = outer_ilt;
295 loop->_next = nullptr;
296 loop->_nest++;
297 assert(loop->_nest <= SHRT_MAX, "sanity");
298 return outer_ilt;
299 }
300
301 // Create a skeleton strip mined outer loop: a Loop head before the
302 // inner strip mined loop, a safepoint and an exit condition guarded
303 // by an opaque node after the inner strip mined loop with a backedge
304 // to the loop head. The inner strip mined loop is left as it is. Only
305 // once loop optimizations are over, do we adjust the inner loop exit
306 // condition to limit its number of iterations, set the outer loop
307 // exit condition and add Phis to the outer loop head. Some loop
308 // optimizations that operate on the inner strip mined loop need to be
309 // aware of the outer strip mined loop: loop unswitching needs to
310 // clone the outer loop as well as the inner, unrolling needs to only
311 // clone the inner loop etc. No optimizations need to change the outer
312 // strip mined loop as it is only a skeleton.
313 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
314 IdealLoopTree* loop, float cl_prob, float le_fcnt,
315 Node*& entry_control, Node*& iffalse) {
316 Node* outer_test = intcon(0);
317 Node *orig = iffalse;
318 iffalse = iffalse->clone();
319 _igvn.register_new_node_with_optimizer(iffalse);
320 set_idom(iffalse, idom(orig), dom_depth(orig));
321
322 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
323 Node *outer_ift = new IfTrueNode (outer_le);
324 Node* outer_iff = orig;
325 _igvn.replace_input_of(outer_iff, 0, outer_le);
326
327 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
328 entry_control = outer_l;
329
330 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
331
332 set_loop(iffalse, outer_ilt);
333 // When this code runs, loop bodies have not yet been populated.
334 const bool body_populated = false;
335 register_control(outer_le, outer_ilt, iffalse, body_populated);
336 register_control(outer_ift, outer_ilt, outer_le, body_populated);
337 set_idom(outer_iff, outer_le, dom_depth(outer_le));
338 _igvn.register_new_node_with_optimizer(outer_l);
339 set_loop(outer_l, outer_ilt);
340 set_idom(outer_l, init_control, dom_depth(init_control)+1);
341
342 return outer_ilt;
343 }
344
345 void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
346 Node* cmp_limit, Node* bol) {
347 assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
348 Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
349 Deoptimization::Reason_loop_limit_check,
350 Op_If);
351 Node* iff = new_predicate_proj->in(0);
352 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
353 bol = _igvn.register_new_node_with_optimizer(bol);
354 set_subtree_ctrl(bol, false);
355 _igvn.replace_input_of(iff, 1, bol);
356
357 #ifndef PRODUCT
358 // report that the loop predication has been actually performed
359 // for this loop
360 if (TraceLoopLimitCheck) {
361 tty->print_cr("Counted Loop Limit Check generated:");
362 debug_only( bol->dump(2); )
363 }
364 #endif
365 }
366
367 Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {
368 // Counted loop head must be a good RegionNode with only 3 not null
369 // control input edges: Self, Entry, LoopBack.
370 if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {
371 return nullptr;
372 }
373 Node *init_control = x->in(LoopNode::EntryControl);
374 Node *back_control = x->in(LoopNode::LoopBackControl);
375 if (init_control == nullptr || back_control == nullptr) { // Partially dead
376 return nullptr;
377 }
378 // Must also check for TOP when looking for a dead loop
379 if (init_control->is_top() || back_control->is_top()) {
380 return nullptr;
381 }
382
383 // Allow funny placement of Safepoint
384 if (back_control->Opcode() == Op_SafePoint) {
385 back_control = back_control->in(TypeFunc::Control);
386 }
387
388 // Controlling test for loop
389 Node *iftrue = back_control;
390 uint iftrue_op = iftrue->Opcode();
391 if (iftrue_op != Op_IfTrue &&
392 iftrue_op != Op_IfFalse) {
393 // I have a weird back-control. Probably the loop-exit test is in
394 // the middle of the loop and I am looking at some trailing control-flow
395 // merge point. To fix this I would have to partially peel the loop.
396 return nullptr; // Obscure back-control
397 }
398
399 // Get boolean guarding loop-back test
400 Node *iff = iftrue->in(0);
401 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {
402 return nullptr;
403 }
404 return iftrue;
405 }
406
407 Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {
408 Node* iftrue = back_control;
409 uint iftrue_op = iftrue->Opcode();
410 Node* iff = iftrue->in(0);
411 BoolNode* test = iff->in(1)->as_Bool();
412 bt = test->_test._test;
413 cl_prob = iff->as_If()->_prob;
414 if (iftrue_op == Op_IfFalse) {
415 bt = BoolTest(bt).negate();
416 cl_prob = 1.0 - cl_prob;
417 }
418 // Get backedge compare
419 Node* cmp = test->in(1);
420 if (!cmp->is_Cmp()) {
421 return nullptr;
422 }
423
424 // Find the trip-counter increment & limit. Limit must be loop invariant.
425 incr = cmp->in(1);
426 limit = cmp->in(2);
427
428 // ---------
429 // need 'loop()' test to tell if limit is loop invariant
430 // ---------
431
432 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
433 Node* tmp = incr; // Then reverse order into the CmpI
434 incr = limit;
435 limit = tmp;
436 bt = BoolTest(bt).commute(); // And commute the exit test
437 }
438 if (is_member(loop, get_ctrl(limit))) { // Limit must be loop-invariant
439 return nullptr;
440 }
441 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
442 return nullptr;
443 }
444 return cmp;
445 }
446
447 Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {
448 if (incr->is_Phi()) {
449 if (incr->as_Phi()->region() != x || incr->req() != 3) {
450 return nullptr; // Not simple trip counter expression
451 }
452 phi_incr = incr;
453 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
454 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
455 return nullptr;
456 }
457 }
458 return incr;
459 }
460
461 Node* PhaseIdealLoop::loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi) {
462 assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
463 // Get merge point
464 xphi = incr->in(1);
465 Node *stride = incr->in(2);
466 if (!stride->is_Con()) { // Oops, swap these
467 if (!xphi->is_Con()) { // Is the other guy a constant?
468 return nullptr; // Nope, unknown stride, bail out
469 }
470 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
471 xphi = stride;
472 stride = tmp;
473 }
474 return stride;
475 }
476
477 PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop) {
478 if (!xphi->is_Phi()) {
479 return nullptr; // Too much math on the trip counter
480 }
481 if (phi_incr != nullptr && phi_incr != xphi) {
482 return nullptr;
483 }
484 PhiNode *phi = xphi->as_Phi();
485
486 // Phi must be of loop header; backedge must wrap to increment
487 if (phi->region() != x) {
488 return nullptr;
489 }
490 return phi;
491 }
492
493 static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {
494 if (final_correction > 0) {
495 if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {
496 return -1;
497 }
498 if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {
499 return 1;
500 }
501 } else {
502 if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {
503 return -1;
504 }
505 if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {
506 return 1;
507 }
508 }
509 return 0;
510 }
511
512 static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {
513 // If the condition is inverted and we will be rolling
514 // through MININT to MAXINT, then bail out.
515 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
516 // Odd stride
517 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
518 // Count down loop rolls through MAXINT
519 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
520 // Count up loop rolls through MININT
521 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
522 return false; // Bail out
523 }
524 return true;
525 }
526
527 Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
528 BasicType bt) {
529 Node* iv_as_long;
530 if (bt == T_LONG) {
531 iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
532 register_new_node(iv_as_long, inner_head);
533 } else {
534 iv_as_long = inner_iv;
535 }
536 Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
537 register_new_node(iv_replacement, inner_head);
538 for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
539 Node* u = iv_to_replace->last_out(i);
540 #ifdef ASSERT
541 if (!is_dominator(inner_head, ctrl_or_self(u))) {
542 assert(u->is_Phi(), "should be a Phi");
543 for (uint j = 1; j < u->req(); j++) {
544 if (u->in(j) == iv_to_replace) {
545 assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
546 }
547 }
548 }
549 #endif
550 _igvn.rehash_node_delayed(u);
551 int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
552 i -= nb;
553 }
554 return iv_replacement;
555 }
556
557 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
558 void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
559 SafePointNode* sfpt) {
560 if (!C->too_many_traps(reason)) {
561 ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
562 register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
563 Node* if_false = new IfFalseNode(parse_predicate);
564 register_control(if_false, _ltree_root, parse_predicate);
565 Node* if_true = new IfTrueNode(parse_predicate);
566 register_control(if_true, loop, parse_predicate);
567
568 int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
569 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
570 const TypePtr* no_memory_effects = nullptr;
571 JVMState* jvms = sfpt->jvms();
572 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
573 no_memory_effects);
574
575 Node* mem = nullptr;
576 Node* i_o = nullptr;
577 if (sfpt->is_Call()) {
578 mem = sfpt->proj_out(TypeFunc::Memory);
579 i_o = sfpt->proj_out(TypeFunc::I_O);
580 } else {
581 mem = sfpt->memory();
582 i_o = sfpt->i_o();
583 }
584
585 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
586 register_new_node(frame, C->start());
587 Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
588 register_new_node(ret, C->start());
589
590 unc->init_req(TypeFunc::Control, if_false);
591 unc->init_req(TypeFunc::I_O, i_o);
592 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
593 unc->init_req(TypeFunc::FramePtr, frame);
594 unc->init_req(TypeFunc::ReturnAdr, ret);
595 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
596 unc->set_cnt(PROB_UNLIKELY_MAG(4));
597 unc->copy_call_debug_info(&_igvn, sfpt);
598
599 for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
600 set_subtree_ctrl(unc->in(i), false);
601 }
602 register_control(unc, _ltree_root, if_false);
603
604 Node* ctrl = new ProjNode(unc, TypeFunc::Control);
605 register_control(ctrl, _ltree_root, unc);
606 Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
607 register_control(halt, _ltree_root, ctrl);
608 _igvn.add_input_to(C->root(), halt);
609
610 _igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
611 set_idom(inner_head, if_true, dom_depth(inner_head));
612 }
613 }
614
615 // Find a safepoint node that dominates the back edge. We need a
616 // SafePointNode so we can use its jvm state to create empty
617 // predicates.
618 static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
619 SafePointNode* safepoint = nullptr;
620 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
621 Node* u = x->fast_out(i);
622 if (u->is_memory_phi()) {
623 Node* m = u->in(LoopNode::LoopBackControl);
624 if (u->adr_type() == TypePtr::BOTTOM) {
625 if (m->is_MergeMem() && mem->is_MergeMem()) {
626 if (m != mem DEBUG_ONLY(|| true)) {
627 // MergeMemStream can modify m, for example to adjust the length to mem.
628 // This is unfortunate, and probably unnecessary. But as it is, we need
629 // to add m to the igvn worklist, else we may have a modified node that
630 // is not on the igvn worklist.
631 phase->igvn()._worklist.push(m);
632 for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
633 if (!mms.is_empty()) {
634 if (mms.memory() != mms.memory2()) {
635 return false;
636 }
637 #ifdef ASSERT
638 if (mms.alias_idx() != Compile::AliasIdxBot) {
639 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
640 }
641 #endif
642 }
643 }
644 }
645 } else if (mem->is_MergeMem()) {
646 if (m != mem->as_MergeMem()->base_memory()) {
647 return false;
648 }
649 } else {
650 return false;
651 }
652 } else {
653 if (mem->is_MergeMem()) {
654 if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {
655 return false;
656 }
657 #ifdef ASSERT
658 mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
659 #endif
660 } else {
661 if (m != mem) {
662 return false;
663 }
664 }
665 }
666 }
667 }
668 return true;
669 }
670
671 SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
672 IfNode* exit_test = back_control->in(0)->as_If();
673 SafePointNode* safepoint = nullptr;
674 if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
675 safepoint = exit_test->in(0)->as_SafePoint();
676 } else {
677 Node* c = back_control;
678 while (c != x && c->Opcode() != Op_SafePoint) {
679 c = idom(c);
680 }
681
682 if (c->Opcode() == Op_SafePoint) {
683 safepoint = c->as_SafePoint();
684 }
685
686 if (safepoint == nullptr) {
687 return nullptr;
688 }
689
690 Node* mem = safepoint->in(TypeFunc::Memory);
691
692 // We can only use that safepoint if there's no side effect between the backedge and the safepoint.
693
694 // mm is the memory state at the safepoint (when it's a MergeMem)
695 // no_side_effect_since_safepoint() goes over the memory state at the backedge. It resets the mm input for each
696 // component of the memory state it encounters so it points to the base memory. Once no_side_effect_since_safepoint()
697 // is done, if no side effect after the safepoint was found, mm should transform to the base memory: the states at
698 // the backedge and safepoint are the same so all components of the memory state at the safepoint should have been
699 // reset.
700 MergeMemNode* mm = nullptr;
701 #ifdef ASSERT
702 if (mem->is_MergeMem()) {
703 mm = mem->clone()->as_MergeMem();
704 _igvn._worklist.push(mm);
705 for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {
706 // Loop invariant memory state won't be reset by no_side_effect_since_safepoint(). Do it here.
707 // Escape Analysis can add state to mm that it doesn't add to the backedge memory Phis, breaking verification
708 // code that relies on mm. Clear that extra state here.
709 if (mms.alias_idx() != Compile::AliasIdxBot &&
710 (loop != get_loop(ctrl_or_self(mms.memory())) ||
711 (mms.adr_type()->isa_oop_ptr() && mms.adr_type()->is_known_instance()))) {
712 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
713 }
714 }
715 }
716 #endif
717 if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
718 safepoint = nullptr;
719 } else {
720 assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
721 }
722 #ifdef ASSERT
723 if (mm != nullptr) {
724 _igvn.remove_dead_node(mm);
725 }
726 #endif
727 }
728 return safepoint;
729 }
730
731 // If the loop has the shape of a counted loop but with a long
732 // induction variable, transform the loop in a loop nest: an inner
733 // loop that iterates for at most max int iterations with an integer
734 // induction variable and an outer loop that iterates over the full
735 // range of long values from the initial loop in (at most) max int
736 // steps. That is:
737 //
738 // x: for (long phi = init; phi < limit; phi += stride) {
739 // // phi := Phi(L, init, incr)
740 // // incr := AddL(phi, longcon(stride))
741 // long incr = phi + stride;
742 // ... use phi and incr ...
743 // }
744 //
745 // OR:
746 //
747 // x: for (long phi = init; (phi += stride) < limit; ) {
748 // // phi := Phi(L, AddL(init, stride), incr)
749 // // incr := AddL(phi, longcon(stride))
750 // long incr = phi + stride;
751 // ... use phi and (phi + stride) ...
752 // }
753 //
754 // ==transform=>
755 //
756 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
757 // assert(stride <= inner_iters_limit); // else abort transform
758 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
759 // outer_head: for (long outer_phi = init;;) {
760 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
761 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
762 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
763 // assert(inner_iters_actual == (int)inner_iters_actual);
764 // int inner_phi, inner_incr;
765 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
766 // // inner_phi := Phi(x, intcon(0), inner_incr)
767 // // inner_incr := AddI(inner_phi, intcon(stride))
768 // inner_incr = inner_phi + stride;
769 // if (inner_incr < inner_iters_actual) {
770 // ... use phi=>(outer_phi+inner_phi) ...
771 // continue;
772 // }
773 // else break;
774 // }
775 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
776 // continue;
777 // else break;
778 // }
779 //
780 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
781 // loops with long range checks transformed to int range checks in the inner loop.
782 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
783 Node* x = loop->_head;
784 // Only for inner loops
785 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
786 return false;
787 }
788
789 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
790 return false;
791 }
792
793 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
794 BasicType bt = x->as_BaseCountedLoop()->bt();
795
796 check_counted_loop_shape(loop, x, bt);
797
798 #ifndef PRODUCT
799 if (bt == T_LONG) {
800 Atomic::inc(&_long_loop_candidates);
801 }
802 #endif
803
804 jlong stride_con_long = head->stride_con();
805 assert(stride_con_long != 0, "missed some peephole opt");
806 // We can't iterate for more than max int at a time.
807 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
808 assert(bt == T_LONG, "only for long loops");
809 return false;
810 }
811 jint stride_con = checked_cast<jint>(stride_con_long);
812 // The number of iterations for the integer count loop: guarantee no
813 // overflow: max_jint - stride_con max. -1 so there's no need for a
814 // loop limit check if the exit test is <= or >=.
815 int iters_limit = max_jint - ABS(stride_con) - 1;
816 #ifdef ASSERT
817 if (bt == T_LONG && StressLongCountedLoop > 0) {
818 iters_limit = iters_limit / StressLongCountedLoop;
819 }
820 #endif
821 // At least 2 iterations so counted loop construction doesn't fail
822 if (iters_limit/ABS(stride_con) < 2) {
823 return false;
824 }
825
826 PhiNode* phi = head->phi()->as_Phi();
827 Node* incr = head->incr();
828
829 Node* back_control = head->in(LoopNode::LoopBackControl);
830
831 // data nodes on back branch not supported
832 if (back_control->outcnt() > 1) {
833 return false;
834 }
835
836 Node* limit = head->limit();
837 // We'll need to use the loop limit before the inner loop is entered
838 if (!is_dominator(get_ctrl(limit), x)) {
839 return false;
840 }
841
842 IfNode* exit_test = head->loopexit();
843
844 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
845
846 Node_List range_checks;
847 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
848
849 if (bt == T_INT) {
850 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
851 if (range_checks.size() == 0) {
852 return false;
853 }
854 }
855
856 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
857 // the inner loop.
858 const Node* init = head->init_trip();
859 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
860 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
861 if (stride_con < 0) {
862 swap(lo, hi);
863 }
864 if (hi->hi_as_long() <= lo->lo_as_long()) {
865 // not a loop after all
866 return false;
867 }
868
869 if (range_checks.size() > 0) {
870 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
871 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
872 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
873 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
874 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
875 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
876 // logic on loop entry and for the outer loop.
877 loop->compute_trip_count(this);
878 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
879 if (head->as_CountedLoop()->trip_count() <= 3) {
880 return false;
881 }
882 } else {
883 loop->compute_profile_trip_cnt(this);
884 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
885 return false;
886 }
887 }
888 }
889
890 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
891 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
892
893 // We need a safepoint to insert Parse Predicates for the inner loop.
894 SafePointNode* safepoint;
895 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
896 // Loop is strip mined: use the safepoint of the outer strip mined loop
897 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
898 assert(outer_loop != nullptr, "no outer loop");
899 safepoint = outer_loop->outer_safepoint();
900 outer_loop->transform_to_counted_loop(&_igvn, this);
901 exit_test = head->loopexit();
902 } else {
903 safepoint = find_safepoint(back_control, x, loop);
904 }
905
906 Node* exit_branch = exit_test->proj_out(false);
907 Node* entry_control = head->in(LoopNode::EntryControl);
908
909 // Clone the control flow of the loop to build an outer loop
910 Node* outer_back_branch = back_control->clone();
911 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
912 Node* inner_exit_branch = exit_branch->clone();
913
914 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
915 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
916
917 const bool body_populated = true;
918 register_control(outer_head, outer_ilt, entry_control, body_populated);
919
920 _igvn.register_new_node_with_optimizer(inner_exit_branch);
921 set_loop(inner_exit_branch, outer_ilt);
922 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
923
924 outer_exit_test->set_req(0, inner_exit_branch);
925 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
926
927 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
928 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
929
930 outer_back_branch->set_req(0, outer_exit_test);
931 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
932
933 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
934 set_idom(x, outer_head, dom_depth(x));
935
936 // add an iv phi to the outer loop and use it to compute the inner
937 // loop iteration limit
938 Node* outer_phi = phi->clone();
939 outer_phi->set_req(0, outer_head);
940 register_new_node(outer_phi, outer_head);
941
942 Node* inner_iters_max = nullptr;
943 if (stride_con > 0) {
944 inner_iters_max = MaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
945 } else {
946 inner_iters_max = MaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
947 }
948
949 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
950 // inner_iters_max may not fit in a signed integer (iterating from
951 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
952 // min.
953 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
954 Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
955
956 Node* inner_iters_actual_int;
957 if (bt == T_LONG) {
958 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
959 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
960 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
961 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
962 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
963 // opts pass, an accurate range of values for the limits is found.
964 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
965 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::UnconditionalDependency);
966 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
967 } else {
968 inner_iters_actual_int = inner_iters_actual;
969 }
970
971 Node* int_zero = intcon(0);
972 if (stride_con < 0) {
973 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
974 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
975 }
976
977 // Clone the iv data nodes as an integer iv
978 Node* int_stride = intcon(stride_con);
979 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
980 Node* inner_incr = new AddINode(inner_phi, int_stride);
981 Node* inner_cmp = nullptr;
982 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
983 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
984 inner_phi->set_req(LoopNode::EntryControl, int_zero);
985 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
986 register_new_node(inner_phi, x);
987 register_new_node(inner_incr, x);
988 register_new_node(inner_cmp, x);
989 register_new_node(inner_bol, x);
990
991 _igvn.replace_input_of(exit_test, 1, inner_bol);
992
993 // Clone inner loop phis to outer loop
994 for (uint i = 0; i < head->outcnt(); i++) {
995 Node* u = head->raw_out(i);
996 if (u->is_Phi() && u != inner_phi && u != phi) {
997 assert(u->in(0) == head, "inconsistent");
998 Node* clone = u->clone();
999 clone->set_req(0, outer_head);
1000 register_new_node(clone, outer_head);
1001 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1002 }
1003 }
1004
1005 // Replace inner loop long iv phi as inner loop int iv phi + outer
1006 // loop iv phi
1007 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1008
1009 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1010
1011 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1012
1013 // Summary of steps from initial loop to loop nest:
1014 //
1015 // == old IR nodes =>
1016 //
1017 // entry_control: {...}
1018 // x:
1019 // for (long phi = init;;) {
1020 // // phi := Phi(x, init, incr)
1021 // // incr := AddL(phi, longcon(stride))
1022 // exit_test:
1023 // if (phi < limit)
1024 // back_control: fallthrough;
1025 // else
1026 // exit_branch: break;
1027 // long incr = phi + stride;
1028 // ... use phi and incr ...
1029 // phi = incr;
1030 // }
1031 //
1032 // == new IR nodes (just before final peel) =>
1033 //
1034 // entry_control: {...}
1035 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1036 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1037 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1038 // outer_head:
1039 // for (long outer_phi = init;;) {
1040 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1041 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1042 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1043 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1044 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1045 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1046 // inner_head: x: //in(1) := outer_head
1047 // int inner_phi;
1048 // for (inner_phi = 0;;) {
1049 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1050 // int inner_incr = inner_phi + stride;
1051 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1052 // exit_test: //exit_test->in(1) := inner_bol;
1053 // if (inner_bol) // WAS (phi < limit)
1054 // back_control: fallthrough;
1055 // else
1056 // inner_exit_branch: break; //exit_branch->clone()
1057 // ... use phi=>(outer_phi+inner_phi) ...
1058 // inner_phi = inner_phi + stride; // inner_incr
1059 // }
1060 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1061 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1062 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1063 // else
1064 // exit_branch: break; //in(0) := outer_exit_test
1065 // }
1066
1067 if (bt == T_INT) {
1068 outer_phi = new ConvI2LNode(outer_phi);
1069 register_new_node(outer_phi, outer_head);
1070 }
1071
1072 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1073 inner_phi, iv_add, inner_head);
1074 // Peel one iteration of the loop and use the safepoint at the end
1075 // of the peeled iteration to insert Parse Predicates. If no well
1076 // positioned safepoint peel to guarantee a safepoint in the outer
1077 // loop.
1078 if (safepoint != nullptr || !loop->_has_call) {
1079 old_new.clear();
1080 do_peeling(loop, old_new);
1081 } else {
1082 C->set_major_progress();
1083 }
1084
1085 if (safepoint != nullptr) {
1086 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1087
1088 if (UseLoopPredicate) {
1089 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
1090 if (UseProfiledLoopPredicate) {
1091 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1092 }
1093 }
1094
1095 // We only want to use the auto-vectorization check as a trap once per bci. And
1096 // PhaseIdealLoop::add_parse_predicate only checks trap limits per method, so
1097 // we do a custom check here.
1098 if (!C->too_many_traps(cloned_sfpt->jvms()->method(), cloned_sfpt->jvms()->bci(), Deoptimization::Reason_auto_vectorization_check)) {
1099 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, inner_head, outer_ilt, cloned_sfpt);
1100 }
1101
1102 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1103 }
1104
1105 #ifndef PRODUCT
1106 if (bt == T_LONG) {
1107 Atomic::inc(&_long_loop_nests);
1108 }
1109 #endif
1110
1111 inner_head->mark_loop_nest_inner_loop();
1112 outer_head->mark_loop_nest_outer_loop();
1113
1114 return true;
1115 }
1116
1117 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1118 Node_List& range_checks) {
1119 const jlong min_iters = 2;
1120 jlong reduced_iters_limit = iters_limit;
1121 jlong original_iters_limit = iters_limit;
1122 for (uint i = 0; i < loop->_body.size(); i++) {
1123 Node* c = loop->_body.at(i);
1124 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1125 IfProjNode* if_proj = c->as_IfProj();
1126 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1127 if (call != nullptr) {
1128 Node* range = nullptr;
1129 Node* offset = nullptr;
1130 jlong scale = 0;
1131 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1132 loop->is_invariant(range) && loop->is_invariant(offset) &&
1133 scale != min_jlong &&
1134 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1135 assert(scale == (jint)scale, "scale should be an int");
1136 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1137 range_checks.push(c);
1138 }
1139 }
1140 }
1141 }
1142
1143 return checked_cast<int>(reduced_iters_limit);
1144 }
1145
1146 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1147 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1148 // first value, followed by A+S (S is the stride), next A+2S, etc. The limit is exclusive, so that the final value B of
1149 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1150
1151 // If |S|>1 the formula for the last value B would require a floor operation, specifically B=floor((Z-sgn(S)-A)/S)*S+A,
1152 // which is B=Z-sgn(S)U for some U in [1,|S|]. So when S>0, i ranges as i:[A,Z) or i:[A,B=Z-U], or else (in reverse)
1153 // as i:(Z,A] or i:[B=Z+U,A]. It will become important to reason about this inclusive range [A,B] or [B,A].
1154
1155 // Within the loop there may be many range checks. Each such range check (R.C.) is of the form 0 <= i*K+L < R, where K
1156 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1157 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1158
1159 // When a long loop over a 64-bit variable i (outer_iv) is decomposed into a series of shorter sub-loops over a 32-bit
1160 // variable j (inner_iv), j ranges over a shorter interval j:[0,B_2] or [0,Z_2) (assuming S > 0), where the limit is
1161 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1162 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1163
1164 // For S<0, j ranges (in reverse!) through j:[-|B_2|,0] or (-|Z_2|,0]. For either sign of S, we can say i=j+C and j
1165 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1166
1167 // The disjoint union of all the C+[A_2,B_2] ranges from the sub-loops must be identical to the whole range [A,B].
1168 // Assuming S>0, the first C must be A itself, and the next C value is the previous C+B_2, plus S. If |S|=1, the next
1169 // C value is also the previous C+Z_2. In each sub-loop, j counts from j=A_2=0 and i counts from C+0 and exits at
1170 // j=B_2 (i=C+B_2), just before it gets to i=C+Z_2. Both i and j count up (from C and 0) if S>0; otherwise they count
1171 // down (from C and 0 again).
1172
1173 // Returning to range checks, we see that each i*K+L <u R expands to (C+j)*K+L <u R, or j*K+Q <u R, where Q=(C*K+L).
1174 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1175 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1176 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1177 // PhaseIdealLoop::add_constraint.)
1178
1179 // We must transform this comparison so that it gets the same answer, but by means of a 32-bit R.C. (using j not i) of
1180 // the form j*K+L_2 <u32 R_2. Note that L_2 and R_2 must be loop-invariant, but only with respect to the sub-loop. Thus, the
1181 // problem reduces to computing values for L_2 and R_2 (for each R.C. in the loop) in the loop header for the sub-loop.
1182 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1183 // values for the 32-bit counter j within which the range checks can be eliminated.
1184
1185 // So, given j*K+Q <u R, we need to find some j*K+L_2 <u32 R_2, where L_2 and R_2 fit in 32 bits, and the 32-bit operations do
1186 // not overflow. We also need to cover the cases where i*K+L (= j*K+Q) overflows to a 64-bit negative, since that is
1187 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1188
1189 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1190
1191 // For each R.C. j*K+Q <u32 R, the range of mathematical values of j*K+Q in the sub-loop is [Q_min, Q_max], where
1192 // Q_min=Q and Q_max=B_2*K+Q (if S>0 and K>0), Q_min=A_2*K+Q and Q_max=Q (if S<0 and K>0),
1193 // Q_min=B_2*K+Q and Q_max=Q if (S>0 and K<0), Q_min=Q and Q_max=A_2*K+Q (if S<0 and K<0)
1194
1195 // Note that the first R.C. value is always Q=(S*K>0 ? Q_min : Q_max). Also Q_{min,max} = Q + {min,max}(A_2*K,B_2*K).
1196 // If S*K>0 then, as the loop iterations progress, each R.C. value i*K+L = j*K+Q goes up from Q=Q_min towards Q_max.
1197 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1198
1199 // Case A: Some Negatives (but no overflow).
1200 // Number line:
1201 // |s64_min . . . 0 . . . s64_max|
1202 // | . Q_min..Q_max . 0 . . . . | s64 negative
1203 // | . . . . R=0 R< R< R< R< | (against R values)
1204 // | . . . Q_min..0..Q_max . . . | small mixed
1205 // | . . . . R R R< R< R< | (against R values)
1206 //
1207 // R values which are out of range (>Q_max+1) are reduced to max(0,Q_max+1). They are marked on the number line as R<.
1208 //
1209 // So, if Q_min <s64 0, then use this test:
1210 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1211 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1212 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1213 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1214
1215 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1216
1217 // Case B: No Negatives.
1218 // Number line:
1219 // |s64_min . . . 0 . . . s64_max|
1220 // | . . . . 0 Q_min..Q_max . . | small positive
1221 // | . . . . R> R R R< R< | (against R values)
1222 // | . . . . 0 . Q_min..Q_max . | s64 positive
1223 // | . . . . R> R> R R R< | (against R values)
1224 //
1225 // R values which are out of range (<Q_min or >Q_max+1) are reduced as marked: R> up to Q_min, R< down to Q_max+1.
1226 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1227 //
1228 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1229 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1230 // More generally:
1231 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1232
1233 // Case C: Overflow in the 64-bit domain
1234 // Number line:
1235 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1236 // | . . . . R> R> R> R> R | (against R values)
1237 //
1238 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1239 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1240 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1241 //
1242 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1243 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1244 // More generally:
1245 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1246 //
1247 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1248 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1249 //
1250 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1251 // clamp(X, L, H) := max(L, min(X, H))
1252 // When degenerately L > H, it returns L not H.
1253 //
1254 // All of the formulas above can be merged into a single one:
1255 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1256 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1257 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1258 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1259 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1260 // replacement R.C.:
1261 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1262 // or equivalently:
1263 // j*K + L_2 <u32 R_2
1264 // where
1265 // L_2 = Q_first - L_clamp
1266 // R_2 = R_clamp - L_clamp
1267 //
1268 // Note on why R is never negative:
1269 //
1270 // Various details of this transformation would break badly if R could be negative, so this transformation only
1271 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1272 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1273 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1274 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1275 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1276 // R.C., but R values from a.length and inside checkIndex are good to go.
1277 //
1278 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1279 Node* inner_iters_actual_int, Node* inner_phi,
1280 Node* iv_add, LoopNode* inner_head) {
1281 Node* long_zero = longcon(0);
1282 Node* int_zero = intcon(0);
1283 Node* long_one = longcon(1);
1284 Node* int_stride = intcon(checked_cast<int>(stride_con));
1285
1286 for (uint i = 0; i < range_checks.size(); i++) {
1287 ProjNode* proj = range_checks.at(i)->as_Proj();
1288 ProjNode* unc_proj = proj->other_if_proj();
1289 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1290 jlong scale = 0;
1291 Node* offset = nullptr;
1292 Node* rc_bol = rc->in(1);
1293 Node* rc_cmp = rc_bol->in(1);
1294 if (rc_cmp->Opcode() == Op_CmpU) {
1295 // could be shared and have already been taken care of
1296 continue;
1297 }
1298 bool short_scale = false;
1299 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1300 assert(ok, "inconsistent: was tested before");
1301 Node* range = rc_cmp->in(2);
1302 Node* c = rc->in(0);
1303 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1304
1305 Node* R = range;
1306 Node* K = longcon(scale);
1307
1308 Node* L = offset;
1309
1310 if (short_scale) {
1311 // This converts:
1312 // (int)i*K + L <u64 R
1313 // with K an int into:
1314 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1315 // to protect against an overflow of (int)i*K
1316 //
1317 // Because if (int)i*K overflows, there are K,L where:
1318 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1319 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1320 // i*(long)K + L <u64 R is true.
1321 //
1322 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1323 //
1324 // It's always true that:
1325 // (int)i*K <u64 (long)max_jint + 1
1326 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1327 // As a consequence:
1328 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1329 // is always false in case of overflow of i*K
1330 //
1331 // Note, there are also K,L where i*K overflows and
1332 // i*K + L <u64 R is true, but
1333 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1334 // So this transformation could cause spurious deoptimizations and failed range check elimination
1335 // (but not incorrect execution) for unlikely corner cases with overflow.
1336 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1337 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1338 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1339 register_new_node(max_range, entry_control);
1340 R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1341 set_subtree_ctrl(R, true);
1342 }
1343
1344 Node* C = outer_phi;
1345
1346 // Start with 64-bit values:
1347 // i*K + L <u64 R
1348 // (C+j)*K + L <u64 R
1349 // j*K + Q <u64 R where Q = Q_first = C*K+L
1350 Node* Q_first = new MulLNode(C, K);
1351 register_new_node(Q_first, entry_control);
1352 Q_first = new AddLNode(Q_first, L);
1353 register_new_node(Q_first, entry_control);
1354
1355 // Compute endpoints of the range of values j*K + Q.
1356 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1357 Node* Q_min = Q_first;
1358
1359 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1360 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1361 register_new_node(B_2, entry_control);
1362 B_2 = new SubINode(B_2, int_stride);
1363 register_new_node(B_2, entry_control);
1364 B_2 = new ConvI2LNode(B_2);
1365 register_new_node(B_2, entry_control);
1366
1367 Node* Q_max = new MulLNode(B_2, K);
1368 register_new_node(Q_max, entry_control);
1369 Q_max = new AddLNode(Q_max, Q_first);
1370 register_new_node(Q_max, entry_control);
1371
1372 if (scale * stride_con < 0) {
1373 swap(Q_min, Q_max);
1374 }
1375 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1376
1377 // L_clamp = Q_min < 0 ? 0 : Q_min
1378 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1379 register_new_node(Q_min_cmp, entry_control);
1380 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1381 register_new_node(Q_min_bool, entry_control);
1382 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1383 register_new_node(L_clamp, entry_control);
1384 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1385
1386 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1387 register_new_node(Q_max_plus_one, entry_control);
1388
1389 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1390 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1391 Node* max_jlong_long = longcon(max_jlong);
1392 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1393 register_new_node(Q_max_cmp, entry_control);
1394 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1395 register_new_node(Q_max_bool, entry_control);
1396 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1397 register_new_node(H_clamp, entry_control);
1398 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1399
1400 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1401 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1402 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1403 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1404 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1405 Node* R_2 = clamp(R, L_clamp, H_clamp);
1406 R_2 = new SubLNode(R_2, L_clamp);
1407 register_new_node(R_2, entry_control);
1408 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1409 register_new_node(R_2, entry_control);
1410
1411 // L_2 = Q_first - L_clamp
1412 // We are subtracting L_clamp from both sides of the <u32 comparison.
1413 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1414 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1415 Node* L_2 = new SubLNode(Q_first, L_clamp);
1416 register_new_node(L_2, entry_control);
1417 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1418 register_new_node(L_2, entry_control);
1419
1420 // Transform the range check using the computed values L_2/R_2
1421 // from: i*K + L <u64 R
1422 // to: j*K + L_2 <u32 R_2
1423 // that is:
1424 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1425 K = intcon(checked_cast<int>(scale));
1426 Node* scaled_iv = new MulINode(inner_phi, K);
1427 register_new_node(scaled_iv, c);
1428 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1429 register_new_node(scaled_iv_plus_offset, c);
1430
1431 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1432 register_new_node(new_rc_cmp, c);
1433
1434 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1435 }
1436 }
1437
1438 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1439 Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1440 set_subtree_ctrl(min, true);
1441 Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1442 set_subtree_ctrl(max, true);
1443 return max;
1444 }
1445
1446 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1447 IfNode* exit_test) {
1448 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1449 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1450 _igvn.register_new_node_with_optimizer(new_inner_head);
1451 _igvn.register_new_node_with_optimizer(new_inner_exit);
1452 loop->_body.push(new_inner_head);
1453 loop->_body.push(new_inner_exit);
1454 loop->_body.yank(head);
1455 loop->_body.yank(exit_test);
1456 set_loop(new_inner_head, loop);
1457 set_loop(new_inner_exit, loop);
1458 set_idom(new_inner_head, idom(head), dom_depth(head));
1459 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1460 lazy_replace(head, new_inner_head);
1461 lazy_replace(exit_test, new_inner_exit);
1462 loop->_head = new_inner_head;
1463 return new_inner_head;
1464 }
1465
1466 #ifdef ASSERT
1467 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1468 Node* back_control = loop_exit_control(x, loop);
1469 assert(back_control != nullptr, "no back control");
1470
1471 BoolTest::mask mask = BoolTest::illegal;
1472 float cl_prob = 0;
1473 Node* incr = nullptr;
1474 Node* limit = nullptr;
1475
1476 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1477 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1478
1479 Node* phi_incr = nullptr;
1480 incr = loop_iv_incr(incr, x, loop, phi_incr);
1481 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1482
1483 Node* xphi = nullptr;
1484 Node* stride = loop_iv_stride(incr, loop, xphi);
1485
1486 assert(stride != nullptr, "no stride");
1487
1488 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1489
1490 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1491
1492 jlong stride_con = stride->get_integer_as_long(bt);
1493
1494 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1495
1496 assert(mask != BoolTest::ne, "unexpected condition");
1497 assert(phi_incr == nullptr, "bad loop shape");
1498 assert(cmp->in(1) == incr, "bad exit test shape");
1499
1500 // Safepoint on backedge not supported
1501 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1502 }
1503 #endif
1504
1505 #ifdef ASSERT
1506 // convert an int counted loop to a long counted to stress handling of
1507 // long counted loops
1508 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1509 Unique_Node_List iv_nodes;
1510 Node_List old_new;
1511 iv_nodes.push(cmp);
1512 bool failed = false;
1513
1514 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1515 Node* n = iv_nodes.at(i);
1516 switch(n->Opcode()) {
1517 case Op_Phi: {
1518 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1519 old_new.map(n->_idx, clone);
1520 break;
1521 }
1522 case Op_CmpI: {
1523 Node* clone = new CmpLNode(nullptr, nullptr);
1524 old_new.map(n->_idx, clone);
1525 break;
1526 }
1527 case Op_AddI: {
1528 Node* clone = new AddLNode(nullptr, nullptr);
1529 old_new.map(n->_idx, clone);
1530 break;
1531 }
1532 case Op_CastII: {
1533 failed = true;
1534 break;
1535 }
1536 default:
1537 DEBUG_ONLY(n->dump());
1538 fatal("unexpected");
1539 }
1540
1541 for (uint i = 1; i < n->req(); i++) {
1542 Node* in = n->in(i);
1543 if (in == nullptr) {
1544 continue;
1545 }
1546 if (loop->is_member(get_loop(get_ctrl(in)))) {
1547 iv_nodes.push(in);
1548 }
1549 }
1550 }
1551
1552 if (failed) {
1553 for (uint i = 0; i < iv_nodes.size(); i++) {
1554 Node* n = iv_nodes.at(i);
1555 Node* clone = old_new[n->_idx];
1556 if (clone != nullptr) {
1557 _igvn.remove_dead_node(clone);
1558 }
1559 }
1560 return false;
1561 }
1562
1563 for (uint i = 0; i < iv_nodes.size(); i++) {
1564 Node* n = iv_nodes.at(i);
1565 Node* clone = old_new[n->_idx];
1566 for (uint i = 1; i < n->req(); i++) {
1567 Node* in = n->in(i);
1568 if (in == nullptr) {
1569 continue;
1570 }
1571 Node* in_clone = old_new[in->_idx];
1572 if (in_clone == nullptr) {
1573 assert(_igvn.type(in)->isa_int(), "");
1574 in_clone = new ConvI2LNode(in);
1575 _igvn.register_new_node_with_optimizer(in_clone);
1576 set_subtree_ctrl(in_clone, false);
1577 }
1578 if (in_clone->in(0) == nullptr) {
1579 in_clone->set_req(0, C->top());
1580 clone->set_req(i, in_clone);
1581 in_clone->set_req(0, nullptr);
1582 } else {
1583 clone->set_req(i, in_clone);
1584 }
1585 }
1586 _igvn.register_new_node_with_optimizer(clone);
1587 }
1588 set_ctrl(old_new[phi->_idx], phi->in(0));
1589
1590 for (uint i = 0; i < iv_nodes.size(); i++) {
1591 Node* n = iv_nodes.at(i);
1592 Node* clone = old_new[n->_idx];
1593 set_subtree_ctrl(clone, false);
1594 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1596 Node* u = n->fast_out(i);
1597 if (iv_nodes.member(u)) {
1598 continue;
1599 }
1600 if (m == nullptr) {
1601 m = new ConvL2INode(clone);
1602 _igvn.register_new_node_with_optimizer(m);
1603 set_subtree_ctrl(m, false);
1604 }
1605 _igvn.rehash_node_delayed(u);
1606 int nb = u->replace_edge(n, m, &_igvn);
1607 --i, imax -= nb;
1608 }
1609 }
1610 return true;
1611 }
1612 #endif
1613
1614 //------------------------------is_counted_loop--------------------------------
1615 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {
1616 PhaseGVN *gvn = &_igvn;
1617
1618 Node* back_control = loop_exit_control(x, loop);
1619 if (back_control == nullptr) {
1620 return false;
1621 }
1622
1623 BoolTest::mask bt = BoolTest::illegal;
1624 float cl_prob = 0;
1625 Node* incr = nullptr;
1626 Node* limit = nullptr;
1627 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1628 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1629 return false; // Avoid pointer & float & 64-bit compares
1630 }
1631
1632 // Trip-counter increment must be commutative & associative.
1633 if (incr->Opcode() == Op_Cast(iv_bt)) {
1634 incr = incr->in(1);
1635 }
1636
1637 Node* phi_incr = nullptr;
1638 incr = loop_iv_incr(incr, x, loop, phi_incr);
1639 if (incr == nullptr) {
1640 return false;
1641 }
1642
1643 Node* trunc1 = nullptr;
1644 Node* trunc2 = nullptr;
1645 const TypeInteger* iv_trunc_t = nullptr;
1646 Node* orig_incr = incr;
1647 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1648 return false; // Funny increment opcode
1649 }
1650 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1651
1652 Node* xphi = nullptr;
1653 Node* stride = loop_iv_stride(incr, loop, xphi);
1654
1655 if (stride == nullptr) {
1656 return false;
1657 }
1658
1659 if (xphi->Opcode() == Op_Cast(iv_bt)) {
1660 xphi = xphi->in(1);
1661 }
1662
1663 // Stride must be constant
1664 jlong stride_con = stride->get_integer_as_long(iv_bt);
1665 assert(stride_con != 0, "missed some peephole opt");
1666
1667 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1668
1669 if (phi == nullptr ||
1670 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1671 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1672 return false;
1673 }
1674
1675 Node* iftrue = back_control;
1676 uint iftrue_op = iftrue->Opcode();
1677 Node* iff = iftrue->in(0);
1678 BoolNode* test = iff->in(1)->as_Bool();
1679
1680 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1681 if (trunc1 != nullptr) {
1682 // When there is a truncation, we must be sure that after the truncation
1683 // the trip counter will end up higher than the limit, otherwise we are looking
1684 // at an endless loop. Can happen with range checks.
1685
1686 // Example:
1687 // int i = 0;
1688 // while (true)
1689 // sum + = array[i];
1690 // i++;
1691 // i = i && 0x7fff;
1692 // }
1693 //
1694 // If the array is shorter than 0x8000 this exits through a AIOOB
1695 // - Counted loop transformation is ok
1696 // If the array is longer then this is an endless loop
1697 // - No transformation can be done.
1698
1699 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1700 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1701 // if the limit can have a higher value than the increment (before the phi)
1702 return false;
1703 }
1704 }
1705
1706 Node *init_trip = phi->in(LoopNode::EntryControl);
1707
1708 // If iv trunc type is smaller than int, check for possible wrap.
1709 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1710 assert(trunc1 != nullptr, "must have found some truncation");
1711
1712 // Get a better type for the phi (filtered thru if's)
1713 const TypeInteger* phi_ft = filtered_type(phi);
1714
1715 // Can iv take on a value that will wrap?
1716 //
1717 // Ensure iv's limit is not within "stride" of the wrap value.
1718 //
1719 // Example for "short" type
1720 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
1721 // If the stride is +10, then the last value of the induction
1722 // variable before the increment (phi_ft->_hi) must be
1723 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
1724 // ensure no truncation occurs after the increment.
1725
1726 if (stride_con > 0) {
1727 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
1728 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
1729 return false; // truncation may occur
1730 }
1731 } else if (stride_con < 0) {
1732 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
1733 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
1734 return false; // truncation may occur
1735 }
1736 }
1737 // No possibility of wrap so truncation can be discarded
1738 // Promote iv type to Int
1739 } else {
1740 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
1741 }
1742
1743 if (!condition_stride_ok(bt, stride_con)) {
1744 return false;
1745 }
1746
1747 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
1748
1749 if (stride_con > 0) {
1750 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
1751 return false; // cyclic loop
1752 }
1753 } else {
1754 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
1755 return false; // cyclic loop
1756 }
1757 }
1758
1759 if (phi_incr != nullptr && bt != BoolTest::ne) {
1760 // check if there is a possibility of IV overflowing after the first increment
1761 if (stride_con > 0) {
1762 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
1763 return false;
1764 }
1765 } else {
1766 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
1767 return false;
1768 }
1769 }
1770 }
1771
1772 // =================================================
1773 // ---- SUCCESS! Found A Trip-Counted Loop! -----
1774 //
1775
1776 if (x->Opcode() == Op_Region) {
1777 // x has not yet been transformed to Loop or LongCountedLoop.
1778 // This should only happen if we are inside an infinite loop.
1779 // It happens like this:
1780 // build_loop_tree -> do not attach infinite loop and nested loops
1781 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
1782 // build_loop_tree -> find and attach infinite and nested loops
1783 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
1784 assert(x->as_Region()->is_in_infinite_subgraph(),
1785 "x can only be a Region and not Loop if inside infinite loop");
1786 // Come back later when Region is transformed to LoopNode
1787 return false;
1788 }
1789
1790 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
1791 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
1792
1793 // ===================================================
1794 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
1795 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
1796 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
1797 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
1798 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
1799 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
1800 //
1801 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
1802 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
1803 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
1804 //
1805 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
1806 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
1807 //
1808 //
1809 // The loop to be converted does not always need to have the often used shape:
1810 //
1811 // i = init
1812 // i = init loop:
1813 // do { ...
1814 // // ... equivalent i+=stride
1815 // i+=stride <==> if (i < limit)
1816 // } while (i < limit); goto loop
1817 // exit:
1818 // ...
1819 //
1820 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
1821 //
1822 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
1823 // in the loop exit check:
1824 //
1825 // i = init
1826 // loop:
1827 // ...
1828 // if (i <= limit)
1829 // i+=stride
1830 // goto loop
1831 // exit:
1832 // ...
1833 //
1834 // Let's define the following terms:
1835 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
1836 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
1837 //
1838 // The iv_pre_i and iv_post_i have the following relation:
1839 // iv_pre_i + stride = iv_post_i
1840 //
1841 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
1842 // iv_post_i < adjusted_limit
1843 //
1844 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
1845 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
1846 // -> adjusted_limit = limit.
1847 // (LE2) iv_post_i <= limit:
1848 // iv_post_i < limit + 1
1849 // -> adjusted limit = limit + 1
1850 // (LE3) iv_pre_i < limit:
1851 // iv_pre_i + stride < limit + stride
1852 // iv_post_i < limit + stride
1853 // -> adjusted_limit = limit + stride
1854 // (LE4) iv_pre_i <= limit:
1855 // iv_pre_i < limit + 1
1856 // iv_pre_i + stride < limit + stride + 1
1857 // iv_post_i < limit + stride + 1
1858 // -> adjusted_limit = limit + stride + 1
1859 //
1860 // Note that:
1861 // (AL) limit <= adjusted_limit.
1862 //
1863 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
1864 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
1865 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
1866 //
1867 // To prove (INV), we require the following two conditions/assumptions:
1868 // (i): adjusted_limit - 1 + stride <= max_int
1869 // (ii): init < limit
1870 //
1871 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
1872 // induction by assuming (i) and (ii).
1873 //
1874 // Proof by Induction
1875 // ------------------
1876 // > Base case (i = 1): We show that (INV) holds after the first iteration:
1877 // min_int <= iv_post_1 = init + stride <= max_int
1878 // Proof:
1879 // First, we note that (ii) implies
1880 // (iii) init <= limit - 1
1881 // max_int >= adjusted_limit - 1 + stride [using (i)]
1882 // >= limit - 1 + stride [using (AL)]
1883 // >= init + stride [using (iii)]
1884 // >= min_int [using stride > 0, no underflow]
1885 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
1886 //
1887 // Note that to prove the base case we need (i) and (ii).
1888 //
1889 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
1890 // min_int <= iv_post_j <= max_int
1891 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
1892 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
1893 // Proof:
1894 // If iv_post_j >= adjusted_limit:
1895 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
1896 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
1897 // If iv_post_j < adjusted_limit:
1898 // First, we note that:
1899 // (iv) iv_post_j <= adjusted_limit - 1
1900 // max_int >= adjusted_limit - 1 + stride [using (i)]
1901 // >= iv_post_j + stride [using (iv)]
1902 // >= min_int [using stride > 0, no underflow]
1903 //
1904 // Note that to prove the step case we only need (i).
1905 //
1906 // Thus, by assuming (i) and (ii), we proved (INV).
1907 //
1908 //
1909 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
1910 //
1911 // (1) Loop Limit Check Predicate for (i):
1912 // Using (i): adjusted_limit - 1 + stride <= max_int
1913 //
1914 // This condition is now restated to use limit instead of adjusted_limit:
1915 //
1916 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
1917 // max_int - stride + 1 >= adjusted_limit
1918 // We can merge the two constants into
1919 // canonicalized_correction = stride - 1
1920 // which gives us
1921 // max_int - canonicalized_correction >= adjusted_limit
1922 //
1923 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
1924 // adjusted_limit = limit + limit_correction
1925 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
1926 // max_int - canonicalized_correction - limit_correction >= limit
1927 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
1928 // (v) final_correction = canonicalized_correction + limit_correction
1929 //
1930 // which gives us:
1931 //
1932 // Final predicate condition:
1933 // max_int - final_correction >= limit
1934 //
1935 // However, we need to be careful that (v) does not over- or underflow.
1936 // We know that:
1937 // canonicalized_correction = stride - 1
1938 // and
1939 // limit_correction <= stride + 1
1940 // and thus
1941 // canonicalized_correction + limit_correction <= 2 * stride
1942 // To prevent an over- or underflow of (v), we must ensure that
1943 // 2 * stride <= max_int
1944 // which can safely be checked without over- or underflow with
1945 // (vi) stride != min_int AND abs(stride) <= max_int / 2
1946 //
1947 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
1948 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
1949 //
1950 // (2) Loop Limit Check Predicate for (ii):
1951 // Using (ii): init < limit
1952 //
1953 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
1954 // (2.1) type(init) < type(limit)
1955 // In this case, we know:
1956 // all possible values of init < all possible values of limit
1957 // and we can skip the predicate.
1958 //
1959 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
1960 // In this case, we do not need to re-check the condition and can skip the predicate.
1961 // This is often found for while- and for-loops which have the following shape:
1962 //
1963 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
1964 // i = init;
1965 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
1966 // do {
1967 // i += stride;
1968 // } while (i < limit);
1969 // }
1970 //
1971 // (2.3) init + stride <= max_int
1972 // In this case, there is no overflow of the iv phi after the first loop iteration.
1973 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
1974 // init < limit
1975 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
1976 // compile time that init + stride <= max_int then we have trivially proven the base case and that
1977 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
1978 // again and can skip the predicate.
1979
1980 // Check (vi) and bail out if the stride is too big.
1981 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
1982 return false;
1983 }
1984
1985 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
1986 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
1987
1988 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
1989 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
1990 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
1991
1992 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
1993 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
1994 const jlong final_correction = canonicalized_correction + limit_correction;
1995
1996 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
1997 Node* init_control = x->in(LoopNode::EntryControl);
1998
1999 // If sov==0, limit's type always satisfies the condition, for
2000 // example, when it is an array length.
2001 if (sov != 0) {
2002 if (sov < 0) {
2003 return false; // Bailout: integer overflow is certain.
2004 }
2005 // (1) Loop Limit Check Predicate is required because we could not statically prove that
2006 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
2007 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2008 const Predicates predicates(init_control);
2009 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2010 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2011 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2012 #ifdef ASSERT
2013 if (TraceLoopLimitCheck) {
2014 tty->print("Missing Loop Limit Check Parse Predicate:");
2015 loop->dump_head();
2016 x->dump(1);
2017 }
2018 #endif
2019 return false;
2020 }
2021
2022 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2023 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2024 return false;
2025 }
2026
2027 Node* cmp_limit;
2028 Node* bol;
2029
2030 if (stride_con > 0) {
2031 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2032 bol = new BoolNode(cmp_limit, BoolTest::le);
2033 } else {
2034 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2035 bol = new BoolNode(cmp_limit, BoolTest::ge);
2036 }
2037
2038 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2039 }
2040
2041 // (2.3)
2042 const bool init_plus_stride_could_overflow =
2043 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2044 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2045 // (2.1)
2046 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2047 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2048
2049 if (init_gte_limit && // (2.1)
2050 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2051 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2052 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2053 // We use the following condition:
2054 // - stride > 0: init < limit
2055 // - stride < 0: init > limit
2056 //
2057 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2058 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2059 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2060
2061 const Predicates predicates(init_control);
2062 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2063 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2064 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2065 #ifdef ASSERT
2066 if (TraceLoopLimitCheck) {
2067 tty->print("Missing Loop Limit Check Parse Predicate:");
2068 loop->dump_head();
2069 x->dump(1);
2070 }
2071 #endif
2072 return false;
2073 }
2074
2075 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2076 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2077 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2078 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2079 return false;
2080 }
2081
2082 Node* cmp_limit;
2083 Node* bol;
2084
2085 if (stride_con > 0) {
2086 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2087 bol = new BoolNode(cmp_limit, BoolTest::lt);
2088 } else {
2089 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2090 bol = new BoolNode(cmp_limit, BoolTest::gt);
2091 }
2092
2093 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2094 }
2095
2096 if (bt == BoolTest::ne) {
2097 // Now we need to canonicalize the loop condition if it is 'ne'.
2098 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2099 if (stride_con > 0) {
2100 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2101 bt = BoolTest::lt;
2102 } else {
2103 assert(stride_con < 0, "must be");
2104 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2105 bt = BoolTest::gt;
2106 }
2107 }
2108
2109 Node* sfpt = nullptr;
2110 if (loop->_child == nullptr) {
2111 sfpt = find_safepoint(back_control, x, loop);
2112 } else {
2113 sfpt = iff->in(0);
2114 if (sfpt->Opcode() != Op_SafePoint) {
2115 sfpt = nullptr;
2116 }
2117 }
2118
2119 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2120 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2121 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2122 iv_bt == T_LONG) &&
2123 sfpt == nullptr) {
2124 // Leaving the safepoint on the backedge and creating a
2125 // CountedLoop will confuse optimizations. We can't move the
2126 // safepoint around because its jvm state wouldn't match a new
2127 // location. Give up on that loop.
2128 return false;
2129 }
2130 if (is_deleteable_safept(backedge_sfpt)) {
2131 lazy_replace(backedge_sfpt, iftrue);
2132 if (loop->_safepts != nullptr) {
2133 loop->_safepts->yank(backedge_sfpt);
2134 }
2135 loop->_tail = iftrue;
2136 }
2137 }
2138
2139
2140 #ifdef ASSERT
2141 if (iv_bt == T_INT &&
2142 !x->as_Loop()->is_loop_nest_inner_loop() &&
2143 StressLongCountedLoop > 0 &&
2144 trunc1 == nullptr &&
2145 convert_to_long_loop(cmp, phi, loop)) {
2146 return false;
2147 }
2148 #endif
2149
2150 Node* adjusted_limit = limit;
2151 if (phi_incr != nullptr) {
2152 // If compare points directly to the phi we need to adjust
2153 // the compare so that it points to the incr. Limit have
2154 // to be adjusted to keep trip count the same and we
2155 // should avoid int overflow.
2156 //
2157 // i = init; do {} while(i++ < limit);
2158 // is converted to
2159 // i = init; do {} while(++i < limit+1);
2160 //
2161 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2162 }
2163
2164 if (includes_limit) {
2165 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2166 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2167 //
2168 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2169 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2170 if (bt == BoolTest::le)
2171 bt = BoolTest::lt;
2172 else if (bt == BoolTest::ge)
2173 bt = BoolTest::gt;
2174 else
2175 ShouldNotReachHere();
2176 }
2177 set_subtree_ctrl(adjusted_limit, false);
2178
2179 // Build a canonical trip test.
2180 // Clone code, as old values may be in use.
2181 incr = incr->clone();
2182 incr->set_req(1,phi);
2183 incr->set_req(2,stride);
2184 incr = _igvn.register_new_node_with_optimizer(incr);
2185 set_early_ctrl(incr, false);
2186 _igvn.rehash_node_delayed(phi);
2187 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2188
2189 // If phi type is more restrictive than Int, raise to
2190 // Int to prevent (almost) infinite recursion in igvn
2191 // which can only handle integer types for constants or minint..maxint.
2192 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2193 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2194 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2195 nphi = _igvn.register_new_node_with_optimizer(nphi);
2196 set_ctrl(nphi, get_ctrl(phi));
2197 _igvn.replace_node(phi, nphi);
2198 phi = nphi->as_Phi();
2199 }
2200 cmp = cmp->clone();
2201 cmp->set_req(1,incr);
2202 cmp->set_req(2, adjusted_limit);
2203 cmp = _igvn.register_new_node_with_optimizer(cmp);
2204 set_ctrl(cmp, iff->in(0));
2205
2206 test = test->clone()->as_Bool();
2207 (*(BoolTest*)&test->_test)._test = bt;
2208 test->set_req(1,cmp);
2209 _igvn.register_new_node_with_optimizer(test);
2210 set_ctrl(test, iff->in(0));
2211
2212 // Replace the old IfNode with a new LoopEndNode
2213 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2214 IfNode *le = lex->as_If();
2215 uint dd = dom_depth(iff);
2216 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2217 set_loop(le, loop);
2218
2219 // Get the loop-exit control
2220 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2221
2222 // Need to swap loop-exit and loop-back control?
2223 if (iftrue_op == Op_IfFalse) {
2224 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2225 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2226
2227 loop->_tail = back_control = ift2;
2228 set_loop(ift2, loop);
2229 set_loop(iff2, get_loop(iffalse));
2230
2231 // Lazy update of 'get_ctrl' mechanism.
2232 lazy_replace(iffalse, iff2);
2233 lazy_replace(iftrue, ift2);
2234
2235 // Swap names
2236 iffalse = iff2;
2237 iftrue = ift2;
2238 } else {
2239 _igvn.rehash_node_delayed(iffalse);
2240 _igvn.rehash_node_delayed(iftrue);
2241 iffalse->set_req_X( 0, le, &_igvn );
2242 iftrue ->set_req_X( 0, le, &_igvn );
2243 }
2244
2245 set_idom(iftrue, le, dd+1);
2246 set_idom(iffalse, le, dd+1);
2247 assert(iff->outcnt() == 0, "should be dead now");
2248 lazy_replace( iff, le ); // fix 'get_ctrl'
2249
2250 Node* entry_control = init_control;
2251 bool strip_mine_loop = iv_bt == T_INT &&
2252 loop->_child == nullptr &&
2253 sfpt != nullptr &&
2254 !loop->_has_call &&
2255 is_deleteable_safept(sfpt);
2256 IdealLoopTree* outer_ilt = nullptr;
2257 if (strip_mine_loop) {
2258 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
2259 cl_prob, le->_fcnt, entry_control,
2260 iffalse);
2261 }
2262
2263 // Now setup a new CountedLoopNode to replace the existing LoopNode
2264 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2265 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2266 // The following assert is approximately true, and defines the intention
2267 // of can_be_counted_loop. It fails, however, because phase->type
2268 // is not yet initialized for this loop and its parts.
2269 //assert(l->can_be_counted_loop(this), "sanity");
2270 _igvn.register_new_node_with_optimizer(l);
2271 set_loop(l, loop);
2272 loop->_head = l;
2273 // Fix all data nodes placed at the old loop head.
2274 // Uses the lazy-update mechanism of 'get_ctrl'.
2275 lazy_replace( x, l );
2276 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2277
2278 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2279 // Check for immediately preceding SafePoint and remove
2280 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2281 if (strip_mine_loop) {
2282 Node* outer_le = outer_ilt->_tail->in(0);
2283 Node* sfpt_clone = sfpt->clone();
2284 sfpt_clone->set_req(0, iffalse);
2285 outer_le->set_req(0, sfpt_clone);
2286
2287 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2288 if (polladdr != nullptr && polladdr->is_Load()) {
2289 // Polling load should be pinned outside inner loop.
2290 Node* new_polladdr = polladdr->clone();
2291 new_polladdr->set_req(0, iffalse);
2292 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2293 set_ctrl(new_polladdr, iffalse);
2294 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2295 }
2296 // When this code runs, loop bodies have not yet been populated.
2297 const bool body_populated = false;
2298 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2299 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2300 }
2301 lazy_replace(sfpt, sfpt->in(TypeFunc::Control));
2302 if (loop->_safepts != nullptr) {
2303 loop->_safepts->yank(sfpt);
2304 }
2305 }
2306 }
2307
2308 #ifdef ASSERT
2309 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2310 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2311 #endif
2312 #ifndef PRODUCT
2313 if (TraceLoopOpts) {
2314 tty->print("Counted ");
2315 loop->dump_head();
2316 }
2317 #endif
2318
2319 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2320
2321 // Capture bounds of the loop in the induction variable Phi before
2322 // subsequent transformation (iteration splitting) obscures the
2323 // bounds
2324 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2325
2326 if (strip_mine_loop) {
2327 l->mark_strip_mined();
2328 l->verify_strip_mined(1);
2329 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2330 loop = outer_ilt;
2331 }
2332
2333 #ifndef PRODUCT
2334 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2335 Atomic::inc(&_long_loop_counted_loops);
2336 }
2337 #endif
2338 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2339 l->mark_loop_nest_outer_loop();
2340 }
2341
2342 return true;
2343 }
2344
2345 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2346 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2347 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2348 const BasicType iv_bt, Node* loop_entry) {
2349 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2350 // successfully find a dominated test with the If node below.
2351 Node* cmp_limit;
2352 Node* bol;
2353 if (stride_con > 0) {
2354 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2355 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2356 } else {
2357 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2358 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2359 }
2360
2361 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2362 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2363 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2364 IfFalseNode* if_false = new IfFalseNode(iff);
2365 IfTrueNode* if_true = new IfTrueNode(iff);
2366 Node* dominated_iff = _igvn.transform(iff);
2367 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2368 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2369
2370 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2371 _igvn.replace_input_of(iff, 0, C->top());
2372 _igvn.replace_input_of(iff, 1, C->top());
2373 return found_dominating_test;
2374 }
2375
2376 //----------------------exact_limit-------------------------------------------
2377 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2378 assert(loop->_head->is_CountedLoop(), "");
2379 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2380 assert(cl->is_valid_counted_loop(T_INT), "");
2381
2382 if (cl->stride_con() == 1 ||
2383 cl->stride_con() == -1 ||
2384 cl->limit()->Opcode() == Op_LoopLimit) {
2385 // Old code has exact limit (it could be incorrect in case of int overflow).
2386 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2387 return cl->limit();
2388 }
2389 Node *limit = nullptr;
2390 #ifdef ASSERT
2391 BoolTest::mask bt = cl->loopexit()->test_trip();
2392 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2393 #endif
2394 if (cl->has_exact_trip_count()) {
2395 // Simple case: loop has constant boundaries.
2396 // Use jlongs to avoid integer overflow.
2397 int stride_con = cl->stride_con();
2398 jlong init_con = cl->init_trip()->get_int();
2399 jlong limit_con = cl->limit()->get_int();
2400 julong trip_cnt = cl->trip_count();
2401 jlong final_con = init_con + trip_cnt*stride_con;
2402 int final_int = (int)final_con;
2403 // The final value should be in integer range since the loop
2404 // is counted and the limit was checked for overflow.
2405 assert(final_con == (jlong)final_int, "final value should be integer");
2406 limit = _igvn.intcon(final_int);
2407 } else {
2408 // Create new LoopLimit node to get exact limit (final iv value).
2409 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2410 register_new_node(limit, cl->in(LoopNode::EntryControl));
2411 }
2412 assert(limit != nullptr, "sanity");
2413 return limit;
2414 }
2415
2416 //------------------------------Ideal------------------------------------------
2417 // Return a node which is more "ideal" than the current node.
2418 // Attempt to convert into a counted-loop.
2419 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2420 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2421 phase->C->set_major_progress();
2422 }
2423 return RegionNode::Ideal(phase, can_reshape);
2424 }
2425
2426 #ifdef ASSERT
2427 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2428 const OuterStripMinedLoopNode* outer = nullptr;
2429 const CountedLoopNode* inner = nullptr;
2430 if (is_strip_mined()) {
2431 if (!is_valid_counted_loop(T_INT)) {
2432 return; // Skip malformed counted loop
2433 }
2434 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2435 inner = as_CountedLoop();
2436 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2437 } else if (is_OuterStripMinedLoop()) {
2438 outer = this->as_OuterStripMinedLoop();
2439 inner = outer->unique_ctrl_out()->as_CountedLoop();
2440 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2441 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2442 }
2443 if (inner != nullptr || outer != nullptr) {
2444 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2445 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2446 Node* outer_le = outer_tail->in(0);
2447 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2448 Node* sfpt = outer_le->in(0);
2449 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2450 Node* inner_out = sfpt->in(0);
2451 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2452 assert(cle == inner->loopexit_or_null(), "mismatch");
2453 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2454 if (has_skeleton) {
2455 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2456 assert(outer->outcnt() == 2, "only control nodes");
2457 } else {
2458 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2459 uint phis = 0;
2460 uint be_loads = 0;
2461 Node* be = inner->in(LoopNode::LoopBackControl);
2462 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2463 Node* u = inner->fast_out(i);
2464 if (u->is_Phi()) {
2465 phis++;
2466 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2467 Node* n = be->fast_out(j);
2468 if (n->is_Load()) {
2469 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2470 do {
2471 n = n->raw_out(0);
2472 } while (!n->is_Phi());
2473 if (n == u) {
2474 be_loads++;
2475 break;
2476 }
2477 }
2478 }
2479 }
2480 }
2481 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2482 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2483 Node* u = outer->fast_out(i);
2484 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2485 }
2486 uint stores = 0;
2487 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2488 Node* u = inner_out->fast_out(i);
2489 if (u->is_Store()) {
2490 stores++;
2491 }
2492 }
2493 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2494 // not guaranteed to be optimized out.
2495 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2496 }
2497 assert(sfpt->outcnt() == 1, "no data node");
2498 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2499 }
2500 }
2501 #endif
2502
2503 //=============================================================================
2504 //------------------------------Ideal------------------------------------------
2505 // Return a node which is more "ideal" than the current node.
2506 // Attempt to convert into a counted-loop.
2507 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2508 return RegionNode::Ideal(phase, can_reshape);
2509 }
2510
2511 //------------------------------dump_spec--------------------------------------
2512 // Dump special per-node info
2513 #ifndef PRODUCT
2514 void CountedLoopNode::dump_spec(outputStream *st) const {
2515 LoopNode::dump_spec(st);
2516 if (stride_is_con()) {
2517 st->print("stride: %d ",stride_con());
2518 }
2519 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2520 if (is_main_loop()) st->print("main of N%d", _idx);
2521 if (is_post_loop()) st->print("post of N%d", _main_idx);
2522 if (is_strip_mined()) st->print(" strip mined");
2523 if (is_multiversion_fast_loop()) { st->print(" multiversion_fast"); }
2524 if (is_multiversion_slow_loop()) { st->print(" multiversion_slow"); }
2525 if (is_multiversion_delayed_slow_loop()) { st->print(" multiversion_delayed_slow"); }
2526 }
2527 #endif
2528
2529 //=============================================================================
2530 jlong BaseCountedLoopEndNode::stride_con() const {
2531 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2532 }
2533
2534
2535 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2536 if (bt == T_INT) {
2537 return new CountedLoopEndNode(control, test, prob, cnt);
2538 }
2539 assert(bt == T_LONG, "unsupported");
2540 return new LongCountedLoopEndNode(control, test, prob, cnt);
2541 }
2542
2543 //=============================================================================
2544 //------------------------------Value-----------------------------------------
2545 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2546 const Type* init_t = phase->type(in(Init));
2547 const Type* limit_t = phase->type(in(Limit));
2548 const Type* stride_t = phase->type(in(Stride));
2549 // Either input is TOP ==> the result is TOP
2550 if (init_t == Type::TOP) return Type::TOP;
2551 if (limit_t == Type::TOP) return Type::TOP;
2552 if (stride_t == Type::TOP) return Type::TOP;
2553
2554 int stride_con = stride_t->is_int()->get_con();
2555 if (stride_con == 1)
2556 return bottom_type(); // Identity
2557
2558 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2559 // Use jlongs to avoid integer overflow.
2560 jlong init_con = init_t->is_int()->get_con();
2561 jlong limit_con = limit_t->is_int()->get_con();
2562 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2563 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2564 jlong final_con = init_con + stride_con*trip_count;
2565 int final_int = (int)final_con;
2566 // The final value should be in integer range in almost all cases,
2567 // since the loop is counted and the limit was checked for overflow.
2568 // There some exceptions, for example:
2569 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2570 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2571 // the loop and its predicates, and we might get constants on one side of the phi that
2572 // would lead to overflows. Such a code path would never lead us to enter the loop
2573 // because of the loop limit overflow check that happens after the LoopLimitNode
2574 // computation with overflow, but before we enter the loop, see JDK-8335747.
2575 if (final_con == (jlong)final_int) {
2576 return TypeInt::make(final_int);
2577 } else {
2578 return bottom_type();
2579 }
2580 }
2581
2582 return bottom_type(); // TypeInt::INT
2583 }
2584
2585 //------------------------------Ideal------------------------------------------
2586 // Return a node which is more "ideal" than the current node.
2587 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2588 if (phase->type(in(Init)) == Type::TOP ||
2589 phase->type(in(Limit)) == Type::TOP ||
2590 phase->type(in(Stride)) == Type::TOP)
2591 return nullptr; // Dead
2592
2593 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2594 if (stride_con == 1)
2595 return nullptr; // Identity
2596
2597 // Delay following optimizations until all loop optimizations
2598 // done to keep Ideal graph simple.
2599 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2600 phase->C->record_for_post_loop_opts_igvn(this);
2601 return nullptr;
2602 }
2603
2604 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2605 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2606 jlong stride_p;
2607 jlong lim, ini;
2608 julong max;
2609 if (stride_con > 0) {
2610 stride_p = stride_con;
2611 lim = limit_t->_hi;
2612 ini = init_t->_lo;
2613 max = (julong)max_jint;
2614 } else {
2615 stride_p = -(jlong)stride_con;
2616 lim = init_t->_hi;
2617 ini = limit_t->_lo;
2618 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2619 }
2620 julong range = lim - ini + stride_p;
2621 if (range <= max) {
2622 // Convert to integer expression if it is not overflow.
2623 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2624 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2625 Node *bias = phase->transform(new AddINode(range, stride_m));
2626 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2627 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2628 return new AddINode(span, in(Init)); // exact limit
2629 }
2630
2631 if (is_power_of_2(stride_p) || // divisor is 2^n
2632 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2633 // Convert to long expression to avoid integer overflow
2634 // and let igvn optimizer convert this division.
2635 //
2636 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2637 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2638 Node* stride = phase->longcon(stride_con);
2639 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2640
2641 Node *range = phase->transform(new SubLNode(limit, init));
2642 Node *bias = phase->transform(new AddLNode(range, stride_m));
2643 Node *span;
2644 if (stride_con > 0 && is_power_of_2(stride_p)) {
2645 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2646 // and avoid generating rounding for division. Zero trip guard should
2647 // guarantee that init < limit but sometimes the guard is missing and
2648 // we can get situation when init > limit. Note, for the empty loop
2649 // optimization zero trip guard is generated explicitly which leaves
2650 // only RCE predicate where exact limit is used and the predicate
2651 // will simply fail forcing recompilation.
2652 Node* neg_stride = phase->longcon(-stride_con);
2653 span = phase->transform(new AndLNode(bias, neg_stride));
2654 } else {
2655 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2656 span = phase->transform(new MulLNode(trip, stride));
2657 }
2658 // Convert back to int
2659 Node *span_int = phase->transform(new ConvL2INode(span));
2660 return new AddINode(span_int, in(Init)); // exact limit
2661 }
2662
2663 return nullptr; // No progress
2664 }
2665
2666 //------------------------------Identity---------------------------------------
2667 // If stride == 1 return limit node.
2668 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2669 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2670 if (stride_con == 1 || stride_con == -1)
2671 return in(Limit);
2672 return this;
2673 }
2674
2675 //=============================================================================
2676 //----------------------match_incr_with_optional_truncation--------------------
2677 // Match increment with optional truncation:
2678 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2679 // Return null for failure. Success returns the increment node.
2680 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2681 const TypeInteger** trunc_type,
2682 BasicType bt) {
2683 // Quick cutouts:
2684 if (expr == nullptr || expr->req() != 3) return nullptr;
2685
2686 Node *t1 = nullptr;
2687 Node *t2 = nullptr;
2688 Node* n1 = expr;
2689 int n1op = n1->Opcode();
2690 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2691
2692 if (bt == T_INT) {
2693 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2694 if (n1op == Op_AndI &&
2695 n1->in(2)->is_Con() &&
2696 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2697 // %%% This check should match any mask of 2**K-1.
2698 t1 = n1;
2699 n1 = t1->in(1);
2700 n1op = n1->Opcode();
2701 trunc_t = TypeInt::CHAR;
2702 } else if (n1op == Op_RShiftI &&
2703 n1->in(1) != nullptr &&
2704 n1->in(1)->Opcode() == Op_LShiftI &&
2705 n1->in(2) == n1->in(1)->in(2) &&
2706 n1->in(2)->is_Con()) {
2707 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2708 // %%% This check should match any shift in [1..31].
2709 if (shift == 16 || shift == 8) {
2710 t1 = n1;
2711 t2 = t1->in(1);
2712 n1 = t2->in(1);
2713 n1op = n1->Opcode();
2714 if (shift == 16) {
2715 trunc_t = TypeInt::SHORT;
2716 } else if (shift == 8) {
2717 trunc_t = TypeInt::BYTE;
2718 }
2719 }
2720 }
2721 }
2722
2723 // If (maybe after stripping) it is an AddI, we won:
2724 if (n1op == Op_Add(bt)) {
2725 *trunc1 = t1;
2726 *trunc2 = t2;
2727 *trunc_type = trunc_t;
2728 return n1;
2729 }
2730
2731 // failed
2732 return nullptr;
2733 }
2734
2735 IfNode* CountedLoopNode::find_multiversion_if_from_multiversion_fast_main_loop() {
2736 assert(is_main_loop() && is_multiversion_fast_loop(), "must be multiversion fast main loop");
2737 CountedLoopEndNode* pre_end = find_pre_loop_end();
2738 if (pre_end == nullptr) { return nullptr; }
2739 Node* pre_entry = pre_end->loopnode()->in(LoopNode::EntryControl);
2740 const Predicates predicates(pre_entry);
2741 IfTrueNode* before_predicates = predicates.entry()->isa_IfTrue();
2742 if (before_predicates != nullptr &&
2743 before_predicates->in(0)->in(1)->is_OpaqueMultiversioning()) {
2744 return before_predicates->in(0)->as_If();
2745 }
2746 return nullptr;
2747 }
2748
2749 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
2750 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
2751 verify_strip_mined(expect_skeleton);
2752 return in(EntryControl)->as_Loop();
2753 }
2754 return this;
2755 }
2756
2757 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
2758 assert(is_strip_mined(), "not a strip mined loop");
2759 Node* c = in(EntryControl);
2760 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
2761 return nullptr;
2762 }
2763 return c->as_OuterStripMinedLoop();
2764 }
2765
2766 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
2767 Node* c = in(LoopBackControl);
2768 if (c == nullptr || c->is_top()) {
2769 return nullptr;
2770 }
2771 return c->as_IfTrue();
2772 }
2773
2774 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
2775 LoopNode* l = outer_loop();
2776 if (l == nullptr) {
2777 return nullptr;
2778 }
2779 return l->outer_loop_tail();
2780 }
2781
2782 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
2783 IfTrueNode* proj = outer_loop_tail();
2784 if (proj == nullptr) {
2785 return nullptr;
2786 }
2787 Node* c = proj->in(0);
2788 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
2789 return nullptr;
2790 }
2791 return c->as_OuterStripMinedLoopEnd();
2792 }
2793
2794 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
2795 LoopNode* l = outer_loop();
2796 if (l == nullptr) {
2797 return nullptr;
2798 }
2799 return l->outer_loop_end();
2800 }
2801
2802 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
2803 IfNode* le = outer_loop_end();
2804 if (le == nullptr) {
2805 return nullptr;
2806 }
2807 Node* c = le->proj_out_or_null(false);
2808 if (c == nullptr) {
2809 return nullptr;
2810 }
2811 return c->as_IfFalse();
2812 }
2813
2814 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
2815 LoopNode* l = outer_loop();
2816 if (l == nullptr) {
2817 return nullptr;
2818 }
2819 return l->outer_loop_exit();
2820 }
2821
2822 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
2823 IfNode* le = outer_loop_end();
2824 if (le == nullptr) {
2825 return nullptr;
2826 }
2827 Node* c = le->in(0);
2828 if (c == nullptr || c->is_top()) {
2829 return nullptr;
2830 }
2831 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
2832 return c->as_SafePoint();
2833 }
2834
2835 SafePointNode* CountedLoopNode::outer_safepoint() const {
2836 LoopNode* l = outer_loop();
2837 if (l == nullptr) {
2838 return nullptr;
2839 }
2840 return l->outer_safepoint();
2841 }
2842
2843 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
2844 Node* ctrl = in(LoopNode::EntryControl);
2845 if (ctrl == nullptr) {
2846 // Dying loop.
2847 return nullptr;
2848 }
2849 if (is_main_loop()) {
2850 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
2851 }
2852 if (is_main_loop() || is_post_loop()) {
2853 AssertionPredicates assertion_predicates(ctrl);
2854 return assertion_predicates.entry();
2855 }
2856 return ctrl;
2857 }
2858
2859
2860 int CountedLoopNode::stride_con() const {
2861 CountedLoopEndNode* cle = loopexit_or_null();
2862 return cle != nullptr ? cle->stride_con() : 0;
2863 }
2864
2865 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
2866 if (bt == T_INT) {
2867 return new CountedLoopNode(entry, backedge);
2868 }
2869 assert(bt == T_LONG, "unsupported");
2870 return new LongCountedLoopNode(entry, backedge);
2871 }
2872
2873 void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
2874 PhaseIdealLoop* iloop) {
2875 Node* cle_out = inner_cle->proj_out(false);
2876 Node* cle_tail = inner_cle->proj_out(true);
2877 if (cle_out->outcnt() > 1) {
2878 // Look for chains of stores that were sunk
2879 // out of the inner loop and are in the outer loop
2880 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
2881 Node* u = cle_out->fast_out(i);
2882 if (u->is_Store()) {
2883 int alias_idx = igvn->C->get_alias_index(u->adr_type());
2884 Node* first = u;
2885 for (;;) {
2886 Node* next = first->in(MemNode::Memory);
2887 if (!next->is_Store() || next->in(0) != cle_out) {
2888 break;
2889 }
2890 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2891 first = next;
2892 }
2893 Node* last = u;
2894 for (;;) {
2895 Node* next = nullptr;
2896 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
2897 Node* uu = last->fast_out(j);
2898 if (uu->is_Store() && uu->in(0) == cle_out) {
2899 assert(next == nullptr, "only one in the outer loop");
2900 next = uu;
2901 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2902 }
2903 }
2904 if (next == nullptr) {
2905 break;
2906 }
2907 last = next;
2908 }
2909 Node* phi = nullptr;
2910 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2911 Node* uu = inner_cl->fast_out(j);
2912 if (uu->is_Phi()) {
2913 Node* be = uu->in(LoopNode::LoopBackControl);
2914 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
2915 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
2916 }
2917 if (be == last || be == first->in(MemNode::Memory)) {
2918 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
2919 assert(phi == nullptr, "only one phi");
2920 phi = uu;
2921 }
2922 }
2923 }
2924 #ifdef ASSERT
2925 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2926 Node* uu = inner_cl->fast_out(j);
2927 if (uu->is_memory_phi()) {
2928 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
2929 assert(phi == uu, "what's that phi?");
2930 } else if (uu->adr_type() == TypePtr::BOTTOM) {
2931 Node* n = uu->in(LoopNode::LoopBackControl);
2932 uint limit = igvn->C->live_nodes();
2933 uint i = 0;
2934 while (n != uu) {
2935 i++;
2936 assert(i < limit, "infinite loop");
2937 if (n->is_Proj()) {
2938 n = n->in(0);
2939 } else if (n->is_SafePoint() || n->is_MemBar()) {
2940 n = n->in(TypeFunc::Memory);
2941 } else if (n->is_Phi()) {
2942 n = n->in(1);
2943 } else if (n->is_MergeMem()) {
2944 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
2945 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
2946 n = n->in(MemNode::Memory);
2947 } else {
2948 n->dump();
2949 ShouldNotReachHere();
2950 }
2951 }
2952 }
2953 }
2954 }
2955 #endif
2956 if (phi == nullptr) {
2957 // If an entire chains was sunk, the
2958 // inner loop has no phi for that memory
2959 // slice, create one for the outer loop
2960 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
2961 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
2962 phi->set_req(LoopNode::LoopBackControl, last);
2963 phi = register_new_node(phi, inner_cl, igvn, iloop);
2964 igvn->replace_input_of(first, MemNode::Memory, phi);
2965 } else {
2966 // Or fix the outer loop fix to include
2967 // that chain of stores.
2968 Node* be = phi->in(LoopNode::LoopBackControl);
2969 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
2970 if (be == first->in(MemNode::Memory)) {
2971 if (be == phi->in(LoopNode::LoopBackControl)) {
2972 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
2973 } else {
2974 igvn->replace_input_of(be, MemNode::Memory, last);
2975 }
2976 } else {
2977 #ifdef ASSERT
2978 if (be == phi->in(LoopNode::LoopBackControl)) {
2979 assert(phi->in(LoopNode::LoopBackControl) == last, "");
2980 } else {
2981 assert(be->in(MemNode::Memory) == last, "");
2982 }
2983 #endif
2984 }
2985 }
2986 }
2987 }
2988 }
2989 }
2990
2991 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
2992 // Look for the outer & inner strip mined loop, reduce number of
2993 // iterations of the inner loop, set exit condition of outer loop,
2994 // construct required phi nodes for outer loop.
2995 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
2996 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
2997 if (LoopStripMiningIter == 0) {
2998 remove_outer_loop_and_safepoint(igvn);
2999 return;
3000 }
3001 if (LoopStripMiningIter == 1) {
3002 transform_to_counted_loop(igvn, nullptr);
3003 return;
3004 }
3005 Node* inner_iv_phi = inner_cl->phi();
3006 if (inner_iv_phi == nullptr) {
3007 IfNode* outer_le = outer_loop_end();
3008 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3009 igvn->replace_node(outer_le, iff);
3010 inner_cl->clear_strip_mined();
3011 return;
3012 }
3013 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3014
3015 int stride = inner_cl->stride_con();
3016 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
3017 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
3018 // stride, the method is guaranteed to return at the next check below.
3019 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
3020 int scaled_iters = (int)scaled_iters_long;
3021 if ((jlong)scaled_iters != scaled_iters_long) {
3022 // Remove outer loop and safepoint (too few iterations)
3023 remove_outer_loop_and_safepoint(igvn);
3024 return;
3025 }
3026 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3027 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3028 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3029 assert(iter_estimate > 0, "broken");
3030 if (iter_estimate <= short_scaled_iters) {
3031 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3032 remove_outer_loop_and_safepoint(igvn);
3033 return;
3034 }
3035 if (iter_estimate <= scaled_iters_long) {
3036 // We would only go through one iteration of
3037 // the outer loop: drop the outer loop but
3038 // keep the safepoint so we don't run for
3039 // too long without a safepoint
3040 IfNode* outer_le = outer_loop_end();
3041 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3042 igvn->replace_node(outer_le, iff);
3043 inner_cl->clear_strip_mined();
3044 return;
3045 }
3046
3047 Node* cle_tail = inner_cle->proj_out(true);
3048 ResourceMark rm;
3049 Node_List old_new;
3050 if (cle_tail->outcnt() > 1) {
3051 // Look for nodes on backedge of inner loop and clone them
3052 Unique_Node_List backedge_nodes;
3053 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3054 Node* u = cle_tail->fast_out(i);
3055 if (u != inner_cl) {
3056 assert(!u->is_CFG(), "control flow on the backedge?");
3057 backedge_nodes.push(u);
3058 }
3059 }
3060 uint last = igvn->C->unique();
3061 for (uint next = 0; next < backedge_nodes.size(); next++) {
3062 Node* n = backedge_nodes.at(next);
3063 old_new.map(n->_idx, n->clone());
3064 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3065 Node* u = n->fast_out(i);
3066 assert(!u->is_CFG(), "broken");
3067 if (u->_idx >= last) {
3068 continue;
3069 }
3070 if (!u->is_Phi()) {
3071 backedge_nodes.push(u);
3072 } else {
3073 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3074 }
3075 }
3076 }
3077 // Put the clones on the outer loop backedge
3078 Node* le_tail = outer_loop_tail();
3079 for (uint next = 0; next < backedge_nodes.size(); next++) {
3080 Node *n = old_new[backedge_nodes.at(next)->_idx];
3081 for (uint i = 1; i < n->req(); i++) {
3082 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3083 n->set_req(i, old_new[n->in(i)->_idx]);
3084 }
3085 }
3086 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3087 n->set_req(0, le_tail);
3088 }
3089 igvn->register_new_node_with_optimizer(n);
3090 }
3091 }
3092
3093 Node* iv_phi = nullptr;
3094 // Make a clone of each phi in the inner loop
3095 // for the outer loop
3096 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3097 Node* u = inner_cl->raw_out(i);
3098 if (u->is_Phi()) {
3099 assert(u->in(0) == inner_cl, "inconsistent");
3100 Node* phi = u->clone();
3101 phi->set_req(0, this);
3102 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3103 if (be != nullptr) {
3104 phi->set_req(LoopNode::LoopBackControl, be);
3105 }
3106 phi = igvn->transform(phi);
3107 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3108 if (u == inner_iv_phi) {
3109 iv_phi = phi;
3110 }
3111 }
3112 }
3113
3114 if (iv_phi != nullptr) {
3115 // Now adjust the inner loop's exit condition
3116 Node* limit = inner_cl->limit();
3117 // If limit < init for stride > 0 (or limit > init for stride < 0),
3118 // the loop body is run only once. Given limit - init (init - limit resp.)
3119 // would be negative, the unsigned comparison below would cause
3120 // the loop body to be run for LoopStripMiningIter.
3121 Node* max = nullptr;
3122 if (stride > 0) {
3123 max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3124 } else {
3125 max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3126 }
3127 // sub is positive and can be larger than the max signed int
3128 // value. Use an unsigned min.
3129 Node* const_iters = igvn->intcon(scaled_iters);
3130 Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3131 // min is the number of iterations for the next inner loop execution:
3132 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3133 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3134
3135 Node* new_limit = nullptr;
3136 if (stride > 0) {
3137 new_limit = igvn->transform(new AddINode(min, iv_phi));
3138 } else {
3139 new_limit = igvn->transform(new SubINode(iv_phi, min));
3140 }
3141 Node* inner_cmp = inner_cle->cmp_node();
3142 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3143 Node* outer_bol = inner_bol;
3144 // cmp node for inner loop may be shared
3145 inner_cmp = inner_cmp->clone();
3146 inner_cmp->set_req(2, new_limit);
3147 inner_bol = inner_bol->clone();
3148 inner_bol->set_req(1, igvn->transform(inner_cmp));
3149 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3150 // Set the outer loop's exit condition too
3151 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3152 } else {
3153 assert(false, "should be able to adjust outer loop");
3154 IfNode* outer_le = outer_loop_end();
3155 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3156 igvn->replace_node(outer_le, iff);
3157 inner_cl->clear_strip_mined();
3158 }
3159 }
3160
3161 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3162 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3163 CountedLoopEndNode* cle = inner_cl->loopexit();
3164 Node* inner_test = cle->in(1);
3165 IfNode* outer_le = outer_loop_end();
3166 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3167 Node* safepoint = outer_safepoint();
3168
3169 fix_sunk_stores(inner_cle, inner_cl, igvn, iloop);
3170
3171 // make counted loop exit test always fail
3172 ConINode* zero = igvn->intcon(0);
3173 if (iloop != nullptr) {
3174 iloop->set_root_as_ctrl(zero);
3175 }
3176 igvn->replace_input_of(cle, 1, zero);
3177 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3178 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3179 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3180 if (iloop == nullptr) {
3181 igvn->replace_node(outer_le, new_end);
3182 } else {
3183 iloop->lazy_replace(outer_le, new_end);
3184 }
3185 // the backedge of the inner loop must be rewired to the new loop end
3186 Node* backedge = cle->proj_out(true);
3187 igvn->replace_input_of(backedge, 0, new_end);
3188 if (iloop != nullptr) {
3189 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3190 }
3191 // make the outer loop go away
3192 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3193 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3194 inner_cl->clear_strip_mined();
3195 if (iloop != nullptr) {
3196 Unique_Node_List wq;
3197 wq.push(safepoint);
3198
3199 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3200 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3201
3202 for (uint i = 0; i < wq.size(); i++) {
3203 Node* n = wq.at(i);
3204 for (uint j = 0; j < n->req(); ++j) {
3205 Node* in = n->in(j);
3206 if (in == nullptr || in->is_CFG()) {
3207 continue;
3208 }
3209 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3210 continue;
3211 }
3212 wq.push(in);
3213 }
3214 assert(!loop->_body.contains(n), "Shouldn't append node to body twice");
3215 loop->_body.push(n);
3216 }
3217 iloop->set_loop(safepoint, loop);
3218 loop->_body.push(safepoint);
3219 iloop->set_loop(safepoint->in(0), loop);
3220 loop->_body.push(safepoint->in(0));
3221 outer_loop_ilt->_tail = igvn->C->top();
3222 }
3223 }
3224
3225 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3226 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3227 Node* outer_sfpt = outer_safepoint();
3228 Node* outer_out = outer_loop_exit();
3229 igvn->replace_node(outer_out, outer_sfpt->in(0));
3230 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3231 inner_cl->clear_strip_mined();
3232 }
3233
3234 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3235 if (iloop == nullptr) {
3236 return igvn->transform(node);
3237 }
3238 iloop->register_new_node(node, ctrl);
3239 return node;
3240 }
3241
3242 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3243 PhaseIdealLoop* iloop) {
3244 if (iloop == nullptr) {
3245 return igvn->transform(node);
3246 }
3247 iloop->register_control(node, iloop->get_loop(loop), idom);
3248 return node;
3249 }
3250
3251 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3252 if (!in(0)) return Type::TOP;
3253 if (phase->type(in(0)) == Type::TOP)
3254 return Type::TOP;
3255
3256 // Until expansion, the loop end condition is not set so this should not constant fold.
3257 if (is_expanded(phase)) {
3258 return IfNode::Value(phase);
3259 }
3260
3261 return TypeTuple::IFBOTH;
3262 }
3263
3264 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3265 // The outer strip mined loop head only has Phi uses after expansion
3266 if (phase->is_IterGVN()) {
3267 Node* backedge = proj_out_or_null(true);
3268 if (backedge != nullptr) {
3269 Node* head = backedge->unique_ctrl_out_or_null();
3270 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3271 if (head->find_out_with(Op_Phi) != nullptr) {
3272 return true;
3273 }
3274 }
3275 }
3276 }
3277 return false;
3278 }
3279
3280 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3281 if (remove_dead_region(phase, can_reshape)) return this;
3282
3283 return nullptr;
3284 }
3285
3286 //------------------------------filtered_type--------------------------------
3287 // Return a type based on condition control flow
3288 // A successful return will be a type that is restricted due
3289 // to a series of dominating if-tests, such as:
3290 // if (i < 10) {
3291 // if (i > 0) {
3292 // here: "i" type is [1..10)
3293 // }
3294 // }
3295 // or a control flow merge
3296 // if (i < 10) {
3297 // do {
3298 // phi( , ) -- at top of loop type is [min_int..10)
3299 // i = ?
3300 // } while ( i < 10)
3301 //
3302 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3303 assert(n && n->bottom_type()->is_int(), "must be int");
3304 const TypeInt* filtered_t = nullptr;
3305 if (!n->is_Phi()) {
3306 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3307 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3308
3309 } else {
3310 Node* phi = n->as_Phi();
3311 Node* region = phi->in(0);
3312 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3313 if (region && region != C->top()) {
3314 for (uint i = 1; i < phi->req(); i++) {
3315 Node* val = phi->in(i);
3316 Node* use_c = region->in(i);
3317 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3318 if (val_t != nullptr) {
3319 if (filtered_t == nullptr) {
3320 filtered_t = val_t;
3321 } else {
3322 filtered_t = filtered_t->meet(val_t)->is_int();
3323 }
3324 }
3325 }
3326 }
3327 }
3328 const TypeInt* n_t = _igvn.type(n)->is_int();
3329 if (filtered_t != nullptr) {
3330 n_t = n_t->join(filtered_t)->is_int();
3331 }
3332 return n_t;
3333 }
3334
3335
3336 //------------------------------filtered_type_from_dominators--------------------------------
3337 // Return a possibly more restrictive type for val based on condition control flow of dominators
3338 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3339 if (val->is_Con()) {
3340 return val->bottom_type()->is_int();
3341 }
3342 uint if_limit = 10; // Max number of dominating if's visited
3343 const TypeInt* rtn_t = nullptr;
3344
3345 if (use_ctrl && use_ctrl != C->top()) {
3346 Node* val_ctrl = get_ctrl(val);
3347 uint val_dom_depth = dom_depth(val_ctrl);
3348 Node* pred = use_ctrl;
3349 uint if_cnt = 0;
3350 while (if_cnt < if_limit) {
3351 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3352 if_cnt++;
3353 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3354 if (if_t != nullptr) {
3355 if (rtn_t == nullptr) {
3356 rtn_t = if_t;
3357 } else {
3358 rtn_t = rtn_t->join(if_t)->is_int();
3359 }
3360 }
3361 }
3362 pred = idom(pred);
3363 if (pred == nullptr || pred == C->top()) {
3364 break;
3365 }
3366 // Stop if going beyond definition block of val
3367 if (dom_depth(pred) < val_dom_depth) {
3368 break;
3369 }
3370 }
3371 }
3372 return rtn_t;
3373 }
3374
3375
3376 //------------------------------dump_spec--------------------------------------
3377 // Dump special per-node info
3378 #ifndef PRODUCT
3379 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3380 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3381 BoolTest bt( test_trip()); // Added this for g++.
3382
3383 st->print("[");
3384 bt.dump_on(st);
3385 st->print("]");
3386 }
3387 st->print(" ");
3388 IfNode::dump_spec(st);
3389 }
3390 #endif
3391
3392 //=============================================================================
3393 //------------------------------is_member--------------------------------------
3394 // Is 'l' a member of 'this'?
3395 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3396 while( l->_nest > _nest ) l = l->_parent;
3397 return l == this;
3398 }
3399
3400 //------------------------------set_nest---------------------------------------
3401 // Set loop tree nesting depth. Accumulate _has_call bits.
3402 int IdealLoopTree::set_nest( uint depth ) {
3403 assert(depth <= SHRT_MAX, "sanity");
3404 _nest = depth;
3405 int bits = _has_call;
3406 if( _child ) bits |= _child->set_nest(depth+1);
3407 if( bits ) _has_call = 1;
3408 if( _next ) bits |= _next ->set_nest(depth );
3409 return bits;
3410 }
3411
3412 //------------------------------split_fall_in----------------------------------
3413 // Split out multiple fall-in edges from the loop header. Move them to a
3414 // private RegionNode before the loop. This becomes the loop landing pad.
3415 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3416 PhaseIterGVN &igvn = phase->_igvn;
3417 uint i;
3418
3419 // Make a new RegionNode to be the landing pad.
3420 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3421 phase->set_loop(landing_pad,_parent);
3422 // If _head was irreducible loop entry, landing_pad may now be too
3423 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3424 // Gather all the fall-in control paths into the landing pad
3425 uint icnt = fall_in_cnt;
3426 uint oreq = _head->req();
3427 for( i = oreq-1; i>0; i-- )
3428 if( !phase->is_member( this, _head->in(i) ) )
3429 landing_pad->set_req(icnt--,_head->in(i));
3430
3431 // Peel off PhiNode edges as well
3432 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3433 Node *oj = _head->fast_out(j);
3434 if( oj->is_Phi() ) {
3435 PhiNode* old_phi = oj->as_Phi();
3436 assert( old_phi->region() == _head, "" );
3437 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3438 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3439 uint icnt = fall_in_cnt;
3440 for( i = oreq-1; i>0; i-- ) {
3441 if( !phase->is_member( this, _head->in(i) ) ) {
3442 p->init_req(icnt--, old_phi->in(i));
3443 // Go ahead and clean out old edges from old phi
3444 old_phi->del_req(i);
3445 }
3446 }
3447 // Search for CSE's here, because ZKM.jar does a lot of
3448 // loop hackery and we need to be a little incremental
3449 // with the CSE to avoid O(N^2) node blow-up.
3450 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3451 if( p2 ) { // Found CSE
3452 p->destruct(&igvn); // Recover useless new node
3453 p = p2; // Use old node
3454 } else {
3455 igvn.register_new_node_with_optimizer(p, old_phi);
3456 }
3457 // Make old Phi refer to new Phi.
3458 old_phi->add_req(p);
3459 // Check for the special case of making the old phi useless and
3460 // disappear it. In JavaGrande I have a case where this useless
3461 // Phi is the loop limit and prevents recognizing a CountedLoop
3462 // which in turn prevents removing an empty loop.
3463 Node *id_old_phi = old_phi->Identity(&igvn);
3464 if( id_old_phi != old_phi ) { // Found a simple identity?
3465 // Note that I cannot call 'replace_node' here, because
3466 // that will yank the edge from old_phi to the Region and
3467 // I'm mid-iteration over the Region's uses.
3468 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3469 Node* use = old_phi->last_out(i);
3470 igvn.rehash_node_delayed(use);
3471 uint uses_found = 0;
3472 for (uint j = 0; j < use->len(); j++) {
3473 if (use->in(j) == old_phi) {
3474 if (j < use->req()) use->set_req (j, id_old_phi);
3475 else use->set_prec(j, id_old_phi);
3476 uses_found++;
3477 }
3478 }
3479 i -= uses_found; // we deleted 1 or more copies of this edge
3480 }
3481 }
3482 igvn._worklist.push(old_phi);
3483 }
3484 }
3485 // Finally clean out the fall-in edges from the RegionNode
3486 for( i = oreq-1; i>0; i-- ) {
3487 if( !phase->is_member( this, _head->in(i) ) ) {
3488 _head->del_req(i);
3489 }
3490 }
3491 igvn.rehash_node_delayed(_head);
3492 // Transform landing pad
3493 igvn.register_new_node_with_optimizer(landing_pad, _head);
3494 // Insert landing pad into the header
3495 _head->add_req(landing_pad);
3496 }
3497
3498 //------------------------------split_outer_loop-------------------------------
3499 // Split out the outermost loop from this shared header.
3500 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3501 PhaseIterGVN &igvn = phase->_igvn;
3502
3503 // Find index of outermost loop; it should also be my tail.
3504 uint outer_idx = 1;
3505 while( _head->in(outer_idx) != _tail ) outer_idx++;
3506
3507 // Make a LoopNode for the outermost loop.
3508 Node *ctl = _head->in(LoopNode::EntryControl);
3509 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3510 outer = igvn.register_new_node_with_optimizer(outer, _head);
3511 phase->set_created_loop_node();
3512
3513 // Outermost loop falls into '_head' loop
3514 _head->set_req(LoopNode::EntryControl, outer);
3515 _head->del_req(outer_idx);
3516 // Split all the Phis up between '_head' loop and 'outer' loop.
3517 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3518 Node *out = _head->fast_out(j);
3519 if( out->is_Phi() ) {
3520 PhiNode *old_phi = out->as_Phi();
3521 assert( old_phi->region() == _head, "" );
3522 Node *phi = PhiNode::make_blank(outer, old_phi);
3523 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3524 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3525 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3526 // Make old Phi point to new Phi on the fall-in path
3527 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3528 old_phi->del_req(outer_idx);
3529 }
3530 }
3531
3532 // Use the new loop head instead of the old shared one
3533 _head = outer;
3534 phase->set_loop(_head, this);
3535 }
3536
3537 //------------------------------fix_parent-------------------------------------
3538 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3539 loop->_parent = parent;
3540 if( loop->_child ) fix_parent( loop->_child, loop );
3541 if( loop->_next ) fix_parent( loop->_next , parent );
3542 }
3543
3544 //------------------------------estimate_path_freq-----------------------------
3545 static float estimate_path_freq( Node *n ) {
3546 // Try to extract some path frequency info
3547 IfNode *iff;
3548 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3549 uint nop = n->Opcode();
3550 if( nop == Op_SafePoint ) { // Skip any safepoint
3551 n = n->in(0);
3552 continue;
3553 }
3554 if( nop == Op_CatchProj ) { // Get count from a prior call
3555 // Assume call does not always throw exceptions: means the call-site
3556 // count is also the frequency of the fall-through path.
3557 assert( n->is_CatchProj(), "" );
3558 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3559 return 0.0f; // Assume call exception path is rare
3560 Node *call = n->in(0)->in(0)->in(0);
3561 assert( call->is_Call(), "expect a call here" );
3562 const JVMState *jvms = ((CallNode*)call)->jvms();
3563 ciMethodData* methodData = jvms->method()->method_data();
3564 if (!methodData->is_mature()) return 0.0f; // No call-site data
3565 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3566 if ((data == nullptr) || !data->is_CounterData()) {
3567 // no call profile available, try call's control input
3568 n = n->in(0);
3569 continue;
3570 }
3571 return data->as_CounterData()->count()/FreqCountInvocations;
3572 }
3573 // See if there's a gating IF test
3574 Node *n_c = n->in(0);
3575 if( !n_c->is_If() ) break; // No estimate available
3576 iff = n_c->as_If();
3577 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3578 // Compute how much count comes on this path
3579 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3580 // Have no count info. Skip dull uncommon-trap like branches.
3581 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3582 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3583 break;
3584 // Skip through never-taken branch; look for a real loop exit.
3585 n = iff->in(0);
3586 }
3587 return 0.0f; // No estimate available
3588 }
3589
3590 //------------------------------merge_many_backedges---------------------------
3591 // Merge all the backedges from the shared header into a private Region.
3592 // Feed that region as the one backedge to this loop.
3593 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3594 uint i;
3595
3596 // Scan for the top 2 hottest backedges
3597 float hotcnt = 0.0f;
3598 float warmcnt = 0.0f;
3599 uint hot_idx = 0;
3600 // Loop starts at 2 because slot 1 is the fall-in path
3601 for( i = 2; i < _head->req(); i++ ) {
3602 float cnt = estimate_path_freq(_head->in(i));
3603 if( cnt > hotcnt ) { // Grab hottest path
3604 warmcnt = hotcnt;
3605 hotcnt = cnt;
3606 hot_idx = i;
3607 } else if( cnt > warmcnt ) { // And 2nd hottest path
3608 warmcnt = cnt;
3609 }
3610 }
3611
3612 // See if the hottest backedge is worthy of being an inner loop
3613 // by being much hotter than the next hottest backedge.
3614 if( hotcnt <= 0.0001 ||
3615 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3616
3617 // Peel out the backedges into a private merge point; peel
3618 // them all except optionally hot_idx.
3619 PhaseIterGVN &igvn = phase->_igvn;
3620
3621 Node *hot_tail = nullptr;
3622 // Make a Region for the merge point
3623 Node *r = new RegionNode(1);
3624 for( i = 2; i < _head->req(); i++ ) {
3625 if( i != hot_idx )
3626 r->add_req( _head->in(i) );
3627 else hot_tail = _head->in(i);
3628 }
3629 igvn.register_new_node_with_optimizer(r, _head);
3630 // Plug region into end of loop _head, followed by hot_tail
3631 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3632 igvn.replace_input_of(_head, 2, r);
3633 if( hot_idx ) _head->add_req(hot_tail);
3634
3635 // Split all the Phis up between '_head' loop and the Region 'r'
3636 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3637 Node *out = _head->fast_out(j);
3638 if( out->is_Phi() ) {
3639 PhiNode* n = out->as_Phi();
3640 igvn.hash_delete(n); // Delete from hash before hacking edges
3641 Node *hot_phi = nullptr;
3642 Node *phi = new PhiNode(r, n->type(), n->adr_type());
3643 // Check all inputs for the ones to peel out
3644 uint j = 1;
3645 for( uint i = 2; i < n->req(); i++ ) {
3646 if( i != hot_idx )
3647 phi->set_req( j++, n->in(i) );
3648 else hot_phi = n->in(i);
3649 }
3650 // Register the phi but do not transform until whole place transforms
3651 igvn.register_new_node_with_optimizer(phi, n);
3652 // Add the merge phi to the old Phi
3653 while( n->req() > 3 ) n->del_req( n->req()-1 );
3654 igvn.replace_input_of(n, 2, phi);
3655 if( hot_idx ) n->add_req(hot_phi);
3656 }
3657 }
3658
3659
3660 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
3661 // of self loop tree. Turn self into a loop headed by _head and with
3662 // tail being the new merge point.
3663 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
3664 phase->set_loop(_tail,ilt); // Adjust tail
3665 _tail = r; // Self's tail is new merge point
3666 phase->set_loop(r,this);
3667 ilt->_child = _child; // New guy has my children
3668 _child = ilt; // Self has new guy as only child
3669 ilt->_parent = this; // new guy has self for parent
3670 ilt->_nest = _nest; // Same nesting depth (for now)
3671
3672 // Starting with 'ilt', look for child loop trees using the same shared
3673 // header. Flatten these out; they will no longer be loops in the end.
3674 IdealLoopTree **pilt = &_child;
3675 while( ilt ) {
3676 if( ilt->_head == _head ) {
3677 uint i;
3678 for( i = 2; i < _head->req(); i++ )
3679 if( _head->in(i) == ilt->_tail )
3680 break; // Still a loop
3681 if( i == _head->req() ) { // No longer a loop
3682 // Flatten ilt. Hang ilt's "_next" list from the end of
3683 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
3684 IdealLoopTree **cp = &ilt->_child;
3685 while( *cp ) cp = &(*cp)->_next; // Find end of child list
3686 *cp = ilt->_next; // Hang next list at end of child list
3687 *pilt = ilt->_child; // Move child up to replace ilt
3688 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
3689 ilt = ilt->_child; // Repeat using new ilt
3690 continue; // do not advance over ilt->_child
3691 }
3692 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
3693 phase->set_loop(_head,ilt);
3694 }
3695 pilt = &ilt->_child; // Advance to next
3696 ilt = *pilt;
3697 }
3698
3699 if( _child ) fix_parent( _child, this );
3700 }
3701
3702 //------------------------------beautify_loops---------------------------------
3703 // Split shared headers and insert loop landing pads.
3704 // Insert a LoopNode to replace the RegionNode.
3705 // Return TRUE if loop tree is structurally changed.
3706 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
3707 bool result = false;
3708 // Cache parts in locals for easy
3709 PhaseIterGVN &igvn = phase->_igvn;
3710
3711 igvn.hash_delete(_head); // Yank from hash before hacking edges
3712
3713 // Check for multiple fall-in paths. Peel off a landing pad if need be.
3714 int fall_in_cnt = 0;
3715 for( uint i = 1; i < _head->req(); i++ )
3716 if( !phase->is_member( this, _head->in(i) ) )
3717 fall_in_cnt++;
3718 assert( fall_in_cnt, "at least 1 fall-in path" );
3719 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
3720 split_fall_in( phase, fall_in_cnt );
3721
3722 // Swap inputs to the _head and all Phis to move the fall-in edge to
3723 // the left.
3724 fall_in_cnt = 1;
3725 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
3726 fall_in_cnt++;
3727 if( fall_in_cnt > 1 ) {
3728 // Since I am just swapping inputs I do not need to update def-use info
3729 Node *tmp = _head->in(1);
3730 igvn.rehash_node_delayed(_head);
3731 _head->set_req( 1, _head->in(fall_in_cnt) );
3732 _head->set_req( fall_in_cnt, tmp );
3733 // Swap also all Phis
3734 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
3735 Node* phi = _head->fast_out(i);
3736 if( phi->is_Phi() ) {
3737 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
3738 tmp = phi->in(1);
3739 phi->set_req( 1, phi->in(fall_in_cnt) );
3740 phi->set_req( fall_in_cnt, tmp );
3741 }
3742 }
3743 }
3744 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
3745 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
3746
3747 // If I am a shared header (multiple backedges), peel off the many
3748 // backedges into a private merge point and use the merge point as
3749 // the one true backedge.
3750 if (_head->req() > 3) {
3751 // Merge the many backedges into a single backedge but leave
3752 // the hottest backedge as separate edge for the following peel.
3753 if (!_irreducible) {
3754 merge_many_backedges( phase );
3755 }
3756
3757 // When recursively beautify my children, split_fall_in can change
3758 // loop tree structure when I am an irreducible loop. Then the head
3759 // of my children has a req() not bigger than 3. Here we need to set
3760 // result to true to catch that case in order to tell the caller to
3761 // rebuild loop tree. See issue JDK-8244407 for details.
3762 result = true;
3763 }
3764
3765 // If I have one hot backedge, peel off myself loop.
3766 // I better be the outermost loop.
3767 if (_head->req() > 3 && !_irreducible) {
3768 split_outer_loop( phase );
3769 result = true;
3770
3771 } else if (!_head->is_Loop() && !_irreducible) {
3772 // Make a new LoopNode to replace the old loop head
3773 Node *l = new LoopNode( _head->in(1), _head->in(2) );
3774 l = igvn.register_new_node_with_optimizer(l, _head);
3775 phase->set_created_loop_node();
3776 // Go ahead and replace _head
3777 phase->_igvn.replace_node( _head, l );
3778 _head = l;
3779 phase->set_loop(_head, this);
3780 }
3781
3782 // Now recursively beautify nested loops
3783 if( _child ) result |= _child->beautify_loops( phase );
3784 if( _next ) result |= _next ->beautify_loops( phase );
3785 return result;
3786 }
3787
3788 //------------------------------allpaths_check_safepts----------------------------
3789 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
3790 // safepoint encountered. Helper for check_safepts.
3791 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
3792 assert(stack.size() == 0, "empty stack");
3793 stack.push(_head);
3794 visited.clear();
3795 visited.set(_head->_idx);
3796 while (stack.size() > 0) {
3797 Node* n = stack.pop();
3798 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3799 // Terminate this path
3800 } else if (n->Opcode() == Op_SafePoint) {
3801 if (_phase->get_loop(n) != this) {
3802 if (_required_safept == nullptr) _required_safept = new Node_List();
3803 // save the first we run into on that path: closest to the tail if the head has a single backedge
3804 _required_safept->push(n);
3805 }
3806 // Terminate this path
3807 } else {
3808 uint start = n->is_Region() ? 1 : 0;
3809 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
3810 for (uint i = start; i < end; i++) {
3811 Node* in = n->in(i);
3812 assert(in->is_CFG(), "must be");
3813 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
3814 stack.push(in);
3815 }
3816 }
3817 }
3818 }
3819 }
3820
3821 //------------------------------check_safepts----------------------------
3822 // Given dominators, try to find loops with calls that must always be
3823 // executed (call dominates loop tail). These loops do not need non-call
3824 // safepoints (ncsfpt).
3825 //
3826 // A complication is that a safepoint in a inner loop may be needed
3827 // by an outer loop. In the following, the inner loop sees it has a
3828 // call (block 3) on every path from the head (block 2) to the
3829 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
3830 // in block 2, _but_ this leaves the outer loop without a safepoint.
3831 //
3832 // entry 0
3833 // |
3834 // v
3835 // outer 1,2 +->1
3836 // | |
3837 // | v
3838 // | 2<---+ ncsfpt in 2
3839 // |_/|\ |
3840 // | v |
3841 // inner 2,3 / 3 | call in 3
3842 // / | |
3843 // v +--+
3844 // exit 4
3845 //
3846 //
3847 // This method creates a list (_required_safept) of ncsfpt nodes that must
3848 // be protected is created for each loop. When a ncsfpt maybe deleted, it
3849 // is first looked for in the lists for the outer loops of the current loop.
3850 //
3851 // The insights into the problem:
3852 // A) counted loops are okay
3853 // B) innermost loops are okay (only an inner loop can delete
3854 // a ncsfpt needed by an outer loop)
3855 // C) a loop is immune from an inner loop deleting a safepoint
3856 // if the loop has a call on the idom-path
3857 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
3858 // idom-path that is not in a nested loop
3859 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
3860 // loop needs to be prevented from deletion by an inner loop
3861 //
3862 // There are two analyses:
3863 // 1) The first, and cheaper one, scans the loop body from
3864 // tail to head following the idom (immediate dominator)
3865 // chain, looking for the cases (C,D,E) above.
3866 // Since inner loops are scanned before outer loops, there is summary
3867 // information about inner loops. Inner loops can be skipped over
3868 // when the tail of an inner loop is encountered.
3869 //
3870 // 2) The second, invoked if the first fails to find a call or ncsfpt on
3871 // the idom path (which is rare), scans all predecessor control paths
3872 // from the tail to the head, terminating a path when a call or sfpt
3873 // is encountered, to find the ncsfpt's that are closest to the tail.
3874 //
3875 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
3876 // Bottom up traversal
3877 IdealLoopTree* ch = _child;
3878 if (_child) _child->check_safepts(visited, stack);
3879 if (_next) _next ->check_safepts(visited, stack);
3880
3881 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
3882 bool has_call = false; // call on dom-path
3883 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
3884 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
3885 if (!_irreducible) {
3886 // Scan the dom-path nodes from tail to head
3887 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
3888 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3889 has_call = true;
3890 _has_sfpt = 1; // Then no need for a safept!
3891 break;
3892 } else if (n->Opcode() == Op_SafePoint) {
3893 if (_phase->get_loop(n) == this) {
3894 has_local_ncsfpt = true;
3895 break;
3896 }
3897 if (nonlocal_ncsfpt == nullptr) {
3898 nonlocal_ncsfpt = n; // save the one closest to the tail
3899 }
3900 } else {
3901 IdealLoopTree* nlpt = _phase->get_loop(n);
3902 if (this != nlpt) {
3903 // If at an inner loop tail, see if the inner loop has already
3904 // recorded seeing a call on the dom-path (and stop.) If not,
3905 // jump to the head of the inner loop.
3906 assert(is_member(nlpt), "nested loop");
3907 Node* tail = nlpt->_tail;
3908 if (tail->in(0)->is_If()) tail = tail->in(0);
3909 if (n == tail) {
3910 // If inner loop has call on dom-path, so does outer loop
3911 if (nlpt->_has_sfpt) {
3912 has_call = true;
3913 _has_sfpt = 1;
3914 break;
3915 }
3916 // Skip to head of inner loop
3917 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
3918 n = nlpt->_head;
3919 if (_head == n) {
3920 // this and nlpt (inner loop) have the same loop head. This should not happen because
3921 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
3922 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
3923 // but then attached in the build_loop_tree afterwards, and so still have unmerged
3924 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
3925 // since we have reached the loop head of this.
3926 assert(_head->as_Region()->is_in_infinite_subgraph(),
3927 "only expect unmerged backedges in infinite loops");
3928 break;
3929 }
3930 }
3931 }
3932 }
3933 }
3934 }
3935 // Record safept's that this loop needs preserved when an
3936 // inner loop attempts to delete it's safepoints.
3937 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
3938 if (nonlocal_ncsfpt != nullptr) {
3939 if (_required_safept == nullptr) _required_safept = new Node_List();
3940 _required_safept->push(nonlocal_ncsfpt);
3941 } else {
3942 // Failed to find a suitable safept on the dom-path. Now use
3943 // an all paths walk from tail to head, looking for safepoints to preserve.
3944 allpaths_check_safepts(visited, stack);
3945 }
3946 }
3947 }
3948 }
3949
3950 //---------------------------is_deleteable_safept----------------------------
3951 // Is safept not required by an outer loop?
3952 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
3953 assert(sfpt->Opcode() == Op_SafePoint, "");
3954 IdealLoopTree* lp = get_loop(sfpt)->_parent;
3955 while (lp != nullptr) {
3956 Node_List* sfpts = lp->_required_safept;
3957 if (sfpts != nullptr) {
3958 for (uint i = 0; i < sfpts->size(); i++) {
3959 if (sfpt == sfpts->at(i))
3960 return false;
3961 }
3962 }
3963 lp = lp->_parent;
3964 }
3965 return true;
3966 }
3967
3968 //---------------------------replace_parallel_iv-------------------------------
3969 // Replace parallel induction variable (parallel to trip counter)
3970 // This optimization looks for patterns similar to:
3971 //
3972 // int a = init2;
3973 // for (int iv = init; iv < limit; iv += stride_con) {
3974 // a += stride_con2;
3975 // }
3976 //
3977 // and transforms it to:
3978 //
3979 // int iv2 = init2
3980 // int iv = init
3981 // loop:
3982 // if (iv >= limit) goto exit
3983 // iv += stride_con
3984 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
3985 // goto loop
3986 // exit:
3987 // ...
3988 //
3989 // Such transformation introduces more optimization opportunities. In this
3990 // particular example, the loop can be eliminated entirely given that
3991 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
3992 // place to only perform this optimization if such a division is exact. This
3993 // example will be transformed into its semantic equivalence:
3994 //
3995 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
3996 //
3997 // which corresponds to the structure of transformed subgraph.
3998 //
3999 // However, if there is a mismatch between types of the loop and the parallel
4000 // induction variable (e.g., a long-typed IV in an int-typed loop), type
4001 // conversions are required:
4002 //
4003 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
4004 //
4005 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
4006 assert(loop->_head->is_CountedLoop(), "");
4007 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4008 if (!cl->is_valid_counted_loop(T_INT)) {
4009 return; // skip malformed counted loop
4010 }
4011 Node *incr = cl->incr();
4012 if (incr == nullptr) {
4013 return; // Dead loop?
4014 }
4015 Node *init = cl->init_trip();
4016 Node *phi = cl->phi();
4017 jlong stride_con = cl->stride_con();
4018
4019 // Visit all children, looking for Phis
4020 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
4021 Node *out = cl->out(i);
4022 // Look for other phis (secondary IVs). Skip dead ones
4023 if (!out->is_Phi() || out == phi || !has_node(out)) {
4024 continue;
4025 }
4026
4027 PhiNode* phi2 = out->as_Phi();
4028 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4029 // Look for induction variables of the form: X += constant
4030 if (phi2->region() != loop->_head ||
4031 incr2->req() != 3 ||
4032 incr2->in(1)->uncast() != phi2 ||
4033 incr2 == incr ||
4034 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4035 !incr2->in(2)->is_Con()) {
4036 continue;
4037 }
4038
4039 if (incr2->in(1)->is_ConstraintCast() &&
4040 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4041 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4042 continue;
4043 }
4044 // Check for parallel induction variable (parallel to trip counter)
4045 // via an affine function. In particular, count-down loops with
4046 // count-up array indices are common. We only RCE references off
4047 // the trip-counter, so we need to convert all these to trip-counter
4048 // expressions.
4049 Node* init2 = phi2->in(LoopNode::EntryControl);
4050
4051 // Determine the basic type of the stride constant (and the iv being incremented).
4052 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4053 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4054
4055 // The ratio of the two strides cannot be represented as an int
4056 // if stride_con2 is min_jint (or min_jlong, respectively) and
4057 // stride_con is -1.
4058 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4059 continue;
4060 }
4061
4062 // The general case here gets a little tricky. We want to find the
4063 // GCD of all possible parallel IV's and make a new IV using this
4064 // GCD for the loop. Then all possible IVs are simple multiples of
4065 // the GCD. In practice, this will cover very few extra loops.
4066 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4067 // where +/-1 is the common case, but other integer multiples are
4068 // also easy to handle.
4069 jlong ratio_con = stride_con2 / stride_con;
4070
4071 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4072 continue;
4073 }
4074
4075 #ifndef PRODUCT
4076 if (TraceLoopOpts) {
4077 tty->print("Parallel IV: %d ", phi2->_idx);
4078 loop->dump_head();
4079 }
4080 #endif
4081
4082 // Convert to using the trip counter. The parallel induction
4083 // variable differs from the trip counter by a loop-invariant
4084 // amount, the difference between their respective initial values.
4085 // It is scaled by the 'ratio_con'.
4086 Node* ratio = integercon(ratio_con, stride_con2_bt);
4087
4088 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4089 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4090
4091 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4092 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4093 set_early_ctrl(ratio_init, false);
4094
4095 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4096 _igvn.register_new_node_with_optimizer(diff, init2);
4097 set_early_ctrl(diff, false);
4098
4099 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4100 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4101 set_ctrl(ratio_idx, cl);
4102
4103 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4104 _igvn.register_new_node_with_optimizer(add);
4105 set_ctrl(add, cl);
4106
4107 _igvn.replace_node( phi2, add );
4108 // Sometimes an induction variable is unused
4109 if (add->outcnt() == 0) {
4110 _igvn.remove_dead_node(add);
4111 }
4112 --i; // deleted this phi; rescan starting with next position
4113 }
4114 }
4115
4116 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4117 BasicType source = _igvn.type(input)->basic_type();
4118 if (source == target) {
4119 return input;
4120 }
4121
4122 Node* converted = ConvertNode::create_convert(source, target, input);
4123 _igvn.register_new_node_with_optimizer(converted, input);
4124 set_early_ctrl(converted, false);
4125
4126 return converted;
4127 }
4128
4129 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4130 Node* keep = nullptr;
4131 if (keep_one) {
4132 // Look for a safepoint on the idom-path.
4133 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4134 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4135 keep = i;
4136 break; // Found one
4137 }
4138 }
4139 }
4140
4141 // Don't remove any safepoints if it is requested to keep a single safepoint and
4142 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4143 // in this case since there's no safepoint dominating all paths in the loop body.
4144 bool prune = !keep_one || keep != nullptr;
4145
4146 // Delete other safepoints in this loop.
4147 Node_List* sfpts = _safepts;
4148 if (prune && sfpts != nullptr) {
4149 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4150 for (uint i = 0; i < sfpts->size(); i++) {
4151 Node* n = sfpts->at(i);
4152 assert(phase->get_loop(n) == this, "");
4153 if (n != keep && phase->is_deleteable_safept(n)) {
4154 phase->lazy_replace(n, n->in(TypeFunc::Control));
4155 }
4156 }
4157 }
4158 }
4159
4160 //------------------------------counted_loop-----------------------------------
4161 // Convert to counted loops where possible
4162 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4163
4164 // For grins, set the inner-loop flag here
4165 if (!_child) {
4166 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4167 }
4168
4169 IdealLoopTree* loop = this;
4170 if (_head->is_CountedLoop() ||
4171 phase->is_counted_loop(_head, loop, T_INT)) {
4172
4173 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4174 // Indicate we do not need a safepoint here
4175 _has_sfpt = 1;
4176 }
4177
4178 // Remove safepoints
4179 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4180 remove_safepoints(phase, keep_one_sfpt);
4181
4182 // Look for induction variables
4183 phase->replace_parallel_iv(this);
4184 } else if (_head->is_LongCountedLoop() ||
4185 phase->is_counted_loop(_head, loop, T_LONG)) {
4186 remove_safepoints(phase, true);
4187 } else {
4188 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4189 if (_parent != nullptr && !_irreducible) {
4190 // Not a counted loop. Keep one safepoint.
4191 bool keep_one_sfpt = true;
4192 remove_safepoints(phase, keep_one_sfpt);
4193 }
4194 }
4195
4196 // Recursively
4197 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4198 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4199 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4200 if (loop->_next) loop->_next ->counted_loop(phase);
4201 }
4202
4203
4204 // The Estimated Loop Clone Size:
4205 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4206 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4207 // respectively, used to ensure that the node usage estimates made are on the
4208 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4209 // possible additional/excessive nodes generated due to data and control flow
4210 // merging, for edges reaching outside the loop.
4211 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4212
4213 precond(0 < factor && factor < 16);
4214
4215 uint const bc = 13;
4216 uint const cc = 17;
4217 uint const sz = _body.size() + (_body.size() + 7) / 2;
4218 uint estimate = factor * (sz + bc) + cc;
4219
4220 assert((estimate - cc) / factor == sz + bc, "overflow");
4221
4222 return estimate + est_loop_flow_merge_sz();
4223 }
4224
4225 // The Estimated Loop (full-) Unroll Size:
4226 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4227 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4228 // node usage estimates made are on the safe side, for the most part. This is
4229 // a "light" version of the loop clone size calculation (above), based on the
4230 // assumption that most of the loop-construct overhead will be unraveled when
4231 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4232 // including an overflow check and returning UINT_MAX in case of an overflow.
4233 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4234
4235 precond(factor > 0);
4236
4237 // Take into account that after unroll conjoined heads and tails will fold.
4238 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4239 uint const cc = 7;
4240 uint const sz = b0 + (b0 + 15) / 16;
4241 uint estimate = factor * sz + cc;
4242
4243 if ((estimate - cc) / factor != sz) {
4244 return UINT_MAX;
4245 }
4246
4247 return estimate + est_loop_flow_merge_sz();
4248 }
4249
4250 // Estimate the growth effect (in nodes) of merging control and data flow when
4251 // cloning a loop body, based on the amount of control and data flow reaching
4252 // outside of the (current) loop body.
4253 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4254
4255 uint ctrl_edge_out_cnt = 0;
4256 uint data_edge_out_cnt = 0;
4257
4258 for (uint i = 0; i < _body.size(); i++) {
4259 Node* node = _body.at(i);
4260 uint outcnt = node->outcnt();
4261
4262 for (uint k = 0; k < outcnt; k++) {
4263 Node* out = node->raw_out(k);
4264 if (out == nullptr) continue;
4265 if (out->is_CFG()) {
4266 if (!is_member(_phase->get_loop(out))) {
4267 ctrl_edge_out_cnt++;
4268 }
4269 } else if (_phase->has_ctrl(out)) {
4270 Node* ctrl = _phase->get_ctrl(out);
4271 assert(ctrl != nullptr, "must be");
4272 assert(ctrl->is_CFG(), "must be");
4273 if (!is_member(_phase->get_loop(ctrl))) {
4274 data_edge_out_cnt++;
4275 }
4276 }
4277 }
4278 }
4279 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4280 // a rather pessimistic estimate for the most part, in particular for some
4281 // complex loops, but still not enough to capture all loops.
4282 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4283 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4284 }
4285 return 0;
4286 }
4287
4288 #ifndef PRODUCT
4289 //------------------------------dump_head--------------------------------------
4290 // Dump 1 liner for loop header info
4291 void IdealLoopTree::dump_head() {
4292 tty->sp(2 * _nest);
4293 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4294 if (_irreducible) tty->print(" IRREDUCIBLE");
4295 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4296 : _head->in(LoopNode::EntryControl);
4297 const Predicates predicates(entry);
4298 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4299 tty->print(" limit_check");
4300 }
4301 if (UseLoopPredicate) {
4302 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4303 tty->print(" profile_predicated");
4304 }
4305 if (predicates.loop_predicate_block()->is_non_empty()) {
4306 tty->print(" predicated");
4307 }
4308 }
4309 if (_head->is_CountedLoop()) {
4310 CountedLoopNode *cl = _head->as_CountedLoop();
4311 tty->print(" counted");
4312
4313 Node* init_n = cl->init_trip();
4314 if (init_n != nullptr && init_n->is_Con())
4315 tty->print(" [%d,", cl->init_trip()->get_int());
4316 else
4317 tty->print(" [int,");
4318 Node* limit_n = cl->limit();
4319 if (limit_n != nullptr && limit_n->is_Con())
4320 tty->print("%d),", cl->limit()->get_int());
4321 else
4322 tty->print("int),");
4323 int stride_con = cl->stride_con();
4324 if (stride_con > 0) tty->print("+");
4325 tty->print("%d", stride_con);
4326
4327 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4328
4329 if (cl->is_pre_loop ()) tty->print(" pre" );
4330 if (cl->is_main_loop()) tty->print(" main");
4331 if (cl->is_post_loop()) tty->print(" post");
4332 if (cl->is_vectorized_loop()) tty->print(" vector");
4333 if (range_checks_present()) tty->print(" rc ");
4334 if (cl->is_multiversion_fast_loop()) { tty->print(" multiversion_fast"); }
4335 if (cl->is_multiversion_slow_loop()) { tty->print(" multiversion_slow"); }
4336 if (cl->is_multiversion_delayed_slow_loop()) { tty->print(" multiversion_delayed_slow"); }
4337 }
4338 if (_has_call) tty->print(" has_call");
4339 if (_has_sfpt) tty->print(" has_sfpt");
4340 if (_rce_candidate) tty->print(" rce");
4341 if (_safepts != nullptr && _safepts->size() > 0) {
4342 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4343 }
4344 if (_required_safept != nullptr && _required_safept->size() > 0) {
4345 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4346 }
4347 if (Verbose) {
4348 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4349 }
4350 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4351 tty->print(" strip_mined");
4352 }
4353 tty->cr();
4354 }
4355
4356 //------------------------------dump-------------------------------------------
4357 // Dump loops by loop tree
4358 void IdealLoopTree::dump() {
4359 dump_head();
4360 if (_child) _child->dump();
4361 if (_next) _next ->dump();
4362 }
4363
4364 #endif
4365
4366 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4367 if (loop == root) {
4368 if (loop->_child != nullptr) {
4369 log->begin_head("loop_tree");
4370 log->end_head();
4371 log_loop_tree_helper(root, loop->_child, log);
4372 log->tail("loop_tree");
4373 assert(loop->_next == nullptr, "what?");
4374 }
4375 } else if (loop != nullptr) {
4376 Node* head = loop->_head;
4377 log->begin_head("loop");
4378 log->print(" idx='%d' ", head->_idx);
4379 if (loop->_irreducible) log->print("irreducible='1' ");
4380 if (head->is_Loop()) {
4381 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4382 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4383 } else if (head->is_CountedLoop()) {
4384 CountedLoopNode* cl = head->as_CountedLoop();
4385 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4386 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4387 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4388 }
4389 log->end_head();
4390 log_loop_tree_helper(root, loop->_child, log);
4391 log->tail("loop");
4392 log_loop_tree_helper(root, loop->_next, log);
4393 }
4394 }
4395
4396 void PhaseIdealLoop::log_loop_tree() {
4397 if (C->log() != nullptr) {
4398 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4399 }
4400 }
4401
4402 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4403 // predicates will be removed during the next round of IGVN.
4404 void PhaseIdealLoop::eliminate_useless_predicates() const {
4405 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4406 return; // No predicates left.
4407 }
4408
4409 EliminateUselessPredicates eliminate_useless_predicates(_igvn, _ltree_root);
4410 eliminate_useless_predicates.eliminate();
4411 }
4412
4413 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4414 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4415 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4416 return;
4417 }
4418 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4419 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4420 IdealLoopTree* lpt = iter.current();
4421 if (lpt->_child == nullptr && lpt->is_counted()) {
4422 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4423 Node* opaque = head->is_canonical_loop_entry();
4424 if (opaque != nullptr) {
4425 useful_zero_trip_guard_opaques_nodes.push(opaque);
4426 }
4427 }
4428 }
4429 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4430 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4431 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4432 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4433 IfNode* iff = opaque->if_node();
4434 IdealLoopTree* loop = get_loop(iff);
4435 while (loop != _ltree_root && loop != nullptr) {
4436 loop = loop->_parent;
4437 }
4438 if (loop == nullptr) {
4439 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4440 // We can't tell if the opaque node is useful or not
4441 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4442 } else {
4443 assert(guarded_loop == nullptr, "");
4444 this->_igvn.replace_node(opaque, opaque->in(1));
4445 }
4446 } else {
4447 assert(guarded_loop != nullptr, "");
4448 }
4449 }
4450 }
4451
4452 void PhaseIdealLoop::eliminate_useless_multiversion_if() {
4453 if (_multiversion_opaque_nodes.size() == 0) {
4454 return;
4455 }
4456
4457 ResourceMark rm;
4458 Unique_Node_List useful_multiversioning_opaque_nodes;
4459
4460 // The OpaqueMultiversioning is only used from the fast main loop in AutoVectorization, to add
4461 // speculative runtime-checks to the multiversion_if. Thus, a OpaqueMultiversioning is only
4462 // useful if it can be found from a fast main loop. If it can not be found from a fast main loop,
4463 // then we cannot ever use that multiversion_if to add more speculative runtime-checks, and hence
4464 // it is useless. If it is still in delayed mode, i.e. has not yet had any runtime-checks added,
4465 // then we can let it constant fold towards the fast loop.
4466 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4467 IdealLoopTree* lpt = iter.current();
4468 if (lpt->_child == nullptr && lpt->is_counted()) {
4469 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4470 if (head->is_main_loop() && head->is_multiversion_fast_loop()) {
4471 // There are fast_loop pre/main/post loops, but the finding traversal starts at the main
4472 // loop, and traverses via the fast pre loop to the multiversion_if.
4473 IfNode* multiversion_if = head->find_multiversion_if_from_multiversion_fast_main_loop();
4474 if (multiversion_if != nullptr) {
4475 useful_multiversioning_opaque_nodes.push(multiversion_if->in(1)->as_OpaqueMultiversioning());
4476 } else {
4477 // We could not find the multiversion_if, and would never find it again. Remove the
4478 // multiversion marking for consistency.
4479 head->set_no_multiversion();
4480 }
4481 }
4482 }
4483 }
4484
4485 for (uint i = 0; i < _multiversion_opaque_nodes.size(); i++) {
4486 OpaqueMultiversioningNode* opaque = _multiversion_opaque_nodes.at(i)->as_OpaqueMultiversioning();
4487 if (!useful_multiversioning_opaque_nodes.member(opaque)) {
4488 if (opaque->is_delayed_slow_loop()) {
4489 // We cannot hack the node directly, otherwise the slow_loop will complain that it cannot
4490 // find the multiversioning opaque node. Instead, we mark the opaque node as useless, and
4491 // it can be constant folded during IGVN.
4492 opaque->mark_useless(_igvn);
4493 }
4494 }
4495 }
4496 }
4497
4498 //------------------------process_expensive_nodes-----------------------------
4499 // Expensive nodes have their control input set to prevent the GVN
4500 // from commoning them and as a result forcing the resulting node to
4501 // be in a more frequent path. Use CFG information here, to change the
4502 // control inputs so that some expensive nodes can be commoned while
4503 // not executed more frequently.
4504 bool PhaseIdealLoop::process_expensive_nodes() {
4505 assert(OptimizeExpensiveOps, "optimization off?");
4506
4507 // Sort nodes to bring similar nodes together
4508 C->sort_expensive_nodes();
4509
4510 bool progress = false;
4511
4512 for (int i = 0; i < C->expensive_count(); ) {
4513 Node* n = C->expensive_node(i);
4514 int start = i;
4515 // Find nodes similar to n
4516 i++;
4517 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4518 int end = i;
4519 // And compare them two by two
4520 for (int j = start; j < end; j++) {
4521 Node* n1 = C->expensive_node(j);
4522 if (is_node_unreachable(n1)) {
4523 continue;
4524 }
4525 for (int k = j+1; k < end; k++) {
4526 Node* n2 = C->expensive_node(k);
4527 if (is_node_unreachable(n2)) {
4528 continue;
4529 }
4530
4531 assert(n1 != n2, "should be pair of nodes");
4532
4533 Node* c1 = n1->in(0);
4534 Node* c2 = n2->in(0);
4535
4536 Node* parent_c1 = c1;
4537 Node* parent_c2 = c2;
4538
4539 // The call to get_early_ctrl_for_expensive() moves the
4540 // expensive nodes up but stops at loops that are in a if
4541 // branch. See whether we can exit the loop and move above the
4542 // If.
4543 if (c1->is_Loop()) {
4544 parent_c1 = c1->in(1);
4545 }
4546 if (c2->is_Loop()) {
4547 parent_c2 = c2->in(1);
4548 }
4549
4550 if (parent_c1 == parent_c2) {
4551 _igvn._worklist.push(n1);
4552 _igvn._worklist.push(n2);
4553 continue;
4554 }
4555
4556 // Look for identical expensive node up the dominator chain.
4557 if (is_dominator(c1, c2)) {
4558 c2 = c1;
4559 } else if (is_dominator(c2, c1)) {
4560 c1 = c2;
4561 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4562 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4563 // Both branches have the same expensive node so move it up
4564 // before the if.
4565 c1 = c2 = idom(parent_c1->in(0));
4566 }
4567 // Do the actual moves
4568 if (n1->in(0) != c1) {
4569 _igvn.replace_input_of(n1, 0, c1);
4570 progress = true;
4571 }
4572 if (n2->in(0) != c2) {
4573 _igvn.replace_input_of(n2, 0, c2);
4574 progress = true;
4575 }
4576 }
4577 }
4578 }
4579
4580 return progress;
4581 }
4582
4583 //=============================================================================
4584 //----------------------------build_and_optimize-------------------------------
4585 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4586 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4587 void PhaseIdealLoop::build_and_optimize() {
4588 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4589
4590 bool do_split_ifs = (_mode == LoopOptsDefault);
4591 bool skip_loop_opts = (_mode == LoopOptsNone);
4592 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4593
4594
4595 int old_progress = C->major_progress();
4596 uint orig_worklist_size = _igvn._worklist.size();
4597
4598 // Reset major-progress flag for the driver's heuristics
4599 C->clear_major_progress();
4600
4601 #ifndef PRODUCT
4602 // Capture for later assert
4603 uint unique = C->unique();
4604 _loop_invokes++;
4605 _loop_work += unique;
4606 #endif
4607
4608 // True if the method has at least 1 irreducible loop
4609 _has_irreducible_loops = false;
4610
4611 _created_loop_node = false;
4612
4613 VectorSet visited;
4614 // Pre-grow the mapping from Nodes to IdealLoopTrees.
4615 _loop_or_ctrl.map(C->unique(), nullptr);
4616 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
4617
4618 // Pre-build the top-level outermost loop tree entry
4619 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
4620 // Do not need a safepoint at the top level
4621 _ltree_root->_has_sfpt = 1;
4622
4623 // Initialize Dominators.
4624 // Checked in clone_loop_predicate() during beautify_loops().
4625 _idom_size = 0;
4626 _idom = nullptr;
4627 _dom_depth = nullptr;
4628 _dom_stk = nullptr;
4629
4630 // Empty pre-order array
4631 allocate_preorders();
4632
4633 // Build a loop tree on the fly. Build a mapping from CFG nodes to
4634 // IdealLoopTree entries. Data nodes are NOT walked.
4635 build_loop_tree();
4636 // Check for bailout, and return
4637 if (C->failing()) {
4638 return;
4639 }
4640
4641 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
4642 // is an exception edge, parsing doesn't set has_loops().
4643 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
4644 // No loops after all
4645 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
4646
4647 // There should always be an outer loop containing the Root and Return nodes.
4648 // If not, we have a degenerate empty program. Bail out in this case.
4649 if (!has_node(C->root())) {
4650 if (!_verify_only) {
4651 C->clear_major_progress();
4652 assert(false, "empty program detected during loop optimization");
4653 C->record_method_not_compilable("empty program detected during loop optimization");
4654 }
4655 return;
4656 }
4657
4658 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4659 // Nothing to do, so get out
4660 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
4661 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
4662 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
4663 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
4664 if (stop_early && !do_expensive_nodes) {
4665 return;
4666 }
4667
4668 // Set loop nesting depth
4669 _ltree_root->set_nest( 0 );
4670
4671 // Split shared headers and insert loop landing pads.
4672 // Do not bother doing this on the Root loop of course.
4673 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
4674 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
4675 if( _ltree_root->_child->beautify_loops( this ) ) {
4676 // Re-build loop tree!
4677 _ltree_root->_child = nullptr;
4678 _loop_or_ctrl.clear();
4679 reallocate_preorders();
4680 build_loop_tree();
4681 // Check for bailout, and return
4682 if (C->failing()) {
4683 return;
4684 }
4685 // Reset loop nesting depth
4686 _ltree_root->set_nest( 0 );
4687
4688 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
4689 }
4690 }
4691
4692 // Build Dominators for elision of null checks & loop finding.
4693 // Since nodes do not have a slot for immediate dominator, make
4694 // a persistent side array for that info indexed on node->_idx.
4695 _idom_size = C->unique();
4696 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
4697 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
4698 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
4699 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
4700
4701 Dominators();
4702
4703 if (!_verify_only) {
4704 // As a side effect, Dominators removed any unreachable CFG paths
4705 // into RegionNodes. It doesn't do this test against Root, so
4706 // we do it here.
4707 for( uint i = 1; i < C->root()->req(); i++ ) {
4708 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
4709 _igvn.delete_input_of(C->root(), i);
4710 i--; // Rerun same iteration on compressed edges
4711 }
4712 }
4713
4714 // Given dominators, try to find inner loops with calls that must
4715 // always be executed (call dominates loop tail). These loops do
4716 // not need a separate safepoint.
4717 Node_List cisstack;
4718 _ltree_root->check_safepts(visited, cisstack);
4719 }
4720
4721 // Walk the DATA nodes and place into loops. Find earliest control
4722 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
4723 // holding the associated IdealLoopTree pointer. For DATA nodes, the
4724 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
4725
4726 // Allocate stack with enough space to avoid frequent realloc
4727 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
4728 Node_Stack nstack(stack_size);
4729
4730 visited.clear();
4731 Node_List worklist;
4732 // Don't need C->root() on worklist since
4733 // it will be processed among C->top() inputs
4734 worklist.push(C->top());
4735 visited.set(C->top()->_idx); // Set C->top() as visited now
4736 build_loop_early( visited, worklist, nstack );
4737
4738 // Given early legal placement, try finding counted loops. This placement
4739 // is good enough to discover most loop invariants.
4740 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
4741 _ltree_root->counted_loop( this );
4742 }
4743
4744 // Find latest loop placement. Find ideal loop placement.
4745 visited.clear();
4746 init_dom_lca_tags();
4747 // Need C->root() on worklist when processing outs
4748 worklist.push(C->root());
4749 NOT_PRODUCT( C->verify_graph_edges(); )
4750 worklist.push(C->top());
4751 build_loop_late( visited, worklist, nstack );
4752 if (C->failing()) { return; }
4753
4754 if (_verify_only) {
4755 C->restore_major_progress(old_progress);
4756 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
4757 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
4758 return;
4759 }
4760
4761 // clear out the dead code after build_loop_late
4762 while (_deadlist.size()) {
4763 _igvn.remove_globally_dead_node(_deadlist.pop());
4764 }
4765
4766 eliminate_useless_zero_trip_guard();
4767 eliminate_useless_multiversion_if();
4768
4769 if (stop_early) {
4770 assert(do_expensive_nodes, "why are we here?");
4771 if (process_expensive_nodes()) {
4772 // If we made some progress when processing expensive nodes then
4773 // the IGVN may modify the graph in a way that will allow us to
4774 // make some more progress: we need to try processing expensive
4775 // nodes again.
4776 C->set_major_progress();
4777 }
4778 return;
4779 }
4780
4781 // Some parser-inserted loop predicates could never be used by loop
4782 // predication or they were moved away from loop during some optimizations.
4783 // For example, peeling. Eliminate them before next loop optimizations.
4784 eliminate_useless_predicates();
4785
4786 #ifndef PRODUCT
4787 C->verify_graph_edges();
4788 if (_verify_me) { // Nested verify pass?
4789 // Check to see if the verify _mode is broken
4790 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
4791 return;
4792 }
4793 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4794 if (TraceLoopOpts && C->has_loops()) {
4795 _ltree_root->dump();
4796 }
4797 #endif
4798
4799 if (skip_loop_opts) {
4800 C->restore_major_progress(old_progress);
4801 return;
4802 }
4803
4804 if (do_max_unroll) {
4805 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4806 IdealLoopTree* lpt = iter.current();
4807 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
4808 lpt->compute_trip_count(this);
4809 if (!lpt->do_one_iteration_loop(this) &&
4810 !lpt->do_remove_empty_loop(this)) {
4811 AutoNodeBudget node_budget(this);
4812 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
4813 lpt->policy_maximally_unroll(this)) {
4814 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4815 do_maximally_unroll(lpt, worklist);
4816 }
4817 }
4818 }
4819 }
4820
4821 C->restore_major_progress(old_progress);
4822 return;
4823 }
4824
4825 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
4826 return;
4827 }
4828
4829 if (ReassociateInvariants && !C->major_progress()) {
4830 // Reassociate invariants and prep for split_thru_phi
4831 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4832 IdealLoopTree* lpt = iter.current();
4833 if (!lpt->is_loop()) {
4834 continue;
4835 }
4836 Node* head = lpt->_head;
4837 if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
4838
4839 // check for vectorized loops, any reassociation of invariants was already done
4840 if (head->is_CountedLoop()) {
4841 if (head->as_CountedLoop()->is_unroll_only()) {
4842 continue;
4843 } else {
4844 AutoNodeBudget node_budget(this);
4845 lpt->reassociate_invariants(this);
4846 }
4847 }
4848 // Because RCE opportunities can be masked by split_thru_phi,
4849 // look for RCE candidates and inhibit split_thru_phi
4850 // on just their loop-phi's for this pass of loop opts
4851 if (SplitIfBlocks && do_split_ifs &&
4852 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
4853 (lpt->policy_range_check(this, true, T_LONG) ||
4854 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
4855 lpt->_rce_candidate = 1; // = true
4856 }
4857 }
4858 }
4859
4860 // Check for aggressive application of split-if and other transforms
4861 // that require basic-block info (like cloning through Phi's)
4862 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
4863 visited.clear();
4864 split_if_with_blocks(visited, nstack);
4865 if (C->failing()) {
4866 return;
4867 }
4868 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4869 }
4870
4871 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
4872 C->set_major_progress();
4873 }
4874
4875 // Perform loop predication before iteration splitting
4876 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
4877 _ltree_root->_child->loop_predication(this);
4878 }
4879
4880 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
4881 if (do_intrinsify_fill()) {
4882 C->set_major_progress();
4883 }
4884 }
4885
4886 // Perform iteration-splitting on inner loops. Split iterations to avoid
4887 // range checks or one-shot null checks.
4888
4889 // If split-if's didn't hack the graph too bad (no CFG changes)
4890 // then do loop opts.
4891 if (C->has_loops() && !C->major_progress()) {
4892 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4893 _ltree_root->_child->iteration_split( this, worklist );
4894 // No verify after peeling! GCM has hoisted code out of the loop.
4895 // After peeling, the hoisted code could sink inside the peeled area.
4896 // The peeling code does not try to recompute the best location for
4897 // all the code before the peeled area, so the verify pass will always
4898 // complain about it.
4899 }
4900
4901 // Check for bailout, and return
4902 if (C->failing()) {
4903 return;
4904 }
4905
4906 // Do verify graph edges in any case
4907 NOT_PRODUCT( C->verify_graph_edges(); );
4908
4909 if (!do_split_ifs) {
4910 // We saw major progress in Split-If to get here. We forced a
4911 // pass with unrolling and not split-if, however more split-if's
4912 // might make progress. If the unrolling didn't make progress
4913 // then the major-progress flag got cleared and we won't try
4914 // another round of Split-If. In particular the ever-common
4915 // instance-of/check-cast pattern requires at least 2 rounds of
4916 // Split-If to clear out.
4917 C->set_major_progress();
4918 }
4919
4920 // Repeat loop optimizations if new loops were seen
4921 if (created_loop_node()) {
4922 C->set_major_progress();
4923 }
4924
4925 // Auto-vectorize main-loop
4926 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
4927 Compile::TracePhase tp(_t_autoVectorize);
4928
4929 // Shared data structures for all AutoVectorizations, to reduce allocations
4930 // of large arrays.
4931 VSharedData vshared;
4932 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4933 IdealLoopTree* lpt = iter.current();
4934 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
4935
4936 if (status == AutoVectorizeStatus::TriedAndFailed) {
4937 // We tried vectorization, but failed. From now on only unroll the loop.
4938 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4939 if (cl->has_passed_slp()) {
4940 C->set_major_progress();
4941 cl->set_notpassed_slp();
4942 cl->mark_do_unroll_only();
4943 }
4944 }
4945 }
4946 }
4947
4948 // Move UnorderedReduction out of counted loop. Can be introduced by AutoVectorization.
4949 if (C->has_loops() && !C->major_progress()) {
4950 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4951 IdealLoopTree* lpt = iter.current();
4952 if (lpt->is_counted() && lpt->is_innermost()) {
4953 move_unordered_reduction_out_of_loop(lpt);
4954 }
4955 }
4956 }
4957
4958 // Keep loop predicates and perform optimizations with them
4959 // until no more loop optimizations could be done.
4960 // After that switch predicates off and do more loop optimizations.
4961 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
4962 C->mark_parse_predicate_nodes_useless(_igvn);
4963 assert(C->parse_predicate_count() == 0, "should be zero now");
4964 if (TraceLoopOpts) {
4965 tty->print_cr("PredicatesOff");
4966 }
4967 C->set_major_progress();
4968 }
4969 }
4970
4971 #ifndef PRODUCT
4972 //------------------------------print_statistics-------------------------------
4973 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
4974 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
4975 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
4976 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
4977 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
4978 void PhaseIdealLoop::print_statistics() {
4979 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d, long loops=%d/%d/%d", _loop_invokes, _loop_work, _long_loop_counted_loops, _long_loop_nests, _long_loop_candidates);
4980 }
4981 #endif
4982
4983 #ifdef ASSERT
4984 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
4985 void PhaseIdealLoop::verify() const {
4986 ResourceMark rm;
4987 int old_progress = C->major_progress();
4988 bool success = true;
4989
4990 PhaseIdealLoop phase_verify(_igvn, this);
4991 if (C->failing_internal()) {
4992 return;
4993 }
4994
4995 // Verify ctrl and idom of every node.
4996 success &= verify_idom_and_nodes(C->root(), &phase_verify);
4997
4998 // Verify loop-tree.
4999 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5000
5001 assert(success, "VerifyLoopOptimizations failed");
5002
5003 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5004 C->restore_major_progress(old_progress);
5005 }
5006
5007 // Perform a BFS starting at n, through all inputs.
5008 // Call verify_idom and verify_node on all nodes of BFS traversal.
5009 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5010 Unique_Node_List worklist;
5011 worklist.push(root);
5012 bool success = true;
5013 for (uint i = 0; i < worklist.size(); i++) {
5014 Node* n = worklist.at(i);
5015 // process node
5016 success &= verify_idom(n, phase_verify);
5017 success &= verify_loop_ctrl(n, phase_verify);
5018 // visit inputs
5019 for (uint j = 0; j < n->req(); j++) {
5020 if (n->in(j) != nullptr) {
5021 worklist.push(n->in(j));
5022 }
5023 }
5024 }
5025 return success;
5026 }
5027
5028 // Verify dominator structure (IDOM).
5029 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5030 // Verify IDOM for all CFG nodes (except root).
5031 if (!n->is_CFG() || n->is_Root()) {
5032 return true; // pass
5033 }
5034
5035 if (n->_idx >= _idom_size) {
5036 tty->print("CFG Node with no idom: ");
5037 n->dump();
5038 return false; // fail
5039 }
5040
5041 Node* id = idom_no_update(n);
5042 Node* id_verify = phase_verify->idom_no_update(n);
5043 if (id != id_verify) {
5044 tty->print("Mismatching idom for node: ");
5045 n->dump();
5046 tty->print(" We have idom: ");
5047 id->dump();
5048 tty->print(" Verify has idom: ");
5049 id_verify->dump();
5050 tty->cr();
5051 return false; // fail
5052 }
5053 return true; // pass
5054 }
5055
5056 // Verify "_loop_or_ctrl": control and loop membership.
5057 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5058 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5059 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5060 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5061 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5062 const uint i = n->_idx;
5063 // The loop-tree was built from def to use (top-down).
5064 // The verification happens from use to def (bottom-up).
5065 // We may thus find nodes during verification that are not in the loop-tree.
5066 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5067 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5068 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5069 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5070 n->dump();
5071 return false; // fail
5072 }
5073 // Not reachable for both.
5074 return true; // pass
5075 }
5076
5077 if (n->is_CFG() == has_ctrl(n)) {
5078 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5079 n->dump();
5080 return false; // fail
5081 }
5082
5083 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5084 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5085 n->dump();
5086 return false; // fail
5087 } else if (has_ctrl(n)) {
5088 assert(phase_verify->has_ctrl(n), "sanity");
5089 // n is a data node.
5090 // Verify that its ctrl is the same.
5091
5092 // Broken part of VerifyLoopOptimizations (A)
5093 // Reason:
5094 // BUG, wrong control set for example in
5095 // PhaseIdealLoop::split_if_with_blocks
5096 // at "set_ctrl(x, new_ctrl);"
5097 /*
5098 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5099 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5100 tty->print("Mismatched control setting for: ");
5101 n->dump();
5102 if( fail++ > 10 ) return;
5103 Node *c = get_ctrl_no_update(n);
5104 tty->print("We have it as: ");
5105 if( c->in(0) ) c->dump();
5106 else tty->print_cr("N%d",c->_idx);
5107 tty->print("Verify thinks: ");
5108 if( loop_verify->has_ctrl(n) )
5109 loop_verify->get_ctrl_no_update(n)->dump();
5110 else
5111 loop_verify->get_loop_idx(n)->dump();
5112 tty->cr();
5113 }
5114 */
5115 return true; // pass
5116 } else {
5117 assert(!phase_verify->has_ctrl(n), "sanity");
5118 // n is a ctrl node.
5119 // Verify that not has_ctrl, and that get_loop_idx is the same.
5120
5121 // Broken part of VerifyLoopOptimizations (B)
5122 // Reason:
5123 // NeverBranch node for example is added to loop outside its scope.
5124 // Once we run build_loop_tree again, it is added to the correct loop.
5125 /*
5126 if (!C->major_progress()) {
5127 // Loop selection can be messed up if we did a major progress
5128 // operation, like split-if. Do not verify in that case.
5129 IdealLoopTree *us = get_loop_idx(n);
5130 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5131 if( us->_head != them->_head || us->_tail != them->_tail ) {
5132 tty->print("Unequals loops for: ");
5133 n->dump();
5134 if( fail++ > 10 ) return;
5135 tty->print("We have it as: ");
5136 us->dump();
5137 tty->print("Verify thinks: ");
5138 them->dump();
5139 tty->cr();
5140 }
5141 }
5142 */
5143 return true; // pass
5144 }
5145 }
5146
5147 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5148 assert(a != nullptr && b != nullptr, "must be");
5149 return a->_head->_idx - b->_head->_idx;
5150 }
5151
5152 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5153 GrowableArray<IdealLoopTree*> children;
5154 IdealLoopTree* child = _child;
5155 while (child != nullptr) {
5156 assert(child->_parent == this, "all must be children of this");
5157 children.insert_sorted<compare_tree>(child);
5158 child = child->_next;
5159 }
5160 return children;
5161 }
5162
5163 // Verify that tree structures match. Because the CFG can change, siblings
5164 // within the loop tree can be reordered. We attempt to deal with that by
5165 // reordering the verify's loop tree if possible.
5166 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5167 assert(_head == loop_verify->_head, "mismatched loop head");
5168 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5169
5170 // Collect the children
5171 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5172 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5173
5174 bool success = true;
5175
5176 // Compare the two children lists
5177 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5178 IdealLoopTree* child = nullptr;
5179 IdealLoopTree* child_verify = nullptr;
5180 // Read from both lists, if possible.
5181 if (i < children.length()) {
5182 child = children.at(i);
5183 }
5184 if (j < children_verify.length()) {
5185 child_verify = children_verify.at(j);
5186 }
5187 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5188 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5189 // We found two non-equal children. Select the smaller one.
5190 if (child->_head->_idx < child_verify->_head->_idx) {
5191 child_verify = nullptr;
5192 } else {
5193 child = nullptr;
5194 }
5195 }
5196 // Process the two children, or potentially log the failure if we only found one.
5197 if (child_verify == nullptr) {
5198 if (child->_irreducible && Compile::current()->major_progress()) {
5199 // Irreducible loops can pick a different header (one of its entries).
5200 } else {
5201 tty->print_cr("We have a loop that verify does not have");
5202 child->dump();
5203 success = false;
5204 }
5205 i++; // step for this
5206 } else if (child == nullptr) {
5207 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5208 // Irreducible loops can pick a different header (one of its entries).
5209 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5210 // Infinite loops do not get attached to the loop-tree on their first visit.
5211 // "this" runs before "loop_verify". It is thus possible that we find the
5212 // infinite loop only for "child_verify". Only finding it with "child" would
5213 // mean that we lost it, which is not ok.
5214 } else {
5215 tty->print_cr("Verify has a loop that we do not have");
5216 child_verify->dump();
5217 success = false;
5218 }
5219 j++; // step for verify
5220 } else {
5221 assert(child->_head == child_verify->_head, "We have both and they are equal");
5222 success &= child->verify_tree(child_verify); // Recursion
5223 i++; // step for this
5224 j++; // step for verify
5225 }
5226 }
5227
5228 // Broken part of VerifyLoopOptimizations (D)
5229 // Reason:
5230 // split_if has to update the _tail, if it is modified. But that is done by
5231 // checking to what loop the iff belongs to. That info can be wrong, and then
5232 // we do not update the _tail correctly.
5233 /*
5234 Node *tail = _tail; // Inline a non-updating version of
5235 while( !tail->in(0) ) // the 'tail()' call.
5236 tail = tail->in(1);
5237 assert( tail == loop->_tail, "mismatched loop tail" );
5238 */
5239
5240 if (_head->is_CountedLoop()) {
5241 CountedLoopNode *cl = _head->as_CountedLoop();
5242
5243 Node* ctrl = cl->init_control();
5244 Node* back = cl->back_control();
5245 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5246 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5247 cl->loopexit(); // assert implied
5248 }
5249
5250 // Broken part of VerifyLoopOptimizations (E)
5251 // Reason:
5252 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5253 // to the loop body. Or maybe they are not added to the correct loop.
5254 // at "Node* x = n->clone();"
5255 /*
5256 // Innermost loops need to verify loop bodies,
5257 // but only if no 'major_progress'
5258 int fail = 0;
5259 if (!Compile::current()->major_progress() && _child == nullptr) {
5260 for( uint i = 0; i < _body.size(); i++ ) {
5261 Node *n = _body.at(i);
5262 if (n->outcnt() == 0) continue; // Ignore dead
5263 uint j;
5264 for( j = 0; j < loop->_body.size(); j++ )
5265 if( loop->_body.at(j) == n )
5266 break;
5267 if( j == loop->_body.size() ) { // Not found in loop body
5268 // Last ditch effort to avoid assertion: Its possible that we
5269 // have some users (so outcnt not zero) but are still dead.
5270 // Try to find from root.
5271 if (Compile::current()->root()->find(n->_idx)) {
5272 fail++;
5273 tty->print("We have that verify does not: ");
5274 n->dump();
5275 }
5276 }
5277 }
5278 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5279 Node *n = loop->_body.at(i2);
5280 if (n->outcnt() == 0) continue; // Ignore dead
5281 uint j;
5282 for( j = 0; j < _body.size(); j++ )
5283 if( _body.at(j) == n )
5284 break;
5285 if( j == _body.size() ) { // Not found in loop body
5286 // Last ditch effort to avoid assertion: Its possible that we
5287 // have some users (so outcnt not zero) but are still dead.
5288 // Try to find from root.
5289 if (Compile::current()->root()->find(n->_idx)) {
5290 fail++;
5291 tty->print("Verify has that we do not: ");
5292 n->dump();
5293 }
5294 }
5295 }
5296 assert( !fail, "loop body mismatch" );
5297 }
5298 */
5299 return success;
5300 }
5301 #endif
5302
5303 //------------------------------set_idom---------------------------------------
5304 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5305 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5306 uint idx = d->_idx;
5307 if (idx >= _idom_size) {
5308 uint newsize = next_power_of_2(idx);
5309 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5310 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5311 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5312 _idom_size = newsize;
5313 }
5314 _idom[idx] = n;
5315 _dom_depth[idx] = dom_depth;
5316 }
5317
5318 //------------------------------recompute_dom_depth---------------------------------------
5319 // The dominator tree is constructed with only parent pointers.
5320 // This recomputes the depth in the tree by first tagging all
5321 // nodes as "no depth yet" marker. The next pass then runs up
5322 // the dom tree from each node marked "no depth yet", and computes
5323 // the depth on the way back down.
5324 void PhaseIdealLoop::recompute_dom_depth() {
5325 uint no_depth_marker = C->unique();
5326 uint i;
5327 // Initialize depth to "no depth yet" and realize all lazy updates
5328 for (i = 0; i < _idom_size; i++) {
5329 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5330 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5331 _dom_depth[i] = no_depth_marker;
5332
5333 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5334 if (_idom[i]->in(0) == nullptr) {
5335 idom(i);
5336 }
5337 }
5338 }
5339 if (_dom_stk == nullptr) {
5340 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5341 if (init_size < 10) init_size = 10;
5342 _dom_stk = new GrowableArray<uint>(init_size);
5343 }
5344 // Compute new depth for each node.
5345 for (i = 0; i < _idom_size; i++) {
5346 uint j = i;
5347 // Run up the dom tree to find a node with a depth
5348 while (_dom_depth[j] == no_depth_marker) {
5349 _dom_stk->push(j);
5350 j = _idom[j]->_idx;
5351 }
5352 // Compute the depth on the way back down this tree branch
5353 uint dd = _dom_depth[j] + 1;
5354 while (_dom_stk->length() > 0) {
5355 uint j = _dom_stk->pop();
5356 _dom_depth[j] = dd;
5357 dd++;
5358 }
5359 }
5360 }
5361
5362 //------------------------------sort-------------------------------------------
5363 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5364 // loop tree, not the root.
5365 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5366 if( !innermost ) return loop; // New innermost loop
5367
5368 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5369 assert( loop_preorder, "not yet post-walked loop" );
5370 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5371 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5372
5373 // Insert at start of list
5374 while( l ) { // Insertion sort based on pre-order
5375 if( l == loop ) return innermost; // Already on list!
5376 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5377 assert( l_preorder, "not yet post-walked l" );
5378 // Check header pre-order number to figure proper nesting
5379 if( loop_preorder > l_preorder )
5380 break; // End of insertion
5381 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5382 // Since I split shared headers, you'd think this could not happen.
5383 // BUT: I must first do the preorder numbering before I can discover I
5384 // have shared headers, so the split headers all get the same preorder
5385 // number as the RegionNode they split from.
5386 if( loop_preorder == l_preorder &&
5387 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5388 break; // Also check for shared headers (same pre#)
5389 pp = &l->_parent; // Chain up list
5390 l = *pp;
5391 }
5392 // Link into list
5393 // Point predecessor to me
5394 *pp = loop;
5395 // Point me to successor
5396 IdealLoopTree *p = loop->_parent;
5397 loop->_parent = l; // Point me to successor
5398 if( p ) sort( p, innermost ); // Insert my parents into list as well
5399 return innermost;
5400 }
5401
5402 //------------------------------build_loop_tree--------------------------------
5403 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5404 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5405 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5406 // tightest enclosing IdealLoopTree for post-walked.
5407 //
5408 // During my forward walk I do a short 1-layer lookahead to see if I can find
5409 // a loop backedge with that doesn't have any work on the backedge. This
5410 // helps me construct nested loops with shared headers better.
5411 //
5412 // Once I've done the forward recursion, I do the post-work. For each child
5413 // I check to see if there is a backedge. Backedges define a loop! I
5414 // insert an IdealLoopTree at the target of the backedge.
5415 //
5416 // During the post-work I also check to see if I have several children
5417 // belonging to different loops. If so, then this Node is a decision point
5418 // where control flow can choose to change loop nests. It is at this
5419 // decision point where I can figure out how loops are nested. At this
5420 // time I can properly order the different loop nests from my children.
5421 // Note that there may not be any backedges at the decision point!
5422 //
5423 // Since the decision point can be far removed from the backedges, I can't
5424 // order my loops at the time I discover them. Thus at the decision point
5425 // I need to inspect loop header pre-order numbers to properly nest my
5426 // loops. This means I need to sort my childrens' loops by pre-order.
5427 // The sort is of size number-of-control-children, which generally limits
5428 // it to size 2 (i.e., I just choose between my 2 target loops).
5429 void PhaseIdealLoop::build_loop_tree() {
5430 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5431 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5432 Node *n = C->root();
5433 bltstack.push(n);
5434 int pre_order = 1;
5435 int stack_size;
5436
5437 while ( ( stack_size = bltstack.length() ) != 0 ) {
5438 n = bltstack.top(); // Leave node on stack
5439 if ( !is_visited(n) ) {
5440 // ---- Pre-pass Work ----
5441 // Pre-walked but not post-walked nodes need a pre_order number.
5442
5443 set_preorder_visited( n, pre_order ); // set as visited
5444
5445 // ---- Scan over children ----
5446 // Scan first over control projections that lead to loop headers.
5447 // This helps us find inner-to-outer loops with shared headers better.
5448
5449 // Scan children's children for loop headers.
5450 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5451 Node* m = n->raw_out(i); // Child
5452 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5453 // Scan over children's children to find loop
5454 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5455 Node* l = m->fast_out(j);
5456 if( is_visited(l) && // Been visited?
5457 !is_postvisited(l) && // But not post-visited
5458 get_preorder(l) < pre_order ) { // And smaller pre-order
5459 // Found! Scan the DFS down this path before doing other paths
5460 bltstack.push(m);
5461 break;
5462 }
5463 }
5464 }
5465 }
5466 pre_order++;
5467 }
5468 else if ( !is_postvisited(n) ) {
5469 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5470 // such as com.sun.rsasign.am::a.
5471 // For non-recursive version, first, process current children.
5472 // On next iteration, check if additional children were added.
5473 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5474 Node* u = n->raw_out(k);
5475 if ( u->is_CFG() && !is_visited(u) ) {
5476 bltstack.push(u);
5477 }
5478 }
5479 if ( bltstack.length() == stack_size ) {
5480 // There were no additional children, post visit node now
5481 (void)bltstack.pop(); // Remove node from stack
5482 pre_order = build_loop_tree_impl(n, pre_order);
5483 // Check for bailout
5484 if (C->failing()) {
5485 return;
5486 }
5487 // Check to grow _preorders[] array for the case when
5488 // build_loop_tree_impl() adds new nodes.
5489 check_grow_preorders();
5490 }
5491 }
5492 else {
5493 (void)bltstack.pop(); // Remove post-visited node from stack
5494 }
5495 }
5496 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5497 }
5498
5499 //------------------------------build_loop_tree_impl---------------------------
5500 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5501 // ---- Post-pass Work ----
5502 // Pre-walked but not post-walked nodes need a pre_order number.
5503
5504 // Tightest enclosing loop for this Node
5505 IdealLoopTree *innermost = nullptr;
5506
5507 // For all children, see if any edge is a backedge. If so, make a loop
5508 // for it. Then find the tightest enclosing loop for the self Node.
5509 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5510 Node* m = n->fast_out(i); // Child
5511 if (n == m) continue; // Ignore control self-cycles
5512 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5513
5514 IdealLoopTree *l; // Child's loop
5515 if (!is_postvisited(m)) { // Child visited but not post-visited?
5516 // Found a backedge
5517 assert(get_preorder(m) < pre_order, "should be backedge");
5518 // Check for the RootNode, which is already a LoopNode and is allowed
5519 // to have multiple "backedges".
5520 if (m == C->root()) { // Found the root?
5521 l = _ltree_root; // Root is the outermost LoopNode
5522 } else { // Else found a nested loop
5523 // Insert a LoopNode to mark this loop.
5524 l = new IdealLoopTree(this, m, n);
5525 } // End of Else found a nested loop
5526 if (!has_loop(m)) { // If 'm' does not already have a loop set
5527 set_loop(m, l); // Set loop header to loop now
5528 }
5529 } else { // Else not a nested loop
5530 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5531 IdealLoopTree* m_loop = get_loop(m);
5532 l = m_loop; // Get previously determined loop
5533 // If successor is header of a loop (nest), move up-loop till it
5534 // is a member of some outer enclosing loop. Since there are no
5535 // shared headers (I've split them already) I only need to go up
5536 // at most 1 level.
5537 while (l && l->_head == m) { // Successor heads loop?
5538 l = l->_parent; // Move up 1 for me
5539 }
5540 // If this loop is not properly parented, then this loop
5541 // has no exit path out, i.e. its an infinite loop.
5542 if (!l) {
5543 // Make loop "reachable" from root so the CFG is reachable. Basically
5544 // insert a bogus loop exit that is never taken. 'm', the loop head,
5545 // points to 'n', one (of possibly many) fall-in paths. There may be
5546 // many backedges as well.
5547
5548 if (!_verify_only) {
5549 // Insert the NeverBranch between 'm' and it's control user.
5550 NeverBranchNode *iff = new NeverBranchNode( m );
5551 _igvn.register_new_node_with_optimizer(iff);
5552 set_loop(iff, m_loop);
5553 Node *if_t = new CProjNode( iff, 0 );
5554 _igvn.register_new_node_with_optimizer(if_t);
5555 set_loop(if_t, m_loop);
5556
5557 Node* cfg = nullptr; // Find the One True Control User of m
5558 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5559 Node* x = m->fast_out(j);
5560 if (x->is_CFG() && x != m && x != iff)
5561 { cfg = x; break; }
5562 }
5563 assert(cfg != nullptr, "must find the control user of m");
5564 uint k = 0; // Probably cfg->in(0)
5565 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5566 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5567
5568 // Now create the never-taken loop exit
5569 Node *if_f = new CProjNode( iff, 1 );
5570 _igvn.register_new_node_with_optimizer(if_f);
5571 set_loop(if_f, _ltree_root);
5572 // Find frame ptr for Halt. Relies on the optimizer
5573 // V-N'ing. Easier and quicker than searching through
5574 // the program structure.
5575 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5576 _igvn.register_new_node_with_optimizer(frame);
5577 // Halt & Catch Fire
5578 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5579 _igvn.register_new_node_with_optimizer(halt);
5580 set_loop(halt, _ltree_root);
5581 _igvn.add_input_to(C->root(), halt);
5582 }
5583 set_loop(C->root(), _ltree_root);
5584 // move to outer most loop with same header
5585 l = m_loop;
5586 while (true) {
5587 IdealLoopTree* next = l->_parent;
5588 if (next == nullptr || next->_head != m) {
5589 break;
5590 }
5591 l = next;
5592 }
5593 // properly insert infinite loop in loop tree
5594 sort(_ltree_root, l);
5595 // fix child link from parent
5596 IdealLoopTree* p = l->_parent;
5597 l->_next = p->_child;
5598 p->_child = l;
5599 // code below needs enclosing loop
5600 l = l->_parent;
5601 }
5602 }
5603 if (is_postvisited(l->_head)) {
5604 // We are currently visiting l, but its head has already been post-visited.
5605 // l is irreducible: we just found a second entry m.
5606 _has_irreducible_loops = true;
5607 RegionNode* secondary_entry = m->as_Region();
5608
5609 if (!secondary_entry->can_be_irreducible_entry()) {
5610 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
5611 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
5612 return pre_order;
5613 }
5614
5615 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
5616 // Since m is a secondary entry to them all.
5617 while( is_postvisited(l->_head) ) {
5618 l->_irreducible = 1; // = true
5619 RegionNode* head = l->_head->as_Region();
5620 if (!head->can_be_irreducible_entry()) {
5621 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
5622 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
5623 return pre_order;
5624 }
5625 l = l->_parent;
5626 // Check for bad CFG here to prevent crash, and bailout of compile
5627 if (l == nullptr) {
5628 #ifndef PRODUCT
5629 if (TraceLoopOpts) {
5630 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
5631 m->dump();
5632 }
5633 #endif
5634 // This is a rare case that we do not want to handle in C2.
5635 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
5636 return pre_order;
5637 }
5638 }
5639 }
5640 if (!_verify_only) {
5641 C->set_has_irreducible_loop(_has_irreducible_loops);
5642 }
5643
5644 // This Node might be a decision point for loops. It is only if
5645 // it's children belong to several different loops. The sort call
5646 // does a trivial amount of work if there is only 1 child or all
5647 // children belong to the same loop. If however, the children
5648 // belong to different loops, the sort call will properly set the
5649 // _parent pointers to show how the loops nest.
5650 //
5651 // In any case, it returns the tightest enclosing loop.
5652 innermost = sort( l, innermost );
5653 }
5654
5655 // Def-use info will have some dead stuff; dead stuff will have no
5656 // loop decided on.
5657
5658 // Am I a loop header? If so fix up my parent's child and next ptrs.
5659 if( innermost && innermost->_head == n ) {
5660 assert( get_loop(n) == innermost, "" );
5661 IdealLoopTree *p = innermost->_parent;
5662 IdealLoopTree *l = innermost;
5663 while (p && l->_head == n) {
5664 l->_next = p->_child; // Put self on parents 'next child'
5665 p->_child = l; // Make self as first child of parent
5666 l = p; // Now walk up the parent chain
5667 p = l->_parent;
5668 }
5669 } else {
5670 // Note that it is possible for a LoopNode to reach here, if the
5671 // backedge has been made unreachable (hence the LoopNode no longer
5672 // denotes a Loop, and will eventually be removed).
5673
5674 // Record tightest enclosing loop for self. Mark as post-visited.
5675 set_loop(n, innermost);
5676 // Also record has_call flag early on
5677 if (innermost) {
5678 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
5679 // Do not count uncommon calls
5680 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
5681 Node *iff = n->in(0)->in(0);
5682 // No any calls for vectorized loops.
5683 if (C->do_superword() ||
5684 !iff->is_If() ||
5685 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
5686 iff->as_If()->_prob >= 0.01) {
5687 innermost->_has_call = 1;
5688 }
5689 }
5690 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
5691 // Disable loop optimizations if the loop has a scalar replaceable
5692 // allocation. This disabling may cause a potential performance lost
5693 // if the allocation is not eliminated for some reason.
5694 innermost->_allow_optimizations = false;
5695 innermost->_has_call = 1; // = true
5696 } else if (n->Opcode() == Op_SafePoint) {
5697 // Record all safepoints in this loop.
5698 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
5699 innermost->_safepts->push(n);
5700 }
5701 }
5702 }
5703
5704 // Flag as post-visited now
5705 set_postvisited(n);
5706 return pre_order;
5707 }
5708
5709 #ifdef ASSERT
5710 //--------------------------verify_regions_in_irreducible_loops----------------
5711 // Iterate down from Root through CFG, verify for every region:
5712 // if it is in an irreducible loop it must be marked as such
5713 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
5714 ResourceMark rm;
5715 if (!_has_irreducible_loops) {
5716 // last build_loop_tree has not found any irreducible loops
5717 // hence no region has to be marked is_in_irreduible_loop
5718 return;
5719 }
5720
5721 RootNode* root = C->root();
5722 Unique_Node_List worklist; // visit all nodes once
5723 worklist.push(root);
5724 bool failure = false;
5725 for (uint i = 0; i < worklist.size(); i++) {
5726 Node* n = worklist.at(i);
5727 if (n->is_Region()) {
5728 RegionNode* region = n->as_Region();
5729 if (is_in_irreducible_loop(region) &&
5730 region->loop_status() == RegionNode::LoopStatus::Reducible) {
5731 failure = true;
5732 tty->print("irreducible! ");
5733 region->dump();
5734 }
5735 }
5736 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
5737 Node* use = n->fast_out(j);
5738 if (use->is_CFG()) {
5739 worklist.push(use); // push if was not pushed before
5740 }
5741 }
5742 }
5743 assert(!failure, "region in irreducible loop was marked as reducible");
5744 }
5745
5746 //---------------------------is_in_irreducible_loop-------------------------
5747 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
5748 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
5749 if (!_has_irreducible_loops) {
5750 return false; // no irreducible loop in graph
5751 }
5752 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
5753 do {
5754 if (l->_irreducible) {
5755 return true; // found it
5756 }
5757 if (l == _ltree_root) {
5758 return false; // reached root, terimnate
5759 }
5760 l = l->_parent;
5761 } while (l != nullptr);
5762 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
5763 // We have "l->_parent == nullptr", which happens only for infinite loops,
5764 // where no parent is attached to the loop. We did not find any irreducible
5765 // loop from this block out to lp. Thus lp only has one entry, and no exit
5766 // (it is infinite and reducible). We can always rewrite an infinite loop
5767 // that is nested inside other loops:
5768 // while(condition) { infinite_loop; }
5769 // with an equivalent program where the infinite loop is an outermost loop
5770 // that is not nested in any loop:
5771 // while(condition) { break; } infinite_loop;
5772 // Thus, we can understand lp as an outermost loop, and can terminate and
5773 // conclude: this block is in no irreducible loop.
5774 return false;
5775 }
5776 #endif
5777
5778 //------------------------------build_loop_early-------------------------------
5779 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
5780 // First pass computes the earliest controlling node possible. This is the
5781 // controlling input with the deepest dominating depth.
5782 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
5783 while (worklist.size() != 0) {
5784 // Use local variables nstack_top_n & nstack_top_i to cache values
5785 // on nstack's top.
5786 Node *nstack_top_n = worklist.pop();
5787 uint nstack_top_i = 0;
5788 //while_nstack_nonempty:
5789 while (true) {
5790 // Get parent node and next input's index from stack's top.
5791 Node *n = nstack_top_n;
5792 uint i = nstack_top_i;
5793 uint cnt = n->req(); // Count of inputs
5794 if (i == 0) { // Pre-process the node.
5795 if( has_node(n) && // Have either loop or control already?
5796 !has_ctrl(n) ) { // Have loop picked out already?
5797 // During "merge_many_backedges" we fold up several nested loops
5798 // into a single loop. This makes the members of the original
5799 // loop bodies pointing to dead loops; they need to move up
5800 // to the new UNION'd larger loop. I set the _head field of these
5801 // dead loops to null and the _parent field points to the owning
5802 // loop. Shades of UNION-FIND algorithm.
5803 IdealLoopTree *ilt;
5804 while( !(ilt = get_loop(n))->_head ) {
5805 // Normally I would use a set_loop here. But in this one special
5806 // case, it is legal (and expected) to change what loop a Node
5807 // belongs to.
5808 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
5809 }
5810 // Remove safepoints ONLY if I've already seen I don't need one.
5811 // (the old code here would yank a 2nd safepoint after seeing a
5812 // first one, even though the 1st did not dominate in the loop body
5813 // and thus could be avoided indefinitely)
5814 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
5815 is_deleteable_safept(n)) {
5816 Node *in = n->in(TypeFunc::Control);
5817 lazy_replace(n,in); // Pull safepoint now
5818 if (ilt->_safepts != nullptr) {
5819 ilt->_safepts->yank(n);
5820 }
5821 // Carry on with the recursion "as if" we are walking
5822 // only the control input
5823 if( !visited.test_set( in->_idx ) ) {
5824 worklist.push(in); // Visit this guy later, using worklist
5825 }
5826 // Get next node from nstack:
5827 // - skip n's inputs processing by setting i > cnt;
5828 // - we also will not call set_early_ctrl(n) since
5829 // has_node(n) == true (see the condition above).
5830 i = cnt + 1;
5831 }
5832 }
5833 } // if (i == 0)
5834
5835 // Visit all inputs
5836 bool done = true; // Assume all n's inputs will be processed
5837 while (i < cnt) {
5838 Node *in = n->in(i);
5839 ++i;
5840 if (in == nullptr) continue;
5841 if (in->pinned() && !in->is_CFG())
5842 set_ctrl(in, in->in(0));
5843 int is_visited = visited.test_set( in->_idx );
5844 if (!has_node(in)) { // No controlling input yet?
5845 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
5846 assert( !is_visited, "visit only once" );
5847 nstack.push(n, i); // Save parent node and next input's index.
5848 nstack_top_n = in; // Process current input now.
5849 nstack_top_i = 0;
5850 done = false; // Not all n's inputs processed.
5851 break; // continue while_nstack_nonempty;
5852 } else if (!is_visited) {
5853 // This guy has a location picked out for him, but has not yet
5854 // been visited. Happens to all CFG nodes, for instance.
5855 // Visit him using the worklist instead of recursion, to break
5856 // cycles. Since he has a location already we do not need to
5857 // find his location before proceeding with the current Node.
5858 worklist.push(in); // Visit this guy later, using worklist
5859 }
5860 }
5861 if (done) {
5862 // All of n's inputs have been processed, complete post-processing.
5863
5864 // Compute earliest point this Node can go.
5865 // CFG, Phi, pinned nodes already know their controlling input.
5866 if (!has_node(n)) {
5867 // Record earliest legal location
5868 set_early_ctrl(n, false);
5869 }
5870 if (nstack.is_empty()) {
5871 // Finished all nodes on stack.
5872 // Process next node on the worklist.
5873 break;
5874 }
5875 // Get saved parent node and next input's index.
5876 nstack_top_n = nstack.node();
5877 nstack_top_i = nstack.index();
5878 nstack.pop();
5879 }
5880 } // while (true)
5881 }
5882 }
5883
5884 //------------------------------dom_lca_internal--------------------------------
5885 // Pair-wise LCA
5886 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
5887 if( !n1 ) return n2; // Handle null original LCA
5888 assert( n1->is_CFG(), "" );
5889 assert( n2->is_CFG(), "" );
5890 // find LCA of all uses
5891 uint d1 = dom_depth(n1);
5892 uint d2 = dom_depth(n2);
5893 while (n1 != n2) {
5894 if (d1 > d2) {
5895 n1 = idom(n1);
5896 d1 = dom_depth(n1);
5897 } else if (d1 < d2) {
5898 n2 = idom(n2);
5899 d2 = dom_depth(n2);
5900 } else {
5901 // Here d1 == d2. Due to edits of the dominator-tree, sections
5902 // of the tree might have the same depth. These sections have
5903 // to be searched more carefully.
5904
5905 // Scan up all the n1's with equal depth, looking for n2.
5906 Node *t1 = idom(n1);
5907 while (dom_depth(t1) == d1) {
5908 if (t1 == n2) return n2;
5909 t1 = idom(t1);
5910 }
5911 // Scan up all the n2's with equal depth, looking for n1.
5912 Node *t2 = idom(n2);
5913 while (dom_depth(t2) == d2) {
5914 if (t2 == n1) return n1;
5915 t2 = idom(t2);
5916 }
5917 // Move up to a new dominator-depth value as well as up the dom-tree.
5918 n1 = t1;
5919 n2 = t2;
5920 d1 = dom_depth(n1);
5921 d2 = dom_depth(n2);
5922 }
5923 }
5924 return n1;
5925 }
5926
5927 //------------------------------compute_idom-----------------------------------
5928 // Locally compute IDOM using dom_lca call. Correct only if the incoming
5929 // IDOMs are correct.
5930 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
5931 assert( region->is_Region(), "" );
5932 Node *LCA = nullptr;
5933 for( uint i = 1; i < region->req(); i++ ) {
5934 if( region->in(i) != C->top() )
5935 LCA = dom_lca( LCA, region->in(i) );
5936 }
5937 return LCA;
5938 }
5939
5940 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
5941 bool had_error = false;
5942 #ifdef ASSERT
5943 if (early != C->root()) {
5944 // Make sure that there's a dominance path from LCA to early
5945 Node* d = LCA;
5946 while (d != early) {
5947 if (d == C->root()) {
5948 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
5949 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
5950 had_error = true;
5951 break;
5952 }
5953 d = idom(d);
5954 }
5955 }
5956 #endif
5957 return had_error;
5958 }
5959
5960
5961 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
5962 // Compute LCA over list of uses
5963 bool had_error = false;
5964 Node *LCA = nullptr;
5965 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
5966 Node* c = n->fast_out(i);
5967 if (_loop_or_ctrl[c->_idx] == nullptr)
5968 continue; // Skip the occasional dead node
5969 if( c->is_Phi() ) { // For Phis, we must land above on the path
5970 for( uint j=1; j<c->req(); j++ ) {// For all inputs
5971 if( c->in(j) == n ) { // Found matching input?
5972 Node *use = c->in(0)->in(j);
5973 if (_verify_only && use->is_top()) continue;
5974 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
5975 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
5976 }
5977 }
5978 } else {
5979 // For CFG data-users, use is in the block just prior
5980 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
5981 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
5982 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
5983 }
5984 }
5985 assert(!had_error, "bad dominance");
5986 return LCA;
5987 }
5988
5989 // Check the shape of the graph at the loop entry. In some cases,
5990 // the shape of the graph does not match the shape outlined below.
5991 // That is caused by the Opaque1 node "protecting" the shape of
5992 // the graph being removed by, for example, the IGVN performed
5993 // in PhaseIdealLoop::build_and_optimize().
5994 //
5995 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
5996 // loop unswitching, and IGVN, or a combination of them) can freely change
5997 // the graph's shape. As a result, the graph shape outlined below cannot
5998 // be guaranteed anymore.
5999 Node* CountedLoopNode::is_canonical_loop_entry() {
6000 if (!is_main_loop() && !is_post_loop()) {
6001 return nullptr;
6002 }
6003 Node* ctrl = skip_assertion_predicates_with_halt();
6004
6005 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6006 return nullptr;
6007 }
6008 Node* iffm = ctrl->in(0);
6009 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6010 return nullptr;
6011 }
6012 Node* bolzm = iffm->in(1);
6013 if (bolzm == nullptr || !bolzm->is_Bool()) {
6014 return nullptr;
6015 }
6016 Node* cmpzm = bolzm->in(1);
6017 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6018 return nullptr;
6019 }
6020
6021 uint input = is_main_loop() ? 2 : 1;
6022 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6023 return nullptr;
6024 }
6025 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6026 #ifdef ASSERT
6027 bool found_opaque = false;
6028 for (uint i = 1; i < cmpzm->req(); i++) {
6029 Node* opnd = cmpzm->in(i);
6030 if (opnd && opnd->is_Opaque1()) {
6031 found_opaque = true;
6032 break;
6033 }
6034 }
6035 assert(found_opaque == res, "wrong pattern");
6036 #endif
6037 return res ? cmpzm->in(input) : nullptr;
6038 }
6039
6040 // Find pre loop end from main loop. Returns nullptr if none.
6041 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6042 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6043 // The loop cannot be optimized if the graph shape at the loop entry is
6044 // inappropriate.
6045 if (is_canonical_loop_entry() == nullptr) {
6046 return nullptr;
6047 }
6048
6049 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6050 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6051 return nullptr;
6052 }
6053 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6054 CountedLoopNode* loop_node = pre_end->loopnode();
6055 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6056 return nullptr;
6057 }
6058 return pre_end;
6059 }
6060
6061 //------------------------------get_late_ctrl----------------------------------
6062 // Compute latest legal control.
6063 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6064 assert(early != nullptr, "early control should not be null");
6065
6066 Node* LCA = compute_lca_of_uses(n, early);
6067 #ifdef ASSERT
6068 if (LCA == C->root() && LCA != early) {
6069 // def doesn't dominate uses so print some useful debugging output
6070 compute_lca_of_uses(n, early, true);
6071 }
6072 #endif
6073
6074 if (n->is_Load() && LCA != early) {
6075 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6076 }
6077
6078 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6079 return LCA;
6080 }
6081
6082 // if this is a load, check for anti-dependent stores
6083 // We use a conservative algorithm to identify potential interfering
6084 // instructions and for rescheduling the load. The users of the memory
6085 // input of this load are examined. Any use which is not a load and is
6086 // dominated by early is considered a potentially interfering store.
6087 // This can produce false positives.
6088 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6089 int load_alias_idx = C->get_alias_index(n->adr_type());
6090 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6091 Unique_Node_List worklist;
6092
6093 Node* mem = n->in(MemNode::Memory);
6094 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6095 Node* s = mem->fast_out(i);
6096 worklist.push(s);
6097 }
6098 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6099 Node* s = worklist.at(i);
6100 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6101 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6102 s->is_Phi()) {
6103 continue;
6104 } else if (s->is_MergeMem()) {
6105 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6106 Node* s1 = s->fast_out(i);
6107 worklist.push(s1);
6108 }
6109 } else {
6110 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6111 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6112 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6113 const TypePtr* adr_type = s->adr_type();
6114 if (s->is_ArrayCopy()) {
6115 // Copy to known instance needs destination type to test for aliasing
6116 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6117 if (dest_type != TypeOopPtr::BOTTOM) {
6118 adr_type = dest_type;
6119 }
6120 }
6121 if (C->can_alias(adr_type, load_alias_idx)) {
6122 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6123 } else if (s->is_CFG() && s->is_Multi()) {
6124 // Look for the memory use of s (that is the use of its memory projection)
6125 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6126 Node* s1 = s->fast_out(i);
6127 assert(s1->is_Proj(), "projection expected");
6128 if (_igvn.type(s1) == Type::MEMORY) {
6129 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6130 Node* s2 = s1->fast_out(j);
6131 worklist.push(s2);
6132 }
6133 }
6134 }
6135 }
6136 }
6137 }
6138 }
6139 // For Phis only consider Region's inputs that were reached by following the memory edges
6140 if (LCA != early) {
6141 for (uint i = 0; i < worklist.size(); i++) {
6142 Node* s = worklist.at(i);
6143 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6144 Node* r = s->in(0);
6145 for (uint j = 1; j < s->req(); j++) {
6146 Node* in = s->in(j);
6147 Node* r_in = r->in(j);
6148 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6149 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6150 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6151 }
6152 }
6153 }
6154 }
6155 }
6156 }
6157 return LCA;
6158 }
6159
6160 // Is CFG node 'dominator' dominating node 'n'?
6161 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6162 if (dominator == n) {
6163 return true;
6164 }
6165 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6166 uint dd = dom_depth(dominator);
6167 while (dom_depth(n) >= dd) {
6168 if (n == dominator) {
6169 return true;
6170 }
6171 n = idom(n);
6172 }
6173 return false;
6174 }
6175
6176 // Is CFG node 'dominator' strictly dominating node 'n'?
6177 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6178 return dominator != n && is_dominator(dominator, n);
6179 }
6180
6181 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6182 // Pair-wise LCA with tags.
6183 // Tag each index with the node 'tag' currently being processed
6184 // before advancing up the dominator chain using idom().
6185 // Later calls that find a match to 'tag' know that this path has already
6186 // been considered in the current LCA (which is input 'n1' by convention).
6187 // Since get_late_ctrl() is only called once for each node, the tag array
6188 // does not need to be cleared between calls to get_late_ctrl().
6189 // Algorithm trades a larger constant factor for better asymptotic behavior
6190 //
6191 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6192 uint d1 = dom_depth(n1);
6193 uint d2 = dom_depth(n2);
6194 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6195
6196 do {
6197 if (d1 > d2) {
6198 // current lca is deeper than n2
6199 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6200 n1 = idom(n1);
6201 d1 = dom_depth(n1);
6202 } else if (d1 < d2) {
6203 // n2 is deeper than current lca
6204 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6205 if (memo == tag) {
6206 return n1; // Return the current LCA
6207 }
6208 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6209 n2 = idom(n2);
6210 d2 = dom_depth(n2);
6211 } else {
6212 // Here d1 == d2. Due to edits of the dominator-tree, sections
6213 // of the tree might have the same depth. These sections have
6214 // to be searched more carefully.
6215
6216 // Scan up all the n1's with equal depth, looking for n2.
6217 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6218 Node *t1 = idom(n1);
6219 while (dom_depth(t1) == d1) {
6220 if (t1 == n2) return n2;
6221 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6222 t1 = idom(t1);
6223 }
6224 // Scan up all the n2's with equal depth, looking for n1.
6225 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6226 Node *t2 = idom(n2);
6227 while (dom_depth(t2) == d2) {
6228 if (t2 == n1) return n1;
6229 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6230 t2 = idom(t2);
6231 }
6232 // Move up to a new dominator-depth value as well as up the dom-tree.
6233 n1 = t1;
6234 n2 = t2;
6235 d1 = dom_depth(n1);
6236 d2 = dom_depth(n2);
6237 }
6238 } while (n1 != n2);
6239 return n1;
6240 }
6241
6242 //------------------------------init_dom_lca_tags------------------------------
6243 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6244 // Intended use does not involve any growth for the array, so it could
6245 // be of fixed size.
6246 void PhaseIdealLoop::init_dom_lca_tags() {
6247 uint limit = C->unique() + 1;
6248 _dom_lca_tags.at_grow(limit, 0);
6249 _dom_lca_tags_round = 0;
6250 #ifdef ASSERT
6251 for (uint i = 0; i < limit; ++i) {
6252 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6253 }
6254 #endif // ASSERT
6255 }
6256
6257 //------------------------------build_loop_late--------------------------------
6258 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6259 // Second pass finds latest legal placement, and ideal loop placement.
6260 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6261 while (worklist.size() != 0) {
6262 Node *n = worklist.pop();
6263 // Only visit once
6264 if (visited.test_set(n->_idx)) continue;
6265 uint cnt = n->outcnt();
6266 uint i = 0;
6267 while (true) {
6268 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6269 // Visit all children
6270 if (i < cnt) {
6271 Node* use = n->raw_out(i);
6272 ++i;
6273 // Check for dead uses. Aggressively prune such junk. It might be
6274 // dead in the global sense, but still have local uses so I cannot
6275 // easily call 'remove_dead_node'.
6276 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6277 // Due to cycles, we might not hit the same fixed point in the verify
6278 // pass as we do in the regular pass. Instead, visit such phis as
6279 // simple uses of the loop head.
6280 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6281 if( !visited.test(use->_idx) )
6282 worklist.push(use);
6283 } else if( !visited.test_set(use->_idx) ) {
6284 nstack.push(n, i); // Save parent and next use's index.
6285 n = use; // Process all children of current use.
6286 cnt = use->outcnt();
6287 i = 0;
6288 }
6289 } else {
6290 // Do not visit around the backedge of loops via data edges.
6291 // push dead code onto a worklist
6292 _deadlist.push(use);
6293 }
6294 } else {
6295 // All of n's children have been processed, complete post-processing.
6296 build_loop_late_post(n);
6297 if (C->failing()) { return; }
6298 if (nstack.is_empty()) {
6299 // Finished all nodes on stack.
6300 // Process next node on the worklist.
6301 break;
6302 }
6303 // Get saved parent node and next use's index. Visit the rest of uses.
6304 n = nstack.node();
6305 cnt = n->outcnt();
6306 i = nstack.index();
6307 nstack.pop();
6308 }
6309 }
6310 }
6311 }
6312
6313 // Verify that no data node is scheduled in the outer loop of a strip
6314 // mined loop.
6315 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6316 #ifdef ASSERT
6317 if (get_loop(least)->_nest == 0) {
6318 return;
6319 }
6320 IdealLoopTree* loop = get_loop(least);
6321 Node* head = loop->_head;
6322 if (head->is_OuterStripMinedLoop() &&
6323 // Verification can't be applied to fully built strip mined loops
6324 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6325 Node* sfpt = head->as_Loop()->outer_safepoint();
6326 ResourceMark rm;
6327 Unique_Node_List wq;
6328 wq.push(sfpt);
6329 for (uint i = 0; i < wq.size(); i++) {
6330 Node *m = wq.at(i);
6331 for (uint i = 1; i < m->req(); i++) {
6332 Node* nn = m->in(i);
6333 if (nn == n) {
6334 return;
6335 }
6336 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6337 wq.push(nn);
6338 }
6339 }
6340 }
6341 ShouldNotReachHere();
6342 }
6343 #endif
6344 }
6345
6346
6347 //------------------------------build_loop_late_post---------------------------
6348 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6349 // Second pass finds latest legal placement, and ideal loop placement.
6350 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6351 build_loop_late_post_work(n, true);
6352 }
6353
6354 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6355 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6356 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6357 // Predication or creating a Loop Limit Check Predicate later.
6358 class DominatedPredicates : public UnifiedPredicateVisitor {
6359 Node* const _dominator;
6360 Node* _earliest_dominated_predicate_entry;
6361 bool _should_continue;
6362 PhaseIdealLoop* const _phase;
6363
6364 public:
6365 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6366 : _dominator(dominator),
6367 _earliest_dominated_predicate_entry(start_node),
6368 _should_continue(true),
6369 _phase(phase) {}
6370 NONCOPYABLE(DominatedPredicates);
6371
6372 bool should_continue() const override {
6373 return _should_continue;
6374 }
6375
6376 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6377 Node* earliest_dominated_predicate_entry() const {
6378 return _earliest_dominated_predicate_entry;
6379 }
6380
6381 void visit_predicate(const Predicate& predicate) override {
6382 Node* entry = predicate.entry();
6383 if (_phase->is_strict_dominator(entry, _dominator)) {
6384 _should_continue = false;
6385 } else {
6386 _earliest_dominated_predicate_entry = entry;
6387 }
6388 }
6389 };
6390
6391 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6392
6393 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6394 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6395 }
6396
6397 #ifdef ASSERT
6398 if (_verify_only && !n->is_CFG()) {
6399 // Check def-use domination.
6400 // We would like to expose this check in product but it appears to be expensive.
6401 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6402 }
6403 #endif
6404
6405 // CFG and pinned nodes already handled
6406 if( n->in(0) ) {
6407 if( n->in(0)->is_top() ) return; // Dead?
6408
6409 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6410 // _must_ be pinned (they have to observe their control edge of course).
6411 // Unlike Stores (which modify an unallocable resource, the memory
6412 // state), Mods/Loads can float around. So free them up.
6413 switch( n->Opcode() ) {
6414 case Op_DivI:
6415 case Op_DivF:
6416 case Op_DivD:
6417 case Op_ModI:
6418 case Op_LoadB: // Same with Loads; they can sink
6419 case Op_LoadUB: // during loop optimizations.
6420 case Op_LoadUS:
6421 case Op_LoadD:
6422 case Op_LoadF:
6423 case Op_LoadI:
6424 case Op_LoadKlass:
6425 case Op_LoadNKlass:
6426 case Op_LoadL:
6427 case Op_LoadS:
6428 case Op_LoadP:
6429 case Op_LoadN:
6430 case Op_LoadRange:
6431 case Op_LoadD_unaligned:
6432 case Op_LoadL_unaligned:
6433 case Op_StrComp: // Does a bunch of load-like effects
6434 case Op_StrEquals:
6435 case Op_StrIndexOf:
6436 case Op_StrIndexOfChar:
6437 case Op_AryEq:
6438 case Op_VectorizedHashCode:
6439 case Op_CountPositives:
6440 pinned = false;
6441 }
6442 if (n->is_CMove() || n->is_ConstraintCast()) {
6443 pinned = false;
6444 }
6445 if( pinned ) {
6446 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6447 if( !chosen_loop->_child ) // Inner loop?
6448 chosen_loop->_body.push(n); // Collect inner loops
6449 return;
6450 }
6451 } else { // No slot zero
6452 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6453 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6454 return;
6455 }
6456 assert(!n->is_CFG() || n->outcnt() == 0, "");
6457 }
6458
6459 // Do I have a "safe range" I can select over?
6460 Node *early = get_ctrl(n);// Early location already computed
6461
6462 // Compute latest point this Node can go
6463 Node *LCA = get_late_ctrl( n, early );
6464 // LCA is null due to uses being dead
6465 if( LCA == nullptr ) {
6466 #ifdef ASSERT
6467 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6468 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6469 }
6470 #endif
6471 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6472 _deadlist.push(n);
6473 return;
6474 }
6475 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6476
6477 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6478 Node *least = legal; // Best legal position so far
6479 while( early != legal ) { // While not at earliest legal
6480 if (legal->is_Start() && !early->is_Root()) {
6481 #ifdef ASSERT
6482 // Bad graph. Print idom path and fail.
6483 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6484 assert(false, "Bad graph detected in build_loop_late");
6485 #endif
6486 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6487 return;
6488 }
6489 // Find least loop nesting depth
6490 legal = idom(legal); // Bump up the IDOM tree
6491 // Check for lower nesting depth
6492 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6493 least = legal;
6494 }
6495 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6496
6497 if (least != early) {
6498 // Move the node above predicates as far up as possible so a
6499 // following pass of Loop Predication doesn't hoist a predicate
6500 // that depends on it above that node.
6501 const PredicateIterator predicate_iterator(least);
6502 DominatedPredicates dominated_predicates(early, least, this);
6503 predicate_iterator.for_each(dominated_predicates);
6504 least = dominated_predicates.earliest_dominated_predicate_entry();
6505 }
6506 // Try not to place code on a loop entry projection
6507 // which can inhibit range check elimination.
6508 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6509 Node* ctrl_out = least->unique_ctrl_out_or_null();
6510 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6511 least == ctrl_out->in(LoopNode::EntryControl) &&
6512 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6513 Node* least_dom = idom(least);
6514 if (get_loop(least_dom)->is_member(get_loop(least))) {
6515 least = least_dom;
6516 }
6517 }
6518 }
6519 // Don't extend live ranges of raw oops
6520 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6521 !n->bottom_type()->isa_rawptr()) {
6522 least = early;
6523 }
6524
6525 #ifdef ASSERT
6526 // Broken part of VerifyLoopOptimizations (F)
6527 // Reason:
6528 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6529 /*
6530 // If verifying, verify that 'verify_me' has a legal location
6531 // and choose it as our location.
6532 if( _verify_me ) {
6533 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6534 Node *legal = LCA;
6535 while( early != legal ) { // While not at earliest legal
6536 if( legal == v_ctrl ) break; // Check for prior good location
6537 legal = idom(legal) ;// Bump up the IDOM tree
6538 }
6539 // Check for prior good location
6540 if( legal == v_ctrl ) least = legal; // Keep prior if found
6541 }
6542 */
6543 #endif
6544
6545 // Assign discovered "here or above" point
6546 least = find_non_split_ctrl(least);
6547 verify_strip_mined_scheduling(n, least);
6548 set_ctrl(n, least);
6549
6550 // Collect inner loop bodies
6551 IdealLoopTree *chosen_loop = get_loop(least);
6552 if( !chosen_loop->_child ) // Inner loop?
6553 chosen_loop->_body.push(n);// Collect inner loops
6554
6555 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6556 _zero_trip_guard_opaque_nodes.push(n);
6557 }
6558
6559 if (!_verify_only && n->Opcode() == Op_OpaqueMultiversioning) {
6560 _multiversion_opaque_nodes.push(n);
6561 }
6562 }
6563
6564 #ifdef ASSERT
6565 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6566 tty->print_cr("%s", msg);
6567 tty->print("n: "); n->dump();
6568 tty->print("early(n): "); early->dump();
6569 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6570 n->in(0) != early && !n->in(0)->is_Root()) {
6571 tty->print("n->in(0): "); n->in(0)->dump();
6572 }
6573 for (uint i = 1; i < n->req(); i++) {
6574 Node* in1 = n->in(i);
6575 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6576 tty->print("n->in(%d): ", i); in1->dump();
6577 Node* in1_early = get_ctrl(in1);
6578 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6579 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6580 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6581 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6582 }
6583 for (uint j = 1; j < in1->req(); j++) {
6584 Node* in2 = in1->in(j);
6585 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6586 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6587 Node* in2_early = get_ctrl(in2);
6588 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6589 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6590 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
6591 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
6592 }
6593 }
6594 }
6595 }
6596 }
6597 tty->cr();
6598 tty->print("LCA(n): "); LCA->dump();
6599 for (uint i = 0; i < n->outcnt(); i++) {
6600 Node* u1 = n->raw_out(i);
6601 if (u1 == n)
6602 continue;
6603 tty->print("n->out(%d): ", i); u1->dump();
6604 if (u1->is_CFG()) {
6605 for (uint j = 0; j < u1->outcnt(); j++) {
6606 Node* u2 = u1->raw_out(j);
6607 if (u2 != u1 && u2 != n && u2->is_CFG()) {
6608 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6609 }
6610 }
6611 } else {
6612 Node* u1_later = get_ctrl(u1);
6613 tty->print("later(n->out(%d)): ", i); u1_later->dump();
6614 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
6615 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
6616 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
6617 }
6618 for (uint j = 0; j < u1->outcnt(); j++) {
6619 Node* u2 = u1->raw_out(j);
6620 if (u2 == n || u2 == u1)
6621 continue;
6622 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6623 if (!u2->is_CFG()) {
6624 Node* u2_later = get_ctrl(u2);
6625 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
6626 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
6627 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
6628 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
6629 }
6630 }
6631 }
6632 }
6633 }
6634 dump_idoms(early, LCA);
6635 tty->cr();
6636 }
6637
6638 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
6639 class RealLCA {
6640 const PhaseIdealLoop* _phase;
6641 Node* _early;
6642 Node* _wrong_lca;
6643 uint _early_index;
6644 int _wrong_lca_index;
6645
6646 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
6647 // is different: Return the previously visited node which must be the real LCA.
6648 // The node lists also contain _early and _wrong_lca, respectively.
6649 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
6650 int early_index = early_with_idoms.size() - 1;
6651 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
6652 bool found_difference = false;
6653 do {
6654 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
6655 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
6656 found_difference = true;
6657 break;
6658 }
6659 early_index--;
6660 wrong_lca_index--;
6661 } while (wrong_lca_index >= 0);
6662
6663 assert(early_index >= 0, "must always find an LCA - cannot be early");
6664 _early_index = early_index;
6665 _wrong_lca_index = wrong_lca_index;
6666 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
6667 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
6668 return real_lca;
6669 }
6670
6671 void dump(Node* real_lca) {
6672 tty->cr();
6673 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
6674 _phase->dump_idom(_early, _early_index + 1);
6675
6676 tty->cr();
6677 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
6678 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
6679
6680 tty->cr();
6681 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
6682 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
6683 if (_wrong_lca_index >= 0) {
6684 tty->print(" (idom[%d])", _wrong_lca_index);
6685 }
6686 tty->print_cr(":");
6687 real_lca->dump();
6688 }
6689
6690 public:
6691 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
6692 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
6693 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
6694 }
6695
6696 void compute_and_dump() {
6697 ResourceMark rm;
6698 Unique_Node_List early_with_idoms;
6699 Unique_Node_List wrong_lca_with_idoms;
6700 early_with_idoms.push(_early);
6701 wrong_lca_with_idoms.push(_wrong_lca);
6702 _phase->get_idoms(_early, 10000, early_with_idoms);
6703 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
6704 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
6705 dump(real_lca);
6706 }
6707 };
6708
6709 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
6710 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
6711 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
6712 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
6713
6714 RealLCA real_lca(this, early, wrong_lca);
6715 real_lca.compute_and_dump();
6716 }
6717 #endif // ASSERT
6718
6719 #ifndef PRODUCT
6720 //------------------------------dump-------------------------------------------
6721 void PhaseIdealLoop::dump() const {
6722 ResourceMark rm;
6723 Node_Stack stack(C->live_nodes() >> 2);
6724 Node_List rpo_list;
6725 VectorSet visited;
6726 visited.set(C->top()->_idx);
6727 rpo(C->root(), stack, visited, rpo_list);
6728 // Dump root loop indexed by last element in PO order
6729 dump(_ltree_root, rpo_list.size(), rpo_list);
6730 }
6731
6732 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
6733 loop->dump_head();
6734
6735 // Now scan for CFG nodes in the same loop
6736 for (uint j = idx; j > 0; j--) {
6737 Node* n = rpo_list[j-1];
6738 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
6739 continue;
6740
6741 if (get_loop(n) != loop) { // Wrong loop nest
6742 if (get_loop(n)->_head == n && // Found nested loop?
6743 get_loop(n)->_parent == loop)
6744 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
6745 continue;
6746 }
6747
6748 // Dump controlling node
6749 tty->sp(2 * loop->_nest);
6750 tty->print("C");
6751 if (n == C->root()) {
6752 n->dump();
6753 } else {
6754 Node* cached_idom = idom_no_update(n);
6755 Node* computed_idom = n->in(0);
6756 if (n->is_Region()) {
6757 computed_idom = compute_idom(n);
6758 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
6759 // any MultiBranch ctrl node), so apply a similar transform to
6760 // the cached idom returned from idom_no_update.
6761 cached_idom = find_non_split_ctrl(cached_idom);
6762 }
6763 tty->print(" ID:%d", computed_idom->_idx);
6764 n->dump();
6765 if (cached_idom != computed_idom) {
6766 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
6767 computed_idom->_idx, cached_idom->_idx);
6768 }
6769 }
6770 // Dump nodes it controls
6771 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
6772 // (k < C->unique() && get_ctrl(find(k)) == n)
6773 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
6774 Node* m = C->root()->find(k);
6775 if (m && m->outcnt() > 0) {
6776 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
6777 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
6778 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
6779 }
6780 tty->sp(2 * loop->_nest + 1);
6781 m->dump();
6782 }
6783 }
6784 }
6785 }
6786 }
6787
6788 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
6789 if (has_ctrl(n)) {
6790 tty->print_cr("No idom for data nodes");
6791 } else {
6792 ResourceMark rm;
6793 Unique_Node_List idoms;
6794 get_idoms(n, count, idoms);
6795 dump_idoms_in_reverse(n, idoms);
6796 }
6797 }
6798
6799 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
6800 Node* next = n;
6801 for (uint i = 0; !next->is_Start() && i < count; i++) {
6802 next = idom(next);
6803 assert(!idoms.member(next), "duplicated idom is not possible");
6804 idoms.push(next);
6805 }
6806 }
6807
6808 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
6809 Node* next;
6810 uint padding = 3;
6811 uint node_index_padding_width = (C->unique() == 0 ? 0 : static_cast<int>(log10(static_cast<double>(C->unique())))) + 1;
6812 for (int i = idom_list.size() - 1; i >= 0; i--) {
6813 if (i == 9 || i == 99) {
6814 padding++;
6815 }
6816 next = idom_list[i];
6817 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
6818 }
6819 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
6820 }
6821 #endif // NOT PRODUCT
6822
6823 // Collect a R-P-O for the whole CFG.
6824 // Result list is in post-order (scan backwards for RPO)
6825 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
6826 stk.push(start, 0);
6827 visited.set(start->_idx);
6828
6829 while (stk.is_nonempty()) {
6830 Node* m = stk.node();
6831 uint idx = stk.index();
6832 if (idx < m->outcnt()) {
6833 stk.set_index(idx + 1);
6834 Node* n = m->raw_out(idx);
6835 if (n->is_CFG() && !visited.test_set(n->_idx)) {
6836 stk.push(n, 0);
6837 }
6838 } else {
6839 rpo_list.push(m);
6840 stk.pop();
6841 }
6842 }
6843 }
6844
6845 ConINode* PhaseIdealLoop::intcon(jint i) {
6846 ConINode* node = _igvn.intcon(i);
6847 set_root_as_ctrl(node);
6848 return node;
6849 }
6850
6851 ConLNode* PhaseIdealLoop::longcon(jlong i) {
6852 ConLNode* node = _igvn.longcon(i);
6853 set_root_as_ctrl(node);
6854 return node;
6855 }
6856
6857 ConNode* PhaseIdealLoop::makecon(const Type* t) {
6858 ConNode* node = _igvn.makecon(t);
6859 set_root_as_ctrl(node);
6860 return node;
6861 }
6862
6863 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
6864 ConNode* node = _igvn.integercon(l, bt);
6865 set_root_as_ctrl(node);
6866 return node;
6867 }
6868
6869 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
6870 ConNode* node = _igvn.zerocon(bt);
6871 set_root_as_ctrl(node);
6872 return node;
6873 }
6874
6875
6876 //=============================================================================
6877 //------------------------------LoopTreeIterator-------------------------------
6878
6879 // Advance to next loop tree using a preorder, left-to-right traversal.
6880 void LoopTreeIterator::next() {
6881 assert(!done(), "must not be done.");
6882 if (_curnt->_child != nullptr) {
6883 _curnt = _curnt->_child;
6884 } else if (_curnt->_next != nullptr) {
6885 _curnt = _curnt->_next;
6886 } else {
6887 while (_curnt != _root && _curnt->_next == nullptr) {
6888 _curnt = _curnt->_parent;
6889 }
6890 if (_curnt == _root) {
6891 _curnt = nullptr;
6892 assert(done(), "must be done.");
6893 } else {
6894 assert(_curnt->_next != nullptr, "must be more to do");
6895 _curnt = _curnt->_next;
6896 }
6897 }
6898 }