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 "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/divnode.hpp"
40 #include "opto/idealGraphPrinter.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/opaquenode.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 }
1091 if (UseProfiledLoopPredicate) {
1092 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1093 }
1094 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1095 }
1096
1097 #ifndef PRODUCT
1098 if (bt == T_LONG) {
1099 Atomic::inc(&_long_loop_nests);
1100 }
1101 #endif
1102
1103 inner_head->mark_loop_nest_inner_loop();
1104 outer_head->mark_loop_nest_outer_loop();
1105
1106 return true;
1107 }
1108
1109 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1110 Node_List& range_checks) {
1111 const jlong min_iters = 2;
1112 jlong reduced_iters_limit = iters_limit;
1113 jlong original_iters_limit = iters_limit;
1114 for (uint i = 0; i < loop->_body.size(); i++) {
1115 Node* c = loop->_body.at(i);
1116 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1117 IfProjNode* if_proj = c->as_IfProj();
1118 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1119 if (call != nullptr) {
1120 Node* range = nullptr;
1121 Node* offset = nullptr;
1122 jlong scale = 0;
1123 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1124 loop->is_invariant(range) && loop->is_invariant(offset) &&
1125 scale != min_jlong &&
1126 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1127 assert(scale == (jint)scale, "scale should be an int");
1128 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1129 range_checks.push(c);
1130 }
1131 }
1132 }
1133 }
1134
1135 return checked_cast<int>(reduced_iters_limit);
1136 }
1137
1138 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1139 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1140 // 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
1141 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1142
1143 // 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,
1144 // 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)
1145 // 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].
1146
1147 // 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
1148 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1149 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1150
1151 // 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
1152 // 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
1153 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1154 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1155
1156 // 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
1157 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1158
1159 // 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].
1160 // 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
1161 // 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
1162 // 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
1163 // down (from C and 0 again).
1164
1165 // 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).
1166 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1167 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1168 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1169 // PhaseIdealLoop::add_constraint.)
1170
1171 // 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
1172 // 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
1173 // 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.
1174 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1175 // values for the 32-bit counter j within which the range checks can be eliminated.
1176
1177 // 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
1178 // 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
1179 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1180
1181 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1182
1183 // 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
1184 // 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),
1185 // 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)
1186
1187 // 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).
1188 // 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.
1189 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1190
1191 // Case A: Some Negatives (but no overflow).
1192 // Number line:
1193 // |s64_min . . . 0 . . . s64_max|
1194 // | . Q_min..Q_max . 0 . . . . | s64 negative
1195 // | . . . . R=0 R< R< R< R< | (against R values)
1196 // | . . . Q_min..0..Q_max . . . | small mixed
1197 // | . . . . R R R< R< R< | (against R values)
1198 //
1199 // 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<.
1200 //
1201 // So, if Q_min <s64 0, then use this test:
1202 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1203 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1204 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1205 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1206
1207 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1208
1209 // Case B: No Negatives.
1210 // Number line:
1211 // |s64_min . . . 0 . . . s64_max|
1212 // | . . . . 0 Q_min..Q_max . . | small positive
1213 // | . . . . R> R R R< R< | (against R values)
1214 // | . . . . 0 . Q_min..Q_max . | s64 positive
1215 // | . . . . R> R> R R R< | (against R values)
1216 //
1217 // 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.
1218 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1219 //
1220 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1221 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1222 // More generally:
1223 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1224
1225 // Case C: Overflow in the 64-bit domain
1226 // Number line:
1227 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1228 // | . . . . R> R> R> R> R | (against R values)
1229 //
1230 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1231 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1232 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1233 //
1234 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1235 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1236 // More generally:
1237 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1238 //
1239 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1240 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1241 //
1242 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1243 // clamp(X, L, H) := max(L, min(X, H))
1244 // When degenerately L > H, it returns L not H.
1245 //
1246 // All of the formulas above can be merged into a single one:
1247 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1248 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1249 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1250 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1251 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1252 // replacement R.C.:
1253 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1254 // or equivalently:
1255 // j*K + L_2 <u32 R_2
1256 // where
1257 // L_2 = Q_first - L_clamp
1258 // R_2 = R_clamp - L_clamp
1259 //
1260 // Note on why R is never negative:
1261 //
1262 // Various details of this transformation would break badly if R could be negative, so this transformation only
1263 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1264 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1265 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1266 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1267 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1268 // R.C., but R values from a.length and inside checkIndex are good to go.
1269 //
1270 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1271 Node* inner_iters_actual_int, Node* inner_phi,
1272 Node* iv_add, LoopNode* inner_head) {
1273 Node* long_zero = longcon(0);
1274 Node* int_zero = intcon(0);
1275 Node* long_one = longcon(1);
1276 Node* int_stride = intcon(checked_cast<int>(stride_con));
1277
1278 for (uint i = 0; i < range_checks.size(); i++) {
1279 ProjNode* proj = range_checks.at(i)->as_Proj();
1280 ProjNode* unc_proj = proj->other_if_proj();
1281 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1282 jlong scale = 0;
1283 Node* offset = nullptr;
1284 Node* rc_bol = rc->in(1);
1285 Node* rc_cmp = rc_bol->in(1);
1286 if (rc_cmp->Opcode() == Op_CmpU) {
1287 // could be shared and have already been taken care of
1288 continue;
1289 }
1290 bool short_scale = false;
1291 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1292 assert(ok, "inconsistent: was tested before");
1293 Node* range = rc_cmp->in(2);
1294 Node* c = rc->in(0);
1295 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1296
1297 Node* R = range;
1298 Node* K = longcon(scale);
1299
1300 Node* L = offset;
1301
1302 if (short_scale) {
1303 // This converts:
1304 // (int)i*K + L <u64 R
1305 // with K an int into:
1306 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1307 // to protect against an overflow of (int)i*K
1308 //
1309 // Because if (int)i*K overflows, there are K,L where:
1310 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1311 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1312 // i*(long)K + L <u64 R is true.
1313 //
1314 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1315 //
1316 // It's always true that:
1317 // (int)i*K <u64 (long)max_jint + 1
1318 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1319 // As a consequence:
1320 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1321 // is always false in case of overflow of i*K
1322 //
1323 // Note, there are also K,L where i*K overflows and
1324 // i*K + L <u64 R is true, but
1325 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1326 // So this transformation could cause spurious deoptimizations and failed range check elimination
1327 // (but not incorrect execution) for unlikely corner cases with overflow.
1328 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1329 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1330 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1331 register_new_node(max_range, entry_control);
1332 R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1333 set_subtree_ctrl(R, true);
1334 }
1335
1336 Node* C = outer_phi;
1337
1338 // Start with 64-bit values:
1339 // i*K + L <u64 R
1340 // (C+j)*K + L <u64 R
1341 // j*K + Q <u64 R where Q = Q_first = C*K+L
1342 Node* Q_first = new MulLNode(C, K);
1343 register_new_node(Q_first, entry_control);
1344 Q_first = new AddLNode(Q_first, L);
1345 register_new_node(Q_first, entry_control);
1346
1347 // Compute endpoints of the range of values j*K + Q.
1348 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1349 Node* Q_min = Q_first;
1350
1351 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1352 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1353 register_new_node(B_2, entry_control);
1354 B_2 = new SubINode(B_2, int_stride);
1355 register_new_node(B_2, entry_control);
1356 B_2 = new ConvI2LNode(B_2);
1357 register_new_node(B_2, entry_control);
1358
1359 Node* Q_max = new MulLNode(B_2, K);
1360 register_new_node(Q_max, entry_control);
1361 Q_max = new AddLNode(Q_max, Q_first);
1362 register_new_node(Q_max, entry_control);
1363
1364 if (scale * stride_con < 0) {
1365 swap(Q_min, Q_max);
1366 }
1367 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1368
1369 // L_clamp = Q_min < 0 ? 0 : Q_min
1370 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1371 register_new_node(Q_min_cmp, entry_control);
1372 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1373 register_new_node(Q_min_bool, entry_control);
1374 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1375 register_new_node(L_clamp, entry_control);
1376 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1377
1378 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1379 register_new_node(Q_max_plus_one, entry_control);
1380
1381 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1382 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1383 Node* max_jlong_long = longcon(max_jlong);
1384 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1385 register_new_node(Q_max_cmp, entry_control);
1386 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1387 register_new_node(Q_max_bool, entry_control);
1388 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1389 register_new_node(H_clamp, entry_control);
1390 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1391
1392 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1393 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1394 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1395 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1396 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1397 Node* R_2 = clamp(R, L_clamp, H_clamp);
1398 R_2 = new SubLNode(R_2, L_clamp);
1399 register_new_node(R_2, entry_control);
1400 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1401 register_new_node(R_2, entry_control);
1402
1403 // L_2 = Q_first - L_clamp
1404 // We are subtracting L_clamp from both sides of the <u32 comparison.
1405 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1406 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1407 Node* L_2 = new SubLNode(Q_first, L_clamp);
1408 register_new_node(L_2, entry_control);
1409 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1410 register_new_node(L_2, entry_control);
1411
1412 // Transform the range check using the computed values L_2/R_2
1413 // from: i*K + L <u64 R
1414 // to: j*K + L_2 <u32 R_2
1415 // that is:
1416 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1417 K = intcon(checked_cast<int>(scale));
1418 Node* scaled_iv = new MulINode(inner_phi, K);
1419 register_new_node(scaled_iv, c);
1420 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1421 register_new_node(scaled_iv_plus_offset, c);
1422
1423 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1424 register_new_node(new_rc_cmp, c);
1425
1426 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1427 }
1428 }
1429
1430 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1431 Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1432 set_subtree_ctrl(min, true);
1433 Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1434 set_subtree_ctrl(max, true);
1435 return max;
1436 }
1437
1438 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1439 IfNode* exit_test) {
1440 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1441 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1442 _igvn.register_new_node_with_optimizer(new_inner_head);
1443 _igvn.register_new_node_with_optimizer(new_inner_exit);
1444 loop->_body.push(new_inner_head);
1445 loop->_body.push(new_inner_exit);
1446 loop->_body.yank(head);
1447 loop->_body.yank(exit_test);
1448 set_loop(new_inner_head, loop);
1449 set_loop(new_inner_exit, loop);
1450 set_idom(new_inner_head, idom(head), dom_depth(head));
1451 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1452 lazy_replace(head, new_inner_head);
1453 lazy_replace(exit_test, new_inner_exit);
1454 loop->_head = new_inner_head;
1455 return new_inner_head;
1456 }
1457
1458 #ifdef ASSERT
1459 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1460 Node* back_control = loop_exit_control(x, loop);
1461 assert(back_control != nullptr, "no back control");
1462
1463 BoolTest::mask mask = BoolTest::illegal;
1464 float cl_prob = 0;
1465 Node* incr = nullptr;
1466 Node* limit = nullptr;
1467
1468 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1469 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1470
1471 Node* phi_incr = nullptr;
1472 incr = loop_iv_incr(incr, x, loop, phi_incr);
1473 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1474
1475 Node* xphi = nullptr;
1476 Node* stride = loop_iv_stride(incr, loop, xphi);
1477
1478 assert(stride != nullptr, "no stride");
1479
1480 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1481
1482 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1483
1484 jlong stride_con = stride->get_integer_as_long(bt);
1485
1486 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1487
1488 assert(mask != BoolTest::ne, "unexpected condition");
1489 assert(phi_incr == nullptr, "bad loop shape");
1490 assert(cmp->in(1) == incr, "bad exit test shape");
1491
1492 // Safepoint on backedge not supported
1493 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1494 }
1495 #endif
1496
1497 #ifdef ASSERT
1498 // convert an int counted loop to a long counted to stress handling of
1499 // long counted loops
1500 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1501 Unique_Node_List iv_nodes;
1502 Node_List old_new;
1503 iv_nodes.push(cmp);
1504 bool failed = false;
1505
1506 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1507 Node* n = iv_nodes.at(i);
1508 switch(n->Opcode()) {
1509 case Op_Phi: {
1510 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1511 old_new.map(n->_idx, clone);
1512 break;
1513 }
1514 case Op_CmpI: {
1515 Node* clone = new CmpLNode(nullptr, nullptr);
1516 old_new.map(n->_idx, clone);
1517 break;
1518 }
1519 case Op_AddI: {
1520 Node* clone = new AddLNode(nullptr, nullptr);
1521 old_new.map(n->_idx, clone);
1522 break;
1523 }
1524 case Op_CastII: {
1525 failed = true;
1526 break;
1527 }
1528 default:
1529 DEBUG_ONLY(n->dump());
1530 fatal("unexpected");
1531 }
1532
1533 for (uint i = 1; i < n->req(); i++) {
1534 Node* in = n->in(i);
1535 if (in == nullptr) {
1536 continue;
1537 }
1538 if (loop->is_member(get_loop(get_ctrl(in)))) {
1539 iv_nodes.push(in);
1540 }
1541 }
1542 }
1543
1544 if (failed) {
1545 for (uint i = 0; i < iv_nodes.size(); i++) {
1546 Node* n = iv_nodes.at(i);
1547 Node* clone = old_new[n->_idx];
1548 if (clone != nullptr) {
1549 _igvn.remove_dead_node(clone);
1550 }
1551 }
1552 return false;
1553 }
1554
1555 for (uint i = 0; i < iv_nodes.size(); i++) {
1556 Node* n = iv_nodes.at(i);
1557 Node* clone = old_new[n->_idx];
1558 for (uint i = 1; i < n->req(); i++) {
1559 Node* in = n->in(i);
1560 if (in == nullptr) {
1561 continue;
1562 }
1563 Node* in_clone = old_new[in->_idx];
1564 if (in_clone == nullptr) {
1565 assert(_igvn.type(in)->isa_int(), "");
1566 in_clone = new ConvI2LNode(in);
1567 _igvn.register_new_node_with_optimizer(in_clone);
1568 set_subtree_ctrl(in_clone, false);
1569 }
1570 if (in_clone->in(0) == nullptr) {
1571 in_clone->set_req(0, C->top());
1572 clone->set_req(i, in_clone);
1573 in_clone->set_req(0, nullptr);
1574 } else {
1575 clone->set_req(i, in_clone);
1576 }
1577 }
1578 _igvn.register_new_node_with_optimizer(clone);
1579 }
1580 set_ctrl(old_new[phi->_idx], phi->in(0));
1581
1582 for (uint i = 0; i < iv_nodes.size(); i++) {
1583 Node* n = iv_nodes.at(i);
1584 Node* clone = old_new[n->_idx];
1585 set_subtree_ctrl(clone, false);
1586 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1587 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1588 Node* u = n->fast_out(i);
1589 if (iv_nodes.member(u)) {
1590 continue;
1591 }
1592 if (m == nullptr) {
1593 m = new ConvL2INode(clone);
1594 _igvn.register_new_node_with_optimizer(m);
1595 set_subtree_ctrl(m, false);
1596 }
1597 _igvn.rehash_node_delayed(u);
1598 int nb = u->replace_edge(n, m, &_igvn);
1599 --i, imax -= nb;
1600 }
1601 }
1602 return true;
1603 }
1604 #endif
1605
1606 //------------------------------is_counted_loop--------------------------------
1607 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {
1608 PhaseGVN *gvn = &_igvn;
1609
1610 Node* back_control = loop_exit_control(x, loop);
1611 if (back_control == nullptr) {
1612 return false;
1613 }
1614
1615 BoolTest::mask bt = BoolTest::illegal;
1616 float cl_prob = 0;
1617 Node* incr = nullptr;
1618 Node* limit = nullptr;
1619 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1620 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1621 return false; // Avoid pointer & float & 64-bit compares
1622 }
1623
1624 // Trip-counter increment must be commutative & associative.
1625 if (incr->Opcode() == Op_Cast(iv_bt)) {
1626 incr = incr->in(1);
1627 }
1628
1629 Node* phi_incr = nullptr;
1630 incr = loop_iv_incr(incr, x, loop, phi_incr);
1631 if (incr == nullptr) {
1632 return false;
1633 }
1634
1635 Node* trunc1 = nullptr;
1636 Node* trunc2 = nullptr;
1637 const TypeInteger* iv_trunc_t = nullptr;
1638 Node* orig_incr = incr;
1639 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1640 return false; // Funny increment opcode
1641 }
1642 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1643
1644 Node* xphi = nullptr;
1645 Node* stride = loop_iv_stride(incr, loop, xphi);
1646
1647 if (stride == nullptr) {
1648 return false;
1649 }
1650
1651 if (xphi->Opcode() == Op_Cast(iv_bt)) {
1652 xphi = xphi->in(1);
1653 }
1654
1655 // Stride must be constant
1656 jlong stride_con = stride->get_integer_as_long(iv_bt);
1657 assert(stride_con != 0, "missed some peephole opt");
1658
1659 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1660
1661 if (phi == nullptr ||
1662 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1663 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1664 return false;
1665 }
1666
1667 Node* iftrue = back_control;
1668 uint iftrue_op = iftrue->Opcode();
1669 Node* iff = iftrue->in(0);
1670 BoolNode* test = iff->in(1)->as_Bool();
1671
1672 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1673 if (trunc1 != nullptr) {
1674 // When there is a truncation, we must be sure that after the truncation
1675 // the trip counter will end up higher than the limit, otherwise we are looking
1676 // at an endless loop. Can happen with range checks.
1677
1678 // Example:
1679 // int i = 0;
1680 // while (true)
1681 // sum + = array[i];
1682 // i++;
1683 // i = i && 0x7fff;
1684 // }
1685 //
1686 // If the array is shorter than 0x8000 this exits through a AIOOB
1687 // - Counted loop transformation is ok
1688 // If the array is longer then this is an endless loop
1689 // - No transformation can be done.
1690
1691 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1692 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1693 // if the limit can have a higher value than the increment (before the phi)
1694 return false;
1695 }
1696 }
1697
1698 Node *init_trip = phi->in(LoopNode::EntryControl);
1699
1700 // If iv trunc type is smaller than int, check for possible wrap.
1701 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1702 assert(trunc1 != nullptr, "must have found some truncation");
1703
1704 // Get a better type for the phi (filtered thru if's)
1705 const TypeInteger* phi_ft = filtered_type(phi);
1706
1707 // Can iv take on a value that will wrap?
1708 //
1709 // Ensure iv's limit is not within "stride" of the wrap value.
1710 //
1711 // Example for "short" type
1712 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
1713 // If the stride is +10, then the last value of the induction
1714 // variable before the increment (phi_ft->_hi) must be
1715 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
1716 // ensure no truncation occurs after the increment.
1717
1718 if (stride_con > 0) {
1719 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
1720 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
1721 return false; // truncation may occur
1722 }
1723 } else if (stride_con < 0) {
1724 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
1725 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
1726 return false; // truncation may occur
1727 }
1728 }
1729 // No possibility of wrap so truncation can be discarded
1730 // Promote iv type to Int
1731 } else {
1732 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
1733 }
1734
1735 if (!condition_stride_ok(bt, stride_con)) {
1736 return false;
1737 }
1738
1739 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
1740
1741 if (stride_con > 0) {
1742 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
1743 return false; // cyclic loop
1744 }
1745 } else {
1746 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
1747 return false; // cyclic loop
1748 }
1749 }
1750
1751 if (phi_incr != nullptr && bt != BoolTest::ne) {
1752 // check if there is a possibility of IV overflowing after the first increment
1753 if (stride_con > 0) {
1754 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
1755 return false;
1756 }
1757 } else {
1758 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
1759 return false;
1760 }
1761 }
1762 }
1763
1764 // =================================================
1765 // ---- SUCCESS! Found A Trip-Counted Loop! -----
1766 //
1767
1768 if (x->Opcode() == Op_Region) {
1769 // x has not yet been transformed to Loop or LongCountedLoop.
1770 // This should only happen if we are inside an infinite loop.
1771 // It happens like this:
1772 // build_loop_tree -> do not attach infinite loop and nested loops
1773 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
1774 // build_loop_tree -> find and attach infinite and nested loops
1775 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
1776 assert(x->as_Region()->is_in_infinite_subgraph(),
1777 "x can only be a Region and not Loop if inside infinite loop");
1778 // Come back later when Region is transformed to LoopNode
1779 return false;
1780 }
1781
1782 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
1783 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
1784
1785 // ===================================================
1786 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
1787 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
1788 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
1789 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
1790 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
1791 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
1792 //
1793 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
1794 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
1795 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
1796 //
1797 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
1798 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
1799 //
1800 //
1801 // The loop to be converted does not always need to have the often used shape:
1802 //
1803 // i = init
1804 // i = init loop:
1805 // do { ...
1806 // // ... equivalent i+=stride
1807 // i+=stride <==> if (i < limit)
1808 // } while (i < limit); goto loop
1809 // exit:
1810 // ...
1811 //
1812 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
1813 //
1814 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
1815 // in the loop exit check:
1816 //
1817 // i = init
1818 // loop:
1819 // ...
1820 // if (i <= limit)
1821 // i+=stride
1822 // goto loop
1823 // exit:
1824 // ...
1825 //
1826 // Let's define the following terms:
1827 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
1828 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
1829 //
1830 // The iv_pre_i and iv_post_i have the following relation:
1831 // iv_pre_i + stride = iv_post_i
1832 //
1833 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
1834 // iv_post_i < adjusted_limit
1835 //
1836 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
1837 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
1838 // -> adjusted_limit = limit.
1839 // (LE2) iv_post_i <= limit:
1840 // iv_post_i < limit + 1
1841 // -> adjusted limit = limit + 1
1842 // (LE3) iv_pre_i < limit:
1843 // iv_pre_i + stride < limit + stride
1844 // iv_post_i < limit + stride
1845 // -> adjusted_limit = limit + stride
1846 // (LE4) iv_pre_i <= limit:
1847 // iv_pre_i < limit + 1
1848 // iv_pre_i + stride < limit + stride + 1
1849 // iv_post_i < limit + stride + 1
1850 // -> adjusted_limit = limit + stride + 1
1851 //
1852 // Note that:
1853 // (AL) limit <= adjusted_limit.
1854 //
1855 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
1856 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
1857 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
1858 //
1859 // To prove (INV), we require the following two conditions/assumptions:
1860 // (i): adjusted_limit - 1 + stride <= max_int
1861 // (ii): init < limit
1862 //
1863 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
1864 // induction by assuming (i) and (ii).
1865 //
1866 // Proof by Induction
1867 // ------------------
1868 // > Base case (i = 1): We show that (INV) holds after the first iteration:
1869 // min_int <= iv_post_1 = init + stride <= max_int
1870 // Proof:
1871 // First, we note that (ii) implies
1872 // (iii) init <= limit - 1
1873 // max_int >= adjusted_limit - 1 + stride [using (i)]
1874 // >= limit - 1 + stride [using (AL)]
1875 // >= init + stride [using (iii)]
1876 // >= min_int [using stride > 0, no underflow]
1877 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
1878 //
1879 // Note that to prove the base case we need (i) and (ii).
1880 //
1881 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
1882 // min_int <= iv_post_j <= max_int
1883 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
1884 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
1885 // Proof:
1886 // If iv_post_j >= adjusted_limit:
1887 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
1888 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
1889 // If iv_post_j < adjusted_limit:
1890 // First, we note that:
1891 // (iv) iv_post_j <= adjusted_limit - 1
1892 // max_int >= adjusted_limit - 1 + stride [using (i)]
1893 // >= iv_post_j + stride [using (iv)]
1894 // >= min_int [using stride > 0, no underflow]
1895 //
1896 // Note that to prove the step case we only need (i).
1897 //
1898 // Thus, by assuming (i) and (ii), we proved (INV).
1899 //
1900 //
1901 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
1902 //
1903 // (1) Loop Limit Check Predicate for (i):
1904 // Using (i): adjusted_limit - 1 + stride <= max_int
1905 //
1906 // This condition is now restated to use limit instead of adjusted_limit:
1907 //
1908 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
1909 // max_int - stride + 1 >= adjusted_limit
1910 // We can merge the two constants into
1911 // canonicalized_correction = stride - 1
1912 // which gives us
1913 // max_int - canonicalized_correction >= adjusted_limit
1914 //
1915 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
1916 // adjusted_limit = limit + limit_correction
1917 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
1918 // max_int - canonicalized_correction - limit_correction >= limit
1919 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
1920 // (v) final_correction = canonicalized_correction + limit_correction
1921 //
1922 // which gives us:
1923 //
1924 // Final predicate condition:
1925 // max_int - final_correction >= limit
1926 //
1927 // However, we need to be careful that (v) does not over- or underflow.
1928 // We know that:
1929 // canonicalized_correction = stride - 1
1930 // and
1931 // limit_correction <= stride + 1
1932 // and thus
1933 // canonicalized_correction + limit_correction <= 2 * stride
1934 // To prevent an over- or underflow of (v), we must ensure that
1935 // 2 * stride <= max_int
1936 // which can safely be checked without over- or underflow with
1937 // (vi) stride != min_int AND abs(stride) <= max_int / 2
1938 //
1939 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
1940 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
1941 //
1942 // (2) Loop Limit Check Predicate for (ii):
1943 // Using (ii): init < limit
1944 //
1945 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
1946 // (2.1) type(init) < type(limit)
1947 // In this case, we know:
1948 // all possible values of init < all possible values of limit
1949 // and we can skip the predicate.
1950 //
1951 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
1952 // In this case, we do not need to re-check the condition and can skip the predicate.
1953 // This is often found for while- and for-loops which have the following shape:
1954 //
1955 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
1956 // i = init;
1957 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
1958 // do {
1959 // i += stride;
1960 // } while (i < limit);
1961 // }
1962 //
1963 // (2.3) init + stride <= max_int
1964 // In this case, there is no overflow of the iv phi after the first loop iteration.
1965 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
1966 // init < limit
1967 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
1968 // compile time that init + stride <= max_int then we have trivially proven the base case and that
1969 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
1970 // again and can skip the predicate.
1971
1972 // Check (vi) and bail out if the stride is too big.
1973 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
1974 return false;
1975 }
1976
1977 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
1978 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
1979
1980 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
1981 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
1982 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
1983
1984 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
1985 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
1986 const jlong final_correction = canonicalized_correction + limit_correction;
1987
1988 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
1989 Node* init_control = x->in(LoopNode::EntryControl);
1990
1991 // If sov==0, limit's type always satisfies the condition, for
1992 // example, when it is an array length.
1993 if (sov != 0) {
1994 if (sov < 0) {
1995 return false; // Bailout: integer overflow is certain.
1996 }
1997 // (1) Loop Limit Check Predicate is required because we could not statically prove that
1998 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
1999 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2000 const Predicates predicates(init_control);
2001 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2002 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2003 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2004 #ifdef ASSERT
2005 if (TraceLoopLimitCheck) {
2006 tty->print("Missing Loop Limit Check Parse Predicate:");
2007 loop->dump_head();
2008 x->dump(1);
2009 }
2010 #endif
2011 return false;
2012 }
2013
2014 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2015 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2016 return false;
2017 }
2018
2019 Node* cmp_limit;
2020 Node* bol;
2021
2022 if (stride_con > 0) {
2023 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2024 bol = new BoolNode(cmp_limit, BoolTest::le);
2025 } else {
2026 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2027 bol = new BoolNode(cmp_limit, BoolTest::ge);
2028 }
2029
2030 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2031 }
2032
2033 // (2.3)
2034 const bool init_plus_stride_could_overflow =
2035 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2036 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2037 // (2.1)
2038 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2039 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2040
2041 if (init_gte_limit && // (2.1)
2042 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2043 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2044 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2045 // We use the following condition:
2046 // - stride > 0: init < limit
2047 // - stride < 0: init > limit
2048 //
2049 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2050 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2051 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2052
2053 const Predicates predicates(init_control);
2054 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2055 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2056 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2057 #ifdef ASSERT
2058 if (TraceLoopLimitCheck) {
2059 tty->print("Missing Loop Limit Check Parse Predicate:");
2060 loop->dump_head();
2061 x->dump(1);
2062 }
2063 #endif
2064 return false;
2065 }
2066
2067 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2068 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2069 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2070 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2071 return false;
2072 }
2073
2074 Node* cmp_limit;
2075 Node* bol;
2076
2077 if (stride_con > 0) {
2078 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2079 bol = new BoolNode(cmp_limit, BoolTest::lt);
2080 } else {
2081 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2082 bol = new BoolNode(cmp_limit, BoolTest::gt);
2083 }
2084
2085 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2086 }
2087
2088 if (bt == BoolTest::ne) {
2089 // Now we need to canonicalize the loop condition if it is 'ne'.
2090 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2091 if (stride_con > 0) {
2092 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2093 bt = BoolTest::lt;
2094 } else {
2095 assert(stride_con < 0, "must be");
2096 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2097 bt = BoolTest::gt;
2098 }
2099 }
2100
2101 Node* sfpt = nullptr;
2102 if (loop->_child == nullptr) {
2103 sfpt = find_safepoint(back_control, x, loop);
2104 } else {
2105 sfpt = iff->in(0);
2106 if (sfpt->Opcode() != Op_SafePoint) {
2107 sfpt = nullptr;
2108 }
2109 }
2110
2111 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2112 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2113 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2114 iv_bt == T_LONG) &&
2115 sfpt == nullptr) {
2116 // Leaving the safepoint on the backedge and creating a
2117 // CountedLoop will confuse optimizations. We can't move the
2118 // safepoint around because its jvm state wouldn't match a new
2119 // location. Give up on that loop.
2120 return false;
2121 }
2122 if (is_deleteable_safept(backedge_sfpt)) {
2123 lazy_replace(backedge_sfpt, iftrue);
2124 if (loop->_safepts != nullptr) {
2125 loop->_safepts->yank(backedge_sfpt);
2126 }
2127 loop->_tail = iftrue;
2128 }
2129 }
2130
2131
2132 #ifdef ASSERT
2133 if (iv_bt == T_INT &&
2134 !x->as_Loop()->is_loop_nest_inner_loop() &&
2135 StressLongCountedLoop > 0 &&
2136 trunc1 == nullptr &&
2137 convert_to_long_loop(cmp, phi, loop)) {
2138 return false;
2139 }
2140 #endif
2141
2142 Node* adjusted_limit = limit;
2143 if (phi_incr != nullptr) {
2144 // If compare points directly to the phi we need to adjust
2145 // the compare so that it points to the incr. Limit have
2146 // to be adjusted to keep trip count the same and we
2147 // should avoid int overflow.
2148 //
2149 // i = init; do {} while(i++ < limit);
2150 // is converted to
2151 // i = init; do {} while(++i < limit+1);
2152 //
2153 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2154 }
2155
2156 if (includes_limit) {
2157 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2158 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2159 //
2160 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2161 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2162 if (bt == BoolTest::le)
2163 bt = BoolTest::lt;
2164 else if (bt == BoolTest::ge)
2165 bt = BoolTest::gt;
2166 else
2167 ShouldNotReachHere();
2168 }
2169 set_subtree_ctrl(adjusted_limit, false);
2170
2171 // Build a canonical trip test.
2172 // Clone code, as old values may be in use.
2173 incr = incr->clone();
2174 incr->set_req(1,phi);
2175 incr->set_req(2,stride);
2176 incr = _igvn.register_new_node_with_optimizer(incr);
2177 set_early_ctrl(incr, false);
2178 _igvn.rehash_node_delayed(phi);
2179 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2180
2181 // If phi type is more restrictive than Int, raise to
2182 // Int to prevent (almost) infinite recursion in igvn
2183 // which can only handle integer types for constants or minint..maxint.
2184 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2185 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2186 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2187 nphi = _igvn.register_new_node_with_optimizer(nphi);
2188 set_ctrl(nphi, get_ctrl(phi));
2189 _igvn.replace_node(phi, nphi);
2190 phi = nphi->as_Phi();
2191 }
2192 cmp = cmp->clone();
2193 cmp->set_req(1,incr);
2194 cmp->set_req(2, adjusted_limit);
2195 cmp = _igvn.register_new_node_with_optimizer(cmp);
2196 set_ctrl(cmp, iff->in(0));
2197
2198 test = test->clone()->as_Bool();
2199 (*(BoolTest*)&test->_test)._test = bt;
2200 test->set_req(1,cmp);
2201 _igvn.register_new_node_with_optimizer(test);
2202 set_ctrl(test, iff->in(0));
2203
2204 // Replace the old IfNode with a new LoopEndNode
2205 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2206 IfNode *le = lex->as_If();
2207 uint dd = dom_depth(iff);
2208 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2209 set_loop(le, loop);
2210
2211 // Get the loop-exit control
2212 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2213
2214 // Need to swap loop-exit and loop-back control?
2215 if (iftrue_op == Op_IfFalse) {
2216 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2217 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2218
2219 loop->_tail = back_control = ift2;
2220 set_loop(ift2, loop);
2221 set_loop(iff2, get_loop(iffalse));
2222
2223 // Lazy update of 'get_ctrl' mechanism.
2224 lazy_replace(iffalse, iff2);
2225 lazy_replace(iftrue, ift2);
2226
2227 // Swap names
2228 iffalse = iff2;
2229 iftrue = ift2;
2230 } else {
2231 _igvn.rehash_node_delayed(iffalse);
2232 _igvn.rehash_node_delayed(iftrue);
2233 iffalse->set_req_X( 0, le, &_igvn );
2234 iftrue ->set_req_X( 0, le, &_igvn );
2235 }
2236
2237 set_idom(iftrue, le, dd+1);
2238 set_idom(iffalse, le, dd+1);
2239 assert(iff->outcnt() == 0, "should be dead now");
2240 lazy_replace( iff, le ); // fix 'get_ctrl'
2241
2242 Node* entry_control = init_control;
2243 bool strip_mine_loop = iv_bt == T_INT &&
2244 loop->_child == nullptr &&
2245 sfpt != nullptr &&
2246 !loop->_has_call &&
2247 is_deleteable_safept(sfpt);
2248 IdealLoopTree* outer_ilt = nullptr;
2249 if (strip_mine_loop) {
2250 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
2251 cl_prob, le->_fcnt, entry_control,
2252 iffalse);
2253 }
2254
2255 // Now setup a new CountedLoopNode to replace the existing LoopNode
2256 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2257 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2258 // The following assert is approximately true, and defines the intention
2259 // of can_be_counted_loop. It fails, however, because phase->type
2260 // is not yet initialized for this loop and its parts.
2261 //assert(l->can_be_counted_loop(this), "sanity");
2262 _igvn.register_new_node_with_optimizer(l);
2263 set_loop(l, loop);
2264 loop->_head = l;
2265 // Fix all data nodes placed at the old loop head.
2266 // Uses the lazy-update mechanism of 'get_ctrl'.
2267 lazy_replace( x, l );
2268 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2269
2270 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2271 // Check for immediately preceding SafePoint and remove
2272 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2273 if (strip_mine_loop) {
2274 Node* outer_le = outer_ilt->_tail->in(0);
2275 Node* sfpt_clone = sfpt->clone();
2276 sfpt_clone->set_req(0, iffalse);
2277 outer_le->set_req(0, sfpt_clone);
2278
2279 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2280 if (polladdr != nullptr && polladdr->is_Load()) {
2281 // Polling load should be pinned outside inner loop.
2282 Node* new_polladdr = polladdr->clone();
2283 new_polladdr->set_req(0, iffalse);
2284 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2285 set_ctrl(new_polladdr, iffalse);
2286 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2287 }
2288 // When this code runs, loop bodies have not yet been populated.
2289 const bool body_populated = false;
2290 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2291 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2292 }
2293 lazy_replace(sfpt, sfpt->in(TypeFunc::Control));
2294 if (loop->_safepts != nullptr) {
2295 loop->_safepts->yank(sfpt);
2296 }
2297 }
2298 }
2299
2300 #ifdef ASSERT
2301 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2302 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2303 #endif
2304 #ifndef PRODUCT
2305 if (TraceLoopOpts) {
2306 tty->print("Counted ");
2307 loop->dump_head();
2308 }
2309 #endif
2310
2311 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2312
2313 // Capture bounds of the loop in the induction variable Phi before
2314 // subsequent transformation (iteration splitting) obscures the
2315 // bounds
2316 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2317
2318 if (strip_mine_loop) {
2319 l->mark_strip_mined();
2320 l->verify_strip_mined(1);
2321 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2322 loop = outer_ilt;
2323 }
2324
2325 #ifndef PRODUCT
2326 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2327 Atomic::inc(&_long_loop_counted_loops);
2328 }
2329 #endif
2330 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2331 l->mark_loop_nest_outer_loop();
2332 }
2333
2334 return true;
2335 }
2336
2337 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2338 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2339 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2340 const BasicType iv_bt, Node* loop_entry) {
2341 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2342 // successfully find a dominated test with the If node below.
2343 Node* cmp_limit;
2344 Node* bol;
2345 if (stride_con > 0) {
2346 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2347 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2348 } else {
2349 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2350 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2351 }
2352
2353 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2354 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2355 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2356 IfFalseNode* if_false = new IfFalseNode(iff);
2357 IfTrueNode* if_true = new IfTrueNode(iff);
2358 Node* dominated_iff = _igvn.transform(iff);
2359 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2360 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2361
2362 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2363 _igvn.replace_input_of(iff, 0, C->top());
2364 _igvn.replace_input_of(iff, 1, C->top());
2365 return found_dominating_test;
2366 }
2367
2368 //----------------------exact_limit-------------------------------------------
2369 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2370 assert(loop->_head->is_CountedLoop(), "");
2371 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2372 assert(cl->is_valid_counted_loop(T_INT), "");
2373
2374 if (cl->stride_con() == 1 ||
2375 cl->stride_con() == -1 ||
2376 cl->limit()->Opcode() == Op_LoopLimit) {
2377 // Old code has exact limit (it could be incorrect in case of int overflow).
2378 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2379 return cl->limit();
2380 }
2381 Node *limit = nullptr;
2382 #ifdef ASSERT
2383 BoolTest::mask bt = cl->loopexit()->test_trip();
2384 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2385 #endif
2386 if (cl->has_exact_trip_count()) {
2387 // Simple case: loop has constant boundaries.
2388 // Use jlongs to avoid integer overflow.
2389 int stride_con = cl->stride_con();
2390 jlong init_con = cl->init_trip()->get_int();
2391 jlong limit_con = cl->limit()->get_int();
2392 julong trip_cnt = cl->trip_count();
2393 jlong final_con = init_con + trip_cnt*stride_con;
2394 int final_int = (int)final_con;
2395 // The final value should be in integer range since the loop
2396 // is counted and the limit was checked for overflow.
2397 assert(final_con == (jlong)final_int, "final value should be integer");
2398 limit = _igvn.intcon(final_int);
2399 } else {
2400 // Create new LoopLimit node to get exact limit (final iv value).
2401 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2402 register_new_node(limit, cl->in(LoopNode::EntryControl));
2403 }
2404 assert(limit != nullptr, "sanity");
2405 return limit;
2406 }
2407
2408 //------------------------------Ideal------------------------------------------
2409 // Return a node which is more "ideal" than the current node.
2410 // Attempt to convert into a counted-loop.
2411 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2412 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2413 phase->C->set_major_progress();
2414 }
2415 return RegionNode::Ideal(phase, can_reshape);
2416 }
2417
2418 #ifdef ASSERT
2419 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2420 const OuterStripMinedLoopNode* outer = nullptr;
2421 const CountedLoopNode* inner = nullptr;
2422 if (is_strip_mined()) {
2423 if (!is_valid_counted_loop(T_INT)) {
2424 return; // Skip malformed counted loop
2425 }
2426 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2427 inner = as_CountedLoop();
2428 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2429 } else if (is_OuterStripMinedLoop()) {
2430 outer = this->as_OuterStripMinedLoop();
2431 inner = outer->unique_ctrl_out()->as_CountedLoop();
2432 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2433 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2434 }
2435 if (inner != nullptr || outer != nullptr) {
2436 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2437 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2438 Node* outer_le = outer_tail->in(0);
2439 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2440 Node* sfpt = outer_le->in(0);
2441 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2442 Node* inner_out = sfpt->in(0);
2443 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2444 assert(cle == inner->loopexit_or_null(), "mismatch");
2445 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2446 if (has_skeleton) {
2447 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2448 assert(outer->outcnt() == 2, "only control nodes");
2449 } else {
2450 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2451 uint phis = 0;
2452 uint be_loads = 0;
2453 Node* be = inner->in(LoopNode::LoopBackControl);
2454 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2455 Node* u = inner->fast_out(i);
2456 if (u->is_Phi()) {
2457 phis++;
2458 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2459 Node* n = be->fast_out(j);
2460 if (n->is_Load()) {
2461 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2462 do {
2463 n = n->raw_out(0);
2464 } while (!n->is_Phi());
2465 if (n == u) {
2466 be_loads++;
2467 break;
2468 }
2469 }
2470 }
2471 }
2472 }
2473 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2474 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2475 Node* u = outer->fast_out(i);
2476 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2477 }
2478 uint stores = 0;
2479 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2480 Node* u = inner_out->fast_out(i);
2481 if (u->is_Store()) {
2482 stores++;
2483 }
2484 }
2485 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2486 // not guaranteed to be optimized out.
2487 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2488 }
2489 assert(sfpt->outcnt() == 1, "no data node");
2490 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2491 }
2492 }
2493 #endif
2494
2495 //=============================================================================
2496 //------------------------------Ideal------------------------------------------
2497 // Return a node which is more "ideal" than the current node.
2498 // Attempt to convert into a counted-loop.
2499 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2500 return RegionNode::Ideal(phase, can_reshape);
2501 }
2502
2503 //------------------------------dump_spec--------------------------------------
2504 // Dump special per-node info
2505 #ifndef PRODUCT
2506 void CountedLoopNode::dump_spec(outputStream *st) const {
2507 LoopNode::dump_spec(st);
2508 if (stride_is_con()) {
2509 st->print("stride: %d ",stride_con());
2510 }
2511 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2512 if (is_main_loop()) st->print("main of N%d", _idx);
2513 if (is_post_loop()) st->print("post of N%d", _main_idx);
2514 if (is_strip_mined()) st->print(" strip mined");
2515 }
2516 #endif
2517
2518 //=============================================================================
2519 jlong BaseCountedLoopEndNode::stride_con() const {
2520 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2521 }
2522
2523
2524 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2525 if (bt == T_INT) {
2526 return new CountedLoopEndNode(control, test, prob, cnt);
2527 }
2528 assert(bt == T_LONG, "unsupported");
2529 return new LongCountedLoopEndNode(control, test, prob, cnt);
2530 }
2531
2532 //=============================================================================
2533 //------------------------------Value-----------------------------------------
2534 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2535 const Type* init_t = phase->type(in(Init));
2536 const Type* limit_t = phase->type(in(Limit));
2537 const Type* stride_t = phase->type(in(Stride));
2538 // Either input is TOP ==> the result is TOP
2539 if (init_t == Type::TOP) return Type::TOP;
2540 if (limit_t == Type::TOP) return Type::TOP;
2541 if (stride_t == Type::TOP) return Type::TOP;
2542
2543 int stride_con = stride_t->is_int()->get_con();
2544 if (stride_con == 1)
2545 return bottom_type(); // Identity
2546
2547 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2548 // Use jlongs to avoid integer overflow.
2549 jlong init_con = init_t->is_int()->get_con();
2550 jlong limit_con = limit_t->is_int()->get_con();
2551 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2552 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2553 jlong final_con = init_con + stride_con*trip_count;
2554 int final_int = (int)final_con;
2555 // The final value should be in integer range in almost all cases,
2556 // since the loop is counted and the limit was checked for overflow.
2557 // There some exceptions, for example:
2558 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2559 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2560 // the loop and its predicates, and we might get constants on one side of the phi that
2561 // would lead to overflows. Such a code path would never lead us to enter the loop
2562 // because of the loop limit overflow check that happens after the LoopLimitNode
2563 // computation with overflow, but before we enter the loop, see JDK-8335747.
2564 if (final_con == (jlong)final_int) {
2565 return TypeInt::make(final_int);
2566 } else {
2567 return bottom_type();
2568 }
2569 }
2570
2571 return bottom_type(); // TypeInt::INT
2572 }
2573
2574 //------------------------------Ideal------------------------------------------
2575 // Return a node which is more "ideal" than the current node.
2576 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2577 if (phase->type(in(Init)) == Type::TOP ||
2578 phase->type(in(Limit)) == Type::TOP ||
2579 phase->type(in(Stride)) == Type::TOP)
2580 return nullptr; // Dead
2581
2582 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2583 if (stride_con == 1)
2584 return nullptr; // Identity
2585
2586 // Delay following optimizations until all loop optimizations
2587 // done to keep Ideal graph simple.
2588 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2589 phase->C->record_for_post_loop_opts_igvn(this);
2590 return nullptr;
2591 }
2592
2593 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2594 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2595 jlong stride_p;
2596 jlong lim, ini;
2597 julong max;
2598 if (stride_con > 0) {
2599 stride_p = stride_con;
2600 lim = limit_t->_hi;
2601 ini = init_t->_lo;
2602 max = (julong)max_jint;
2603 } else {
2604 stride_p = -(jlong)stride_con;
2605 lim = init_t->_hi;
2606 ini = limit_t->_lo;
2607 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2608 }
2609 julong range = lim - ini + stride_p;
2610 if (range <= max) {
2611 // Convert to integer expression if it is not overflow.
2612 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2613 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2614 Node *bias = phase->transform(new AddINode(range, stride_m));
2615 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2616 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2617 return new AddINode(span, in(Init)); // exact limit
2618 }
2619
2620 if (is_power_of_2(stride_p) || // divisor is 2^n
2621 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2622 // Convert to long expression to avoid integer overflow
2623 // and let igvn optimizer convert this division.
2624 //
2625 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2626 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2627 Node* stride = phase->longcon(stride_con);
2628 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2629
2630 Node *range = phase->transform(new SubLNode(limit, init));
2631 Node *bias = phase->transform(new AddLNode(range, stride_m));
2632 Node *span;
2633 if (stride_con > 0 && is_power_of_2(stride_p)) {
2634 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2635 // and avoid generating rounding for division. Zero trip guard should
2636 // guarantee that init < limit but sometimes the guard is missing and
2637 // we can get situation when init > limit. Note, for the empty loop
2638 // optimization zero trip guard is generated explicitly which leaves
2639 // only RCE predicate where exact limit is used and the predicate
2640 // will simply fail forcing recompilation.
2641 Node* neg_stride = phase->longcon(-stride_con);
2642 span = phase->transform(new AndLNode(bias, neg_stride));
2643 } else {
2644 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2645 span = phase->transform(new MulLNode(trip, stride));
2646 }
2647 // Convert back to int
2648 Node *span_int = phase->transform(new ConvL2INode(span));
2649 return new AddINode(span_int, in(Init)); // exact limit
2650 }
2651
2652 return nullptr; // No progress
2653 }
2654
2655 //------------------------------Identity---------------------------------------
2656 // If stride == 1 return limit node.
2657 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2658 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2659 if (stride_con == 1 || stride_con == -1)
2660 return in(Limit);
2661 return this;
2662 }
2663
2664 //=============================================================================
2665 //----------------------match_incr_with_optional_truncation--------------------
2666 // Match increment with optional truncation:
2667 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2668 // Return null for failure. Success returns the increment node.
2669 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2670 const TypeInteger** trunc_type,
2671 BasicType bt) {
2672 // Quick cutouts:
2673 if (expr == nullptr || expr->req() != 3) return nullptr;
2674
2675 Node *t1 = nullptr;
2676 Node *t2 = nullptr;
2677 Node* n1 = expr;
2678 int n1op = n1->Opcode();
2679 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2680
2681 if (bt == T_INT) {
2682 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2683 if (n1op == Op_AndI &&
2684 n1->in(2)->is_Con() &&
2685 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2686 // %%% This check should match any mask of 2**K-1.
2687 t1 = n1;
2688 n1 = t1->in(1);
2689 n1op = n1->Opcode();
2690 trunc_t = TypeInt::CHAR;
2691 } else if (n1op == Op_RShiftI &&
2692 n1->in(1) != nullptr &&
2693 n1->in(1)->Opcode() == Op_LShiftI &&
2694 n1->in(2) == n1->in(1)->in(2) &&
2695 n1->in(2)->is_Con()) {
2696 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2697 // %%% This check should match any shift in [1..31].
2698 if (shift == 16 || shift == 8) {
2699 t1 = n1;
2700 t2 = t1->in(1);
2701 n1 = t2->in(1);
2702 n1op = n1->Opcode();
2703 if (shift == 16) {
2704 trunc_t = TypeInt::SHORT;
2705 } else if (shift == 8) {
2706 trunc_t = TypeInt::BYTE;
2707 }
2708 }
2709 }
2710 }
2711
2712 // If (maybe after stripping) it is an AddI, we won:
2713 if (n1op == Op_Add(bt)) {
2714 *trunc1 = t1;
2715 *trunc2 = t2;
2716 *trunc_type = trunc_t;
2717 return n1;
2718 }
2719
2720 // failed
2721 return nullptr;
2722 }
2723
2724 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
2725 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
2726 verify_strip_mined(expect_skeleton);
2727 return in(EntryControl)->as_Loop();
2728 }
2729 return this;
2730 }
2731
2732 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
2733 assert(is_strip_mined(), "not a strip mined loop");
2734 Node* c = in(EntryControl);
2735 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
2736 return nullptr;
2737 }
2738 return c->as_OuterStripMinedLoop();
2739 }
2740
2741 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
2742 Node* c = in(LoopBackControl);
2743 if (c == nullptr || c->is_top()) {
2744 return nullptr;
2745 }
2746 return c->as_IfTrue();
2747 }
2748
2749 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
2750 LoopNode* l = outer_loop();
2751 if (l == nullptr) {
2752 return nullptr;
2753 }
2754 return l->outer_loop_tail();
2755 }
2756
2757 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
2758 IfTrueNode* proj = outer_loop_tail();
2759 if (proj == nullptr) {
2760 return nullptr;
2761 }
2762 Node* c = proj->in(0);
2763 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
2764 return nullptr;
2765 }
2766 return c->as_OuterStripMinedLoopEnd();
2767 }
2768
2769 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
2770 LoopNode* l = outer_loop();
2771 if (l == nullptr) {
2772 return nullptr;
2773 }
2774 return l->outer_loop_end();
2775 }
2776
2777 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
2778 IfNode* le = outer_loop_end();
2779 if (le == nullptr) {
2780 return nullptr;
2781 }
2782 Node* c = le->proj_out_or_null(false);
2783 if (c == nullptr) {
2784 return nullptr;
2785 }
2786 return c->as_IfFalse();
2787 }
2788
2789 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
2790 LoopNode* l = outer_loop();
2791 if (l == nullptr) {
2792 return nullptr;
2793 }
2794 return l->outer_loop_exit();
2795 }
2796
2797 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
2798 IfNode* le = outer_loop_end();
2799 if (le == nullptr) {
2800 return nullptr;
2801 }
2802 Node* c = le->in(0);
2803 if (c == nullptr || c->is_top()) {
2804 return nullptr;
2805 }
2806 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
2807 return c->as_SafePoint();
2808 }
2809
2810 SafePointNode* CountedLoopNode::outer_safepoint() const {
2811 LoopNode* l = outer_loop();
2812 if (l == nullptr) {
2813 return nullptr;
2814 }
2815 return l->outer_safepoint();
2816 }
2817
2818 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
2819 Node* ctrl = in(LoopNode::EntryControl);
2820 if (ctrl == nullptr) {
2821 // Dying loop.
2822 return nullptr;
2823 }
2824 if (is_main_loop()) {
2825 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
2826 }
2827 if (is_main_loop() || is_post_loop()) {
2828 AssertionPredicates assertion_predicates(ctrl);
2829 return assertion_predicates.entry();
2830 }
2831 return ctrl;
2832 }
2833
2834
2835 int CountedLoopNode::stride_con() const {
2836 CountedLoopEndNode* cle = loopexit_or_null();
2837 return cle != nullptr ? cle->stride_con() : 0;
2838 }
2839
2840 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
2841 if (bt == T_INT) {
2842 return new CountedLoopNode(entry, backedge);
2843 }
2844 assert(bt == T_LONG, "unsupported");
2845 return new LongCountedLoopNode(entry, backedge);
2846 }
2847
2848 void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
2849 PhaseIdealLoop* iloop) {
2850 Node* cle_out = inner_cle->proj_out(false);
2851 Node* cle_tail = inner_cle->proj_out(true);
2852 if (cle_out->outcnt() > 1) {
2853 // Look for chains of stores that were sunk
2854 // out of the inner loop and are in the outer loop
2855 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
2856 Node* u = cle_out->fast_out(i);
2857 if (u->is_Store()) {
2858 int alias_idx = igvn->C->get_alias_index(u->adr_type());
2859 Node* first = u;
2860 for (;;) {
2861 Node* next = first->in(MemNode::Memory);
2862 if (!next->is_Store() || next->in(0) != cle_out) {
2863 break;
2864 }
2865 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2866 first = next;
2867 }
2868 Node* last = u;
2869 for (;;) {
2870 Node* next = nullptr;
2871 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
2872 Node* uu = last->fast_out(j);
2873 if (uu->is_Store() && uu->in(0) == cle_out) {
2874 assert(next == nullptr, "only one in the outer loop");
2875 next = uu;
2876 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2877 }
2878 }
2879 if (next == nullptr) {
2880 break;
2881 }
2882 last = next;
2883 }
2884 Node* phi = nullptr;
2885 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2886 Node* uu = inner_cl->fast_out(j);
2887 if (uu->is_Phi()) {
2888 Node* be = uu->in(LoopNode::LoopBackControl);
2889 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
2890 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
2891 }
2892 if (be == last || be == first->in(MemNode::Memory)) {
2893 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
2894 assert(phi == nullptr, "only one phi");
2895 phi = uu;
2896 }
2897 }
2898 }
2899 #ifdef ASSERT
2900 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2901 Node* uu = inner_cl->fast_out(j);
2902 if (uu->is_memory_phi()) {
2903 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
2904 assert(phi == uu, "what's that phi?");
2905 } else if (uu->adr_type() == TypePtr::BOTTOM) {
2906 Node* n = uu->in(LoopNode::LoopBackControl);
2907 uint limit = igvn->C->live_nodes();
2908 uint i = 0;
2909 while (n != uu) {
2910 i++;
2911 assert(i < limit, "infinite loop");
2912 if (n->is_Proj()) {
2913 n = n->in(0);
2914 } else if (n->is_SafePoint() || n->is_MemBar()) {
2915 n = n->in(TypeFunc::Memory);
2916 } else if (n->is_Phi()) {
2917 n = n->in(1);
2918 } else if (n->is_MergeMem()) {
2919 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
2920 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
2921 n = n->in(MemNode::Memory);
2922 } else {
2923 n->dump();
2924 ShouldNotReachHere();
2925 }
2926 }
2927 }
2928 }
2929 }
2930 #endif
2931 if (phi == nullptr) {
2932 // If an entire chains was sunk, the
2933 // inner loop has no phi for that memory
2934 // slice, create one for the outer loop
2935 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
2936 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
2937 phi->set_req(LoopNode::LoopBackControl, last);
2938 phi = register_new_node(phi, inner_cl, igvn, iloop);
2939 igvn->replace_input_of(first, MemNode::Memory, phi);
2940 } else {
2941 // Or fix the outer loop fix to include
2942 // that chain of stores.
2943 Node* be = phi->in(LoopNode::LoopBackControl);
2944 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
2945 if (be == first->in(MemNode::Memory)) {
2946 if (be == phi->in(LoopNode::LoopBackControl)) {
2947 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
2948 } else {
2949 igvn->replace_input_of(be, MemNode::Memory, last);
2950 }
2951 } else {
2952 #ifdef ASSERT
2953 if (be == phi->in(LoopNode::LoopBackControl)) {
2954 assert(phi->in(LoopNode::LoopBackControl) == last, "");
2955 } else {
2956 assert(be->in(MemNode::Memory) == last, "");
2957 }
2958 #endif
2959 }
2960 }
2961 }
2962 }
2963 }
2964 }
2965
2966 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
2967 // Look for the outer & inner strip mined loop, reduce number of
2968 // iterations of the inner loop, set exit condition of outer loop,
2969 // construct required phi nodes for outer loop.
2970 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
2971 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
2972 if (LoopStripMiningIter == 0) {
2973 remove_outer_loop_and_safepoint(igvn);
2974 return;
2975 }
2976 if (LoopStripMiningIter == 1) {
2977 transform_to_counted_loop(igvn, nullptr);
2978 return;
2979 }
2980 Node* inner_iv_phi = inner_cl->phi();
2981 if (inner_iv_phi == nullptr) {
2982 IfNode* outer_le = outer_loop_end();
2983 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
2984 igvn->replace_node(outer_le, iff);
2985 inner_cl->clear_strip_mined();
2986 return;
2987 }
2988 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
2989
2990 int stride = inner_cl->stride_con();
2991 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
2992 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
2993 // stride, the method is guaranteed to return at the next check below.
2994 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
2995 int scaled_iters = (int)scaled_iters_long;
2996 if ((jlong)scaled_iters != scaled_iters_long) {
2997 // Remove outer loop and safepoint (too few iterations)
2998 remove_outer_loop_and_safepoint(igvn);
2999 return;
3000 }
3001 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3002 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3003 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3004 assert(iter_estimate > 0, "broken");
3005 if (iter_estimate <= short_scaled_iters) {
3006 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3007 remove_outer_loop_and_safepoint(igvn);
3008 return;
3009 }
3010 if (iter_estimate <= scaled_iters_long) {
3011 // We would only go through one iteration of
3012 // the outer loop: drop the outer loop but
3013 // keep the safepoint so we don't run for
3014 // too long without a safepoint
3015 IfNode* outer_le = outer_loop_end();
3016 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3017 igvn->replace_node(outer_le, iff);
3018 inner_cl->clear_strip_mined();
3019 return;
3020 }
3021
3022 Node* cle_tail = inner_cle->proj_out(true);
3023 ResourceMark rm;
3024 Node_List old_new;
3025 if (cle_tail->outcnt() > 1) {
3026 // Look for nodes on backedge of inner loop and clone them
3027 Unique_Node_List backedge_nodes;
3028 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3029 Node* u = cle_tail->fast_out(i);
3030 if (u != inner_cl) {
3031 assert(!u->is_CFG(), "control flow on the backedge?");
3032 backedge_nodes.push(u);
3033 }
3034 }
3035 uint last = igvn->C->unique();
3036 for (uint next = 0; next < backedge_nodes.size(); next++) {
3037 Node* n = backedge_nodes.at(next);
3038 old_new.map(n->_idx, n->clone());
3039 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3040 Node* u = n->fast_out(i);
3041 assert(!u->is_CFG(), "broken");
3042 if (u->_idx >= last) {
3043 continue;
3044 }
3045 if (!u->is_Phi()) {
3046 backedge_nodes.push(u);
3047 } else {
3048 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3049 }
3050 }
3051 }
3052 // Put the clones on the outer loop backedge
3053 Node* le_tail = outer_loop_tail();
3054 for (uint next = 0; next < backedge_nodes.size(); next++) {
3055 Node *n = old_new[backedge_nodes.at(next)->_idx];
3056 for (uint i = 1; i < n->req(); i++) {
3057 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3058 n->set_req(i, old_new[n->in(i)->_idx]);
3059 }
3060 }
3061 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3062 n->set_req(0, le_tail);
3063 }
3064 igvn->register_new_node_with_optimizer(n);
3065 }
3066 }
3067
3068 Node* iv_phi = nullptr;
3069 // Make a clone of each phi in the inner loop
3070 // for the outer loop
3071 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3072 Node* u = inner_cl->raw_out(i);
3073 if (u->is_Phi()) {
3074 assert(u->in(0) == inner_cl, "inconsistent");
3075 Node* phi = u->clone();
3076 phi->set_req(0, this);
3077 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3078 if (be != nullptr) {
3079 phi->set_req(LoopNode::LoopBackControl, be);
3080 }
3081 phi = igvn->transform(phi);
3082 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3083 if (u == inner_iv_phi) {
3084 iv_phi = phi;
3085 }
3086 }
3087 }
3088
3089 if (iv_phi != nullptr) {
3090 // Now adjust the inner loop's exit condition
3091 Node* limit = inner_cl->limit();
3092 // If limit < init for stride > 0 (or limit > init for stride < 0),
3093 // the loop body is run only once. Given limit - init (init - limit resp.)
3094 // would be negative, the unsigned comparison below would cause
3095 // the loop body to be run for LoopStripMiningIter.
3096 Node* max = nullptr;
3097 if (stride > 0) {
3098 max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3099 } else {
3100 max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3101 }
3102 // sub is positive and can be larger than the max signed int
3103 // value. Use an unsigned min.
3104 Node* const_iters = igvn->intcon(scaled_iters);
3105 Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3106 // min is the number of iterations for the next inner loop execution:
3107 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3108 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3109
3110 Node* new_limit = nullptr;
3111 if (stride > 0) {
3112 new_limit = igvn->transform(new AddINode(min, iv_phi));
3113 } else {
3114 new_limit = igvn->transform(new SubINode(iv_phi, min));
3115 }
3116 Node* inner_cmp = inner_cle->cmp_node();
3117 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3118 Node* outer_bol = inner_bol;
3119 // cmp node for inner loop may be shared
3120 inner_cmp = inner_cmp->clone();
3121 inner_cmp->set_req(2, new_limit);
3122 inner_bol = inner_bol->clone();
3123 inner_bol->set_req(1, igvn->transform(inner_cmp));
3124 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3125 // Set the outer loop's exit condition too
3126 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3127 } else {
3128 assert(false, "should be able to adjust outer loop");
3129 IfNode* outer_le = outer_loop_end();
3130 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3131 igvn->replace_node(outer_le, iff);
3132 inner_cl->clear_strip_mined();
3133 }
3134 }
3135
3136 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3137 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3138 CountedLoopEndNode* cle = inner_cl->loopexit();
3139 Node* inner_test = cle->in(1);
3140 IfNode* outer_le = outer_loop_end();
3141 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3142 Node* safepoint = outer_safepoint();
3143
3144 fix_sunk_stores(inner_cle, inner_cl, igvn, iloop);
3145
3146 // make counted loop exit test always fail
3147 ConINode* zero = igvn->intcon(0);
3148 if (iloop != nullptr) {
3149 iloop->set_root_as_ctrl(zero);
3150 }
3151 igvn->replace_input_of(cle, 1, zero);
3152 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3153 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3154 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3155 if (iloop == nullptr) {
3156 igvn->replace_node(outer_le, new_end);
3157 } else {
3158 iloop->lazy_replace(outer_le, new_end);
3159 }
3160 // the backedge of the inner loop must be rewired to the new loop end
3161 Node* backedge = cle->proj_out(true);
3162 igvn->replace_input_of(backedge, 0, new_end);
3163 if (iloop != nullptr) {
3164 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3165 }
3166 // make the outer loop go away
3167 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3168 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3169 inner_cl->clear_strip_mined();
3170 if (iloop != nullptr) {
3171 Unique_Node_List wq;
3172 wq.push(safepoint);
3173
3174 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3175 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3176
3177 for (uint i = 0; i < wq.size(); i++) {
3178 Node* n = wq.at(i);
3179 for (uint j = 0; j < n->req(); ++j) {
3180 Node* in = n->in(j);
3181 if (in == nullptr || in->is_CFG()) {
3182 continue;
3183 }
3184 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3185 continue;
3186 }
3187 assert(!loop->_body.contains(in), "");
3188 loop->_body.push(in);
3189 wq.push(in);
3190 }
3191 }
3192 iloop->set_loop(safepoint, loop);
3193 loop->_body.push(safepoint);
3194 iloop->set_loop(safepoint->in(0), loop);
3195 loop->_body.push(safepoint->in(0));
3196 outer_loop_ilt->_tail = igvn->C->top();
3197 }
3198 }
3199
3200 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3201 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3202 Node* outer_sfpt = outer_safepoint();
3203 Node* outer_out = outer_loop_exit();
3204 igvn->replace_node(outer_out, outer_sfpt->in(0));
3205 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3206 inner_cl->clear_strip_mined();
3207 }
3208
3209 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3210 if (iloop == nullptr) {
3211 return igvn->transform(node);
3212 }
3213 iloop->register_new_node(node, ctrl);
3214 return node;
3215 }
3216
3217 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3218 PhaseIdealLoop* iloop) {
3219 if (iloop == nullptr) {
3220 return igvn->transform(node);
3221 }
3222 iloop->register_control(node, iloop->get_loop(loop), idom);
3223 return node;
3224 }
3225
3226 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3227 if (!in(0)) return Type::TOP;
3228 if (phase->type(in(0)) == Type::TOP)
3229 return Type::TOP;
3230
3231 // Until expansion, the loop end condition is not set so this should not constant fold.
3232 if (is_expanded(phase)) {
3233 return IfNode::Value(phase);
3234 }
3235
3236 return TypeTuple::IFBOTH;
3237 }
3238
3239 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3240 // The outer strip mined loop head only has Phi uses after expansion
3241 if (phase->is_IterGVN()) {
3242 Node* backedge = proj_out_or_null(true);
3243 if (backedge != nullptr) {
3244 Node* head = backedge->unique_ctrl_out_or_null();
3245 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3246 if (head->find_out_with(Op_Phi) != nullptr) {
3247 return true;
3248 }
3249 }
3250 }
3251 }
3252 return false;
3253 }
3254
3255 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3256 if (remove_dead_region(phase, can_reshape)) return this;
3257
3258 return nullptr;
3259 }
3260
3261 //------------------------------filtered_type--------------------------------
3262 // Return a type based on condition control flow
3263 // A successful return will be a type that is restricted due
3264 // to a series of dominating if-tests, such as:
3265 // if (i < 10) {
3266 // if (i > 0) {
3267 // here: "i" type is [1..10)
3268 // }
3269 // }
3270 // or a control flow merge
3271 // if (i < 10) {
3272 // do {
3273 // phi( , ) -- at top of loop type is [min_int..10)
3274 // i = ?
3275 // } while ( i < 10)
3276 //
3277 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3278 assert(n && n->bottom_type()->is_int(), "must be int");
3279 const TypeInt* filtered_t = nullptr;
3280 if (!n->is_Phi()) {
3281 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3282 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3283
3284 } else {
3285 Node* phi = n->as_Phi();
3286 Node* region = phi->in(0);
3287 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3288 if (region && region != C->top()) {
3289 for (uint i = 1; i < phi->req(); i++) {
3290 Node* val = phi->in(i);
3291 Node* use_c = region->in(i);
3292 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3293 if (val_t != nullptr) {
3294 if (filtered_t == nullptr) {
3295 filtered_t = val_t;
3296 } else {
3297 filtered_t = filtered_t->meet(val_t)->is_int();
3298 }
3299 }
3300 }
3301 }
3302 }
3303 const TypeInt* n_t = _igvn.type(n)->is_int();
3304 if (filtered_t != nullptr) {
3305 n_t = n_t->join(filtered_t)->is_int();
3306 }
3307 return n_t;
3308 }
3309
3310
3311 //------------------------------filtered_type_from_dominators--------------------------------
3312 // Return a possibly more restrictive type for val based on condition control flow of dominators
3313 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3314 if (val->is_Con()) {
3315 return val->bottom_type()->is_int();
3316 }
3317 uint if_limit = 10; // Max number of dominating if's visited
3318 const TypeInt* rtn_t = nullptr;
3319
3320 if (use_ctrl && use_ctrl != C->top()) {
3321 Node* val_ctrl = get_ctrl(val);
3322 uint val_dom_depth = dom_depth(val_ctrl);
3323 Node* pred = use_ctrl;
3324 uint if_cnt = 0;
3325 while (if_cnt < if_limit) {
3326 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3327 if_cnt++;
3328 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3329 if (if_t != nullptr) {
3330 if (rtn_t == nullptr) {
3331 rtn_t = if_t;
3332 } else {
3333 rtn_t = rtn_t->join(if_t)->is_int();
3334 }
3335 }
3336 }
3337 pred = idom(pred);
3338 if (pred == nullptr || pred == C->top()) {
3339 break;
3340 }
3341 // Stop if going beyond definition block of val
3342 if (dom_depth(pred) < val_dom_depth) {
3343 break;
3344 }
3345 }
3346 }
3347 return rtn_t;
3348 }
3349
3350
3351 //------------------------------dump_spec--------------------------------------
3352 // Dump special per-node info
3353 #ifndef PRODUCT
3354 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3355 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3356 BoolTest bt( test_trip()); // Added this for g++.
3357
3358 st->print("[");
3359 bt.dump_on(st);
3360 st->print("]");
3361 }
3362 st->print(" ");
3363 IfNode::dump_spec(st);
3364 }
3365 #endif
3366
3367 //=============================================================================
3368 //------------------------------is_member--------------------------------------
3369 // Is 'l' a member of 'this'?
3370 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3371 while( l->_nest > _nest ) l = l->_parent;
3372 return l == this;
3373 }
3374
3375 //------------------------------set_nest---------------------------------------
3376 // Set loop tree nesting depth. Accumulate _has_call bits.
3377 int IdealLoopTree::set_nest( uint depth ) {
3378 assert(depth <= SHRT_MAX, "sanity");
3379 _nest = depth;
3380 int bits = _has_call;
3381 if( _child ) bits |= _child->set_nest(depth+1);
3382 if( bits ) _has_call = 1;
3383 if( _next ) bits |= _next ->set_nest(depth );
3384 return bits;
3385 }
3386
3387 //------------------------------split_fall_in----------------------------------
3388 // Split out multiple fall-in edges from the loop header. Move them to a
3389 // private RegionNode before the loop. This becomes the loop landing pad.
3390 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3391 PhaseIterGVN &igvn = phase->_igvn;
3392 uint i;
3393
3394 // Make a new RegionNode to be the landing pad.
3395 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3396 phase->set_loop(landing_pad,_parent);
3397 // If _head was irreducible loop entry, landing_pad may now be too
3398 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3399 // Gather all the fall-in control paths into the landing pad
3400 uint icnt = fall_in_cnt;
3401 uint oreq = _head->req();
3402 for( i = oreq-1; i>0; i-- )
3403 if( !phase->is_member( this, _head->in(i) ) )
3404 landing_pad->set_req(icnt--,_head->in(i));
3405
3406 // Peel off PhiNode edges as well
3407 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3408 Node *oj = _head->fast_out(j);
3409 if( oj->is_Phi() ) {
3410 PhiNode* old_phi = oj->as_Phi();
3411 assert( old_phi->region() == _head, "" );
3412 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3413 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3414 uint icnt = fall_in_cnt;
3415 for( i = oreq-1; i>0; i-- ) {
3416 if( !phase->is_member( this, _head->in(i) ) ) {
3417 p->init_req(icnt--, old_phi->in(i));
3418 // Go ahead and clean out old edges from old phi
3419 old_phi->del_req(i);
3420 }
3421 }
3422 // Search for CSE's here, because ZKM.jar does a lot of
3423 // loop hackery and we need to be a little incremental
3424 // with the CSE to avoid O(N^2) node blow-up.
3425 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3426 if( p2 ) { // Found CSE
3427 p->destruct(&igvn); // Recover useless new node
3428 p = p2; // Use old node
3429 } else {
3430 igvn.register_new_node_with_optimizer(p, old_phi);
3431 }
3432 // Make old Phi refer to new Phi.
3433 old_phi->add_req(p);
3434 // Check for the special case of making the old phi useless and
3435 // disappear it. In JavaGrande I have a case where this useless
3436 // Phi is the loop limit and prevents recognizing a CountedLoop
3437 // which in turn prevents removing an empty loop.
3438 Node *id_old_phi = old_phi->Identity(&igvn);
3439 if( id_old_phi != old_phi ) { // Found a simple identity?
3440 // Note that I cannot call 'replace_node' here, because
3441 // that will yank the edge from old_phi to the Region and
3442 // I'm mid-iteration over the Region's uses.
3443 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3444 Node* use = old_phi->last_out(i);
3445 igvn.rehash_node_delayed(use);
3446 uint uses_found = 0;
3447 for (uint j = 0; j < use->len(); j++) {
3448 if (use->in(j) == old_phi) {
3449 if (j < use->req()) use->set_req (j, id_old_phi);
3450 else use->set_prec(j, id_old_phi);
3451 uses_found++;
3452 }
3453 }
3454 i -= uses_found; // we deleted 1 or more copies of this edge
3455 }
3456 }
3457 igvn._worklist.push(old_phi);
3458 }
3459 }
3460 // Finally clean out the fall-in edges from the RegionNode
3461 for( i = oreq-1; i>0; i-- ) {
3462 if( !phase->is_member( this, _head->in(i) ) ) {
3463 _head->del_req(i);
3464 }
3465 }
3466 igvn.rehash_node_delayed(_head);
3467 // Transform landing pad
3468 igvn.register_new_node_with_optimizer(landing_pad, _head);
3469 // Insert landing pad into the header
3470 _head->add_req(landing_pad);
3471 }
3472
3473 //------------------------------split_outer_loop-------------------------------
3474 // Split out the outermost loop from this shared header.
3475 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3476 PhaseIterGVN &igvn = phase->_igvn;
3477
3478 // Find index of outermost loop; it should also be my tail.
3479 uint outer_idx = 1;
3480 while( _head->in(outer_idx) != _tail ) outer_idx++;
3481
3482 // Make a LoopNode for the outermost loop.
3483 Node *ctl = _head->in(LoopNode::EntryControl);
3484 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3485 outer = igvn.register_new_node_with_optimizer(outer, _head);
3486 phase->set_created_loop_node();
3487
3488 // Outermost loop falls into '_head' loop
3489 _head->set_req(LoopNode::EntryControl, outer);
3490 _head->del_req(outer_idx);
3491 // Split all the Phis up between '_head' loop and 'outer' loop.
3492 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3493 Node *out = _head->fast_out(j);
3494 if( out->is_Phi() ) {
3495 PhiNode *old_phi = out->as_Phi();
3496 assert( old_phi->region() == _head, "" );
3497 Node *phi = PhiNode::make_blank(outer, old_phi);
3498 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3499 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3500 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3501 // Make old Phi point to new Phi on the fall-in path
3502 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3503 old_phi->del_req(outer_idx);
3504 }
3505 }
3506
3507 // Use the new loop head instead of the old shared one
3508 _head = outer;
3509 phase->set_loop(_head, this);
3510 }
3511
3512 //------------------------------fix_parent-------------------------------------
3513 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3514 loop->_parent = parent;
3515 if( loop->_child ) fix_parent( loop->_child, loop );
3516 if( loop->_next ) fix_parent( loop->_next , parent );
3517 }
3518
3519 //------------------------------estimate_path_freq-----------------------------
3520 static float estimate_path_freq( Node *n ) {
3521 // Try to extract some path frequency info
3522 IfNode *iff;
3523 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3524 uint nop = n->Opcode();
3525 if( nop == Op_SafePoint ) { // Skip any safepoint
3526 n = n->in(0);
3527 continue;
3528 }
3529 if( nop == Op_CatchProj ) { // Get count from a prior call
3530 // Assume call does not always throw exceptions: means the call-site
3531 // count is also the frequency of the fall-through path.
3532 assert( n->is_CatchProj(), "" );
3533 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3534 return 0.0f; // Assume call exception path is rare
3535 Node *call = n->in(0)->in(0)->in(0);
3536 assert( call->is_Call(), "expect a call here" );
3537 const JVMState *jvms = ((CallNode*)call)->jvms();
3538 ciMethodData* methodData = jvms->method()->method_data();
3539 if (!methodData->is_mature()) return 0.0f; // No call-site data
3540 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3541 if ((data == nullptr) || !data->is_CounterData()) {
3542 // no call profile available, try call's control input
3543 n = n->in(0);
3544 continue;
3545 }
3546 return data->as_CounterData()->count()/FreqCountInvocations;
3547 }
3548 // See if there's a gating IF test
3549 Node *n_c = n->in(0);
3550 if( !n_c->is_If() ) break; // No estimate available
3551 iff = n_c->as_If();
3552 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3553 // Compute how much count comes on this path
3554 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3555 // Have no count info. Skip dull uncommon-trap like branches.
3556 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3557 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3558 break;
3559 // Skip through never-taken branch; look for a real loop exit.
3560 n = iff->in(0);
3561 }
3562 return 0.0f; // No estimate available
3563 }
3564
3565 //------------------------------merge_many_backedges---------------------------
3566 // Merge all the backedges from the shared header into a private Region.
3567 // Feed that region as the one backedge to this loop.
3568 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3569 uint i;
3570
3571 // Scan for the top 2 hottest backedges
3572 float hotcnt = 0.0f;
3573 float warmcnt = 0.0f;
3574 uint hot_idx = 0;
3575 // Loop starts at 2 because slot 1 is the fall-in path
3576 for( i = 2; i < _head->req(); i++ ) {
3577 float cnt = estimate_path_freq(_head->in(i));
3578 if( cnt > hotcnt ) { // Grab hottest path
3579 warmcnt = hotcnt;
3580 hotcnt = cnt;
3581 hot_idx = i;
3582 } else if( cnt > warmcnt ) { // And 2nd hottest path
3583 warmcnt = cnt;
3584 }
3585 }
3586
3587 // See if the hottest backedge is worthy of being an inner loop
3588 // by being much hotter than the next hottest backedge.
3589 if( hotcnt <= 0.0001 ||
3590 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3591
3592 // Peel out the backedges into a private merge point; peel
3593 // them all except optionally hot_idx.
3594 PhaseIterGVN &igvn = phase->_igvn;
3595
3596 Node *hot_tail = nullptr;
3597 // Make a Region for the merge point
3598 Node *r = new RegionNode(1);
3599 for( i = 2; i < _head->req(); i++ ) {
3600 if( i != hot_idx )
3601 r->add_req( _head->in(i) );
3602 else hot_tail = _head->in(i);
3603 }
3604 igvn.register_new_node_with_optimizer(r, _head);
3605 // Plug region into end of loop _head, followed by hot_tail
3606 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3607 igvn.replace_input_of(_head, 2, r);
3608 if( hot_idx ) _head->add_req(hot_tail);
3609
3610 // Split all the Phis up between '_head' loop and the Region 'r'
3611 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3612 Node *out = _head->fast_out(j);
3613 if( out->is_Phi() ) {
3614 PhiNode* n = out->as_Phi();
3615 igvn.hash_delete(n); // Delete from hash before hacking edges
3616 Node *hot_phi = nullptr;
3617 Node *phi = new PhiNode(r, n->type(), n->adr_type());
3618 // Check all inputs for the ones to peel out
3619 uint j = 1;
3620 for( uint i = 2; i < n->req(); i++ ) {
3621 if( i != hot_idx )
3622 phi->set_req( j++, n->in(i) );
3623 else hot_phi = n->in(i);
3624 }
3625 // Register the phi but do not transform until whole place transforms
3626 igvn.register_new_node_with_optimizer(phi, n);
3627 // Add the merge phi to the old Phi
3628 while( n->req() > 3 ) n->del_req( n->req()-1 );
3629 igvn.replace_input_of(n, 2, phi);
3630 if( hot_idx ) n->add_req(hot_phi);
3631 }
3632 }
3633
3634
3635 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
3636 // of self loop tree. Turn self into a loop headed by _head and with
3637 // tail being the new merge point.
3638 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
3639 phase->set_loop(_tail,ilt); // Adjust tail
3640 _tail = r; // Self's tail is new merge point
3641 phase->set_loop(r,this);
3642 ilt->_child = _child; // New guy has my children
3643 _child = ilt; // Self has new guy as only child
3644 ilt->_parent = this; // new guy has self for parent
3645 ilt->_nest = _nest; // Same nesting depth (for now)
3646
3647 // Starting with 'ilt', look for child loop trees using the same shared
3648 // header. Flatten these out; they will no longer be loops in the end.
3649 IdealLoopTree **pilt = &_child;
3650 while( ilt ) {
3651 if( ilt->_head == _head ) {
3652 uint i;
3653 for( i = 2; i < _head->req(); i++ )
3654 if( _head->in(i) == ilt->_tail )
3655 break; // Still a loop
3656 if( i == _head->req() ) { // No longer a loop
3657 // Flatten ilt. Hang ilt's "_next" list from the end of
3658 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
3659 IdealLoopTree **cp = &ilt->_child;
3660 while( *cp ) cp = &(*cp)->_next; // Find end of child list
3661 *cp = ilt->_next; // Hang next list at end of child list
3662 *pilt = ilt->_child; // Move child up to replace ilt
3663 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
3664 ilt = ilt->_child; // Repeat using new ilt
3665 continue; // do not advance over ilt->_child
3666 }
3667 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
3668 phase->set_loop(_head,ilt);
3669 }
3670 pilt = &ilt->_child; // Advance to next
3671 ilt = *pilt;
3672 }
3673
3674 if( _child ) fix_parent( _child, this );
3675 }
3676
3677 //------------------------------beautify_loops---------------------------------
3678 // Split shared headers and insert loop landing pads.
3679 // Insert a LoopNode to replace the RegionNode.
3680 // Return TRUE if loop tree is structurally changed.
3681 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
3682 bool result = false;
3683 // Cache parts in locals for easy
3684 PhaseIterGVN &igvn = phase->_igvn;
3685
3686 igvn.hash_delete(_head); // Yank from hash before hacking edges
3687
3688 // Check for multiple fall-in paths. Peel off a landing pad if need be.
3689 int fall_in_cnt = 0;
3690 for( uint i = 1; i < _head->req(); i++ )
3691 if( !phase->is_member( this, _head->in(i) ) )
3692 fall_in_cnt++;
3693 assert( fall_in_cnt, "at least 1 fall-in path" );
3694 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
3695 split_fall_in( phase, fall_in_cnt );
3696
3697 // Swap inputs to the _head and all Phis to move the fall-in edge to
3698 // the left.
3699 fall_in_cnt = 1;
3700 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
3701 fall_in_cnt++;
3702 if( fall_in_cnt > 1 ) {
3703 // Since I am just swapping inputs I do not need to update def-use info
3704 Node *tmp = _head->in(1);
3705 igvn.rehash_node_delayed(_head);
3706 _head->set_req( 1, _head->in(fall_in_cnt) );
3707 _head->set_req( fall_in_cnt, tmp );
3708 // Swap also all Phis
3709 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
3710 Node* phi = _head->fast_out(i);
3711 if( phi->is_Phi() ) {
3712 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
3713 tmp = phi->in(1);
3714 phi->set_req( 1, phi->in(fall_in_cnt) );
3715 phi->set_req( fall_in_cnt, tmp );
3716 }
3717 }
3718 }
3719 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
3720 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
3721
3722 // If I am a shared header (multiple backedges), peel off the many
3723 // backedges into a private merge point and use the merge point as
3724 // the one true backedge.
3725 if (_head->req() > 3) {
3726 // Merge the many backedges into a single backedge but leave
3727 // the hottest backedge as separate edge for the following peel.
3728 if (!_irreducible) {
3729 merge_many_backedges( phase );
3730 }
3731
3732 // When recursively beautify my children, split_fall_in can change
3733 // loop tree structure when I am an irreducible loop. Then the head
3734 // of my children has a req() not bigger than 3. Here we need to set
3735 // result to true to catch that case in order to tell the caller to
3736 // rebuild loop tree. See issue JDK-8244407 for details.
3737 result = true;
3738 }
3739
3740 // If I have one hot backedge, peel off myself loop.
3741 // I better be the outermost loop.
3742 if (_head->req() > 3 && !_irreducible) {
3743 split_outer_loop( phase );
3744 result = true;
3745
3746 } else if (!_head->is_Loop() && !_irreducible) {
3747 // Make a new LoopNode to replace the old loop head
3748 Node *l = new LoopNode( _head->in(1), _head->in(2) );
3749 l = igvn.register_new_node_with_optimizer(l, _head);
3750 phase->set_created_loop_node();
3751 // Go ahead and replace _head
3752 phase->_igvn.replace_node( _head, l );
3753 _head = l;
3754 phase->set_loop(_head, this);
3755 }
3756
3757 // Now recursively beautify nested loops
3758 if( _child ) result |= _child->beautify_loops( phase );
3759 if( _next ) result |= _next ->beautify_loops( phase );
3760 return result;
3761 }
3762
3763 //------------------------------allpaths_check_safepts----------------------------
3764 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
3765 // safepoint encountered. Helper for check_safepts.
3766 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
3767 assert(stack.size() == 0, "empty stack");
3768 stack.push(_head);
3769 visited.clear();
3770 visited.set(_head->_idx);
3771 while (stack.size() > 0) {
3772 Node* n = stack.pop();
3773 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3774 // Terminate this path
3775 } else if (n->Opcode() == Op_SafePoint) {
3776 if (_phase->get_loop(n) != this) {
3777 if (_required_safept == nullptr) _required_safept = new Node_List();
3778 // save the first we run into on that path: closest to the tail if the head has a single backedge
3779 _required_safept->push(n);
3780 }
3781 // Terminate this path
3782 } else {
3783 uint start = n->is_Region() ? 1 : 0;
3784 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
3785 for (uint i = start; i < end; i++) {
3786 Node* in = n->in(i);
3787 assert(in->is_CFG(), "must be");
3788 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
3789 stack.push(in);
3790 }
3791 }
3792 }
3793 }
3794 }
3795
3796 //------------------------------check_safepts----------------------------
3797 // Given dominators, try to find loops with calls that must always be
3798 // executed (call dominates loop tail). These loops do not need non-call
3799 // safepoints (ncsfpt).
3800 //
3801 // A complication is that a safepoint in a inner loop may be needed
3802 // by an outer loop. In the following, the inner loop sees it has a
3803 // call (block 3) on every path from the head (block 2) to the
3804 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
3805 // in block 2, _but_ this leaves the outer loop without a safepoint.
3806 //
3807 // entry 0
3808 // |
3809 // v
3810 // outer 1,2 +->1
3811 // | |
3812 // | v
3813 // | 2<---+ ncsfpt in 2
3814 // |_/|\ |
3815 // | v |
3816 // inner 2,3 / 3 | call in 3
3817 // / | |
3818 // v +--+
3819 // exit 4
3820 //
3821 //
3822 // This method creates a list (_required_safept) of ncsfpt nodes that must
3823 // be protected is created for each loop. When a ncsfpt maybe deleted, it
3824 // is first looked for in the lists for the outer loops of the current loop.
3825 //
3826 // The insights into the problem:
3827 // A) counted loops are okay
3828 // B) innermost loops are okay (only an inner loop can delete
3829 // a ncsfpt needed by an outer loop)
3830 // C) a loop is immune from an inner loop deleting a safepoint
3831 // if the loop has a call on the idom-path
3832 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
3833 // idom-path that is not in a nested loop
3834 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
3835 // loop needs to be prevented from deletion by an inner loop
3836 //
3837 // There are two analyses:
3838 // 1) The first, and cheaper one, scans the loop body from
3839 // tail to head following the idom (immediate dominator)
3840 // chain, looking for the cases (C,D,E) above.
3841 // Since inner loops are scanned before outer loops, there is summary
3842 // information about inner loops. Inner loops can be skipped over
3843 // when the tail of an inner loop is encountered.
3844 //
3845 // 2) The second, invoked if the first fails to find a call or ncsfpt on
3846 // the idom path (which is rare), scans all predecessor control paths
3847 // from the tail to the head, terminating a path when a call or sfpt
3848 // is encountered, to find the ncsfpt's that are closest to the tail.
3849 //
3850 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
3851 // Bottom up traversal
3852 IdealLoopTree* ch = _child;
3853 if (_child) _child->check_safepts(visited, stack);
3854 if (_next) _next ->check_safepts(visited, stack);
3855
3856 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
3857 bool has_call = false; // call on dom-path
3858 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
3859 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
3860 if (!_irreducible) {
3861 // Scan the dom-path nodes from tail to head
3862 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
3863 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3864 has_call = true;
3865 _has_sfpt = 1; // Then no need for a safept!
3866 break;
3867 } else if (n->Opcode() == Op_SafePoint) {
3868 if (_phase->get_loop(n) == this) {
3869 has_local_ncsfpt = true;
3870 break;
3871 }
3872 if (nonlocal_ncsfpt == nullptr) {
3873 nonlocal_ncsfpt = n; // save the one closest to the tail
3874 }
3875 } else {
3876 IdealLoopTree* nlpt = _phase->get_loop(n);
3877 if (this != nlpt) {
3878 // If at an inner loop tail, see if the inner loop has already
3879 // recorded seeing a call on the dom-path (and stop.) If not,
3880 // jump to the head of the inner loop.
3881 assert(is_member(nlpt), "nested loop");
3882 Node* tail = nlpt->_tail;
3883 if (tail->in(0)->is_If()) tail = tail->in(0);
3884 if (n == tail) {
3885 // If inner loop has call on dom-path, so does outer loop
3886 if (nlpt->_has_sfpt) {
3887 has_call = true;
3888 _has_sfpt = 1;
3889 break;
3890 }
3891 // Skip to head of inner loop
3892 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
3893 n = nlpt->_head;
3894 if (_head == n) {
3895 // this and nlpt (inner loop) have the same loop head. This should not happen because
3896 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
3897 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
3898 // but then attached in the build_loop_tree afterwards, and so still have unmerged
3899 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
3900 // since we have reached the loop head of this.
3901 assert(_head->as_Region()->is_in_infinite_subgraph(),
3902 "only expect unmerged backedges in infinite loops");
3903 break;
3904 }
3905 }
3906 }
3907 }
3908 }
3909 }
3910 // Record safept's that this loop needs preserved when an
3911 // inner loop attempts to delete it's safepoints.
3912 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
3913 if (nonlocal_ncsfpt != nullptr) {
3914 if (_required_safept == nullptr) _required_safept = new Node_List();
3915 _required_safept->push(nonlocal_ncsfpt);
3916 } else {
3917 // Failed to find a suitable safept on the dom-path. Now use
3918 // an all paths walk from tail to head, looking for safepoints to preserve.
3919 allpaths_check_safepts(visited, stack);
3920 }
3921 }
3922 }
3923 }
3924
3925 //---------------------------is_deleteable_safept----------------------------
3926 // Is safept not required by an outer loop?
3927 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
3928 assert(sfpt->Opcode() == Op_SafePoint, "");
3929 IdealLoopTree* lp = get_loop(sfpt)->_parent;
3930 while (lp != nullptr) {
3931 Node_List* sfpts = lp->_required_safept;
3932 if (sfpts != nullptr) {
3933 for (uint i = 0; i < sfpts->size(); i++) {
3934 if (sfpt == sfpts->at(i))
3935 return false;
3936 }
3937 }
3938 lp = lp->_parent;
3939 }
3940 return true;
3941 }
3942
3943 //---------------------------replace_parallel_iv-------------------------------
3944 // Replace parallel induction variable (parallel to trip counter)
3945 // This optimization looks for patterns similar to:
3946 //
3947 // int a = init2;
3948 // for (int iv = init; iv < limit; iv += stride_con) {
3949 // a += stride_con2;
3950 // }
3951 //
3952 // and transforms it to:
3953 //
3954 // int iv2 = init2
3955 // int iv = init
3956 // loop:
3957 // if (iv >= limit) goto exit
3958 // iv += stride_con
3959 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
3960 // goto loop
3961 // exit:
3962 // ...
3963 //
3964 // Such transformation introduces more optimization opportunities. In this
3965 // particular example, the loop can be eliminated entirely given that
3966 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
3967 // place to only perform this optimization if such a division is exact. This
3968 // example will be transformed into its semantic equivalence:
3969 //
3970 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
3971 //
3972 // which corresponds to the structure of transformed subgraph.
3973 //
3974 // However, if there is a mismatch between types of the loop and the parallel
3975 // induction variable (e.g., a long-typed IV in an int-typed loop), type
3976 // conversions are required:
3977 //
3978 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
3979 //
3980 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
3981 assert(loop->_head->is_CountedLoop(), "");
3982 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3983 if (!cl->is_valid_counted_loop(T_INT)) {
3984 return; // skip malformed counted loop
3985 }
3986 Node *incr = cl->incr();
3987 if (incr == nullptr) {
3988 return; // Dead loop?
3989 }
3990 Node *init = cl->init_trip();
3991 Node *phi = cl->phi();
3992 jlong stride_con = cl->stride_con();
3993
3994 // Visit all children, looking for Phis
3995 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
3996 Node *out = cl->out(i);
3997 // Look for other phis (secondary IVs). Skip dead ones
3998 if (!out->is_Phi() || out == phi || !has_node(out)) {
3999 continue;
4000 }
4001
4002 PhiNode* phi2 = out->as_Phi();
4003 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4004 // Look for induction variables of the form: X += constant
4005 if (phi2->region() != loop->_head ||
4006 incr2->req() != 3 ||
4007 incr2->in(1)->uncast() != phi2 ||
4008 incr2 == incr ||
4009 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4010 !incr2->in(2)->is_Con()) {
4011 continue;
4012 }
4013
4014 if (incr2->in(1)->is_ConstraintCast() &&
4015 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4016 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4017 continue;
4018 }
4019 // Check for parallel induction variable (parallel to trip counter)
4020 // via an affine function. In particular, count-down loops with
4021 // count-up array indices are common. We only RCE references off
4022 // the trip-counter, so we need to convert all these to trip-counter
4023 // expressions.
4024 Node* init2 = phi2->in(LoopNode::EntryControl);
4025
4026 // Determine the basic type of the stride constant (and the iv being incremented).
4027 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4028 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4029
4030 // The ratio of the two strides cannot be represented as an int
4031 // if stride_con2 is min_jint (or min_jlong, respectively) and
4032 // stride_con is -1.
4033 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4034 continue;
4035 }
4036
4037 // The general case here gets a little tricky. We want to find the
4038 // GCD of all possible parallel IV's and make a new IV using this
4039 // GCD for the loop. Then all possible IVs are simple multiples of
4040 // the GCD. In practice, this will cover very few extra loops.
4041 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4042 // where +/-1 is the common case, but other integer multiples are
4043 // also easy to handle.
4044 jlong ratio_con = stride_con2 / stride_con;
4045
4046 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4047 continue;
4048 }
4049
4050 #ifndef PRODUCT
4051 if (TraceLoopOpts) {
4052 tty->print("Parallel IV: %d ", phi2->_idx);
4053 loop->dump_head();
4054 }
4055 #endif
4056
4057 // Convert to using the trip counter. The parallel induction
4058 // variable differs from the trip counter by a loop-invariant
4059 // amount, the difference between their respective initial values.
4060 // It is scaled by the 'ratio_con'.
4061 Node* ratio = integercon(ratio_con, stride_con2_bt);
4062
4063 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4064 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4065
4066 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4067 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4068 set_early_ctrl(ratio_init, false);
4069
4070 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4071 _igvn.register_new_node_with_optimizer(diff, init2);
4072 set_early_ctrl(diff, false);
4073
4074 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4075 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4076 set_ctrl(ratio_idx, cl);
4077
4078 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4079 _igvn.register_new_node_with_optimizer(add);
4080 set_ctrl(add, cl);
4081
4082 _igvn.replace_node( phi2, add );
4083 // Sometimes an induction variable is unused
4084 if (add->outcnt() == 0) {
4085 _igvn.remove_dead_node(add);
4086 }
4087 --i; // deleted this phi; rescan starting with next position
4088 }
4089 }
4090
4091 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4092 BasicType source = _igvn.type(input)->basic_type();
4093 if (source == target) {
4094 return input;
4095 }
4096
4097 Node* converted = ConvertNode::create_convert(source, target, input);
4098 _igvn.register_new_node_with_optimizer(converted, input);
4099 set_early_ctrl(converted, false);
4100
4101 return converted;
4102 }
4103
4104 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4105 Node* keep = nullptr;
4106 if (keep_one) {
4107 // Look for a safepoint on the idom-path.
4108 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4109 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4110 keep = i;
4111 break; // Found one
4112 }
4113 }
4114 }
4115
4116 // Don't remove any safepoints if it is requested to keep a single safepoint and
4117 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4118 // in this case since there's no safepoint dominating all paths in the loop body.
4119 bool prune = !keep_one || keep != nullptr;
4120
4121 // Delete other safepoints in this loop.
4122 Node_List* sfpts = _safepts;
4123 if (prune && sfpts != nullptr) {
4124 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4125 for (uint i = 0; i < sfpts->size(); i++) {
4126 Node* n = sfpts->at(i);
4127 assert(phase->get_loop(n) == this, "");
4128 if (n != keep && phase->is_deleteable_safept(n)) {
4129 phase->lazy_replace(n, n->in(TypeFunc::Control));
4130 }
4131 }
4132 }
4133 }
4134
4135 //------------------------------counted_loop-----------------------------------
4136 // Convert to counted loops where possible
4137 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4138
4139 // For grins, set the inner-loop flag here
4140 if (!_child) {
4141 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4142 }
4143
4144 IdealLoopTree* loop = this;
4145 if (_head->is_CountedLoop() ||
4146 phase->is_counted_loop(_head, loop, T_INT)) {
4147
4148 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4149 // Indicate we do not need a safepoint here
4150 _has_sfpt = 1;
4151 }
4152
4153 // Remove safepoints
4154 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4155 remove_safepoints(phase, keep_one_sfpt);
4156
4157 // Look for induction variables
4158 phase->replace_parallel_iv(this);
4159 } else if (_head->is_LongCountedLoop() ||
4160 phase->is_counted_loop(_head, loop, T_LONG)) {
4161 remove_safepoints(phase, true);
4162 } else {
4163 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4164 if (_parent != nullptr && !_irreducible) {
4165 // Not a counted loop. Keep one safepoint.
4166 bool keep_one_sfpt = true;
4167 remove_safepoints(phase, keep_one_sfpt);
4168 }
4169 }
4170
4171 // Recursively
4172 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4173 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4174 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4175 if (loop->_next) loop->_next ->counted_loop(phase);
4176 }
4177
4178
4179 // The Estimated Loop Clone Size:
4180 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4181 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4182 // respectively, used to ensure that the node usage estimates made are on the
4183 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4184 // possible additional/excessive nodes generated due to data and control flow
4185 // merging, for edges reaching outside the loop.
4186 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4187
4188 precond(0 < factor && factor < 16);
4189
4190 uint const bc = 13;
4191 uint const cc = 17;
4192 uint const sz = _body.size() + (_body.size() + 7) / 2;
4193 uint estimate = factor * (sz + bc) + cc;
4194
4195 assert((estimate - cc) / factor == sz + bc, "overflow");
4196
4197 return estimate + est_loop_flow_merge_sz();
4198 }
4199
4200 // The Estimated Loop (full-) Unroll Size:
4201 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4202 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4203 // node usage estimates made are on the safe side, for the most part. This is
4204 // a "light" version of the loop clone size calculation (above), based on the
4205 // assumption that most of the loop-construct overhead will be unraveled when
4206 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4207 // including an overflow check and returning UINT_MAX in case of an overflow.
4208 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4209
4210 precond(factor > 0);
4211
4212 // Take into account that after unroll conjoined heads and tails will fold.
4213 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4214 uint const cc = 7;
4215 uint const sz = b0 + (b0 + 15) / 16;
4216 uint estimate = factor * sz + cc;
4217
4218 if ((estimate - cc) / factor != sz) {
4219 return UINT_MAX;
4220 }
4221
4222 return estimate + est_loop_flow_merge_sz();
4223 }
4224
4225 // Estimate the growth effect (in nodes) of merging control and data flow when
4226 // cloning a loop body, based on the amount of control and data flow reaching
4227 // outside of the (current) loop body.
4228 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4229
4230 uint ctrl_edge_out_cnt = 0;
4231 uint data_edge_out_cnt = 0;
4232
4233 for (uint i = 0; i < _body.size(); i++) {
4234 Node* node = _body.at(i);
4235 uint outcnt = node->outcnt();
4236
4237 for (uint k = 0; k < outcnt; k++) {
4238 Node* out = node->raw_out(k);
4239 if (out == nullptr) continue;
4240 if (out->is_CFG()) {
4241 if (!is_member(_phase->get_loop(out))) {
4242 ctrl_edge_out_cnt++;
4243 }
4244 } else if (_phase->has_ctrl(out)) {
4245 Node* ctrl = _phase->get_ctrl(out);
4246 assert(ctrl != nullptr, "must be");
4247 assert(ctrl->is_CFG(), "must be");
4248 if (!is_member(_phase->get_loop(ctrl))) {
4249 data_edge_out_cnt++;
4250 }
4251 }
4252 }
4253 }
4254 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4255 // a rather pessimistic estimate for the most part, in particular for some
4256 // complex loops, but still not enough to capture all loops.
4257 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4258 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4259 }
4260 return 0;
4261 }
4262
4263 #ifndef PRODUCT
4264 //------------------------------dump_head--------------------------------------
4265 // Dump 1 liner for loop header info
4266 void IdealLoopTree::dump_head() {
4267 tty->sp(2 * _nest);
4268 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4269 if (_irreducible) tty->print(" IRREDUCIBLE");
4270 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4271 : _head->in(LoopNode::EntryControl);
4272 const Predicates predicates(entry);
4273 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4274 tty->print(" limit_check");
4275 }
4276 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4277 tty->print(" profile_predicated");
4278 }
4279 if (UseLoopPredicate && predicates.loop_predicate_block()->is_non_empty()) {
4280 tty->print(" predicated");
4281 }
4282 if (_head->is_CountedLoop()) {
4283 CountedLoopNode *cl = _head->as_CountedLoop();
4284 tty->print(" counted");
4285
4286 Node* init_n = cl->init_trip();
4287 if (init_n != nullptr && init_n->is_Con())
4288 tty->print(" [%d,", cl->init_trip()->get_int());
4289 else
4290 tty->print(" [int,");
4291 Node* limit_n = cl->limit();
4292 if (limit_n != nullptr && limit_n->is_Con())
4293 tty->print("%d),", cl->limit()->get_int());
4294 else
4295 tty->print("int),");
4296 int stride_con = cl->stride_con();
4297 if (stride_con > 0) tty->print("+");
4298 tty->print("%d", stride_con);
4299
4300 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4301
4302 if (cl->is_pre_loop ()) tty->print(" pre" );
4303 if (cl->is_main_loop()) tty->print(" main");
4304 if (cl->is_post_loop()) tty->print(" post");
4305 if (cl->is_vectorized_loop()) tty->print(" vector");
4306 if (range_checks_present()) tty->print(" rc ");
4307 }
4308 if (_has_call) tty->print(" has_call");
4309 if (_has_sfpt) tty->print(" has_sfpt");
4310 if (_rce_candidate) tty->print(" rce");
4311 if (_safepts != nullptr && _safepts->size() > 0) {
4312 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4313 }
4314 if (_required_safept != nullptr && _required_safept->size() > 0) {
4315 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4316 }
4317 if (Verbose) {
4318 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4319 }
4320 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4321 tty->print(" strip_mined");
4322 }
4323 tty->cr();
4324 }
4325
4326 //------------------------------dump-------------------------------------------
4327 // Dump loops by loop tree
4328 void IdealLoopTree::dump() {
4329 dump_head();
4330 if (_child) _child->dump();
4331 if (_next) _next ->dump();
4332 }
4333
4334 #endif
4335
4336 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4337 if (loop == root) {
4338 if (loop->_child != nullptr) {
4339 log->begin_head("loop_tree");
4340 log->end_head();
4341 log_loop_tree_helper(root, loop->_child, log);
4342 log->tail("loop_tree");
4343 assert(loop->_next == nullptr, "what?");
4344 }
4345 } else if (loop != nullptr) {
4346 Node* head = loop->_head;
4347 log->begin_head("loop");
4348 log->print(" idx='%d' ", head->_idx);
4349 if (loop->_irreducible) log->print("irreducible='1' ");
4350 if (head->is_Loop()) {
4351 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4352 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4353 } else if (head->is_CountedLoop()) {
4354 CountedLoopNode* cl = head->as_CountedLoop();
4355 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4356 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4357 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4358 }
4359 log->end_head();
4360 log_loop_tree_helper(root, loop->_child, log);
4361 log->tail("loop");
4362 log_loop_tree_helper(root, loop->_next, log);
4363 }
4364 }
4365
4366 void PhaseIdealLoop::log_loop_tree() {
4367 if (C->log() != nullptr) {
4368 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4369 }
4370 }
4371
4372 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4373 // predicates will be removed during the next round of IGVN.
4374 void PhaseIdealLoop::eliminate_useless_predicates() {
4375 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4376 return; // No predicates left.
4377 }
4378
4379 eliminate_useless_parse_predicates();
4380 eliminate_useless_template_assertion_predicates();
4381 }
4382
4383 // Eliminate all Parse Predicates that do not belong to a loop anymore by marking them useless. These will be removed
4384 // during the next round of IGVN.
4385 void PhaseIdealLoop::eliminate_useless_parse_predicates() {
4386 mark_all_parse_predicates_useless();
4387 if (C->has_loops()) {
4388 mark_loop_associated_parse_predicates_useful();
4389 }
4390 add_useless_parse_predicates_to_igvn_worklist();
4391 }
4392
4393 void PhaseIdealLoop::mark_all_parse_predicates_useless() const {
4394 for (int i = 0; i < C->parse_predicate_count(); i++) {
4395 C->parse_predicate(i)->mark_useless();
4396 }
4397 }
4398
4399 void PhaseIdealLoop::mark_loop_associated_parse_predicates_useful() {
4400 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4401 IdealLoopTree* loop = iterator.current();
4402 if (loop->can_apply_loop_predication()) {
4403 mark_useful_parse_predicates_for_loop(loop);
4404 }
4405 }
4406 }
4407
4408 // This visitor marks all visited Parse Predicates useful.
4409 class ParsePredicateUsefulMarker : public PredicateVisitor {
4410 public:
4411 using PredicateVisitor::visit;
4412
4413 void visit(const ParsePredicate& parse_predicate) override {
4414 parse_predicate.head()->mark_useful();
4415 }
4416 };
4417
4418 void PhaseIdealLoop::mark_useful_parse_predicates_for_loop(IdealLoopTree* loop) {
4419 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4420 const PredicateIterator predicate_iterator(entry);
4421 ParsePredicateUsefulMarker useful_marker;
4422 predicate_iterator.for_each(useful_marker);
4423 }
4424
4425 void PhaseIdealLoop::add_useless_parse_predicates_to_igvn_worklist() {
4426 for (int i = 0; i < C->parse_predicate_count(); i++) {
4427 ParsePredicateNode* parse_predicate_node = C->parse_predicate(i);
4428 if (parse_predicate_node->is_useless()) {
4429 _igvn._worklist.push(parse_predicate_node);
4430 }
4431 }
4432 }
4433
4434
4435 // Eliminate all Template Assertion Predicates that do not belong to their originally associated loop anymore by
4436 // replacing the OpaqueTemplateAssertionPredicate node of the If node with true. These nodes will be removed during the
4437 // next round of IGVN.
4438 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates() {
4439 Unique_Node_List useful_predicates;
4440 if (C->has_loops()) {
4441 collect_useful_template_assertion_predicates(useful_predicates);
4442 }
4443 eliminate_useless_template_assertion_predicates(useful_predicates);
4444 }
4445
4446 void PhaseIdealLoop::collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4447 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4448 IdealLoopTree* loop = iterator.current();
4449 if (loop->can_apply_loop_predication()) {
4450 collect_useful_template_assertion_predicates_for_loop(loop, useful_predicates);
4451 }
4452 }
4453 }
4454
4455 void PhaseIdealLoop::collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop,
4456 Unique_Node_List &useful_predicates) {
4457 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4458 const Predicates predicates(entry);
4459 if (UseProfiledLoopPredicate) {
4460 const PredicateBlock* profiled_loop_predicate_block = predicates.profiled_loop_predicate_block();
4461 if (profiled_loop_predicate_block->has_parse_predicate()) {
4462 ParsePredicateSuccessProj* parse_predicate_proj = profiled_loop_predicate_block->parse_predicate_success_proj();
4463 get_opaque_template_assertion_predicate_nodes(parse_predicate_proj, useful_predicates);
4464 }
4465 }
4466
4467 if (UseLoopPredicate) {
4468 const PredicateBlock* loop_predicate_block = predicates.loop_predicate_block();
4469 if (loop_predicate_block->has_parse_predicate()) {
4470 ParsePredicateSuccessProj* parse_predicate_proj = loop_predicate_block->parse_predicate_success_proj();
4471 get_opaque_template_assertion_predicate_nodes(parse_predicate_proj, useful_predicates);
4472 }
4473 }
4474 }
4475
4476 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4477 for (int i = C->template_assertion_predicate_count(); i > 0; i--) {
4478 OpaqueTemplateAssertionPredicateNode* opaque_node =
4479 C->template_assertion_predicate_opaq_node(i - 1)->as_OpaqueTemplateAssertionPredicate();
4480 if (!useful_predicates.member(opaque_node)) { // not in the useful list
4481 ConINode* one = intcon(1);
4482 _igvn.replace_node(opaque_node, one);
4483 }
4484 }
4485 }
4486
4487 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4488 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4489 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4490 return;
4491 }
4492 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4493 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4494 IdealLoopTree* lpt = iter.current();
4495 if (lpt->_child == nullptr && lpt->is_counted()) {
4496 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4497 Node* opaque = head->is_canonical_loop_entry();
4498 if (opaque != nullptr) {
4499 useful_zero_trip_guard_opaques_nodes.push(opaque);
4500 }
4501 }
4502 }
4503 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4504 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4505 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4506 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4507 IfNode* iff = opaque->if_node();
4508 IdealLoopTree* loop = get_loop(iff);
4509 while (loop != _ltree_root && loop != nullptr) {
4510 loop = loop->_parent;
4511 }
4512 if (loop == nullptr) {
4513 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4514 // We can't tell if the opaque node is useful or not
4515 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4516 } else {
4517 assert(guarded_loop == nullptr, "");
4518 this->_igvn.replace_node(opaque, opaque->in(1));
4519 }
4520 } else {
4521 assert(guarded_loop != nullptr, "");
4522 }
4523 }
4524 }
4525
4526 //------------------------process_expensive_nodes-----------------------------
4527 // Expensive nodes have their control input set to prevent the GVN
4528 // from commoning them and as a result forcing the resulting node to
4529 // be in a more frequent path. Use CFG information here, to change the
4530 // control inputs so that some expensive nodes can be commoned while
4531 // not executed more frequently.
4532 bool PhaseIdealLoop::process_expensive_nodes() {
4533 assert(OptimizeExpensiveOps, "optimization off?");
4534
4535 // Sort nodes to bring similar nodes together
4536 C->sort_expensive_nodes();
4537
4538 bool progress = false;
4539
4540 for (int i = 0; i < C->expensive_count(); ) {
4541 Node* n = C->expensive_node(i);
4542 int start = i;
4543 // Find nodes similar to n
4544 i++;
4545 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4546 int end = i;
4547 // And compare them two by two
4548 for (int j = start; j < end; j++) {
4549 Node* n1 = C->expensive_node(j);
4550 if (is_node_unreachable(n1)) {
4551 continue;
4552 }
4553 for (int k = j+1; k < end; k++) {
4554 Node* n2 = C->expensive_node(k);
4555 if (is_node_unreachable(n2)) {
4556 continue;
4557 }
4558
4559 assert(n1 != n2, "should be pair of nodes");
4560
4561 Node* c1 = n1->in(0);
4562 Node* c2 = n2->in(0);
4563
4564 Node* parent_c1 = c1;
4565 Node* parent_c2 = c2;
4566
4567 // The call to get_early_ctrl_for_expensive() moves the
4568 // expensive nodes up but stops at loops that are in a if
4569 // branch. See whether we can exit the loop and move above the
4570 // If.
4571 if (c1->is_Loop()) {
4572 parent_c1 = c1->in(1);
4573 }
4574 if (c2->is_Loop()) {
4575 parent_c2 = c2->in(1);
4576 }
4577
4578 if (parent_c1 == parent_c2) {
4579 _igvn._worklist.push(n1);
4580 _igvn._worklist.push(n2);
4581 continue;
4582 }
4583
4584 // Look for identical expensive node up the dominator chain.
4585 if (is_dominator(c1, c2)) {
4586 c2 = c1;
4587 } else if (is_dominator(c2, c1)) {
4588 c1 = c2;
4589 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4590 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4591 // Both branches have the same expensive node so move it up
4592 // before the if.
4593 c1 = c2 = idom(parent_c1->in(0));
4594 }
4595 // Do the actual moves
4596 if (n1->in(0) != c1) {
4597 _igvn.replace_input_of(n1, 0, c1);
4598 progress = true;
4599 }
4600 if (n2->in(0) != c2) {
4601 _igvn.replace_input_of(n2, 0, c2);
4602 progress = true;
4603 }
4604 }
4605 }
4606 }
4607
4608 return progress;
4609 }
4610
4611 //=============================================================================
4612 //----------------------------build_and_optimize-------------------------------
4613 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4614 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4615 void PhaseIdealLoop::build_and_optimize() {
4616 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4617
4618 bool do_split_ifs = (_mode == LoopOptsDefault);
4619 bool skip_loop_opts = (_mode == LoopOptsNone);
4620 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4621
4622
4623 int old_progress = C->major_progress();
4624 uint orig_worklist_size = _igvn._worklist.size();
4625
4626 // Reset major-progress flag for the driver's heuristics
4627 C->clear_major_progress();
4628
4629 #ifndef PRODUCT
4630 // Capture for later assert
4631 uint unique = C->unique();
4632 _loop_invokes++;
4633 _loop_work += unique;
4634 #endif
4635
4636 // True if the method has at least 1 irreducible loop
4637 _has_irreducible_loops = false;
4638
4639 _created_loop_node = false;
4640
4641 VectorSet visited;
4642 // Pre-grow the mapping from Nodes to IdealLoopTrees.
4643 _loop_or_ctrl.map(C->unique(), nullptr);
4644 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
4645
4646 // Pre-build the top-level outermost loop tree entry
4647 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
4648 // Do not need a safepoint at the top level
4649 _ltree_root->_has_sfpt = 1;
4650
4651 // Initialize Dominators.
4652 // Checked in clone_loop_predicate() during beautify_loops().
4653 _idom_size = 0;
4654 _idom = nullptr;
4655 _dom_depth = nullptr;
4656 _dom_stk = nullptr;
4657
4658 // Empty pre-order array
4659 allocate_preorders();
4660
4661 // Build a loop tree on the fly. Build a mapping from CFG nodes to
4662 // IdealLoopTree entries. Data nodes are NOT walked.
4663 build_loop_tree();
4664 // Check for bailout, and return
4665 if (C->failing()) {
4666 return;
4667 }
4668
4669 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
4670 // is an exception edge, parsing doesn't set has_loops().
4671 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
4672 // No loops after all
4673 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
4674
4675 // There should always be an outer loop containing the Root and Return nodes.
4676 // If not, we have a degenerate empty program. Bail out in this case.
4677 if (!has_node(C->root())) {
4678 if (!_verify_only) {
4679 C->clear_major_progress();
4680 assert(false, "empty program detected during loop optimization");
4681 C->record_method_not_compilable("empty program detected during loop optimization");
4682 }
4683 return;
4684 }
4685
4686 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4687 // Nothing to do, so get out
4688 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
4689 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
4690 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
4691 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
4692 if (stop_early && !do_expensive_nodes) {
4693 return;
4694 }
4695
4696 // Set loop nesting depth
4697 _ltree_root->set_nest( 0 );
4698
4699 // Split shared headers and insert loop landing pads.
4700 // Do not bother doing this on the Root loop of course.
4701 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
4702 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
4703 if( _ltree_root->_child->beautify_loops( this ) ) {
4704 // Re-build loop tree!
4705 _ltree_root->_child = nullptr;
4706 _loop_or_ctrl.clear();
4707 reallocate_preorders();
4708 build_loop_tree();
4709 // Check for bailout, and return
4710 if (C->failing()) {
4711 return;
4712 }
4713 // Reset loop nesting depth
4714 _ltree_root->set_nest( 0 );
4715
4716 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
4717 }
4718 }
4719
4720 // Build Dominators for elision of null checks & loop finding.
4721 // Since nodes do not have a slot for immediate dominator, make
4722 // a persistent side array for that info indexed on node->_idx.
4723 _idom_size = C->unique();
4724 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
4725 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
4726 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
4727 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
4728
4729 Dominators();
4730
4731 if (!_verify_only) {
4732 // As a side effect, Dominators removed any unreachable CFG paths
4733 // into RegionNodes. It doesn't do this test against Root, so
4734 // we do it here.
4735 for( uint i = 1; i < C->root()->req(); i++ ) {
4736 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
4737 _igvn.delete_input_of(C->root(), i);
4738 i--; // Rerun same iteration on compressed edges
4739 }
4740 }
4741
4742 // Given dominators, try to find inner loops with calls that must
4743 // always be executed (call dominates loop tail). These loops do
4744 // not need a separate safepoint.
4745 Node_List cisstack;
4746 _ltree_root->check_safepts(visited, cisstack);
4747 }
4748
4749 // Walk the DATA nodes and place into loops. Find earliest control
4750 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
4751 // holding the associated IdealLoopTree pointer. For DATA nodes, the
4752 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
4753
4754 // Allocate stack with enough space to avoid frequent realloc
4755 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
4756 Node_Stack nstack(stack_size);
4757
4758 visited.clear();
4759 Node_List worklist;
4760 // Don't need C->root() on worklist since
4761 // it will be processed among C->top() inputs
4762 worklist.push(C->top());
4763 visited.set(C->top()->_idx); // Set C->top() as visited now
4764 build_loop_early( visited, worklist, nstack );
4765
4766 // Given early legal placement, try finding counted loops. This placement
4767 // is good enough to discover most loop invariants.
4768 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
4769 _ltree_root->counted_loop( this );
4770 }
4771
4772 // Find latest loop placement. Find ideal loop placement.
4773 visited.clear();
4774 init_dom_lca_tags();
4775 // Need C->root() on worklist when processing outs
4776 worklist.push(C->root());
4777 NOT_PRODUCT( C->verify_graph_edges(); )
4778 worklist.push(C->top());
4779 build_loop_late( visited, worklist, nstack );
4780 if (C->failing()) { return; }
4781
4782 if (_verify_only) {
4783 C->restore_major_progress(old_progress);
4784 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
4785 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
4786 return;
4787 }
4788
4789 // clear out the dead code after build_loop_late
4790 while (_deadlist.size()) {
4791 _igvn.remove_globally_dead_node(_deadlist.pop());
4792 }
4793
4794 eliminate_useless_zero_trip_guard();
4795
4796 if (stop_early) {
4797 assert(do_expensive_nodes, "why are we here?");
4798 if (process_expensive_nodes()) {
4799 // If we made some progress when processing expensive nodes then
4800 // the IGVN may modify the graph in a way that will allow us to
4801 // make some more progress: we need to try processing expensive
4802 // nodes again.
4803 C->set_major_progress();
4804 }
4805 return;
4806 }
4807
4808 // Some parser-inserted loop predicates could never be used by loop
4809 // predication or they were moved away from loop during some optimizations.
4810 // For example, peeling. Eliminate them before next loop optimizations.
4811 eliminate_useless_predicates();
4812
4813 #ifndef PRODUCT
4814 C->verify_graph_edges();
4815 if (_verify_me) { // Nested verify pass?
4816 // Check to see if the verify _mode is broken
4817 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
4818 return;
4819 }
4820 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4821 if (TraceLoopOpts && C->has_loops()) {
4822 _ltree_root->dump();
4823 }
4824 #endif
4825
4826 if (skip_loop_opts) {
4827 C->restore_major_progress(old_progress);
4828 return;
4829 }
4830
4831 if (do_max_unroll) {
4832 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4833 IdealLoopTree* lpt = iter.current();
4834 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
4835 lpt->compute_trip_count(this);
4836 if (!lpt->do_one_iteration_loop(this) &&
4837 !lpt->do_remove_empty_loop(this)) {
4838 AutoNodeBudget node_budget(this);
4839 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
4840 lpt->policy_maximally_unroll(this)) {
4841 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4842 do_maximally_unroll(lpt, worklist);
4843 }
4844 }
4845 }
4846 }
4847
4848 C->restore_major_progress(old_progress);
4849 return;
4850 }
4851
4852 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
4853 return;
4854 }
4855
4856 if (ReassociateInvariants && !C->major_progress()) {
4857 // Reassociate invariants and prep for split_thru_phi
4858 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4859 IdealLoopTree* lpt = iter.current();
4860 if (!lpt->is_loop()) {
4861 continue;
4862 }
4863 Node* head = lpt->_head;
4864 if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
4865
4866 // check for vectorized loops, any reassociation of invariants was already done
4867 if (head->is_CountedLoop()) {
4868 if (head->as_CountedLoop()->is_unroll_only()) {
4869 continue;
4870 } else {
4871 AutoNodeBudget node_budget(this);
4872 lpt->reassociate_invariants(this);
4873 }
4874 }
4875 // Because RCE opportunities can be masked by split_thru_phi,
4876 // look for RCE candidates and inhibit split_thru_phi
4877 // on just their loop-phi's for this pass of loop opts
4878 if (SplitIfBlocks && do_split_ifs &&
4879 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
4880 (lpt->policy_range_check(this, true, T_LONG) ||
4881 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
4882 lpt->_rce_candidate = 1; // = true
4883 }
4884 }
4885 }
4886
4887 // Check for aggressive application of split-if and other transforms
4888 // that require basic-block info (like cloning through Phi's)
4889 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
4890 visited.clear();
4891 split_if_with_blocks(visited, nstack);
4892 if (C->failing()) {
4893 return;
4894 }
4895 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4896 }
4897
4898 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
4899 C->set_major_progress();
4900 }
4901
4902 // Perform loop predication before iteration splitting
4903 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
4904 _ltree_root->_child->loop_predication(this);
4905 }
4906
4907 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
4908 if (do_intrinsify_fill()) {
4909 C->set_major_progress();
4910 }
4911 }
4912
4913 // Perform iteration-splitting on inner loops. Split iterations to avoid
4914 // range checks or one-shot null checks.
4915
4916 // If split-if's didn't hack the graph too bad (no CFG changes)
4917 // then do loop opts.
4918 if (C->has_loops() && !C->major_progress()) {
4919 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4920 _ltree_root->_child->iteration_split( this, worklist );
4921 // No verify after peeling! GCM has hoisted code out of the loop.
4922 // After peeling, the hoisted code could sink inside the peeled area.
4923 // The peeling code does not try to recompute the best location for
4924 // all the code before the peeled area, so the verify pass will always
4925 // complain about it.
4926 }
4927
4928 // Check for bailout, and return
4929 if (C->failing()) {
4930 return;
4931 }
4932
4933 // Do verify graph edges in any case
4934 NOT_PRODUCT( C->verify_graph_edges(); );
4935
4936 if (!do_split_ifs) {
4937 // We saw major progress in Split-If to get here. We forced a
4938 // pass with unrolling and not split-if, however more split-if's
4939 // might make progress. If the unrolling didn't make progress
4940 // then the major-progress flag got cleared and we won't try
4941 // another round of Split-If. In particular the ever-common
4942 // instance-of/check-cast pattern requires at least 2 rounds of
4943 // Split-If to clear out.
4944 C->set_major_progress();
4945 }
4946
4947 // Repeat loop optimizations if new loops were seen
4948 if (created_loop_node()) {
4949 C->set_major_progress();
4950 }
4951
4952 // Keep loop predicates and perform optimizations with them
4953 // until no more loop optimizations could be done.
4954 // After that switch predicates off and do more loop optimizations.
4955 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
4956 C->mark_parse_predicate_nodes_useless(_igvn);
4957 assert(C->parse_predicate_count() == 0, "should be zero now");
4958 if (TraceLoopOpts) {
4959 tty->print_cr("PredicatesOff");
4960 }
4961 C->set_major_progress();
4962 }
4963
4964 // Auto-vectorize main-loop
4965 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
4966 Compile::TracePhase tp(_t_autoVectorize);
4967
4968 // Shared data structures for all AutoVectorizations, to reduce allocations
4969 // of large arrays.
4970 VSharedData vshared;
4971 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4972 IdealLoopTree* lpt = iter.current();
4973 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
4974
4975 if (status == AutoVectorizeStatus::TriedAndFailed) {
4976 // We tried vectorization, but failed. From now on only unroll the loop.
4977 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4978 if (cl->has_passed_slp()) {
4979 C->set_major_progress();
4980 cl->set_notpassed_slp();
4981 cl->mark_do_unroll_only();
4982 }
4983 }
4984 }
4985 }
4986
4987 // Move UnorderedReduction out of counted loop. Can be introduced by AutoVectorization.
4988 if (C->has_loops() && !C->major_progress()) {
4989 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4990 IdealLoopTree* lpt = iter.current();
4991 if (lpt->is_counted() && lpt->is_innermost()) {
4992 move_unordered_reduction_out_of_loop(lpt);
4993 }
4994 }
4995 }
4996 }
4997
4998 #ifndef PRODUCT
4999 //------------------------------print_statistics-------------------------------
5000 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
5001 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5002 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5003 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5004 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5005 void PhaseIdealLoop::print_statistics() {
5006 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);
5007 }
5008 #endif
5009
5010 #ifdef ASSERT
5011 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5012 void PhaseIdealLoop::verify() const {
5013 ResourceMark rm;
5014 int old_progress = C->major_progress();
5015 bool success = true;
5016
5017 PhaseIdealLoop phase_verify(_igvn, this);
5018 if (C->failing_internal()) {
5019 return;
5020 }
5021
5022 // Verify ctrl and idom of every node.
5023 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5024
5025 // Verify loop-tree.
5026 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5027
5028 assert(success, "VerifyLoopOptimizations failed");
5029
5030 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5031 C->restore_major_progress(old_progress);
5032 }
5033
5034 // Perform a BFS starting at n, through all inputs.
5035 // Call verify_idom and verify_node on all nodes of BFS traversal.
5036 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5037 Unique_Node_List worklist;
5038 worklist.push(root);
5039 bool success = true;
5040 for (uint i = 0; i < worklist.size(); i++) {
5041 Node* n = worklist.at(i);
5042 // process node
5043 success &= verify_idom(n, phase_verify);
5044 success &= verify_loop_ctrl(n, phase_verify);
5045 // visit inputs
5046 for (uint j = 0; j < n->req(); j++) {
5047 if (n->in(j) != nullptr) {
5048 worklist.push(n->in(j));
5049 }
5050 }
5051 }
5052 return success;
5053 }
5054
5055 // Verify dominator structure (IDOM).
5056 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5057 // Verify IDOM for all CFG nodes (except root).
5058 if (!n->is_CFG() || n->is_Root()) {
5059 return true; // pass
5060 }
5061
5062 if (n->_idx >= _idom_size) {
5063 tty->print("CFG Node with no idom: ");
5064 n->dump();
5065 return false; // fail
5066 }
5067
5068 Node* id = idom_no_update(n);
5069 Node* id_verify = phase_verify->idom_no_update(n);
5070 if (id != id_verify) {
5071 tty->print("Mismatching idom for node: ");
5072 n->dump();
5073 tty->print(" We have idom: ");
5074 id->dump();
5075 tty->print(" Verify has idom: ");
5076 id_verify->dump();
5077 tty->cr();
5078 return false; // fail
5079 }
5080 return true; // pass
5081 }
5082
5083 // Verify "_loop_or_ctrl": control and loop membership.
5084 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5085 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5086 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5087 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5088 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5089 const uint i = n->_idx;
5090 // The loop-tree was built from def to use (top-down).
5091 // The verification happens from use to def (bottom-up).
5092 // We may thus find nodes during verification that are not in the loop-tree.
5093 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5094 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5095 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5096 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5097 n->dump();
5098 return false; // fail
5099 }
5100 // Not reachable for both.
5101 return true; // pass
5102 }
5103
5104 if (n->is_CFG() == has_ctrl(n)) {
5105 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5106 n->dump();
5107 return false; // fail
5108 }
5109
5110 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5111 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5112 n->dump();
5113 return false; // fail
5114 } else if (has_ctrl(n)) {
5115 assert(phase_verify->has_ctrl(n), "sanity");
5116 // n is a data node.
5117 // Verify that its ctrl is the same.
5118
5119 // Broken part of VerifyLoopOptimizations (A)
5120 // Reason:
5121 // BUG, wrong control set for example in
5122 // PhaseIdealLoop::split_if_with_blocks
5123 // at "set_ctrl(x, new_ctrl);"
5124 /*
5125 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5126 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5127 tty->print("Mismatched control setting for: ");
5128 n->dump();
5129 if( fail++ > 10 ) return;
5130 Node *c = get_ctrl_no_update(n);
5131 tty->print("We have it as: ");
5132 if( c->in(0) ) c->dump();
5133 else tty->print_cr("N%d",c->_idx);
5134 tty->print("Verify thinks: ");
5135 if( loop_verify->has_ctrl(n) )
5136 loop_verify->get_ctrl_no_update(n)->dump();
5137 else
5138 loop_verify->get_loop_idx(n)->dump();
5139 tty->cr();
5140 }
5141 */
5142 return true; // pass
5143 } else {
5144 assert(!phase_verify->has_ctrl(n), "sanity");
5145 // n is a ctrl node.
5146 // Verify that not has_ctrl, and that get_loop_idx is the same.
5147
5148 // Broken part of VerifyLoopOptimizations (B)
5149 // Reason:
5150 // NeverBranch node for example is added to loop outside its scope.
5151 // Once we run build_loop_tree again, it is added to the correct loop.
5152 /*
5153 if (!C->major_progress()) {
5154 // Loop selection can be messed up if we did a major progress
5155 // operation, like split-if. Do not verify in that case.
5156 IdealLoopTree *us = get_loop_idx(n);
5157 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5158 if( us->_head != them->_head || us->_tail != them->_tail ) {
5159 tty->print("Unequals loops for: ");
5160 n->dump();
5161 if( fail++ > 10 ) return;
5162 tty->print("We have it as: ");
5163 us->dump();
5164 tty->print("Verify thinks: ");
5165 them->dump();
5166 tty->cr();
5167 }
5168 }
5169 */
5170 return true; // pass
5171 }
5172 }
5173
5174 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5175 assert(a != nullptr && b != nullptr, "must be");
5176 return a->_head->_idx - b->_head->_idx;
5177 }
5178
5179 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5180 GrowableArray<IdealLoopTree*> children;
5181 IdealLoopTree* child = _child;
5182 while (child != nullptr) {
5183 assert(child->_parent == this, "all must be children of this");
5184 children.insert_sorted<compare_tree>(child);
5185 child = child->_next;
5186 }
5187 return children;
5188 }
5189
5190 // Verify that tree structures match. Because the CFG can change, siblings
5191 // within the loop tree can be reordered. We attempt to deal with that by
5192 // reordering the verify's loop tree if possible.
5193 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5194 assert(_head == loop_verify->_head, "mismatched loop head");
5195 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5196
5197 // Collect the children
5198 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5199 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5200
5201 bool success = true;
5202
5203 // Compare the two children lists
5204 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5205 IdealLoopTree* child = nullptr;
5206 IdealLoopTree* child_verify = nullptr;
5207 // Read from both lists, if possible.
5208 if (i < children.length()) {
5209 child = children.at(i);
5210 }
5211 if (j < children_verify.length()) {
5212 child_verify = children_verify.at(j);
5213 }
5214 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5215 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5216 // We found two non-equal children. Select the smaller one.
5217 if (child->_head->_idx < child_verify->_head->_idx) {
5218 child_verify = nullptr;
5219 } else {
5220 child = nullptr;
5221 }
5222 }
5223 // Process the two children, or potentially log the failure if we only found one.
5224 if (child_verify == nullptr) {
5225 if (child->_irreducible && Compile::current()->major_progress()) {
5226 // Irreducible loops can pick a different header (one of its entries).
5227 } else {
5228 tty->print_cr("We have a loop that verify does not have");
5229 child->dump();
5230 success = false;
5231 }
5232 i++; // step for this
5233 } else if (child == nullptr) {
5234 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5235 // Irreducible loops can pick a different header (one of its entries).
5236 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5237 // Infinite loops do not get attached to the loop-tree on their first visit.
5238 // "this" runs before "loop_verify". It is thus possible that we find the
5239 // infinite loop only for "child_verify". Only finding it with "child" would
5240 // mean that we lost it, which is not ok.
5241 } else {
5242 tty->print_cr("Verify has a loop that we do not have");
5243 child_verify->dump();
5244 success = false;
5245 }
5246 j++; // step for verify
5247 } else {
5248 assert(child->_head == child_verify->_head, "We have both and they are equal");
5249 success &= child->verify_tree(child_verify); // Recursion
5250 i++; // step for this
5251 j++; // step for verify
5252 }
5253 }
5254
5255 // Broken part of VerifyLoopOptimizations (D)
5256 // Reason:
5257 // split_if has to update the _tail, if it is modified. But that is done by
5258 // checking to what loop the iff belongs to. That info can be wrong, and then
5259 // we do not update the _tail correctly.
5260 /*
5261 Node *tail = _tail; // Inline a non-updating version of
5262 while( !tail->in(0) ) // the 'tail()' call.
5263 tail = tail->in(1);
5264 assert( tail == loop->_tail, "mismatched loop tail" );
5265 */
5266
5267 if (_head->is_CountedLoop()) {
5268 CountedLoopNode *cl = _head->as_CountedLoop();
5269
5270 Node* ctrl = cl->init_control();
5271 Node* back = cl->back_control();
5272 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5273 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5274 cl->loopexit(); // assert implied
5275 }
5276
5277 // Broken part of VerifyLoopOptimizations (E)
5278 // Reason:
5279 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5280 // to the loop body. Or maybe they are not added to the correct loop.
5281 // at "Node* x = n->clone();"
5282 /*
5283 // Innermost loops need to verify loop bodies,
5284 // but only if no 'major_progress'
5285 int fail = 0;
5286 if (!Compile::current()->major_progress() && _child == nullptr) {
5287 for( uint i = 0; i < _body.size(); i++ ) {
5288 Node *n = _body.at(i);
5289 if (n->outcnt() == 0) continue; // Ignore dead
5290 uint j;
5291 for( j = 0; j < loop->_body.size(); j++ )
5292 if( loop->_body.at(j) == n )
5293 break;
5294 if( j == loop->_body.size() ) { // Not found in loop body
5295 // Last ditch effort to avoid assertion: Its possible that we
5296 // have some users (so outcnt not zero) but are still dead.
5297 // Try to find from root.
5298 if (Compile::current()->root()->find(n->_idx)) {
5299 fail++;
5300 tty->print("We have that verify does not: ");
5301 n->dump();
5302 }
5303 }
5304 }
5305 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5306 Node *n = loop->_body.at(i2);
5307 if (n->outcnt() == 0) continue; // Ignore dead
5308 uint j;
5309 for( j = 0; j < _body.size(); j++ )
5310 if( _body.at(j) == n )
5311 break;
5312 if( j == _body.size() ) { // Not found in loop body
5313 // Last ditch effort to avoid assertion: Its possible that we
5314 // have some users (so outcnt not zero) but are still dead.
5315 // Try to find from root.
5316 if (Compile::current()->root()->find(n->_idx)) {
5317 fail++;
5318 tty->print("Verify has that we do not: ");
5319 n->dump();
5320 }
5321 }
5322 }
5323 assert( !fail, "loop body mismatch" );
5324 }
5325 */
5326 return success;
5327 }
5328 #endif
5329
5330 //------------------------------set_idom---------------------------------------
5331 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5332 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5333 uint idx = d->_idx;
5334 if (idx >= _idom_size) {
5335 uint newsize = next_power_of_2(idx);
5336 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5337 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5338 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5339 _idom_size = newsize;
5340 }
5341 _idom[idx] = n;
5342 _dom_depth[idx] = dom_depth;
5343 }
5344
5345 //------------------------------recompute_dom_depth---------------------------------------
5346 // The dominator tree is constructed with only parent pointers.
5347 // This recomputes the depth in the tree by first tagging all
5348 // nodes as "no depth yet" marker. The next pass then runs up
5349 // the dom tree from each node marked "no depth yet", and computes
5350 // the depth on the way back down.
5351 void PhaseIdealLoop::recompute_dom_depth() {
5352 uint no_depth_marker = C->unique();
5353 uint i;
5354 // Initialize depth to "no depth yet" and realize all lazy updates
5355 for (i = 0; i < _idom_size; i++) {
5356 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5357 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5358 _dom_depth[i] = no_depth_marker;
5359
5360 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5361 if (_idom[i]->in(0) == nullptr) {
5362 idom(i);
5363 }
5364 }
5365 }
5366 if (_dom_stk == nullptr) {
5367 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5368 if (init_size < 10) init_size = 10;
5369 _dom_stk = new GrowableArray<uint>(init_size);
5370 }
5371 // Compute new depth for each node.
5372 for (i = 0; i < _idom_size; i++) {
5373 uint j = i;
5374 // Run up the dom tree to find a node with a depth
5375 while (_dom_depth[j] == no_depth_marker) {
5376 _dom_stk->push(j);
5377 j = _idom[j]->_idx;
5378 }
5379 // Compute the depth on the way back down this tree branch
5380 uint dd = _dom_depth[j] + 1;
5381 while (_dom_stk->length() > 0) {
5382 uint j = _dom_stk->pop();
5383 _dom_depth[j] = dd;
5384 dd++;
5385 }
5386 }
5387 }
5388
5389 //------------------------------sort-------------------------------------------
5390 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5391 // loop tree, not the root.
5392 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5393 if( !innermost ) return loop; // New innermost loop
5394
5395 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5396 assert( loop_preorder, "not yet post-walked loop" );
5397 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5398 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5399
5400 // Insert at start of list
5401 while( l ) { // Insertion sort based on pre-order
5402 if( l == loop ) return innermost; // Already on list!
5403 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5404 assert( l_preorder, "not yet post-walked l" );
5405 // Check header pre-order number to figure proper nesting
5406 if( loop_preorder > l_preorder )
5407 break; // End of insertion
5408 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5409 // Since I split shared headers, you'd think this could not happen.
5410 // BUT: I must first do the preorder numbering before I can discover I
5411 // have shared headers, so the split headers all get the same preorder
5412 // number as the RegionNode they split from.
5413 if( loop_preorder == l_preorder &&
5414 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5415 break; // Also check for shared headers (same pre#)
5416 pp = &l->_parent; // Chain up list
5417 l = *pp;
5418 }
5419 // Link into list
5420 // Point predecessor to me
5421 *pp = loop;
5422 // Point me to successor
5423 IdealLoopTree *p = loop->_parent;
5424 loop->_parent = l; // Point me to successor
5425 if( p ) sort( p, innermost ); // Insert my parents into list as well
5426 return innermost;
5427 }
5428
5429 //------------------------------build_loop_tree--------------------------------
5430 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5431 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5432 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5433 // tightest enclosing IdealLoopTree for post-walked.
5434 //
5435 // During my forward walk I do a short 1-layer lookahead to see if I can find
5436 // a loop backedge with that doesn't have any work on the backedge. This
5437 // helps me construct nested loops with shared headers better.
5438 //
5439 // Once I've done the forward recursion, I do the post-work. For each child
5440 // I check to see if there is a backedge. Backedges define a loop! I
5441 // insert an IdealLoopTree at the target of the backedge.
5442 //
5443 // During the post-work I also check to see if I have several children
5444 // belonging to different loops. If so, then this Node is a decision point
5445 // where control flow can choose to change loop nests. It is at this
5446 // decision point where I can figure out how loops are nested. At this
5447 // time I can properly order the different loop nests from my children.
5448 // Note that there may not be any backedges at the decision point!
5449 //
5450 // Since the decision point can be far removed from the backedges, I can't
5451 // order my loops at the time I discover them. Thus at the decision point
5452 // I need to inspect loop header pre-order numbers to properly nest my
5453 // loops. This means I need to sort my childrens' loops by pre-order.
5454 // The sort is of size number-of-control-children, which generally limits
5455 // it to size 2 (i.e., I just choose between my 2 target loops).
5456 void PhaseIdealLoop::build_loop_tree() {
5457 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5458 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5459 Node *n = C->root();
5460 bltstack.push(n);
5461 int pre_order = 1;
5462 int stack_size;
5463
5464 while ( ( stack_size = bltstack.length() ) != 0 ) {
5465 n = bltstack.top(); // Leave node on stack
5466 if ( !is_visited(n) ) {
5467 // ---- Pre-pass Work ----
5468 // Pre-walked but not post-walked nodes need a pre_order number.
5469
5470 set_preorder_visited( n, pre_order ); // set as visited
5471
5472 // ---- Scan over children ----
5473 // Scan first over control projections that lead to loop headers.
5474 // This helps us find inner-to-outer loops with shared headers better.
5475
5476 // Scan children's children for loop headers.
5477 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5478 Node* m = n->raw_out(i); // Child
5479 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5480 // Scan over children's children to find loop
5481 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5482 Node* l = m->fast_out(j);
5483 if( is_visited(l) && // Been visited?
5484 !is_postvisited(l) && // But not post-visited
5485 get_preorder(l) < pre_order ) { // And smaller pre-order
5486 // Found! Scan the DFS down this path before doing other paths
5487 bltstack.push(m);
5488 break;
5489 }
5490 }
5491 }
5492 }
5493 pre_order++;
5494 }
5495 else if ( !is_postvisited(n) ) {
5496 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5497 // such as com.sun.rsasign.am::a.
5498 // For non-recursive version, first, process current children.
5499 // On next iteration, check if additional children were added.
5500 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5501 Node* u = n->raw_out(k);
5502 if ( u->is_CFG() && !is_visited(u) ) {
5503 bltstack.push(u);
5504 }
5505 }
5506 if ( bltstack.length() == stack_size ) {
5507 // There were no additional children, post visit node now
5508 (void)bltstack.pop(); // Remove node from stack
5509 pre_order = build_loop_tree_impl(n, pre_order);
5510 // Check for bailout
5511 if (C->failing()) {
5512 return;
5513 }
5514 // Check to grow _preorders[] array for the case when
5515 // build_loop_tree_impl() adds new nodes.
5516 check_grow_preorders();
5517 }
5518 }
5519 else {
5520 (void)bltstack.pop(); // Remove post-visited node from stack
5521 }
5522 }
5523 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5524 }
5525
5526 //------------------------------build_loop_tree_impl---------------------------
5527 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5528 // ---- Post-pass Work ----
5529 // Pre-walked but not post-walked nodes need a pre_order number.
5530
5531 // Tightest enclosing loop for this Node
5532 IdealLoopTree *innermost = nullptr;
5533
5534 // For all children, see if any edge is a backedge. If so, make a loop
5535 // for it. Then find the tightest enclosing loop for the self Node.
5536 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5537 Node* m = n->fast_out(i); // Child
5538 if (n == m) continue; // Ignore control self-cycles
5539 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5540
5541 IdealLoopTree *l; // Child's loop
5542 if (!is_postvisited(m)) { // Child visited but not post-visited?
5543 // Found a backedge
5544 assert(get_preorder(m) < pre_order, "should be backedge");
5545 // Check for the RootNode, which is already a LoopNode and is allowed
5546 // to have multiple "backedges".
5547 if (m == C->root()) { // Found the root?
5548 l = _ltree_root; // Root is the outermost LoopNode
5549 } else { // Else found a nested loop
5550 // Insert a LoopNode to mark this loop.
5551 l = new IdealLoopTree(this, m, n);
5552 } // End of Else found a nested loop
5553 if (!has_loop(m)) { // If 'm' does not already have a loop set
5554 set_loop(m, l); // Set loop header to loop now
5555 }
5556 } else { // Else not a nested loop
5557 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5558 IdealLoopTree* m_loop = get_loop(m);
5559 l = m_loop; // Get previously determined loop
5560 // If successor is header of a loop (nest), move up-loop till it
5561 // is a member of some outer enclosing loop. Since there are no
5562 // shared headers (I've split them already) I only need to go up
5563 // at most 1 level.
5564 while (l && l->_head == m) { // Successor heads loop?
5565 l = l->_parent; // Move up 1 for me
5566 }
5567 // If this loop is not properly parented, then this loop
5568 // has no exit path out, i.e. its an infinite loop.
5569 if (!l) {
5570 // Make loop "reachable" from root so the CFG is reachable. Basically
5571 // insert a bogus loop exit that is never taken. 'm', the loop head,
5572 // points to 'n', one (of possibly many) fall-in paths. There may be
5573 // many backedges as well.
5574
5575 if (!_verify_only) {
5576 // Insert the NeverBranch between 'm' and it's control user.
5577 NeverBranchNode *iff = new NeverBranchNode( m );
5578 _igvn.register_new_node_with_optimizer(iff);
5579 set_loop(iff, m_loop);
5580 Node *if_t = new CProjNode( iff, 0 );
5581 _igvn.register_new_node_with_optimizer(if_t);
5582 set_loop(if_t, m_loop);
5583
5584 Node* cfg = nullptr; // Find the One True Control User of m
5585 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5586 Node* x = m->fast_out(j);
5587 if (x->is_CFG() && x != m && x != iff)
5588 { cfg = x; break; }
5589 }
5590 assert(cfg != nullptr, "must find the control user of m");
5591 uint k = 0; // Probably cfg->in(0)
5592 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5593 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5594
5595 // Now create the never-taken loop exit
5596 Node *if_f = new CProjNode( iff, 1 );
5597 _igvn.register_new_node_with_optimizer(if_f);
5598 set_loop(if_f, _ltree_root);
5599 // Find frame ptr for Halt. Relies on the optimizer
5600 // V-N'ing. Easier and quicker than searching through
5601 // the program structure.
5602 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5603 _igvn.register_new_node_with_optimizer(frame);
5604 // Halt & Catch Fire
5605 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5606 _igvn.register_new_node_with_optimizer(halt);
5607 set_loop(halt, _ltree_root);
5608 _igvn.add_input_to(C->root(), halt);
5609 }
5610 set_loop(C->root(), _ltree_root);
5611 // move to outer most loop with same header
5612 l = m_loop;
5613 while (true) {
5614 IdealLoopTree* next = l->_parent;
5615 if (next == nullptr || next->_head != m) {
5616 break;
5617 }
5618 l = next;
5619 }
5620 // properly insert infinite loop in loop tree
5621 sort(_ltree_root, l);
5622 // fix child link from parent
5623 IdealLoopTree* p = l->_parent;
5624 l->_next = p->_child;
5625 p->_child = l;
5626 // code below needs enclosing loop
5627 l = l->_parent;
5628 }
5629 }
5630 if (is_postvisited(l->_head)) {
5631 // We are currently visiting l, but its head has already been post-visited.
5632 // l is irreducible: we just found a second entry m.
5633 _has_irreducible_loops = true;
5634 RegionNode* secondary_entry = m->as_Region();
5635 DEBUG_ONLY(secondary_entry->verify_can_be_irreducible_entry();)
5636
5637 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
5638 // Since m is a secondary entry to them all.
5639 while( is_postvisited(l->_head) ) {
5640 l->_irreducible = 1; // = true
5641 RegionNode* head = l->_head->as_Region();
5642 DEBUG_ONLY(head->verify_can_be_irreducible_entry();)
5643 l = l->_parent;
5644 // Check for bad CFG here to prevent crash, and bailout of compile
5645 if (l == nullptr) {
5646 #ifndef PRODUCT
5647 if (TraceLoopOpts) {
5648 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
5649 m->dump();
5650 }
5651 #endif
5652 // This is a rare case that we do not want to handle in C2.
5653 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
5654 return pre_order;
5655 }
5656 }
5657 }
5658 if (!_verify_only) {
5659 C->set_has_irreducible_loop(_has_irreducible_loops);
5660 }
5661
5662 // This Node might be a decision point for loops. It is only if
5663 // it's children belong to several different loops. The sort call
5664 // does a trivial amount of work if there is only 1 child or all
5665 // children belong to the same loop. If however, the children
5666 // belong to different loops, the sort call will properly set the
5667 // _parent pointers to show how the loops nest.
5668 //
5669 // In any case, it returns the tightest enclosing loop.
5670 innermost = sort( l, innermost );
5671 }
5672
5673 // Def-use info will have some dead stuff; dead stuff will have no
5674 // loop decided on.
5675
5676 // Am I a loop header? If so fix up my parent's child and next ptrs.
5677 if( innermost && innermost->_head == n ) {
5678 assert( get_loop(n) == innermost, "" );
5679 IdealLoopTree *p = innermost->_parent;
5680 IdealLoopTree *l = innermost;
5681 while (p && l->_head == n) {
5682 l->_next = p->_child; // Put self on parents 'next child'
5683 p->_child = l; // Make self as first child of parent
5684 l = p; // Now walk up the parent chain
5685 p = l->_parent;
5686 }
5687 } else {
5688 // Note that it is possible for a LoopNode to reach here, if the
5689 // backedge has been made unreachable (hence the LoopNode no longer
5690 // denotes a Loop, and will eventually be removed).
5691
5692 // Record tightest enclosing loop for self. Mark as post-visited.
5693 set_loop(n, innermost);
5694 // Also record has_call flag early on
5695 if (innermost) {
5696 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
5697 // Do not count uncommon calls
5698 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
5699 Node *iff = n->in(0)->in(0);
5700 // No any calls for vectorized loops.
5701 if (C->do_superword() ||
5702 !iff->is_If() ||
5703 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
5704 iff->as_If()->_prob >= 0.01) {
5705 innermost->_has_call = 1;
5706 }
5707 }
5708 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
5709 // Disable loop optimizations if the loop has a scalar replaceable
5710 // allocation. This disabling may cause a potential performance lost
5711 // if the allocation is not eliminated for some reason.
5712 innermost->_allow_optimizations = false;
5713 innermost->_has_call = 1; // = true
5714 } else if (n->Opcode() == Op_SafePoint) {
5715 // Record all safepoints in this loop.
5716 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
5717 innermost->_safepts->push(n);
5718 }
5719 }
5720 }
5721
5722 // Flag as post-visited now
5723 set_postvisited(n);
5724 return pre_order;
5725 }
5726
5727 #ifdef ASSERT
5728 //--------------------------verify_regions_in_irreducible_loops----------------
5729 // Iterate down from Root through CFG, verify for every region:
5730 // if it is in an irreducible loop it must be marked as such
5731 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
5732 ResourceMark rm;
5733 if (!_has_irreducible_loops) {
5734 // last build_loop_tree has not found any irreducible loops
5735 // hence no region has to be marked is_in_irreduible_loop
5736 return;
5737 }
5738
5739 RootNode* root = C->root();
5740 Unique_Node_List worklist; // visit all nodes once
5741 worklist.push(root);
5742 bool failure = false;
5743 for (uint i = 0; i < worklist.size(); i++) {
5744 Node* n = worklist.at(i);
5745 if (n->is_Region()) {
5746 RegionNode* region = n->as_Region();
5747 if (is_in_irreducible_loop(region) &&
5748 region->loop_status() == RegionNode::LoopStatus::Reducible) {
5749 failure = true;
5750 tty->print("irreducible! ");
5751 region->dump();
5752 }
5753 }
5754 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
5755 Node* use = n->fast_out(j);
5756 if (use->is_CFG()) {
5757 worklist.push(use); // push if was not pushed before
5758 }
5759 }
5760 }
5761 assert(!failure, "region in irreducible loop was marked as reducible");
5762 }
5763
5764 //---------------------------is_in_irreducible_loop-------------------------
5765 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
5766 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
5767 if (!_has_irreducible_loops) {
5768 return false; // no irreducible loop in graph
5769 }
5770 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
5771 do {
5772 if (l->_irreducible) {
5773 return true; // found it
5774 }
5775 if (l == _ltree_root) {
5776 return false; // reached root, terimnate
5777 }
5778 l = l->_parent;
5779 } while (l != nullptr);
5780 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
5781 // We have "l->_parent == nullptr", which happens only for infinite loops,
5782 // where no parent is attached to the loop. We did not find any irreducible
5783 // loop from this block out to lp. Thus lp only has one entry, and no exit
5784 // (it is infinite and reducible). We can always rewrite an infinite loop
5785 // that is nested inside other loops:
5786 // while(condition) { infinite_loop; }
5787 // with an equivalent program where the infinite loop is an outermost loop
5788 // that is not nested in any loop:
5789 // while(condition) { break; } infinite_loop;
5790 // Thus, we can understand lp as an outermost loop, and can terminate and
5791 // conclude: this block is in no irreducible loop.
5792 return false;
5793 }
5794 #endif
5795
5796 //------------------------------build_loop_early-------------------------------
5797 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
5798 // First pass computes the earliest controlling node possible. This is the
5799 // controlling input with the deepest dominating depth.
5800 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
5801 while (worklist.size() != 0) {
5802 // Use local variables nstack_top_n & nstack_top_i to cache values
5803 // on nstack's top.
5804 Node *nstack_top_n = worklist.pop();
5805 uint nstack_top_i = 0;
5806 //while_nstack_nonempty:
5807 while (true) {
5808 // Get parent node and next input's index from stack's top.
5809 Node *n = nstack_top_n;
5810 uint i = nstack_top_i;
5811 uint cnt = n->req(); // Count of inputs
5812 if (i == 0) { // Pre-process the node.
5813 if( has_node(n) && // Have either loop or control already?
5814 !has_ctrl(n) ) { // Have loop picked out already?
5815 // During "merge_many_backedges" we fold up several nested loops
5816 // into a single loop. This makes the members of the original
5817 // loop bodies pointing to dead loops; they need to move up
5818 // to the new UNION'd larger loop. I set the _head field of these
5819 // dead loops to null and the _parent field points to the owning
5820 // loop. Shades of UNION-FIND algorithm.
5821 IdealLoopTree *ilt;
5822 while( !(ilt = get_loop(n))->_head ) {
5823 // Normally I would use a set_loop here. But in this one special
5824 // case, it is legal (and expected) to change what loop a Node
5825 // belongs to.
5826 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
5827 }
5828 // Remove safepoints ONLY if I've already seen I don't need one.
5829 // (the old code here would yank a 2nd safepoint after seeing a
5830 // first one, even though the 1st did not dominate in the loop body
5831 // and thus could be avoided indefinitely)
5832 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
5833 is_deleteable_safept(n)) {
5834 Node *in = n->in(TypeFunc::Control);
5835 lazy_replace(n,in); // Pull safepoint now
5836 if (ilt->_safepts != nullptr) {
5837 ilt->_safepts->yank(n);
5838 }
5839 // Carry on with the recursion "as if" we are walking
5840 // only the control input
5841 if( !visited.test_set( in->_idx ) ) {
5842 worklist.push(in); // Visit this guy later, using worklist
5843 }
5844 // Get next node from nstack:
5845 // - skip n's inputs processing by setting i > cnt;
5846 // - we also will not call set_early_ctrl(n) since
5847 // has_node(n) == true (see the condition above).
5848 i = cnt + 1;
5849 }
5850 }
5851 } // if (i == 0)
5852
5853 // Visit all inputs
5854 bool done = true; // Assume all n's inputs will be processed
5855 while (i < cnt) {
5856 Node *in = n->in(i);
5857 ++i;
5858 if (in == nullptr) continue;
5859 if (in->pinned() && !in->is_CFG())
5860 set_ctrl(in, in->in(0));
5861 int is_visited = visited.test_set( in->_idx );
5862 if (!has_node(in)) { // No controlling input yet?
5863 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
5864 assert( !is_visited, "visit only once" );
5865 nstack.push(n, i); // Save parent node and next input's index.
5866 nstack_top_n = in; // Process current input now.
5867 nstack_top_i = 0;
5868 done = false; // Not all n's inputs processed.
5869 break; // continue while_nstack_nonempty;
5870 } else if (!is_visited) {
5871 // This guy has a location picked out for him, but has not yet
5872 // been visited. Happens to all CFG nodes, for instance.
5873 // Visit him using the worklist instead of recursion, to break
5874 // cycles. Since he has a location already we do not need to
5875 // find his location before proceeding with the current Node.
5876 worklist.push(in); // Visit this guy later, using worklist
5877 }
5878 }
5879 if (done) {
5880 // All of n's inputs have been processed, complete post-processing.
5881
5882 // Compute earliest point this Node can go.
5883 // CFG, Phi, pinned nodes already know their controlling input.
5884 if (!has_node(n)) {
5885 // Record earliest legal location
5886 set_early_ctrl(n, false);
5887 }
5888 if (nstack.is_empty()) {
5889 // Finished all nodes on stack.
5890 // Process next node on the worklist.
5891 break;
5892 }
5893 // Get saved parent node and next input's index.
5894 nstack_top_n = nstack.node();
5895 nstack_top_i = nstack.index();
5896 nstack.pop();
5897 }
5898 } // while (true)
5899 }
5900 }
5901
5902 //------------------------------dom_lca_internal--------------------------------
5903 // Pair-wise LCA
5904 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
5905 if( !n1 ) return n2; // Handle null original LCA
5906 assert( n1->is_CFG(), "" );
5907 assert( n2->is_CFG(), "" );
5908 // find LCA of all uses
5909 uint d1 = dom_depth(n1);
5910 uint d2 = dom_depth(n2);
5911 while (n1 != n2) {
5912 if (d1 > d2) {
5913 n1 = idom(n1);
5914 d1 = dom_depth(n1);
5915 } else if (d1 < d2) {
5916 n2 = idom(n2);
5917 d2 = dom_depth(n2);
5918 } else {
5919 // Here d1 == d2. Due to edits of the dominator-tree, sections
5920 // of the tree might have the same depth. These sections have
5921 // to be searched more carefully.
5922
5923 // Scan up all the n1's with equal depth, looking for n2.
5924 Node *t1 = idom(n1);
5925 while (dom_depth(t1) == d1) {
5926 if (t1 == n2) return n2;
5927 t1 = idom(t1);
5928 }
5929 // Scan up all the n2's with equal depth, looking for n1.
5930 Node *t2 = idom(n2);
5931 while (dom_depth(t2) == d2) {
5932 if (t2 == n1) return n1;
5933 t2 = idom(t2);
5934 }
5935 // Move up to a new dominator-depth value as well as up the dom-tree.
5936 n1 = t1;
5937 n2 = t2;
5938 d1 = dom_depth(n1);
5939 d2 = dom_depth(n2);
5940 }
5941 }
5942 return n1;
5943 }
5944
5945 //------------------------------compute_idom-----------------------------------
5946 // Locally compute IDOM using dom_lca call. Correct only if the incoming
5947 // IDOMs are correct.
5948 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
5949 assert( region->is_Region(), "" );
5950 Node *LCA = nullptr;
5951 for( uint i = 1; i < region->req(); i++ ) {
5952 if( region->in(i) != C->top() )
5953 LCA = dom_lca( LCA, region->in(i) );
5954 }
5955 return LCA;
5956 }
5957
5958 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
5959 bool had_error = false;
5960 #ifdef ASSERT
5961 if (early != C->root()) {
5962 // Make sure that there's a dominance path from LCA to early
5963 Node* d = LCA;
5964 while (d != early) {
5965 if (d == C->root()) {
5966 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
5967 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
5968 had_error = true;
5969 break;
5970 }
5971 d = idom(d);
5972 }
5973 }
5974 #endif
5975 return had_error;
5976 }
5977
5978
5979 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
5980 // Compute LCA over list of uses
5981 bool had_error = false;
5982 Node *LCA = nullptr;
5983 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
5984 Node* c = n->fast_out(i);
5985 if (_loop_or_ctrl[c->_idx] == nullptr)
5986 continue; // Skip the occasional dead node
5987 if( c->is_Phi() ) { // For Phis, we must land above on the path
5988 for( uint j=1; j<c->req(); j++ ) {// For all inputs
5989 if( c->in(j) == n ) { // Found matching input?
5990 Node *use = c->in(0)->in(j);
5991 if (_verify_only && use->is_top()) continue;
5992 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
5993 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
5994 }
5995 }
5996 } else {
5997 // For CFG data-users, use is in the block just prior
5998 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
5999 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6000 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6001 }
6002 }
6003 assert(!had_error, "bad dominance");
6004 return LCA;
6005 }
6006
6007 // Check the shape of the graph at the loop entry. In some cases,
6008 // the shape of the graph does not match the shape outlined below.
6009 // That is caused by the Opaque1 node "protecting" the shape of
6010 // the graph being removed by, for example, the IGVN performed
6011 // in PhaseIdealLoop::build_and_optimize().
6012 //
6013 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6014 // loop unswitching, and IGVN, or a combination of them) can freely change
6015 // the graph's shape. As a result, the graph shape outlined below cannot
6016 // be guaranteed anymore.
6017 Node* CountedLoopNode::is_canonical_loop_entry() {
6018 if (!is_main_loop() && !is_post_loop()) {
6019 return nullptr;
6020 }
6021 Node* ctrl = skip_assertion_predicates_with_halt();
6022
6023 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6024 return nullptr;
6025 }
6026 Node* iffm = ctrl->in(0);
6027 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6028 return nullptr;
6029 }
6030 Node* bolzm = iffm->in(1);
6031 if (bolzm == nullptr || !bolzm->is_Bool()) {
6032 return nullptr;
6033 }
6034 Node* cmpzm = bolzm->in(1);
6035 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6036 return nullptr;
6037 }
6038
6039 uint input = is_main_loop() ? 2 : 1;
6040 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6041 return nullptr;
6042 }
6043 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6044 #ifdef ASSERT
6045 bool found_opaque = false;
6046 for (uint i = 1; i < cmpzm->req(); i++) {
6047 Node* opnd = cmpzm->in(i);
6048 if (opnd && opnd->is_Opaque1()) {
6049 found_opaque = true;
6050 break;
6051 }
6052 }
6053 assert(found_opaque == res, "wrong pattern");
6054 #endif
6055 return res ? cmpzm->in(input) : nullptr;
6056 }
6057
6058 // Find pre loop end from main loop. Returns nullptr if none.
6059 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6060 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6061 // The loop cannot be optimized if the graph shape at the loop entry is
6062 // inappropriate.
6063 if (is_canonical_loop_entry() == nullptr) {
6064 return nullptr;
6065 }
6066
6067 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6068 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6069 return nullptr;
6070 }
6071 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6072 CountedLoopNode* loop_node = pre_end->loopnode();
6073 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6074 return nullptr;
6075 }
6076 return pre_end;
6077 }
6078
6079 //------------------------------get_late_ctrl----------------------------------
6080 // Compute latest legal control.
6081 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6082 assert(early != nullptr, "early control should not be null");
6083
6084 Node* LCA = compute_lca_of_uses(n, early);
6085 #ifdef ASSERT
6086 if (LCA == C->root() && LCA != early) {
6087 // def doesn't dominate uses so print some useful debugging output
6088 compute_lca_of_uses(n, early, true);
6089 }
6090 #endif
6091
6092 if (n->is_Load() && LCA != early) {
6093 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6094 }
6095
6096 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6097 return LCA;
6098 }
6099
6100 // if this is a load, check for anti-dependent stores
6101 // We use a conservative algorithm to identify potential interfering
6102 // instructions and for rescheduling the load. The users of the memory
6103 // input of this load are examined. Any use which is not a load and is
6104 // dominated by early is considered a potentially interfering store.
6105 // This can produce false positives.
6106 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6107 int load_alias_idx = C->get_alias_index(n->adr_type());
6108 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6109 Unique_Node_List worklist;
6110
6111 Node* mem = n->in(MemNode::Memory);
6112 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6113 Node* s = mem->fast_out(i);
6114 worklist.push(s);
6115 }
6116 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6117 Node* s = worklist.at(i);
6118 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6119 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6120 s->is_Phi()) {
6121 continue;
6122 } else if (s->is_MergeMem()) {
6123 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6124 Node* s1 = s->fast_out(i);
6125 worklist.push(s1);
6126 }
6127 } else {
6128 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6129 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6130 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6131 const TypePtr* adr_type = s->adr_type();
6132 if (s->is_ArrayCopy()) {
6133 // Copy to known instance needs destination type to test for aliasing
6134 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6135 if (dest_type != TypeOopPtr::BOTTOM) {
6136 adr_type = dest_type;
6137 }
6138 }
6139 if (C->can_alias(adr_type, load_alias_idx)) {
6140 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6141 } else if (s->is_CFG() && s->is_Multi()) {
6142 // Look for the memory use of s (that is the use of its memory projection)
6143 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6144 Node* s1 = s->fast_out(i);
6145 assert(s1->is_Proj(), "projection expected");
6146 if (_igvn.type(s1) == Type::MEMORY) {
6147 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6148 Node* s2 = s1->fast_out(j);
6149 worklist.push(s2);
6150 }
6151 }
6152 }
6153 }
6154 }
6155 }
6156 }
6157 // For Phis only consider Region's inputs that were reached by following the memory edges
6158 if (LCA != early) {
6159 for (uint i = 0; i < worklist.size(); i++) {
6160 Node* s = worklist.at(i);
6161 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6162 Node* r = s->in(0);
6163 for (uint j = 1; j < s->req(); j++) {
6164 Node* in = s->in(j);
6165 Node* r_in = r->in(j);
6166 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6167 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6168 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6169 }
6170 }
6171 }
6172 }
6173 }
6174 }
6175 return LCA;
6176 }
6177
6178 // Is CFG node 'dominator' dominating node 'n'?
6179 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6180 if (dominator == n) {
6181 return true;
6182 }
6183 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6184 uint dd = dom_depth(dominator);
6185 while (dom_depth(n) >= dd) {
6186 if (n == dominator) {
6187 return true;
6188 }
6189 n = idom(n);
6190 }
6191 return false;
6192 }
6193
6194 // Is CFG node 'dominator' strictly dominating node 'n'?
6195 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6196 return dominator != n && is_dominator(dominator, n);
6197 }
6198
6199 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6200 // Pair-wise LCA with tags.
6201 // Tag each index with the node 'tag' currently being processed
6202 // before advancing up the dominator chain using idom().
6203 // Later calls that find a match to 'tag' know that this path has already
6204 // been considered in the current LCA (which is input 'n1' by convention).
6205 // Since get_late_ctrl() is only called once for each node, the tag array
6206 // does not need to be cleared between calls to get_late_ctrl().
6207 // Algorithm trades a larger constant factor for better asymptotic behavior
6208 //
6209 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6210 uint d1 = dom_depth(n1);
6211 uint d2 = dom_depth(n2);
6212 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6213
6214 do {
6215 if (d1 > d2) {
6216 // current lca is deeper than n2
6217 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6218 n1 = idom(n1);
6219 d1 = dom_depth(n1);
6220 } else if (d1 < d2) {
6221 // n2 is deeper than current lca
6222 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6223 if (memo == tag) {
6224 return n1; // Return the current LCA
6225 }
6226 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6227 n2 = idom(n2);
6228 d2 = dom_depth(n2);
6229 } else {
6230 // Here d1 == d2. Due to edits of the dominator-tree, sections
6231 // of the tree might have the same depth. These sections have
6232 // to be searched more carefully.
6233
6234 // Scan up all the n1's with equal depth, looking for n2.
6235 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6236 Node *t1 = idom(n1);
6237 while (dom_depth(t1) == d1) {
6238 if (t1 == n2) return n2;
6239 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6240 t1 = idom(t1);
6241 }
6242 // Scan up all the n2's with equal depth, looking for n1.
6243 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6244 Node *t2 = idom(n2);
6245 while (dom_depth(t2) == d2) {
6246 if (t2 == n1) return n1;
6247 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6248 t2 = idom(t2);
6249 }
6250 // Move up to a new dominator-depth value as well as up the dom-tree.
6251 n1 = t1;
6252 n2 = t2;
6253 d1 = dom_depth(n1);
6254 d2 = dom_depth(n2);
6255 }
6256 } while (n1 != n2);
6257 return n1;
6258 }
6259
6260 //------------------------------init_dom_lca_tags------------------------------
6261 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6262 // Intended use does not involve any growth for the array, so it could
6263 // be of fixed size.
6264 void PhaseIdealLoop::init_dom_lca_tags() {
6265 uint limit = C->unique() + 1;
6266 _dom_lca_tags.at_grow(limit, 0);
6267 _dom_lca_tags_round = 0;
6268 #ifdef ASSERT
6269 for (uint i = 0; i < limit; ++i) {
6270 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6271 }
6272 #endif // ASSERT
6273 }
6274
6275 //------------------------------build_loop_late--------------------------------
6276 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6277 // Second pass finds latest legal placement, and ideal loop placement.
6278 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6279 while (worklist.size() != 0) {
6280 Node *n = worklist.pop();
6281 // Only visit once
6282 if (visited.test_set(n->_idx)) continue;
6283 uint cnt = n->outcnt();
6284 uint i = 0;
6285 while (true) {
6286 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6287 // Visit all children
6288 if (i < cnt) {
6289 Node* use = n->raw_out(i);
6290 ++i;
6291 // Check for dead uses. Aggressively prune such junk. It might be
6292 // dead in the global sense, but still have local uses so I cannot
6293 // easily call 'remove_dead_node'.
6294 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6295 // Due to cycles, we might not hit the same fixed point in the verify
6296 // pass as we do in the regular pass. Instead, visit such phis as
6297 // simple uses of the loop head.
6298 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6299 if( !visited.test(use->_idx) )
6300 worklist.push(use);
6301 } else if( !visited.test_set(use->_idx) ) {
6302 nstack.push(n, i); // Save parent and next use's index.
6303 n = use; // Process all children of current use.
6304 cnt = use->outcnt();
6305 i = 0;
6306 }
6307 } else {
6308 // Do not visit around the backedge of loops via data edges.
6309 // push dead code onto a worklist
6310 _deadlist.push(use);
6311 }
6312 } else {
6313 // All of n's children have been processed, complete post-processing.
6314 build_loop_late_post(n);
6315 if (C->failing()) { return; }
6316 if (nstack.is_empty()) {
6317 // Finished all nodes on stack.
6318 // Process next node on the worklist.
6319 break;
6320 }
6321 // Get saved parent node and next use's index. Visit the rest of uses.
6322 n = nstack.node();
6323 cnt = n->outcnt();
6324 i = nstack.index();
6325 nstack.pop();
6326 }
6327 }
6328 }
6329 }
6330
6331 // Verify that no data node is scheduled in the outer loop of a strip
6332 // mined loop.
6333 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6334 #ifdef ASSERT
6335 if (get_loop(least)->_nest == 0) {
6336 return;
6337 }
6338 IdealLoopTree* loop = get_loop(least);
6339 Node* head = loop->_head;
6340 if (head->is_OuterStripMinedLoop() &&
6341 // Verification can't be applied to fully built strip mined loops
6342 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6343 Node* sfpt = head->as_Loop()->outer_safepoint();
6344 ResourceMark rm;
6345 Unique_Node_List wq;
6346 wq.push(sfpt);
6347 for (uint i = 0; i < wq.size(); i++) {
6348 Node *m = wq.at(i);
6349 for (uint i = 1; i < m->req(); i++) {
6350 Node* nn = m->in(i);
6351 if (nn == n) {
6352 return;
6353 }
6354 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6355 wq.push(nn);
6356 }
6357 }
6358 }
6359 ShouldNotReachHere();
6360 }
6361 #endif
6362 }
6363
6364
6365 //------------------------------build_loop_late_post---------------------------
6366 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6367 // Second pass finds latest legal placement, and ideal loop placement.
6368 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6369 build_loop_late_post_work(n, true);
6370 }
6371
6372 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6373 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6374 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6375 // Predication or creating a Loop Limit Check Predicate later.
6376 class DominatedPredicates : public UnifiedPredicateVisitor {
6377 Node* const _dominator;
6378 Node* _earliest_dominated_predicate_entry;
6379 bool _should_continue;
6380 PhaseIdealLoop* const _phase;
6381
6382 public:
6383 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6384 : _dominator(dominator),
6385 _earliest_dominated_predicate_entry(start_node),
6386 _should_continue(true),
6387 _phase(phase) {}
6388 NONCOPYABLE(DominatedPredicates);
6389
6390 bool should_continue() const override {
6391 return _should_continue;
6392 }
6393
6394 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6395 Node* earliest_dominated_predicate_entry() const {
6396 return _earliest_dominated_predicate_entry;
6397 }
6398
6399 void visit_predicate(const Predicate& predicate) override {
6400 Node* entry = predicate.entry();
6401 if (_phase->is_strict_dominator(entry, _dominator)) {
6402 _should_continue = false;
6403 } else {
6404 _earliest_dominated_predicate_entry = entry;
6405 }
6406 }
6407 };
6408
6409 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6410
6411 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6412 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6413 }
6414
6415 #ifdef ASSERT
6416 if (_verify_only && !n->is_CFG()) {
6417 // Check def-use domination.
6418 // We would like to expose this check in product but it appears to be expensive.
6419 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6420 }
6421 #endif
6422
6423 // CFG and pinned nodes already handled
6424 if( n->in(0) ) {
6425 if( n->in(0)->is_top() ) return; // Dead?
6426
6427 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6428 // _must_ be pinned (they have to observe their control edge of course).
6429 // Unlike Stores (which modify an unallocable resource, the memory
6430 // state), Mods/Loads can float around. So free them up.
6431 switch( n->Opcode() ) {
6432 case Op_DivI:
6433 case Op_DivF:
6434 case Op_DivD:
6435 case Op_ModI:
6436 case Op_ModF:
6437 case Op_ModD:
6438 case Op_LoadB: // Same with Loads; they can sink
6439 case Op_LoadUB: // during loop optimizations.
6440 case Op_LoadUS:
6441 case Op_LoadD:
6442 case Op_LoadF:
6443 case Op_LoadI:
6444 case Op_LoadKlass:
6445 case Op_LoadNKlass:
6446 case Op_LoadL:
6447 case Op_LoadS:
6448 case Op_LoadP:
6449 case Op_LoadN:
6450 case Op_LoadRange:
6451 case Op_LoadD_unaligned:
6452 case Op_LoadL_unaligned:
6453 case Op_StrComp: // Does a bunch of load-like effects
6454 case Op_StrEquals:
6455 case Op_StrIndexOf:
6456 case Op_StrIndexOfChar:
6457 case Op_AryEq:
6458 case Op_VectorizedHashCode:
6459 case Op_CountPositives:
6460 pinned = false;
6461 }
6462 if (n->is_CMove() || n->is_ConstraintCast()) {
6463 pinned = false;
6464 }
6465 if( pinned ) {
6466 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6467 if( !chosen_loop->_child ) // Inner loop?
6468 chosen_loop->_body.push(n); // Collect inner loops
6469 return;
6470 }
6471 } else { // No slot zero
6472 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6473 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6474 return;
6475 }
6476 assert(!n->is_CFG() || n->outcnt() == 0, "");
6477 }
6478
6479 // Do I have a "safe range" I can select over?
6480 Node *early = get_ctrl(n);// Early location already computed
6481
6482 // Compute latest point this Node can go
6483 Node *LCA = get_late_ctrl( n, early );
6484 // LCA is null due to uses being dead
6485 if( LCA == nullptr ) {
6486 #ifdef ASSERT
6487 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6488 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6489 }
6490 #endif
6491 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6492 _deadlist.push(n);
6493 return;
6494 }
6495 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6496
6497 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6498 Node *least = legal; // Best legal position so far
6499 while( early != legal ) { // While not at earliest legal
6500 if (legal->is_Start() && !early->is_Root()) {
6501 #ifdef ASSERT
6502 // Bad graph. Print idom path and fail.
6503 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6504 assert(false, "Bad graph detected in build_loop_late");
6505 #endif
6506 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6507 return;
6508 }
6509 // Find least loop nesting depth
6510 legal = idom(legal); // Bump up the IDOM tree
6511 // Check for lower nesting depth
6512 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6513 least = legal;
6514 }
6515 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6516
6517 if (least != early) {
6518 // Move the node above predicates as far up as possible so a
6519 // following pass of Loop Predication doesn't hoist a predicate
6520 // that depends on it above that node.
6521 const PredicateIterator predicate_iterator(least);
6522 DominatedPredicates dominated_predicates(early, least, this);
6523 predicate_iterator.for_each(dominated_predicates);
6524 least = dominated_predicates.earliest_dominated_predicate_entry();
6525 }
6526 // Try not to place code on a loop entry projection
6527 // which can inhibit range check elimination.
6528 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6529 Node* ctrl_out = least->unique_ctrl_out_or_null();
6530 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6531 least == ctrl_out->in(LoopNode::EntryControl) &&
6532 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6533 Node* least_dom = idom(least);
6534 if (get_loop(least_dom)->is_member(get_loop(least))) {
6535 least = least_dom;
6536 }
6537 }
6538 }
6539 // Don't extend live ranges of raw oops
6540 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6541 !n->bottom_type()->isa_rawptr()) {
6542 least = early;
6543 }
6544
6545 #ifdef ASSERT
6546 // Broken part of VerifyLoopOptimizations (F)
6547 // Reason:
6548 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6549 /*
6550 // If verifying, verify that 'verify_me' has a legal location
6551 // and choose it as our location.
6552 if( _verify_me ) {
6553 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6554 Node *legal = LCA;
6555 while( early != legal ) { // While not at earliest legal
6556 if( legal == v_ctrl ) break; // Check for prior good location
6557 legal = idom(legal) ;// Bump up the IDOM tree
6558 }
6559 // Check for prior good location
6560 if( legal == v_ctrl ) least = legal; // Keep prior if found
6561 }
6562 */
6563 #endif
6564
6565 // Assign discovered "here or above" point
6566 least = find_non_split_ctrl(least);
6567 verify_strip_mined_scheduling(n, least);
6568 set_ctrl(n, least);
6569
6570 // Collect inner loop bodies
6571 IdealLoopTree *chosen_loop = get_loop(least);
6572 if( !chosen_loop->_child ) // Inner loop?
6573 chosen_loop->_body.push(n);// Collect inner loops
6574
6575 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6576 _zero_trip_guard_opaque_nodes.push(n);
6577 }
6578
6579 }
6580
6581 #ifdef ASSERT
6582 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6583 tty->print_cr("%s", msg);
6584 tty->print("n: "); n->dump();
6585 tty->print("early(n): "); early->dump();
6586 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6587 n->in(0) != early && !n->in(0)->is_Root()) {
6588 tty->print("n->in(0): "); n->in(0)->dump();
6589 }
6590 for (uint i = 1; i < n->req(); i++) {
6591 Node* in1 = n->in(i);
6592 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6593 tty->print("n->in(%d): ", i); in1->dump();
6594 Node* in1_early = get_ctrl(in1);
6595 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6596 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6597 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6598 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6599 }
6600 for (uint j = 1; j < in1->req(); j++) {
6601 Node* in2 = in1->in(j);
6602 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6603 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6604 Node* in2_early = get_ctrl(in2);
6605 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6606 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6607 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
6608 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
6609 }
6610 }
6611 }
6612 }
6613 }
6614 tty->cr();
6615 tty->print("LCA(n): "); LCA->dump();
6616 for (uint i = 0; i < n->outcnt(); i++) {
6617 Node* u1 = n->raw_out(i);
6618 if (u1 == n)
6619 continue;
6620 tty->print("n->out(%d): ", i); u1->dump();
6621 if (u1->is_CFG()) {
6622 for (uint j = 0; j < u1->outcnt(); j++) {
6623 Node* u2 = u1->raw_out(j);
6624 if (u2 != u1 && u2 != n && u2->is_CFG()) {
6625 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6626 }
6627 }
6628 } else {
6629 Node* u1_later = get_ctrl(u1);
6630 tty->print("later(n->out(%d)): ", i); u1_later->dump();
6631 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
6632 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
6633 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
6634 }
6635 for (uint j = 0; j < u1->outcnt(); j++) {
6636 Node* u2 = u1->raw_out(j);
6637 if (u2 == n || u2 == u1)
6638 continue;
6639 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6640 if (!u2->is_CFG()) {
6641 Node* u2_later = get_ctrl(u2);
6642 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
6643 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
6644 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
6645 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
6646 }
6647 }
6648 }
6649 }
6650 }
6651 dump_idoms(early, LCA);
6652 tty->cr();
6653 }
6654
6655 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
6656 class RealLCA {
6657 const PhaseIdealLoop* _phase;
6658 Node* _early;
6659 Node* _wrong_lca;
6660 uint _early_index;
6661 int _wrong_lca_index;
6662
6663 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
6664 // is different: Return the previously visited node which must be the real LCA.
6665 // The node lists also contain _early and _wrong_lca, respectively.
6666 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
6667 int early_index = early_with_idoms.size() - 1;
6668 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
6669 bool found_difference = false;
6670 do {
6671 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
6672 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
6673 found_difference = true;
6674 break;
6675 }
6676 early_index--;
6677 wrong_lca_index--;
6678 } while (wrong_lca_index >= 0);
6679
6680 assert(early_index >= 0, "must always find an LCA - cannot be early");
6681 _early_index = early_index;
6682 _wrong_lca_index = wrong_lca_index;
6683 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
6684 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
6685 return real_lca;
6686 }
6687
6688 void dump(Node* real_lca) {
6689 tty->cr();
6690 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
6691 _phase->dump_idom(_early, _early_index + 1);
6692
6693 tty->cr();
6694 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
6695 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
6696
6697 tty->cr();
6698 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
6699 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
6700 if (_wrong_lca_index >= 0) {
6701 tty->print(" (idom[%d])", _wrong_lca_index);
6702 }
6703 tty->print_cr(":");
6704 real_lca->dump();
6705 }
6706
6707 public:
6708 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
6709 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
6710 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
6711 }
6712
6713 void compute_and_dump() {
6714 ResourceMark rm;
6715 Unique_Node_List early_with_idoms;
6716 Unique_Node_List wrong_lca_with_idoms;
6717 early_with_idoms.push(_early);
6718 wrong_lca_with_idoms.push(_wrong_lca);
6719 _phase->get_idoms(_early, 10000, early_with_idoms);
6720 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
6721 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
6722 dump(real_lca);
6723 }
6724 };
6725
6726 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
6727 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
6728 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
6729 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
6730
6731 RealLCA real_lca(this, early, wrong_lca);
6732 real_lca.compute_and_dump();
6733 }
6734 #endif // ASSERT
6735
6736 #ifndef PRODUCT
6737 //------------------------------dump-------------------------------------------
6738 void PhaseIdealLoop::dump() const {
6739 ResourceMark rm;
6740 Node_Stack stack(C->live_nodes() >> 2);
6741 Node_List rpo_list;
6742 VectorSet visited;
6743 visited.set(C->top()->_idx);
6744 rpo(C->root(), stack, visited, rpo_list);
6745 // Dump root loop indexed by last element in PO order
6746 dump(_ltree_root, rpo_list.size(), rpo_list);
6747 }
6748
6749 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
6750 loop->dump_head();
6751
6752 // Now scan for CFG nodes in the same loop
6753 for (uint j = idx; j > 0; j--) {
6754 Node* n = rpo_list[j-1];
6755 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
6756 continue;
6757
6758 if (get_loop(n) != loop) { // Wrong loop nest
6759 if (get_loop(n)->_head == n && // Found nested loop?
6760 get_loop(n)->_parent == loop)
6761 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
6762 continue;
6763 }
6764
6765 // Dump controlling node
6766 tty->sp(2 * loop->_nest);
6767 tty->print("C");
6768 if (n == C->root()) {
6769 n->dump();
6770 } else {
6771 Node* cached_idom = idom_no_update(n);
6772 Node* computed_idom = n->in(0);
6773 if (n->is_Region()) {
6774 computed_idom = compute_idom(n);
6775 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
6776 // any MultiBranch ctrl node), so apply a similar transform to
6777 // the cached idom returned from idom_no_update.
6778 cached_idom = find_non_split_ctrl(cached_idom);
6779 }
6780 tty->print(" ID:%d", computed_idom->_idx);
6781 n->dump();
6782 if (cached_idom != computed_idom) {
6783 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
6784 computed_idom->_idx, cached_idom->_idx);
6785 }
6786 }
6787 // Dump nodes it controls
6788 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
6789 // (k < C->unique() && get_ctrl(find(k)) == n)
6790 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
6791 Node* m = C->root()->find(k);
6792 if (m && m->outcnt() > 0) {
6793 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
6794 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
6795 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
6796 }
6797 tty->sp(2 * loop->_nest + 1);
6798 m->dump();
6799 }
6800 }
6801 }
6802 }
6803 }
6804
6805 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
6806 if (has_ctrl(n)) {
6807 tty->print_cr("No idom for data nodes");
6808 } else {
6809 ResourceMark rm;
6810 Unique_Node_List idoms;
6811 get_idoms(n, count, idoms);
6812 dump_idoms_in_reverse(n, idoms);
6813 }
6814 }
6815
6816 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
6817 Node* next = n;
6818 for (uint i = 0; !next->is_Start() && i < count; i++) {
6819 next = idom(next);
6820 assert(!idoms.member(next), "duplicated idom is not possible");
6821 idoms.push(next);
6822 }
6823 }
6824
6825 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
6826 Node* next;
6827 uint padding = 3;
6828 uint node_index_padding_width = static_cast<int>(log10(static_cast<double>(C->unique()))) + 1;
6829 for (int i = idom_list.size() - 1; i >= 0; i--) {
6830 if (i == 9 || i == 99) {
6831 padding++;
6832 }
6833 next = idom_list[i];
6834 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
6835 }
6836 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
6837 }
6838 #endif // NOT PRODUCT
6839
6840 // Collect a R-P-O for the whole CFG.
6841 // Result list is in post-order (scan backwards for RPO)
6842 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
6843 stk.push(start, 0);
6844 visited.set(start->_idx);
6845
6846 while (stk.is_nonempty()) {
6847 Node* m = stk.node();
6848 uint idx = stk.index();
6849 if (idx < m->outcnt()) {
6850 stk.set_index(idx + 1);
6851 Node* n = m->raw_out(idx);
6852 if (n->is_CFG() && !visited.test_set(n->_idx)) {
6853 stk.push(n, 0);
6854 }
6855 } else {
6856 rpo_list.push(m);
6857 stk.pop();
6858 }
6859 }
6860 }
6861
6862 ConINode* PhaseIdealLoop::intcon(jint i) {
6863 ConINode* node = _igvn.intcon(i);
6864 set_root_as_ctrl(node);
6865 return node;
6866 }
6867
6868 ConLNode* PhaseIdealLoop::longcon(jlong i) {
6869 ConLNode* node = _igvn.longcon(i);
6870 set_root_as_ctrl(node);
6871 return node;
6872 }
6873
6874 ConNode* PhaseIdealLoop::makecon(const Type* t) {
6875 ConNode* node = _igvn.makecon(t);
6876 set_root_as_ctrl(node);
6877 return node;
6878 }
6879
6880 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
6881 ConNode* node = _igvn.integercon(l, bt);
6882 set_root_as_ctrl(node);
6883 return node;
6884 }
6885
6886 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
6887 ConNode* node = _igvn.zerocon(bt);
6888 set_root_as_ctrl(node);
6889 return node;
6890 }
6891
6892
6893 //=============================================================================
6894 //------------------------------LoopTreeIterator-------------------------------
6895
6896 // Advance to next loop tree using a preorder, left-to-right traversal.
6897 void LoopTreeIterator::next() {
6898 assert(!done(), "must not be done.");
6899 if (_curnt->_child != nullptr) {
6900 _curnt = _curnt->_child;
6901 } else if (_curnt->_next != nullptr) {
6902 _curnt = _curnt->_next;
6903 } else {
6904 while (_curnt != _root && _curnt->_next == nullptr) {
6905 _curnt = _curnt->_parent;
6906 }
6907 if (_curnt == _root) {
6908 _curnt = nullptr;
6909 assert(done(), "must be done.");
6910 } else {
6911 assert(_curnt->_next != nullptr, "must be more to do");
6912 _curnt = _curnt->_next;
6913 }
6914 }
6915 }