1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciMethodData.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/connode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/mulnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/predicates.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/vectorization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/checkedCast.hpp"
50 #include "utilities/powerOfTwo.hpp"
51
52 //=============================================================================
53 //--------------------------is_cloop_ind_var-----------------------------------
54 // Determine if a node is a counted loop induction variable.
55 // NOTE: The method is declared in "node.hpp".
56 bool Node::is_cloop_ind_var() const {
57 return (is_Phi() &&
58 as_Phi()->region()->is_CountedLoop() &&
59 as_Phi()->region()->as_CountedLoop()->phi() == this);
60 }
61
62 //=============================================================================
63 //------------------------------dump_spec--------------------------------------
64 // Dump special per-node info
65 #ifndef PRODUCT
66 void LoopNode::dump_spec(outputStream *st) const {
67 RegionNode::dump_spec(st);
68 if (is_inner_loop()) st->print( "inner " );
69 if (is_partial_peel_loop()) st->print( "partial_peel " );
70 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
71 }
72 #endif
73
74 //------------------------------is_valid_counted_loop-------------------------
75 bool LoopNode::is_valid_counted_loop(BasicType bt) const {
76 if (is_BaseCountedLoop() && as_BaseCountedLoop()->bt() == bt) {
77 BaseCountedLoopNode* l = as_BaseCountedLoop();
78 BaseCountedLoopEndNode* le = l->loopexit_or_null();
79 if (le != nullptr &&
80 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
81 Node* phi = l->phi();
82 Node* exit = le->proj_out_or_null(0 /* false */);
83 if (exit != nullptr && exit->Opcode() == Op_IfFalse &&
84 phi != nullptr && phi->is_Phi() &&
85 phi->in(LoopNode::LoopBackControl) == l->incr() &&
86 le->loopnode() == l && le->stride_is_con()) {
87 return true;
88 }
89 }
90 }
91 return false;
92 }
93
94 //------------------------------get_early_ctrl---------------------------------
95 // Compute earliest legal control
96 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
97 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
98 uint i;
99 Node *early;
100 if (n->in(0) && !n->is_expensive()) {
101 early = n->in(0);
102 if (!early->is_CFG()) // Might be a non-CFG multi-def
103 early = get_ctrl(early); // So treat input as a straight data input
104 i = 1;
105 } else {
106 early = get_ctrl(n->in(1));
107 i = 2;
108 }
109 uint e_d = dom_depth(early);
110 assert( early, "" );
111 for (; i < n->req(); i++) {
112 Node *cin = get_ctrl(n->in(i));
113 assert( cin, "" );
114 // Keep deepest dominator depth
115 uint c_d = dom_depth(cin);
116 if (c_d > e_d) { // Deeper guy?
117 early = cin; // Keep deepest found so far
118 e_d = c_d;
119 } else if (c_d == e_d && // Same depth?
120 early != cin) { // If not equal, must use slower algorithm
121 // If same depth but not equal, one _must_ dominate the other
122 // and we want the deeper (i.e., dominated) guy.
123 Node *n1 = early;
124 Node *n2 = cin;
125 while (1) {
126 n1 = idom(n1); // Walk up until break cycle
127 n2 = idom(n2);
128 if (n1 == cin || // Walked early up to cin
129 dom_depth(n2) < c_d)
130 break; // early is deeper; keep him
131 if (n2 == early || // Walked cin up to early
132 dom_depth(n1) < c_d) {
133 early = cin; // cin is deeper; keep him
134 break;
135 }
136 }
137 e_d = dom_depth(early); // Reset depth register cache
138 }
139 }
140
141 // Return earliest legal location
142 assert(early == find_non_split_ctrl(early), "unexpected early control");
143
144 if (n->is_expensive() && !_verify_only && !_verify_me) {
145 assert(n->in(0), "should have control input");
146 early = get_early_ctrl_for_expensive(n, early);
147 }
148
149 return early;
150 }
151
152 //------------------------------get_early_ctrl_for_expensive---------------------------------
153 // Move node up the dominator tree as high as legal while still beneficial
154 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
155 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
156 assert(OptimizeExpensiveOps, "optimization off?");
157
158 Node* ctl = n->in(0);
159 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
160 uint min_dom_depth = dom_depth(earliest);
161 #ifdef ASSERT
162 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
163 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
164 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
165 }
166 #endif
167 if (dom_depth(ctl) < min_dom_depth) {
168 return earliest;
169 }
170
171 while (true) {
172 Node* next = ctl;
173 // Moving the node out of a loop on the projection of an If
174 // confuses Loop Predication. So, once we hit a loop in an If branch
175 // that doesn't branch to an UNC, we stop. The code that process
176 // expensive nodes will notice the loop and skip over it to try to
177 // move the node further up.
178 if (ctl->is_CountedLoop() && ctl->in(1) != nullptr && ctl->in(1)->in(0) != nullptr && ctl->in(1)->in(0)->is_If()) {
179 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern()) {
180 break;
181 }
182 next = idom(ctl->in(1)->in(0));
183 } else if (ctl->is_Proj()) {
184 // We only move it up along a projection if the projection is
185 // the single control projection for its parent: same code path,
186 // if it's a If with UNC or fallthrough of a call.
187 Node* parent_ctl = ctl->in(0);
188 if (parent_ctl == nullptr) {
189 break;
190 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != nullptr) {
191 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
192 } else if (parent_ctl->is_If()) {
193 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern()) {
194 break;
195 }
196 assert(idom(ctl) == parent_ctl, "strange");
197 next = idom(parent_ctl);
198 } else if (ctl->is_CatchProj()) {
199 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
200 break;
201 }
202 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
203 next = parent_ctl->in(0)->in(0)->in(0);
204 } else {
205 // Check if parent control has a single projection (this
206 // control is the only possible successor of the parent
207 // control). If so, we can try to move the node above the
208 // parent control.
209 int nb_ctl_proj = 0;
210 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
211 Node *p = parent_ctl->fast_out(i);
212 if (p->is_Proj() && p->is_CFG()) {
213 nb_ctl_proj++;
214 if (nb_ctl_proj > 1) {
215 break;
216 }
217 }
218 }
219
220 if (nb_ctl_proj > 1) {
221 break;
222 }
223 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
224 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
225 assert(idom(ctl) == parent_ctl, "strange");
226 next = idom(parent_ctl);
227 }
228 } else {
229 next = idom(ctl);
230 }
231 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
232 break;
233 }
234 ctl = next;
235 }
236
237 if (ctl != n->in(0)) {
238 _igvn.replace_input_of(n, 0, ctl);
239 _igvn.hash_insert(n);
240 }
241
242 return ctl;
243 }
244
245
246 //------------------------------set_early_ctrl---------------------------------
247 // Set earliest legal control
248 void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {
249 Node *early = get_early_ctrl(n);
250
251 // Record earliest legal location
252 set_ctrl(n, early);
253 IdealLoopTree *loop = get_loop(early);
254 if (update_body && loop->_child == nullptr) {
255 loop->_body.push(n);
256 }
257 }
258
259 //------------------------------set_subtree_ctrl-------------------------------
260 // set missing _ctrl entries on new nodes
261 void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {
262 // Already set? Get out.
263 if (_loop_or_ctrl[n->_idx]) return;
264 // Recursively set _loop_or_ctrl array to indicate where the Node goes
265 uint i;
266 for (i = 0; i < n->req(); ++i) {
267 Node *m = n->in(i);
268 if (m && m != C->root()) {
269 set_subtree_ctrl(m, update_body);
270 }
271 }
272
273 // Fixup self
274 set_early_ctrl(n, update_body);
275 }
276
277 IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {
278 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
279 IdealLoopTree* parent = loop->_parent;
280 IdealLoopTree* sibling = parent->_child;
281 if (sibling == loop) {
282 parent->_child = outer_ilt;
283 } else {
284 while (sibling->_next != loop) {
285 sibling = sibling->_next;
286 }
287 sibling->_next = outer_ilt;
288 }
289 outer_ilt->_next = loop->_next;
290 outer_ilt->_parent = parent;
291 outer_ilt->_child = loop;
292 outer_ilt->_nest = loop->_nest;
293 loop->_parent = outer_ilt;
294 loop->_next = nullptr;
295 loop->_nest++;
296 assert(loop->_nest <= SHRT_MAX, "sanity");
297 return outer_ilt;
298 }
299
300 // Create a skeleton strip mined outer loop: a Loop head before the
301 // inner strip mined loop, a safepoint and an exit condition guarded
302 // by an opaque node after the inner strip mined loop with a backedge
303 // to the loop head. The inner strip mined loop is left as it is. Only
304 // once loop optimizations are over, do we adjust the inner loop exit
305 // condition to limit its number of iterations, set the outer loop
306 // exit condition and add Phis to the outer loop head. Some loop
307 // optimizations that operate on the inner strip mined loop need to be
308 // aware of the outer strip mined loop: loop unswitching needs to
309 // clone the outer loop as well as the inner, unrolling needs to only
310 // clone the inner loop etc. No optimizations need to change the outer
311 // strip mined loop as it is only a skeleton.
312 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
313 IdealLoopTree* loop, float cl_prob, float le_fcnt,
314 Node*& entry_control, Node*& iffalse) {
315 Node* outer_test = intcon(0);
316 Node *orig = iffalse;
317 iffalse = iffalse->clone();
318 _igvn.register_new_node_with_optimizer(iffalse);
319 set_idom(iffalse, idom(orig), dom_depth(orig));
320
321 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
322 Node *outer_ift = new IfTrueNode (outer_le);
323 Node* outer_iff = orig;
324 _igvn.replace_input_of(outer_iff, 0, outer_le);
325
326 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
327 entry_control = outer_l;
328
329 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
330
331 set_loop(iffalse, outer_ilt);
332 // When this code runs, loop bodies have not yet been populated.
333 const bool body_populated = false;
334 register_control(outer_le, outer_ilt, iffalse, body_populated);
335 register_control(outer_ift, outer_ilt, outer_le, body_populated);
336 set_idom(outer_iff, outer_le, dom_depth(outer_le));
337 _igvn.register_new_node_with_optimizer(outer_l);
338 set_loop(outer_l, outer_ilt);
339 set_idom(outer_l, init_control, dom_depth(init_control)+1);
340
341 return outer_ilt;
342 }
343
344 void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
345 Node* cmp_limit, Node* bol) {
346 assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
347 Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
348 Deoptimization::Reason_loop_limit_check,
349 Op_If);
350 Node* iff = new_predicate_proj->in(0);
351 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
352 bol = _igvn.register_new_node_with_optimizer(bol);
353 set_subtree_ctrl(bol, false);
354 _igvn.replace_input_of(iff, 1, bol);
355
356 #ifndef PRODUCT
357 // report that the loop predication has been actually performed
358 // for this loop
359 if (TraceLoopLimitCheck) {
360 tty->print_cr("Counted Loop Limit Check generated:");
361 debug_only( bol->dump(2); )
362 }
363 #endif
364 }
365
366 Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {
367 // Counted loop head must be a good RegionNode with only 3 not null
368 // control input edges: Self, Entry, LoopBack.
369 if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {
370 return nullptr;
371 }
372 Node *init_control = x->in(LoopNode::EntryControl);
373 Node *back_control = x->in(LoopNode::LoopBackControl);
374 if (init_control == nullptr || back_control == nullptr) { // Partially dead
375 return nullptr;
376 }
377 // Must also check for TOP when looking for a dead loop
378 if (init_control->is_top() || back_control->is_top()) {
379 return nullptr;
380 }
381
382 // Allow funny placement of Safepoint
383 if (back_control->Opcode() == Op_SafePoint) {
384 back_control = back_control->in(TypeFunc::Control);
385 }
386
387 // Controlling test for loop
388 Node *iftrue = back_control;
389 uint iftrue_op = iftrue->Opcode();
390 if (iftrue_op != Op_IfTrue &&
391 iftrue_op != Op_IfFalse) {
392 // I have a weird back-control. Probably the loop-exit test is in
393 // the middle of the loop and I am looking at some trailing control-flow
394 // merge point. To fix this I would have to partially peel the loop.
395 return nullptr; // Obscure back-control
396 }
397
398 // Get boolean guarding loop-back test
399 Node *iff = iftrue->in(0);
400 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {
401 return nullptr;
402 }
403 return iftrue;
404 }
405
406 Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {
407 Node* iftrue = back_control;
408 uint iftrue_op = iftrue->Opcode();
409 Node* iff = iftrue->in(0);
410 BoolNode* test = iff->in(1)->as_Bool();
411 bt = test->_test._test;
412 cl_prob = iff->as_If()->_prob;
413 if (iftrue_op == Op_IfFalse) {
414 bt = BoolTest(bt).negate();
415 cl_prob = 1.0 - cl_prob;
416 }
417 // Get backedge compare
418 Node* cmp = test->in(1);
419 if (!cmp->is_Cmp()) {
420 return nullptr;
421 }
422
423 // Find the trip-counter increment & limit. Limit must be loop invariant.
424 incr = cmp->in(1);
425 limit = cmp->in(2);
426
427 // ---------
428 // need 'loop()' test to tell if limit is loop invariant
429 // ---------
430
431 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
432 Node* tmp = incr; // Then reverse order into the CmpI
433 incr = limit;
434 limit = tmp;
435 bt = BoolTest(bt).commute(); // And commute the exit test
436 }
437 if (is_member(loop, get_ctrl(limit))) { // Limit must be loop-invariant
438 return nullptr;
439 }
440 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
441 return nullptr;
442 }
443 return cmp;
444 }
445
446 Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {
447 if (incr->is_Phi()) {
448 if (incr->as_Phi()->region() != x || incr->req() != 3) {
449 return nullptr; // Not simple trip counter expression
450 }
451 phi_incr = incr;
452 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
453 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
454 return nullptr;
455 }
456 }
457 return incr;
458 }
459
460 Node* PhaseIdealLoop::loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi) {
461 assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
462 // Get merge point
463 xphi = incr->in(1);
464 Node *stride = incr->in(2);
465 if (!stride->is_Con()) { // Oops, swap these
466 if (!xphi->is_Con()) { // Is the other guy a constant?
467 return nullptr; // Nope, unknown stride, bail out
468 }
469 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
470 xphi = stride;
471 stride = tmp;
472 }
473 return stride;
474 }
475
476 PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop) {
477 if (!xphi->is_Phi()) {
478 return nullptr; // Too much math on the trip counter
479 }
480 if (phi_incr != nullptr && phi_incr != xphi) {
481 return nullptr;
482 }
483 PhiNode *phi = xphi->as_Phi();
484
485 // Phi must be of loop header; backedge must wrap to increment
486 if (phi->region() != x) {
487 return nullptr;
488 }
489 return phi;
490 }
491
492 static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {
493 if (final_correction > 0) {
494 if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {
495 return -1;
496 }
497 if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {
498 return 1;
499 }
500 } else {
501 if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {
502 return -1;
503 }
504 if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {
505 return 1;
506 }
507 }
508 return 0;
509 }
510
511 static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {
512 // If the condition is inverted and we will be rolling
513 // through MININT to MAXINT, then bail out.
514 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
515 // Odd stride
516 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
517 // Count down loop rolls through MAXINT
518 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
519 // Count up loop rolls through MININT
520 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
521 return false; // Bail out
522 }
523 return true;
524 }
525
526 Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
527 BasicType bt) {
528 Node* iv_as_long;
529 if (bt == T_LONG) {
530 iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
531 register_new_node(iv_as_long, inner_head);
532 } else {
533 iv_as_long = inner_iv;
534 }
535 Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
536 register_new_node(iv_replacement, inner_head);
537 for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
538 Node* u = iv_to_replace->last_out(i);
539 #ifdef ASSERT
540 if (!is_dominator(inner_head, ctrl_or_self(u))) {
541 assert(u->is_Phi(), "should be a Phi");
542 for (uint j = 1; j < u->req(); j++) {
543 if (u->in(j) == iv_to_replace) {
544 assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
545 }
546 }
547 }
548 #endif
549 _igvn.rehash_node_delayed(u);
550 int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
551 i -= nb;
552 }
553 return iv_replacement;
554 }
555
556 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
557 void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
558 SafePointNode* sfpt) {
559 if (!C->too_many_traps(reason)) {
560 ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
561 register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
562 Node* if_false = new IfFalseNode(parse_predicate);
563 register_control(if_false, _ltree_root, parse_predicate);
564 Node* if_true = new IfTrueNode(parse_predicate);
565 register_control(if_true, loop, parse_predicate);
566
567 int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
568 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
569 const TypePtr* no_memory_effects = nullptr;
570 JVMState* jvms = sfpt->jvms();
571 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
572 no_memory_effects);
573
574 Node* mem = nullptr;
575 Node* i_o = nullptr;
576 if (sfpt->is_Call()) {
577 mem = sfpt->proj_out(TypeFunc::Memory);
578 i_o = sfpt->proj_out(TypeFunc::I_O);
579 } else {
580 mem = sfpt->memory();
581 i_o = sfpt->i_o();
582 }
583
584 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
585 register_new_node(frame, C->start());
586 Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
587 register_new_node(ret, C->start());
588
589 unc->init_req(TypeFunc::Control, if_false);
590 unc->init_req(TypeFunc::I_O, i_o);
591 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
592 unc->init_req(TypeFunc::FramePtr, frame);
593 unc->init_req(TypeFunc::ReturnAdr, ret);
594 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
595 unc->set_cnt(PROB_UNLIKELY_MAG(4));
596 unc->copy_call_debug_info(&_igvn, sfpt);
597
598 for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
599 set_subtree_ctrl(unc->in(i), false);
600 }
601 register_control(unc, _ltree_root, if_false);
602
603 Node* ctrl = new ProjNode(unc, TypeFunc::Control);
604 register_control(ctrl, _ltree_root, unc);
605 Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
606 register_control(halt, _ltree_root, ctrl);
607 _igvn.add_input_to(C->root(), halt);
608
609 _igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
610 set_idom(inner_head, if_true, dom_depth(inner_head));
611 }
612 }
613
614 // Find a safepoint node that dominates the back edge. We need a
615 // SafePointNode so we can use its jvm state to create empty
616 // predicates.
617 static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
618 SafePointNode* safepoint = nullptr;
619 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
620 Node* u = x->fast_out(i);
621 if (u->is_memory_phi()) {
622 Node* m = u->in(LoopNode::LoopBackControl);
623 if (u->adr_type() == TypePtr::BOTTOM) {
624 if (m->is_MergeMem() && mem->is_MergeMem()) {
625 if (m != mem DEBUG_ONLY(|| true)) {
626 // MergeMemStream can modify m, for example to adjust the length to mem.
627 // This is unfortunate, and probably unnecessary. But as it is, we need
628 // to add m to the igvn worklist, else we may have a modified node that
629 // is not on the igvn worklist.
630 phase->igvn()._worklist.push(m);
631 for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
632 if (!mms.is_empty()) {
633 if (mms.memory() != mms.memory2()) {
634 return false;
635 }
636 #ifdef ASSERT
637 if (mms.alias_idx() != Compile::AliasIdxBot) {
638 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
639 }
640 #endif
641 }
642 }
643 }
644 } else if (mem->is_MergeMem()) {
645 if (m != mem->as_MergeMem()->base_memory()) {
646 return false;
647 }
648 } else {
649 return false;
650 }
651 } else {
652 if (mem->is_MergeMem()) {
653 if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {
654 return false;
655 }
656 #ifdef ASSERT
657 mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
658 #endif
659 } else {
660 if (m != mem) {
661 return false;
662 }
663 }
664 }
665 }
666 }
667 return true;
668 }
669
670 SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
671 IfNode* exit_test = back_control->in(0)->as_If();
672 SafePointNode* safepoint = nullptr;
673 if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
674 safepoint = exit_test->in(0)->as_SafePoint();
675 } else {
676 Node* c = back_control;
677 while (c != x && c->Opcode() != Op_SafePoint) {
678 c = idom(c);
679 }
680
681 if (c->Opcode() == Op_SafePoint) {
682 safepoint = c->as_SafePoint();
683 }
684
685 if (safepoint == nullptr) {
686 return nullptr;
687 }
688
689 Node* mem = safepoint->in(TypeFunc::Memory);
690
691 // We can only use that safepoint if there's no side effect between the backedge and the safepoint.
692
693 // mm is the memory state at the safepoint (when it's a MergeMem)
694 // no_side_effect_since_safepoint() goes over the memory state at the backedge. It resets the mm input for each
695 // component of the memory state it encounters so it points to the base memory. Once no_side_effect_since_safepoint()
696 // is done, if no side effect after the safepoint was found, mm should transform to the base memory: the states at
697 // the backedge and safepoint are the same so all components of the memory state at the safepoint should have been
698 // reset.
699 MergeMemNode* mm = nullptr;
700 #ifdef ASSERT
701 if (mem->is_MergeMem()) {
702 mm = mem->clone()->as_MergeMem();
703 _igvn._worklist.push(mm);
704 for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {
705 // Loop invariant memory state won't be reset by no_side_effect_since_safepoint(). Do it here.
706 // Escape Analysis can add state to mm that it doesn't add to the backedge memory Phis, breaking verification
707 // code that relies on mm. Clear that extra state here.
708 if (mms.alias_idx() != Compile::AliasIdxBot &&
709 (loop != get_loop(ctrl_or_self(mms.memory())) ||
710 (mms.adr_type()->isa_oop_ptr() && mms.adr_type()->is_known_instance()))) {
711 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
712 }
713 }
714 }
715 #endif
716 if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
717 safepoint = nullptr;
718 } else {
719 assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
720 }
721 #ifdef ASSERT
722 if (mm != nullptr) {
723 _igvn.remove_dead_node(mm);
724 }
725 #endif
726 }
727 return safepoint;
728 }
729
730 // If the loop has the shape of a counted loop but with a long
731 // induction variable, transform the loop in a loop nest: an inner
732 // loop that iterates for at most max int iterations with an integer
733 // induction variable and an outer loop that iterates over the full
734 // range of long values from the initial loop in (at most) max int
735 // steps. That is:
736 //
737 // x: for (long phi = init; phi < limit; phi += stride) {
738 // // phi := Phi(L, init, incr)
739 // // incr := AddL(phi, longcon(stride))
740 // long incr = phi + stride;
741 // ... use phi and incr ...
742 // }
743 //
744 // OR:
745 //
746 // x: for (long phi = init; (phi += stride) < limit; ) {
747 // // phi := Phi(L, AddL(init, stride), incr)
748 // // incr := AddL(phi, longcon(stride))
749 // long incr = phi + stride;
750 // ... use phi and (phi + stride) ...
751 // }
752 //
753 // ==transform=>
754 //
755 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
756 // assert(stride <= inner_iters_limit); // else abort transform
757 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
758 // outer_head: for (long outer_phi = init;;) {
759 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
760 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
761 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
762 // assert(inner_iters_actual == (int)inner_iters_actual);
763 // int inner_phi, inner_incr;
764 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
765 // // inner_phi := Phi(x, intcon(0), inner_incr)
766 // // inner_incr := AddI(inner_phi, intcon(stride))
767 // inner_incr = inner_phi + stride;
768 // if (inner_incr < inner_iters_actual) {
769 // ... use phi=>(outer_phi+inner_phi) ...
770 // continue;
771 // }
772 // else break;
773 // }
774 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
775 // continue;
776 // else break;
777 // }
778 //
779 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
780 // loops with long range checks transformed to int range checks in the inner loop.
781 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
782 Node* x = loop->_head;
783 // Only for inner loops
784 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
785 return false;
786 }
787
788 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
789 return false;
790 }
791
792 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
793 BasicType bt = x->as_BaseCountedLoop()->bt();
794
795 check_counted_loop_shape(loop, x, bt);
796
797 #ifndef PRODUCT
798 if (bt == T_LONG) {
799 Atomic::inc(&_long_loop_candidates);
800 }
801 #endif
802
803 jlong stride_con_long = head->stride_con();
804 assert(stride_con_long != 0, "missed some peephole opt");
805 // We can't iterate for more than max int at a time.
806 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
807 assert(bt == T_LONG, "only for long loops");
808 return false;
809 }
810 jint stride_con = checked_cast<jint>(stride_con_long);
811 // The number of iterations for the integer count loop: guarantee no
812 // overflow: max_jint - stride_con max. -1 so there's no need for a
813 // loop limit check if the exit test is <= or >=.
814 int iters_limit = max_jint - ABS(stride_con) - 1;
815 #ifdef ASSERT
816 if (bt == T_LONG && StressLongCountedLoop > 0) {
817 iters_limit = iters_limit / StressLongCountedLoop;
818 }
819 #endif
820 // At least 2 iterations so counted loop construction doesn't fail
821 if (iters_limit/ABS(stride_con) < 2) {
822 return false;
823 }
824
825 PhiNode* phi = head->phi()->as_Phi();
826 Node* incr = head->incr();
827
828 Node* back_control = head->in(LoopNode::LoopBackControl);
829
830 // data nodes on back branch not supported
831 if (back_control->outcnt() > 1) {
832 return false;
833 }
834
835 Node* limit = head->limit();
836 // We'll need to use the loop limit before the inner loop is entered
837 if (!is_dominator(get_ctrl(limit), x)) {
838 return false;
839 }
840
841 IfNode* exit_test = head->loopexit();
842
843 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
844
845 Node_List range_checks;
846 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
847
848 if (bt == T_INT) {
849 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
850 if (range_checks.size() == 0) {
851 return false;
852 }
853 }
854
855 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
856 // the inner loop.
857 const Node* init = head->init_trip();
858 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
859 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
860 if (stride_con < 0) {
861 swap(lo, hi);
862 }
863 if (hi->hi_as_long() <= lo->lo_as_long()) {
864 // not a loop after all
865 return false;
866 }
867
868 if (range_checks.size() > 0) {
869 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
870 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
871 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
872 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
873 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
874 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
875 // logic on loop entry and for the outer loop.
876 loop->compute_trip_count(this);
877 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
878 if (head->as_CountedLoop()->trip_count() <= 3) {
879 return false;
880 }
881 } else {
882 loop->compute_profile_trip_cnt(this);
883 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
884 return false;
885 }
886 }
887 }
888
889 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
890 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
891
892 // We need a safepoint to insert Parse Predicates for the inner loop.
893 SafePointNode* safepoint;
894 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
895 // Loop is strip mined: use the safepoint of the outer strip mined loop
896 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
897 assert(outer_loop != nullptr, "no outer loop");
898 safepoint = outer_loop->outer_safepoint();
899 outer_loop->transform_to_counted_loop(&_igvn, this);
900 exit_test = head->loopexit();
901 } else {
902 safepoint = find_safepoint(back_control, x, loop);
903 }
904
905 Node* exit_branch = exit_test->proj_out(false);
906 Node* entry_control = head->in(LoopNode::EntryControl);
907
908 // Clone the control flow of the loop to build an outer loop
909 Node* outer_back_branch = back_control->clone();
910 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
911 Node* inner_exit_branch = exit_branch->clone();
912
913 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
914 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
915
916 const bool body_populated = true;
917 register_control(outer_head, outer_ilt, entry_control, body_populated);
918
919 _igvn.register_new_node_with_optimizer(inner_exit_branch);
920 set_loop(inner_exit_branch, outer_ilt);
921 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
922
923 outer_exit_test->set_req(0, inner_exit_branch);
924 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
925
926 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
927 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
928
929 outer_back_branch->set_req(0, outer_exit_test);
930 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
931
932 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
933 set_idom(x, outer_head, dom_depth(x));
934
935 // add an iv phi to the outer loop and use it to compute the inner
936 // loop iteration limit
937 Node* outer_phi = phi->clone();
938 outer_phi->set_req(0, outer_head);
939 register_new_node(outer_phi, outer_head);
940
941 Node* inner_iters_max = nullptr;
942 if (stride_con > 0) {
943 inner_iters_max = MaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
944 } else {
945 inner_iters_max = MaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
946 }
947
948 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
949 // inner_iters_max may not fit in a signed integer (iterating from
950 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
951 // min.
952 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
953 Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
954
955 Node* inner_iters_actual_int;
956 if (bt == T_LONG) {
957 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
958 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
959 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
960 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
961 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
962 // opts pass, an accurate range of values for the limits is found.
963 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
964 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::UnconditionalDependency);
965 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
966 } else {
967 inner_iters_actual_int = inner_iters_actual;
968 }
969
970 Node* int_zero = intcon(0);
971 if (stride_con < 0) {
972 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
973 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
974 }
975
976 // Clone the iv data nodes as an integer iv
977 Node* int_stride = intcon(stride_con);
978 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
979 Node* inner_incr = new AddINode(inner_phi, int_stride);
980 Node* inner_cmp = nullptr;
981 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
982 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
983 inner_phi->set_req(LoopNode::EntryControl, int_zero);
984 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
985 register_new_node(inner_phi, x);
986 register_new_node(inner_incr, x);
987 register_new_node(inner_cmp, x);
988 register_new_node(inner_bol, x);
989
990 _igvn.replace_input_of(exit_test, 1, inner_bol);
991
992 // Clone inner loop phis to outer loop
993 for (uint i = 0; i < head->outcnt(); i++) {
994 Node* u = head->raw_out(i);
995 if (u->is_Phi() && u != inner_phi && u != phi) {
996 assert(u->in(0) == head, "inconsistent");
997 Node* clone = u->clone();
998 clone->set_req(0, outer_head);
999 register_new_node(clone, outer_head);
1000 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1001 }
1002 }
1003
1004 // Replace inner loop long iv phi as inner loop int iv phi + outer
1005 // loop iv phi
1006 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1007
1008 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1009
1010 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1011
1012 // Summary of steps from initial loop to loop nest:
1013 //
1014 // == old IR nodes =>
1015 //
1016 // entry_control: {...}
1017 // x:
1018 // for (long phi = init;;) {
1019 // // phi := Phi(x, init, incr)
1020 // // incr := AddL(phi, longcon(stride))
1021 // exit_test:
1022 // if (phi < limit)
1023 // back_control: fallthrough;
1024 // else
1025 // exit_branch: break;
1026 // long incr = phi + stride;
1027 // ... use phi and incr ...
1028 // phi = incr;
1029 // }
1030 //
1031 // == new IR nodes (just before final peel) =>
1032 //
1033 // entry_control: {...}
1034 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1035 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1036 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1037 // outer_head:
1038 // for (long outer_phi = init;;) {
1039 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1040 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1041 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1042 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1043 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1044 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1045 // inner_head: x: //in(1) := outer_head
1046 // int inner_phi;
1047 // for (inner_phi = 0;;) {
1048 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1049 // int inner_incr = inner_phi + stride;
1050 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1051 // exit_test: //exit_test->in(1) := inner_bol;
1052 // if (inner_bol) // WAS (phi < limit)
1053 // back_control: fallthrough;
1054 // else
1055 // inner_exit_branch: break; //exit_branch->clone()
1056 // ... use phi=>(outer_phi+inner_phi) ...
1057 // inner_phi = inner_phi + stride; // inner_incr
1058 // }
1059 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1060 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1061 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1062 // else
1063 // exit_branch: break; //in(0) := outer_exit_test
1064 // }
1065
1066 if (bt == T_INT) {
1067 outer_phi = new ConvI2LNode(outer_phi);
1068 register_new_node(outer_phi, outer_head);
1069 }
1070
1071 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1072 inner_phi, iv_add, inner_head);
1073 // Peel one iteration of the loop and use the safepoint at the end
1074 // of the peeled iteration to insert Parse Predicates. If no well
1075 // positioned safepoint peel to guarantee a safepoint in the outer
1076 // loop.
1077 if (safepoint != nullptr || !loop->_has_call) {
1078 old_new.clear();
1079 do_peeling(loop, old_new);
1080 } else {
1081 C->set_major_progress();
1082 }
1083
1084 if (safepoint != nullptr) {
1085 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1086
1087 if (UseLoopPredicate) {
1088 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
1089 }
1090 if (UseProfiledLoopPredicate) {
1091 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1092 }
1093 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1094 }
1095
1096 #ifndef PRODUCT
1097 if (bt == T_LONG) {
1098 Atomic::inc(&_long_loop_nests);
1099 }
1100 #endif
1101
1102 inner_head->mark_loop_nest_inner_loop();
1103 outer_head->mark_loop_nest_outer_loop();
1104
1105 return true;
1106 }
1107
1108 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1109 Node_List& range_checks) {
1110 const jlong min_iters = 2;
1111 jlong reduced_iters_limit = iters_limit;
1112 jlong original_iters_limit = iters_limit;
1113 for (uint i = 0; i < loop->_body.size(); i++) {
1114 Node* c = loop->_body.at(i);
1115 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1116 IfProjNode* if_proj = c->as_IfProj();
1117 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1118 if (call != nullptr) {
1119 Node* range = nullptr;
1120 Node* offset = nullptr;
1121 jlong scale = 0;
1122 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1123 loop->is_invariant(range) && loop->is_invariant(offset) &&
1124 scale != min_jlong &&
1125 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1126 assert(scale == (jint)scale, "scale should be an int");
1127 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1128 range_checks.push(c);
1129 }
1130 }
1131 }
1132 }
1133
1134 return checked_cast<int>(reduced_iters_limit);
1135 }
1136
1137 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1138 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1139 // 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
1140 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1141
1142 // 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,
1143 // 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)
1144 // 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].
1145
1146 // 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
1147 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1148 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1149
1150 // 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
1151 // 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
1152 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1153 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1154
1155 // 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
1156 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1157
1158 // 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].
1159 // 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
1160 // 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
1161 // 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
1162 // down (from C and 0 again).
1163
1164 // 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).
1165 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1166 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1167 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1168 // PhaseIdealLoop::add_constraint.)
1169
1170 // 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
1171 // 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
1172 // 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.
1173 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1174 // values for the 32-bit counter j within which the range checks can be eliminated.
1175
1176 // 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
1177 // 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
1178 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1179
1180 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1181
1182 // 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
1183 // 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),
1184 // 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)
1185
1186 // 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).
1187 // 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.
1188 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1189
1190 // Case A: Some Negatives (but no overflow).
1191 // Number line:
1192 // |s64_min . . . 0 . . . s64_max|
1193 // | . Q_min..Q_max . 0 . . . . | s64 negative
1194 // | . . . . R=0 R< R< R< R< | (against R values)
1195 // | . . . Q_min..0..Q_max . . . | small mixed
1196 // | . . . . R R R< R< R< | (against R values)
1197 //
1198 // 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<.
1199 //
1200 // So, if Q_min <s64 0, then use this test:
1201 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1202 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1203 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1204 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1205
1206 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1207
1208 // Case B: No Negatives.
1209 // Number line:
1210 // |s64_min . . . 0 . . . s64_max|
1211 // | . . . . 0 Q_min..Q_max . . | small positive
1212 // | . . . . R> R R R< R< | (against R values)
1213 // | . . . . 0 . Q_min..Q_max . | s64 positive
1214 // | . . . . R> R> R R R< | (against R values)
1215 //
1216 // 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.
1217 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1218 //
1219 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1220 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1221 // More generally:
1222 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1223
1224 // Case C: Overflow in the 64-bit domain
1225 // Number line:
1226 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1227 // | . . . . R> R> R> R> R | (against R values)
1228 //
1229 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1230 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1231 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1232 //
1233 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1234 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1235 // More generally:
1236 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1237 //
1238 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1239 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1240 //
1241 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1242 // clamp(X, L, H) := max(L, min(X, H))
1243 // When degenerately L > H, it returns L not H.
1244 //
1245 // All of the formulas above can be merged into a single one:
1246 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1247 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1248 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1249 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1250 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1251 // replacement R.C.:
1252 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1253 // or equivalently:
1254 // j*K + L_2 <u32 R_2
1255 // where
1256 // L_2 = Q_first - L_clamp
1257 // R_2 = R_clamp - L_clamp
1258 //
1259 // Note on why R is never negative:
1260 //
1261 // Various details of this transformation would break badly if R could be negative, so this transformation only
1262 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1263 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1264 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1265 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1266 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1267 // R.C., but R values from a.length and inside checkIndex are good to go.
1268 //
1269 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1270 Node* inner_iters_actual_int, Node* inner_phi,
1271 Node* iv_add, LoopNode* inner_head) {
1272 Node* long_zero = longcon(0);
1273 Node* int_zero = intcon(0);
1274 Node* long_one = longcon(1);
1275 Node* int_stride = intcon(checked_cast<int>(stride_con));
1276
1277 for (uint i = 0; i < range_checks.size(); i++) {
1278 ProjNode* proj = range_checks.at(i)->as_Proj();
1279 ProjNode* unc_proj = proj->other_if_proj();
1280 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1281 jlong scale = 0;
1282 Node* offset = nullptr;
1283 Node* rc_bol = rc->in(1);
1284 Node* rc_cmp = rc_bol->in(1);
1285 if (rc_cmp->Opcode() == Op_CmpU) {
1286 // could be shared and have already been taken care of
1287 continue;
1288 }
1289 bool short_scale = false;
1290 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1291 assert(ok, "inconsistent: was tested before");
1292 Node* range = rc_cmp->in(2);
1293 Node* c = rc->in(0);
1294 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1295
1296 Node* R = range;
1297 Node* K = longcon(scale);
1298
1299 Node* L = offset;
1300
1301 if (short_scale) {
1302 // This converts:
1303 // (int)i*K + L <u64 R
1304 // with K an int into:
1305 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1306 // to protect against an overflow of (int)i*K
1307 //
1308 // Because if (int)i*K overflows, there are K,L where:
1309 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1310 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1311 // i*(long)K + L <u64 R is true.
1312 //
1313 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1314 //
1315 // It's always true that:
1316 // (int)i*K <u64 (long)max_jint + 1
1317 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1318 // As a consequence:
1319 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1320 // is always false in case of overflow of i*K
1321 //
1322 // Note, there are also K,L where i*K overflows and
1323 // i*K + L <u64 R is true, but
1324 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1325 // So this transformation could cause spurious deoptimizations and failed range check elimination
1326 // (but not incorrect execution) for unlikely corner cases with overflow.
1327 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1328 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1329 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1330 register_new_node(max_range, entry_control);
1331 R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1332 set_subtree_ctrl(R, true);
1333 }
1334
1335 Node* C = outer_phi;
1336
1337 // Start with 64-bit values:
1338 // i*K + L <u64 R
1339 // (C+j)*K + L <u64 R
1340 // j*K + Q <u64 R where Q = Q_first = C*K+L
1341 Node* Q_first = new MulLNode(C, K);
1342 register_new_node(Q_first, entry_control);
1343 Q_first = new AddLNode(Q_first, L);
1344 register_new_node(Q_first, entry_control);
1345
1346 // Compute endpoints of the range of values j*K + Q.
1347 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1348 Node* Q_min = Q_first;
1349
1350 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1351 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1352 register_new_node(B_2, entry_control);
1353 B_2 = new SubINode(B_2, int_stride);
1354 register_new_node(B_2, entry_control);
1355 B_2 = new ConvI2LNode(B_2);
1356 register_new_node(B_2, entry_control);
1357
1358 Node* Q_max = new MulLNode(B_2, K);
1359 register_new_node(Q_max, entry_control);
1360 Q_max = new AddLNode(Q_max, Q_first);
1361 register_new_node(Q_max, entry_control);
1362
1363 if (scale * stride_con < 0) {
1364 swap(Q_min, Q_max);
1365 }
1366 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1367
1368 // L_clamp = Q_min < 0 ? 0 : Q_min
1369 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1370 register_new_node(Q_min_cmp, entry_control);
1371 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1372 register_new_node(Q_min_bool, entry_control);
1373 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1374 register_new_node(L_clamp, entry_control);
1375 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1376
1377 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1378 register_new_node(Q_max_plus_one, entry_control);
1379
1380 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1381 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1382 Node* max_jlong_long = longcon(max_jlong);
1383 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1384 register_new_node(Q_max_cmp, entry_control);
1385 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1386 register_new_node(Q_max_bool, entry_control);
1387 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1388 register_new_node(H_clamp, entry_control);
1389 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1390
1391 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1392 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1393 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1394 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1395 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1396 Node* R_2 = clamp(R, L_clamp, H_clamp);
1397 R_2 = new SubLNode(R_2, L_clamp);
1398 register_new_node(R_2, entry_control);
1399 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1400 register_new_node(R_2, entry_control);
1401
1402 // L_2 = Q_first - L_clamp
1403 // We are subtracting L_clamp from both sides of the <u32 comparison.
1404 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1405 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1406 Node* L_2 = new SubLNode(Q_first, L_clamp);
1407 register_new_node(L_2, entry_control);
1408 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1409 register_new_node(L_2, entry_control);
1410
1411 // Transform the range check using the computed values L_2/R_2
1412 // from: i*K + L <u64 R
1413 // to: j*K + L_2 <u32 R_2
1414 // that is:
1415 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1416 K = intcon(checked_cast<int>(scale));
1417 Node* scaled_iv = new MulINode(inner_phi, K);
1418 register_new_node(scaled_iv, c);
1419 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1420 register_new_node(scaled_iv_plus_offset, c);
1421
1422 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1423 register_new_node(new_rc_cmp, c);
1424
1425 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1426 }
1427 }
1428
1429 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1430 Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1431 set_subtree_ctrl(min, true);
1432 Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1433 set_subtree_ctrl(max, true);
1434 return max;
1435 }
1436
1437 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1438 IfNode* exit_test) {
1439 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1440 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1441 _igvn.register_new_node_with_optimizer(new_inner_head);
1442 _igvn.register_new_node_with_optimizer(new_inner_exit);
1443 loop->_body.push(new_inner_head);
1444 loop->_body.push(new_inner_exit);
1445 loop->_body.yank(head);
1446 loop->_body.yank(exit_test);
1447 set_loop(new_inner_head, loop);
1448 set_loop(new_inner_exit, loop);
1449 set_idom(new_inner_head, idom(head), dom_depth(head));
1450 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1451 lazy_replace(head, new_inner_head);
1452 lazy_replace(exit_test, new_inner_exit);
1453 loop->_head = new_inner_head;
1454 return new_inner_head;
1455 }
1456
1457 #ifdef ASSERT
1458 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1459 Node* back_control = loop_exit_control(x, loop);
1460 assert(back_control != nullptr, "no back control");
1461
1462 BoolTest::mask mask = BoolTest::illegal;
1463 float cl_prob = 0;
1464 Node* incr = nullptr;
1465 Node* limit = nullptr;
1466
1467 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1468 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1469
1470 Node* phi_incr = nullptr;
1471 incr = loop_iv_incr(incr, x, loop, phi_incr);
1472 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1473
1474 Node* xphi = nullptr;
1475 Node* stride = loop_iv_stride(incr, loop, xphi);
1476
1477 assert(stride != nullptr, "no stride");
1478
1479 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1480
1481 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1482
1483 jlong stride_con = stride->get_integer_as_long(bt);
1484
1485 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1486
1487 assert(mask != BoolTest::ne, "unexpected condition");
1488 assert(phi_incr == nullptr, "bad loop shape");
1489 assert(cmp->in(1) == incr, "bad exit test shape");
1490
1491 // Safepoint on backedge not supported
1492 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1493 }
1494 #endif
1495
1496 #ifdef ASSERT
1497 // convert an int counted loop to a long counted to stress handling of
1498 // long counted loops
1499 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1500 Unique_Node_List iv_nodes;
1501 Node_List old_new;
1502 iv_nodes.push(cmp);
1503 bool failed = false;
1504
1505 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1506 Node* n = iv_nodes.at(i);
1507 switch(n->Opcode()) {
1508 case Op_Phi: {
1509 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1510 old_new.map(n->_idx, clone);
1511 break;
1512 }
1513 case Op_CmpI: {
1514 Node* clone = new CmpLNode(nullptr, nullptr);
1515 old_new.map(n->_idx, clone);
1516 break;
1517 }
1518 case Op_AddI: {
1519 Node* clone = new AddLNode(nullptr, nullptr);
1520 old_new.map(n->_idx, clone);
1521 break;
1522 }
1523 case Op_CastII: {
1524 failed = true;
1525 break;
1526 }
1527 default:
1528 DEBUG_ONLY(n->dump());
1529 fatal("unexpected");
1530 }
1531
1532 for (uint i = 1; i < n->req(); i++) {
1533 Node* in = n->in(i);
1534 if (in == nullptr) {
1535 continue;
1536 }
1537 if (loop->is_member(get_loop(get_ctrl(in)))) {
1538 iv_nodes.push(in);
1539 }
1540 }
1541 }
1542
1543 if (failed) {
1544 for (uint i = 0; i < iv_nodes.size(); i++) {
1545 Node* n = iv_nodes.at(i);
1546 Node* clone = old_new[n->_idx];
1547 if (clone != nullptr) {
1548 _igvn.remove_dead_node(clone);
1549 }
1550 }
1551 return false;
1552 }
1553
1554 for (uint i = 0; i < iv_nodes.size(); i++) {
1555 Node* n = iv_nodes.at(i);
1556 Node* clone = old_new[n->_idx];
1557 for (uint i = 1; i < n->req(); i++) {
1558 Node* in = n->in(i);
1559 if (in == nullptr) {
1560 continue;
1561 }
1562 Node* in_clone = old_new[in->_idx];
1563 if (in_clone == nullptr) {
1564 assert(_igvn.type(in)->isa_int(), "");
1565 in_clone = new ConvI2LNode(in);
1566 _igvn.register_new_node_with_optimizer(in_clone);
1567 set_subtree_ctrl(in_clone, false);
1568 }
1569 if (in_clone->in(0) == nullptr) {
1570 in_clone->set_req(0, C->top());
1571 clone->set_req(i, in_clone);
1572 in_clone->set_req(0, nullptr);
1573 } else {
1574 clone->set_req(i, in_clone);
1575 }
1576 }
1577 _igvn.register_new_node_with_optimizer(clone);
1578 }
1579 set_ctrl(old_new[phi->_idx], phi->in(0));
1580
1581 for (uint i = 0; i < iv_nodes.size(); i++) {
1582 Node* n = iv_nodes.at(i);
1583 Node* clone = old_new[n->_idx];
1584 set_subtree_ctrl(clone, false);
1585 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1586 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1587 Node* u = n->fast_out(i);
1588 if (iv_nodes.member(u)) {
1589 continue;
1590 }
1591 if (m == nullptr) {
1592 m = new ConvL2INode(clone);
1593 _igvn.register_new_node_with_optimizer(m);
1594 set_subtree_ctrl(m, false);
1595 }
1596 _igvn.rehash_node_delayed(u);
1597 int nb = u->replace_edge(n, m, &_igvn);
1598 --i, imax -= nb;
1599 }
1600 }
1601 return true;
1602 }
1603 #endif
1604
1605 //------------------------------is_counted_loop--------------------------------
1606 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {
1607 PhaseGVN *gvn = &_igvn;
1608
1609 Node* back_control = loop_exit_control(x, loop);
1610 if (back_control == nullptr) {
1611 return false;
1612 }
1613
1614 BoolTest::mask bt = BoolTest::illegal;
1615 float cl_prob = 0;
1616 Node* incr = nullptr;
1617 Node* limit = nullptr;
1618 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1619 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1620 return false; // Avoid pointer & float & 64-bit compares
1621 }
1622
1623 // Trip-counter increment must be commutative & associative.
1624 if (incr->Opcode() == Op_Cast(iv_bt)) {
1625 incr = incr->in(1);
1626 }
1627
1628 Node* phi_incr = nullptr;
1629 incr = loop_iv_incr(incr, x, loop, phi_incr);
1630 if (incr == nullptr) {
1631 return false;
1632 }
1633
1634 Node* trunc1 = nullptr;
1635 Node* trunc2 = nullptr;
1636 const TypeInteger* iv_trunc_t = nullptr;
1637 Node* orig_incr = incr;
1638 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1639 return false; // Funny increment opcode
1640 }
1641 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1642
1643 Node* xphi = nullptr;
1644 Node* stride = loop_iv_stride(incr, loop, xphi);
1645
1646 if (stride == nullptr) {
1647 return false;
1648 }
1649
1650 if (xphi->Opcode() == Op_Cast(iv_bt)) {
1651 xphi = xphi->in(1);
1652 }
1653
1654 // Stride must be constant
1655 jlong stride_con = stride->get_integer_as_long(iv_bt);
1656 assert(stride_con != 0, "missed some peephole opt");
1657
1658 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1659
1660 if (phi == nullptr ||
1661 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1662 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1663 return false;
1664 }
1665
1666 Node* iftrue = back_control;
1667 uint iftrue_op = iftrue->Opcode();
1668 Node* iff = iftrue->in(0);
1669 BoolNode* test = iff->in(1)->as_Bool();
1670
1671 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1672 if (trunc1 != nullptr) {
1673 // When there is a truncation, we must be sure that after the truncation
1674 // the trip counter will end up higher than the limit, otherwise we are looking
1675 // at an endless loop. Can happen with range checks.
1676
1677 // Example:
1678 // int i = 0;
1679 // while (true)
1680 // sum + = array[i];
1681 // i++;
1682 // i = i && 0x7fff;
1683 // }
1684 //
1685 // If the array is shorter than 0x8000 this exits through a AIOOB
1686 // - Counted loop transformation is ok
1687 // If the array is longer then this is an endless loop
1688 // - No transformation can be done.
1689
1690 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1691 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1692 // if the limit can have a higher value than the increment (before the phi)
1693 return false;
1694 }
1695 }
1696
1697 Node *init_trip = phi->in(LoopNode::EntryControl);
1698
1699 // If iv trunc type is smaller than int, check for possible wrap.
1700 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1701 assert(trunc1 != nullptr, "must have found some truncation");
1702
1703 // Get a better type for the phi (filtered thru if's)
1704 const TypeInteger* phi_ft = filtered_type(phi);
1705
1706 // Can iv take on a value that will wrap?
1707 //
1708 // Ensure iv's limit is not within "stride" of the wrap value.
1709 //
1710 // Example for "short" type
1711 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
1712 // If the stride is +10, then the last value of the induction
1713 // variable before the increment (phi_ft->_hi) must be
1714 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
1715 // ensure no truncation occurs after the increment.
1716
1717 if (stride_con > 0) {
1718 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
1719 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
1720 return false; // truncation may occur
1721 }
1722 } else if (stride_con < 0) {
1723 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
1724 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
1725 return false; // truncation may occur
1726 }
1727 }
1728 // No possibility of wrap so truncation can be discarded
1729 // Promote iv type to Int
1730 } else {
1731 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
1732 }
1733
1734 if (!condition_stride_ok(bt, stride_con)) {
1735 return false;
1736 }
1737
1738 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
1739
1740 if (stride_con > 0) {
1741 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
1742 return false; // cyclic loop
1743 }
1744 } else {
1745 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
1746 return false; // cyclic loop
1747 }
1748 }
1749
1750 if (phi_incr != nullptr && bt != BoolTest::ne) {
1751 // check if there is a possibility of IV overflowing after the first increment
1752 if (stride_con > 0) {
1753 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
1754 return false;
1755 }
1756 } else {
1757 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
1758 return false;
1759 }
1760 }
1761 }
1762
1763 // =================================================
1764 // ---- SUCCESS! Found A Trip-Counted Loop! -----
1765 //
1766
1767 if (x->Opcode() == Op_Region) {
1768 // x has not yet been transformed to Loop or LongCountedLoop.
1769 // This should only happen if we are inside an infinite loop.
1770 // It happens like this:
1771 // build_loop_tree -> do not attach infinite loop and nested loops
1772 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
1773 // build_loop_tree -> find and attach infinite and nested loops
1774 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
1775 assert(x->as_Region()->is_in_infinite_subgraph(),
1776 "x can only be a Region and not Loop if inside infinite loop");
1777 // Come back later when Region is transformed to LoopNode
1778 return false;
1779 }
1780
1781 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
1782 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
1783
1784 // ===================================================
1785 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
1786 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
1787 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
1788 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
1789 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
1790 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
1791 //
1792 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
1793 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
1794 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
1795 //
1796 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
1797 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
1798 //
1799 //
1800 // The loop to be converted does not always need to have the often used shape:
1801 //
1802 // i = init
1803 // i = init loop:
1804 // do { ...
1805 // // ... equivalent i+=stride
1806 // i+=stride <==> if (i < limit)
1807 // } while (i < limit); goto loop
1808 // exit:
1809 // ...
1810 //
1811 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
1812 //
1813 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
1814 // in the loop exit check:
1815 //
1816 // i = init
1817 // loop:
1818 // ...
1819 // if (i <= limit)
1820 // i+=stride
1821 // goto loop
1822 // exit:
1823 // ...
1824 //
1825 // Let's define the following terms:
1826 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
1827 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
1828 //
1829 // The iv_pre_i and iv_post_i have the following relation:
1830 // iv_pre_i + stride = iv_post_i
1831 //
1832 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
1833 // iv_post_i < adjusted_limit
1834 //
1835 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
1836 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
1837 // -> adjusted_limit = limit.
1838 // (LE2) iv_post_i <= limit:
1839 // iv_post_i < limit + 1
1840 // -> adjusted limit = limit + 1
1841 // (LE3) iv_pre_i < limit:
1842 // iv_pre_i + stride < limit + stride
1843 // iv_post_i < limit + stride
1844 // -> adjusted_limit = limit + stride
1845 // (LE4) iv_pre_i <= limit:
1846 // iv_pre_i < limit + 1
1847 // iv_pre_i + stride < limit + stride + 1
1848 // iv_post_i < limit + stride + 1
1849 // -> adjusted_limit = limit + stride + 1
1850 //
1851 // Note that:
1852 // (AL) limit <= adjusted_limit.
1853 //
1854 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
1855 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
1856 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
1857 //
1858 // To prove (INV), we require the following two conditions/assumptions:
1859 // (i): adjusted_limit - 1 + stride <= max_int
1860 // (ii): init < limit
1861 //
1862 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
1863 // induction by assuming (i) and (ii).
1864 //
1865 // Proof by Induction
1866 // ------------------
1867 // > Base case (i = 1): We show that (INV) holds after the first iteration:
1868 // min_int <= iv_post_1 = init + stride <= max_int
1869 // Proof:
1870 // First, we note that (ii) implies
1871 // (iii) init <= limit - 1
1872 // max_int >= adjusted_limit - 1 + stride [using (i)]
1873 // >= limit - 1 + stride [using (AL)]
1874 // >= init + stride [using (iii)]
1875 // >= min_int [using stride > 0, no underflow]
1876 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
1877 //
1878 // Note that to prove the base case we need (i) and (ii).
1879 //
1880 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
1881 // min_int <= iv_post_j <= max_int
1882 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
1883 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
1884 // Proof:
1885 // If iv_post_j >= adjusted_limit:
1886 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
1887 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
1888 // If iv_post_j < adjusted_limit:
1889 // First, we note that:
1890 // (iv) iv_post_j <= adjusted_limit - 1
1891 // max_int >= adjusted_limit - 1 + stride [using (i)]
1892 // >= iv_post_j + stride [using (iv)]
1893 // >= min_int [using stride > 0, no underflow]
1894 //
1895 // Note that to prove the step case we only need (i).
1896 //
1897 // Thus, by assuming (i) and (ii), we proved (INV).
1898 //
1899 //
1900 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
1901 //
1902 // (1) Loop Limit Check Predicate for (i):
1903 // Using (i): adjusted_limit - 1 + stride <= max_int
1904 //
1905 // This condition is now restated to use limit instead of adjusted_limit:
1906 //
1907 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
1908 // max_int - stride + 1 >= adjusted_limit
1909 // We can merge the two constants into
1910 // canonicalized_correction = stride - 1
1911 // which gives us
1912 // max_int - canonicalized_correction >= adjusted_limit
1913 //
1914 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
1915 // adjusted_limit = limit + limit_correction
1916 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
1917 // max_int - canonicalized_correction - limit_correction >= limit
1918 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
1919 // (v) final_correction = canonicalized_correction + limit_correction
1920 //
1921 // which gives us:
1922 //
1923 // Final predicate condition:
1924 // max_int - final_correction >= limit
1925 //
1926 // However, we need to be careful that (v) does not over- or underflow.
1927 // We know that:
1928 // canonicalized_correction = stride - 1
1929 // and
1930 // limit_correction <= stride + 1
1931 // and thus
1932 // canonicalized_correction + limit_correction <= 2 * stride
1933 // To prevent an over- or underflow of (v), we must ensure that
1934 // 2 * stride <= max_int
1935 // which can safely be checked without over- or underflow with
1936 // (vi) stride != min_int AND abs(stride) <= max_int / 2
1937 //
1938 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
1939 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
1940 //
1941 // (2) Loop Limit Check Predicate for (ii):
1942 // Using (ii): init < limit
1943 //
1944 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
1945 // (2.1) type(init) < type(limit)
1946 // In this case, we know:
1947 // all possible values of init < all possible values of limit
1948 // and we can skip the predicate.
1949 //
1950 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
1951 // In this case, we do not need to re-check the condition and can skip the predicate.
1952 // This is often found for while- and for-loops which have the following shape:
1953 //
1954 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
1955 // i = init;
1956 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
1957 // do {
1958 // i += stride;
1959 // } while (i < limit);
1960 // }
1961 //
1962 // (2.3) init + stride <= max_int
1963 // In this case, there is no overflow of the iv phi after the first loop iteration.
1964 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
1965 // init < limit
1966 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
1967 // compile time that init + stride <= max_int then we have trivially proven the base case and that
1968 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
1969 // again and can skip the predicate.
1970
1971 // Check (vi) and bail out if the stride is too big.
1972 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
1973 return false;
1974 }
1975
1976 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
1977 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
1978
1979 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
1980 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
1981 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
1982
1983 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
1984 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
1985 const jlong final_correction = canonicalized_correction + limit_correction;
1986
1987 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
1988 Node* init_control = x->in(LoopNode::EntryControl);
1989
1990 // If sov==0, limit's type always satisfies the condition, for
1991 // example, when it is an array length.
1992 if (sov != 0) {
1993 if (sov < 0) {
1994 return false; // Bailout: integer overflow is certain.
1995 }
1996 // (1) Loop Limit Check Predicate is required because we could not statically prove that
1997 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
1998 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
1999 const Predicates predicates(init_control);
2000 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2001 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2002 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2003 #ifdef ASSERT
2004 if (TraceLoopLimitCheck) {
2005 tty->print("Missing Loop Limit Check Parse Predicate:");
2006 loop->dump_head();
2007 x->dump(1);
2008 }
2009 #endif
2010 return false;
2011 }
2012
2013 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2014 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2015 return false;
2016 }
2017
2018 Node* cmp_limit;
2019 Node* bol;
2020
2021 if (stride_con > 0) {
2022 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2023 bol = new BoolNode(cmp_limit, BoolTest::le);
2024 } else {
2025 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2026 bol = new BoolNode(cmp_limit, BoolTest::ge);
2027 }
2028
2029 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2030 }
2031
2032 // (2.3)
2033 const bool init_plus_stride_could_overflow =
2034 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2035 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2036 // (2.1)
2037 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2038 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2039
2040 if (init_gte_limit && // (2.1)
2041 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2042 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2043 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2044 // We use the following condition:
2045 // - stride > 0: init < limit
2046 // - stride < 0: init > limit
2047 //
2048 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2049 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2050 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2051
2052 const Predicates predicates(init_control);
2053 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2054 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2055 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2056 #ifdef ASSERT
2057 if (TraceLoopLimitCheck) {
2058 tty->print("Missing Loop Limit Check Parse Predicate:");
2059 loop->dump_head();
2060 x->dump(1);
2061 }
2062 #endif
2063 return false;
2064 }
2065
2066 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2067 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2068 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2069 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2070 return false;
2071 }
2072
2073 Node* cmp_limit;
2074 Node* bol;
2075
2076 if (stride_con > 0) {
2077 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2078 bol = new BoolNode(cmp_limit, BoolTest::lt);
2079 } else {
2080 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2081 bol = new BoolNode(cmp_limit, BoolTest::gt);
2082 }
2083
2084 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2085 }
2086
2087 if (bt == BoolTest::ne) {
2088 // Now we need to canonicalize the loop condition if it is 'ne'.
2089 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2090 if (stride_con > 0) {
2091 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2092 bt = BoolTest::lt;
2093 } else {
2094 assert(stride_con < 0, "must be");
2095 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2096 bt = BoolTest::gt;
2097 }
2098 }
2099
2100 Node* sfpt = nullptr;
2101 if (loop->_child == nullptr) {
2102 sfpt = find_safepoint(back_control, x, loop);
2103 } else {
2104 sfpt = iff->in(0);
2105 if (sfpt->Opcode() != Op_SafePoint) {
2106 sfpt = nullptr;
2107 }
2108 }
2109
2110 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2111 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2112 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2113 iv_bt == T_LONG) &&
2114 sfpt == nullptr) {
2115 // Leaving the safepoint on the backedge and creating a
2116 // CountedLoop will confuse optimizations. We can't move the
2117 // safepoint around because its jvm state wouldn't match a new
2118 // location. Give up on that loop.
2119 return false;
2120 }
2121 if (is_deleteable_safept(backedge_sfpt)) {
2122 lazy_replace(backedge_sfpt, iftrue);
2123 if (loop->_safepts != nullptr) {
2124 loop->_safepts->yank(backedge_sfpt);
2125 }
2126 loop->_tail = iftrue;
2127 }
2128 }
2129
2130
2131 #ifdef ASSERT
2132 if (iv_bt == T_INT &&
2133 !x->as_Loop()->is_loop_nest_inner_loop() &&
2134 StressLongCountedLoop > 0 &&
2135 trunc1 == nullptr &&
2136 convert_to_long_loop(cmp, phi, loop)) {
2137 return false;
2138 }
2139 #endif
2140
2141 Node* adjusted_limit = limit;
2142 if (phi_incr != nullptr) {
2143 // If compare points directly to the phi we need to adjust
2144 // the compare so that it points to the incr. Limit have
2145 // to be adjusted to keep trip count the same and we
2146 // should avoid int overflow.
2147 //
2148 // i = init; do {} while(i++ < limit);
2149 // is converted to
2150 // i = init; do {} while(++i < limit+1);
2151 //
2152 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2153 }
2154
2155 if (includes_limit) {
2156 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2157 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2158 //
2159 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2160 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2161 if (bt == BoolTest::le)
2162 bt = BoolTest::lt;
2163 else if (bt == BoolTest::ge)
2164 bt = BoolTest::gt;
2165 else
2166 ShouldNotReachHere();
2167 }
2168 set_subtree_ctrl(adjusted_limit, false);
2169
2170 // Build a canonical trip test.
2171 // Clone code, as old values may be in use.
2172 incr = incr->clone();
2173 incr->set_req(1,phi);
2174 incr->set_req(2,stride);
2175 incr = _igvn.register_new_node_with_optimizer(incr);
2176 set_early_ctrl(incr, false);
2177 _igvn.rehash_node_delayed(phi);
2178 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2179
2180 // If phi type is more restrictive than Int, raise to
2181 // Int to prevent (almost) infinite recursion in igvn
2182 // which can only handle integer types for constants or minint..maxint.
2183 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2184 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2185 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2186 nphi = _igvn.register_new_node_with_optimizer(nphi);
2187 set_ctrl(nphi, get_ctrl(phi));
2188 _igvn.replace_node(phi, nphi);
2189 phi = nphi->as_Phi();
2190 }
2191 cmp = cmp->clone();
2192 cmp->set_req(1,incr);
2193 cmp->set_req(2, adjusted_limit);
2194 cmp = _igvn.register_new_node_with_optimizer(cmp);
2195 set_ctrl(cmp, iff->in(0));
2196
2197 test = test->clone()->as_Bool();
2198 (*(BoolTest*)&test->_test)._test = bt;
2199 test->set_req(1,cmp);
2200 _igvn.register_new_node_with_optimizer(test);
2201 set_ctrl(test, iff->in(0));
2202
2203 // Replace the old IfNode with a new LoopEndNode
2204 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2205 IfNode *le = lex->as_If();
2206 uint dd = dom_depth(iff);
2207 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2208 set_loop(le, loop);
2209
2210 // Get the loop-exit control
2211 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2212
2213 // Need to swap loop-exit and loop-back control?
2214 if (iftrue_op == Op_IfFalse) {
2215 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2216 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2217
2218 loop->_tail = back_control = ift2;
2219 set_loop(ift2, loop);
2220 set_loop(iff2, get_loop(iffalse));
2221
2222 // Lazy update of 'get_ctrl' mechanism.
2223 lazy_replace(iffalse, iff2);
2224 lazy_replace(iftrue, ift2);
2225
2226 // Swap names
2227 iffalse = iff2;
2228 iftrue = ift2;
2229 } else {
2230 _igvn.rehash_node_delayed(iffalse);
2231 _igvn.rehash_node_delayed(iftrue);
2232 iffalse->set_req_X( 0, le, &_igvn );
2233 iftrue ->set_req_X( 0, le, &_igvn );
2234 }
2235
2236 set_idom(iftrue, le, dd+1);
2237 set_idom(iffalse, le, dd+1);
2238 assert(iff->outcnt() == 0, "should be dead now");
2239 lazy_replace( iff, le ); // fix 'get_ctrl'
2240
2241 Node* entry_control = init_control;
2242 bool strip_mine_loop = iv_bt == T_INT &&
2243 loop->_child == nullptr &&
2244 sfpt != nullptr &&
2245 !loop->_has_call &&
2246 is_deleteable_safept(sfpt);
2247 IdealLoopTree* outer_ilt = nullptr;
2248 if (strip_mine_loop) {
2249 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
2250 cl_prob, le->_fcnt, entry_control,
2251 iffalse);
2252 }
2253
2254 // Now setup a new CountedLoopNode to replace the existing LoopNode
2255 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2256 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2257 // The following assert is approximately true, and defines the intention
2258 // of can_be_counted_loop. It fails, however, because phase->type
2259 // is not yet initialized for this loop and its parts.
2260 //assert(l->can_be_counted_loop(this), "sanity");
2261 _igvn.register_new_node_with_optimizer(l);
2262 set_loop(l, loop);
2263 loop->_head = l;
2264 // Fix all data nodes placed at the old loop head.
2265 // Uses the lazy-update mechanism of 'get_ctrl'.
2266 lazy_replace( x, l );
2267 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2268
2269 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2270 // Check for immediately preceding SafePoint and remove
2271 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2272 if (strip_mine_loop) {
2273 Node* outer_le = outer_ilt->_tail->in(0);
2274 Node* sfpt_clone = sfpt->clone();
2275 sfpt_clone->set_req(0, iffalse);
2276 outer_le->set_req(0, sfpt_clone);
2277
2278 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2279 if (polladdr != nullptr && polladdr->is_Load()) {
2280 // Polling load should be pinned outside inner loop.
2281 Node* new_polladdr = polladdr->clone();
2282 new_polladdr->set_req(0, iffalse);
2283 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2284 set_ctrl(new_polladdr, iffalse);
2285 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2286 }
2287 // When this code runs, loop bodies have not yet been populated.
2288 const bool body_populated = false;
2289 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2290 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2291 }
2292 lazy_replace(sfpt, sfpt->in(TypeFunc::Control));
2293 if (loop->_safepts != nullptr) {
2294 loop->_safepts->yank(sfpt);
2295 }
2296 }
2297 }
2298
2299 #ifdef ASSERT
2300 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2301 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2302 #endif
2303 #ifndef PRODUCT
2304 if (TraceLoopOpts) {
2305 tty->print("Counted ");
2306 loop->dump_head();
2307 }
2308 #endif
2309
2310 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2311
2312 // Capture bounds of the loop in the induction variable Phi before
2313 // subsequent transformation (iteration splitting) obscures the
2314 // bounds
2315 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2316
2317 if (strip_mine_loop) {
2318 l->mark_strip_mined();
2319 l->verify_strip_mined(1);
2320 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2321 loop = outer_ilt;
2322 }
2323
2324 #ifndef PRODUCT
2325 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2326 Atomic::inc(&_long_loop_counted_loops);
2327 }
2328 #endif
2329 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2330 l->mark_loop_nest_outer_loop();
2331 }
2332
2333 return true;
2334 }
2335
2336 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2337 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2338 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2339 const BasicType iv_bt, Node* loop_entry) {
2340 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2341 // successfully find a dominated test with the If node below.
2342 Node* cmp_limit;
2343 Node* bol;
2344 if (stride_con > 0) {
2345 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2346 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2347 } else {
2348 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2349 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2350 }
2351
2352 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2353 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2354 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2355 IfFalseNode* if_false = new IfFalseNode(iff);
2356 IfTrueNode* if_true = new IfTrueNode(iff);
2357 Node* dominated_iff = _igvn.transform(iff);
2358 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2359 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2360
2361 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2362 _igvn.replace_input_of(iff, 0, C->top());
2363 _igvn.replace_input_of(iff, 1, C->top());
2364 return found_dominating_test;
2365 }
2366
2367 //----------------------exact_limit-------------------------------------------
2368 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2369 assert(loop->_head->is_CountedLoop(), "");
2370 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2371 assert(cl->is_valid_counted_loop(T_INT), "");
2372
2373 if (cl->stride_con() == 1 ||
2374 cl->stride_con() == -1 ||
2375 cl->limit()->Opcode() == Op_LoopLimit) {
2376 // Old code has exact limit (it could be incorrect in case of int overflow).
2377 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2378 return cl->limit();
2379 }
2380 Node *limit = nullptr;
2381 #ifdef ASSERT
2382 BoolTest::mask bt = cl->loopexit()->test_trip();
2383 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2384 #endif
2385 if (cl->has_exact_trip_count()) {
2386 // Simple case: loop has constant boundaries.
2387 // Use jlongs to avoid integer overflow.
2388 int stride_con = cl->stride_con();
2389 jlong init_con = cl->init_trip()->get_int();
2390 jlong limit_con = cl->limit()->get_int();
2391 julong trip_cnt = cl->trip_count();
2392 jlong final_con = init_con + trip_cnt*stride_con;
2393 int final_int = (int)final_con;
2394 // The final value should be in integer range since the loop
2395 // is counted and the limit was checked for overflow.
2396 assert(final_con == (jlong)final_int, "final value should be integer");
2397 limit = _igvn.intcon(final_int);
2398 } else {
2399 // Create new LoopLimit node to get exact limit (final iv value).
2400 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2401 register_new_node(limit, cl->in(LoopNode::EntryControl));
2402 }
2403 assert(limit != nullptr, "sanity");
2404 return limit;
2405 }
2406
2407 //------------------------------Ideal------------------------------------------
2408 // Return a node which is more "ideal" than the current node.
2409 // Attempt to convert into a counted-loop.
2410 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2411 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2412 phase->C->set_major_progress();
2413 }
2414 return RegionNode::Ideal(phase, can_reshape);
2415 }
2416
2417 #ifdef ASSERT
2418 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2419 const OuterStripMinedLoopNode* outer = nullptr;
2420 const CountedLoopNode* inner = nullptr;
2421 if (is_strip_mined()) {
2422 if (!is_valid_counted_loop(T_INT)) {
2423 return; // Skip malformed counted loop
2424 }
2425 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2426 inner = as_CountedLoop();
2427 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2428 } else if (is_OuterStripMinedLoop()) {
2429 outer = this->as_OuterStripMinedLoop();
2430 inner = outer->unique_ctrl_out()->as_CountedLoop();
2431 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2432 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2433 }
2434 if (inner != nullptr || outer != nullptr) {
2435 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2436 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2437 Node* outer_le = outer_tail->in(0);
2438 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2439 Node* sfpt = outer_le->in(0);
2440 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2441 Node* inner_out = sfpt->in(0);
2442 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2443 assert(cle == inner->loopexit_or_null(), "mismatch");
2444 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2445 if (has_skeleton) {
2446 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2447 assert(outer->outcnt() == 2, "only control nodes");
2448 } else {
2449 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2450 uint phis = 0;
2451 uint be_loads = 0;
2452 Node* be = inner->in(LoopNode::LoopBackControl);
2453 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2454 Node* u = inner->fast_out(i);
2455 if (u->is_Phi()) {
2456 phis++;
2457 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2458 Node* n = be->fast_out(j);
2459 if (n->is_Load()) {
2460 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2461 do {
2462 n = n->raw_out(0);
2463 } while (!n->is_Phi());
2464 if (n == u) {
2465 be_loads++;
2466 break;
2467 }
2468 }
2469 }
2470 }
2471 }
2472 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2473 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2474 Node* u = outer->fast_out(i);
2475 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2476 }
2477 uint stores = 0;
2478 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2479 Node* u = inner_out->fast_out(i);
2480 if (u->is_Store()) {
2481 stores++;
2482 }
2483 }
2484 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2485 // not guaranteed to be optimized out.
2486 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2487 }
2488 assert(sfpt->outcnt() == 1, "no data node");
2489 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2490 }
2491 }
2492 #endif
2493
2494 //=============================================================================
2495 //------------------------------Ideal------------------------------------------
2496 // Return a node which is more "ideal" than the current node.
2497 // Attempt to convert into a counted-loop.
2498 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2499 return RegionNode::Ideal(phase, can_reshape);
2500 }
2501
2502 //------------------------------dump_spec--------------------------------------
2503 // Dump special per-node info
2504 #ifndef PRODUCT
2505 void CountedLoopNode::dump_spec(outputStream *st) const {
2506 LoopNode::dump_spec(st);
2507 if (stride_is_con()) {
2508 st->print("stride: %d ",stride_con());
2509 }
2510 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2511 if (is_main_loop()) st->print("main of N%d", _idx);
2512 if (is_post_loop()) st->print("post of N%d", _main_idx);
2513 if (is_strip_mined()) st->print(" strip mined");
2514 }
2515 #endif
2516
2517 //=============================================================================
2518 jlong BaseCountedLoopEndNode::stride_con() const {
2519 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2520 }
2521
2522
2523 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2524 if (bt == T_INT) {
2525 return new CountedLoopEndNode(control, test, prob, cnt);
2526 }
2527 assert(bt == T_LONG, "unsupported");
2528 return new LongCountedLoopEndNode(control, test, prob, cnt);
2529 }
2530
2531 //=============================================================================
2532 //------------------------------Value-----------------------------------------
2533 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2534 const Type* init_t = phase->type(in(Init));
2535 const Type* limit_t = phase->type(in(Limit));
2536 const Type* stride_t = phase->type(in(Stride));
2537 // Either input is TOP ==> the result is TOP
2538 if (init_t == Type::TOP) return Type::TOP;
2539 if (limit_t == Type::TOP) return Type::TOP;
2540 if (stride_t == Type::TOP) return Type::TOP;
2541
2542 int stride_con = stride_t->is_int()->get_con();
2543 if (stride_con == 1)
2544 return bottom_type(); // Identity
2545
2546 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2547 // Use jlongs to avoid integer overflow.
2548 jlong init_con = init_t->is_int()->get_con();
2549 jlong limit_con = limit_t->is_int()->get_con();
2550 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2551 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2552 jlong final_con = init_con + stride_con*trip_count;
2553 int final_int = (int)final_con;
2554 // The final value should be in integer range in almost all cases,
2555 // since the loop is counted and the limit was checked for overflow.
2556 // There some exceptions, for example:
2557 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2558 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2559 // the loop and its predicates, and we might get constants on one side of the phi that
2560 // would lead to overflows. Such a code path would never lead us to enter the loop
2561 // because of the loop limit overflow check that happens after the LoopLimitNode
2562 // computation with overflow, but before we enter the loop, see JDK-8335747.
2563 if (final_con == (jlong)final_int) {
2564 return TypeInt::make(final_int);
2565 } else {
2566 return bottom_type();
2567 }
2568 }
2569
2570 return bottom_type(); // TypeInt::INT
2571 }
2572
2573 //------------------------------Ideal------------------------------------------
2574 // Return a node which is more "ideal" than the current node.
2575 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2576 if (phase->type(in(Init)) == Type::TOP ||
2577 phase->type(in(Limit)) == Type::TOP ||
2578 phase->type(in(Stride)) == Type::TOP)
2579 return nullptr; // Dead
2580
2581 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2582 if (stride_con == 1)
2583 return nullptr; // Identity
2584
2585 // Delay following optimizations until all loop optimizations
2586 // done to keep Ideal graph simple.
2587 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2588 phase->C->record_for_post_loop_opts_igvn(this);
2589 return nullptr;
2590 }
2591
2592 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2593 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2594 jlong stride_p;
2595 jlong lim, ini;
2596 julong max;
2597 if (stride_con > 0) {
2598 stride_p = stride_con;
2599 lim = limit_t->_hi;
2600 ini = init_t->_lo;
2601 max = (julong)max_jint;
2602 } else {
2603 stride_p = -(jlong)stride_con;
2604 lim = init_t->_hi;
2605 ini = limit_t->_lo;
2606 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2607 }
2608 julong range = lim - ini + stride_p;
2609 if (range <= max) {
2610 // Convert to integer expression if it is not overflow.
2611 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2612 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2613 Node *bias = phase->transform(new AddINode(range, stride_m));
2614 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2615 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2616 return new AddINode(span, in(Init)); // exact limit
2617 }
2618
2619 if (is_power_of_2(stride_p) || // divisor is 2^n
2620 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2621 // Convert to long expression to avoid integer overflow
2622 // and let igvn optimizer convert this division.
2623 //
2624 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2625 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2626 Node* stride = phase->longcon(stride_con);
2627 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2628
2629 Node *range = phase->transform(new SubLNode(limit, init));
2630 Node *bias = phase->transform(new AddLNode(range, stride_m));
2631 Node *span;
2632 if (stride_con > 0 && is_power_of_2(stride_p)) {
2633 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2634 // and avoid generating rounding for division. Zero trip guard should
2635 // guarantee that init < limit but sometimes the guard is missing and
2636 // we can get situation when init > limit. Note, for the empty loop
2637 // optimization zero trip guard is generated explicitly which leaves
2638 // only RCE predicate where exact limit is used and the predicate
2639 // will simply fail forcing recompilation.
2640 Node* neg_stride = phase->longcon(-stride_con);
2641 span = phase->transform(new AndLNode(bias, neg_stride));
2642 } else {
2643 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2644 span = phase->transform(new MulLNode(trip, stride));
2645 }
2646 // Convert back to int
2647 Node *span_int = phase->transform(new ConvL2INode(span));
2648 return new AddINode(span_int, in(Init)); // exact limit
2649 }
2650
2651 return nullptr; // No progress
2652 }
2653
2654 //------------------------------Identity---------------------------------------
2655 // If stride == 1 return limit node.
2656 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2657 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2658 if (stride_con == 1 || stride_con == -1)
2659 return in(Limit);
2660 return this;
2661 }
2662
2663 //=============================================================================
2664 //----------------------match_incr_with_optional_truncation--------------------
2665 // Match increment with optional truncation:
2666 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2667 // Return null for failure. Success returns the increment node.
2668 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2669 const TypeInteger** trunc_type,
2670 BasicType bt) {
2671 // Quick cutouts:
2672 if (expr == nullptr || expr->req() != 3) return nullptr;
2673
2674 Node *t1 = nullptr;
2675 Node *t2 = nullptr;
2676 Node* n1 = expr;
2677 int n1op = n1->Opcode();
2678 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2679
2680 if (bt == T_INT) {
2681 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2682 if (n1op == Op_AndI &&
2683 n1->in(2)->is_Con() &&
2684 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2685 // %%% This check should match any mask of 2**K-1.
2686 t1 = n1;
2687 n1 = t1->in(1);
2688 n1op = n1->Opcode();
2689 trunc_t = TypeInt::CHAR;
2690 } else if (n1op == Op_RShiftI &&
2691 n1->in(1) != nullptr &&
2692 n1->in(1)->Opcode() == Op_LShiftI &&
2693 n1->in(2) == n1->in(1)->in(2) &&
2694 n1->in(2)->is_Con()) {
2695 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2696 // %%% This check should match any shift in [1..31].
2697 if (shift == 16 || shift == 8) {
2698 t1 = n1;
2699 t2 = t1->in(1);
2700 n1 = t2->in(1);
2701 n1op = n1->Opcode();
2702 if (shift == 16) {
2703 trunc_t = TypeInt::SHORT;
2704 } else if (shift == 8) {
2705 trunc_t = TypeInt::BYTE;
2706 }
2707 }
2708 }
2709 }
2710
2711 // If (maybe after stripping) it is an AddI, we won:
2712 if (n1op == Op_Add(bt)) {
2713 *trunc1 = t1;
2714 *trunc2 = t2;
2715 *trunc_type = trunc_t;
2716 return n1;
2717 }
2718
2719 // failed
2720 return nullptr;
2721 }
2722
2723 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
2724 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
2725 verify_strip_mined(expect_skeleton);
2726 return in(EntryControl)->as_Loop();
2727 }
2728 return this;
2729 }
2730
2731 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
2732 assert(is_strip_mined(), "not a strip mined loop");
2733 Node* c = in(EntryControl);
2734 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
2735 return nullptr;
2736 }
2737 return c->as_OuterStripMinedLoop();
2738 }
2739
2740 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
2741 Node* c = in(LoopBackControl);
2742 if (c == nullptr || c->is_top()) {
2743 return nullptr;
2744 }
2745 return c->as_IfTrue();
2746 }
2747
2748 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
2749 LoopNode* l = outer_loop();
2750 if (l == nullptr) {
2751 return nullptr;
2752 }
2753 return l->outer_loop_tail();
2754 }
2755
2756 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
2757 IfTrueNode* proj = outer_loop_tail();
2758 if (proj == nullptr) {
2759 return nullptr;
2760 }
2761 Node* c = proj->in(0);
2762 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
2763 return nullptr;
2764 }
2765 return c->as_OuterStripMinedLoopEnd();
2766 }
2767
2768 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
2769 LoopNode* l = outer_loop();
2770 if (l == nullptr) {
2771 return nullptr;
2772 }
2773 return l->outer_loop_end();
2774 }
2775
2776 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
2777 IfNode* le = outer_loop_end();
2778 if (le == nullptr) {
2779 return nullptr;
2780 }
2781 Node* c = le->proj_out_or_null(false);
2782 if (c == nullptr) {
2783 return nullptr;
2784 }
2785 return c->as_IfFalse();
2786 }
2787
2788 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
2789 LoopNode* l = outer_loop();
2790 if (l == nullptr) {
2791 return nullptr;
2792 }
2793 return l->outer_loop_exit();
2794 }
2795
2796 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
2797 IfNode* le = outer_loop_end();
2798 if (le == nullptr) {
2799 return nullptr;
2800 }
2801 Node* c = le->in(0);
2802 if (c == nullptr || c->is_top()) {
2803 return nullptr;
2804 }
2805 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
2806 return c->as_SafePoint();
2807 }
2808
2809 SafePointNode* CountedLoopNode::outer_safepoint() const {
2810 LoopNode* l = outer_loop();
2811 if (l == nullptr) {
2812 return nullptr;
2813 }
2814 return l->outer_safepoint();
2815 }
2816
2817 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
2818 Node* ctrl = in(LoopNode::EntryControl);
2819 if (ctrl == nullptr) {
2820 // Dying loop.
2821 return nullptr;
2822 }
2823 if (is_main_loop()) {
2824 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
2825 }
2826 if (is_main_loop() || is_post_loop()) {
2827 AssertionPredicates assertion_predicates(ctrl);
2828 return assertion_predicates.entry();
2829 }
2830 return ctrl;
2831 }
2832
2833
2834 int CountedLoopNode::stride_con() const {
2835 CountedLoopEndNode* cle = loopexit_or_null();
2836 return cle != nullptr ? cle->stride_con() : 0;
2837 }
2838
2839 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
2840 if (bt == T_INT) {
2841 return new CountedLoopNode(entry, backedge);
2842 }
2843 assert(bt == T_LONG, "unsupported");
2844 return new LongCountedLoopNode(entry, backedge);
2845 }
2846
2847 void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
2848 PhaseIdealLoop* iloop) {
2849 Node* cle_out = inner_cle->proj_out(false);
2850 Node* cle_tail = inner_cle->proj_out(true);
2851 if (cle_out->outcnt() > 1) {
2852 // Look for chains of stores that were sunk
2853 // out of the inner loop and are in the outer loop
2854 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
2855 Node* u = cle_out->fast_out(i);
2856 if (u->is_Store()) {
2857 int alias_idx = igvn->C->get_alias_index(u->adr_type());
2858 Node* first = u;
2859 for (;;) {
2860 Node* next = first->in(MemNode::Memory);
2861 if (!next->is_Store() || next->in(0) != cle_out) {
2862 break;
2863 }
2864 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2865 first = next;
2866 }
2867 Node* last = u;
2868 for (;;) {
2869 Node* next = nullptr;
2870 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
2871 Node* uu = last->fast_out(j);
2872 if (uu->is_Store() && uu->in(0) == cle_out) {
2873 assert(next == nullptr, "only one in the outer loop");
2874 next = uu;
2875 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2876 }
2877 }
2878 if (next == nullptr) {
2879 break;
2880 }
2881 last = next;
2882 }
2883 Node* phi = nullptr;
2884 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2885 Node* uu = inner_cl->fast_out(j);
2886 if (uu->is_Phi()) {
2887 Node* be = uu->in(LoopNode::LoopBackControl);
2888 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
2889 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
2890 }
2891 if (be == last || be == first->in(MemNode::Memory)) {
2892 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
2893 assert(phi == nullptr, "only one phi");
2894 phi = uu;
2895 }
2896 }
2897 }
2898 #ifdef ASSERT
2899 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2900 Node* uu = inner_cl->fast_out(j);
2901 if (uu->is_memory_phi()) {
2902 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
2903 assert(phi == uu, "what's that phi?");
2904 } else if (uu->adr_type() == TypePtr::BOTTOM) {
2905 Node* n = uu->in(LoopNode::LoopBackControl);
2906 uint limit = igvn->C->live_nodes();
2907 uint i = 0;
2908 while (n != uu) {
2909 i++;
2910 assert(i < limit, "infinite loop");
2911 if (n->is_Proj()) {
2912 n = n->in(0);
2913 } else if (n->is_SafePoint() || n->is_MemBar()) {
2914 n = n->in(TypeFunc::Memory);
2915 } else if (n->is_Phi()) {
2916 n = n->in(1);
2917 } else if (n->is_MergeMem()) {
2918 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
2919 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
2920 n = n->in(MemNode::Memory);
2921 } else {
2922 n->dump();
2923 ShouldNotReachHere();
2924 }
2925 }
2926 }
2927 }
2928 }
2929 #endif
2930 if (phi == nullptr) {
2931 // If an entire chains was sunk, the
2932 // inner loop has no phi for that memory
2933 // slice, create one for the outer loop
2934 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
2935 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
2936 phi->set_req(LoopNode::LoopBackControl, last);
2937 phi = register_new_node(phi, inner_cl, igvn, iloop);
2938 igvn->replace_input_of(first, MemNode::Memory, phi);
2939 } else {
2940 // Or fix the outer loop fix to include
2941 // that chain of stores.
2942 Node* be = phi->in(LoopNode::LoopBackControl);
2943 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
2944 if (be == first->in(MemNode::Memory)) {
2945 if (be == phi->in(LoopNode::LoopBackControl)) {
2946 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
2947 } else {
2948 igvn->replace_input_of(be, MemNode::Memory, last);
2949 }
2950 } else {
2951 #ifdef ASSERT
2952 if (be == phi->in(LoopNode::LoopBackControl)) {
2953 assert(phi->in(LoopNode::LoopBackControl) == last, "");
2954 } else {
2955 assert(be->in(MemNode::Memory) == last, "");
2956 }
2957 #endif
2958 }
2959 }
2960 }
2961 }
2962 }
2963 }
2964
2965 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
2966 // Look for the outer & inner strip mined loop, reduce number of
2967 // iterations of the inner loop, set exit condition of outer loop,
2968 // construct required phi nodes for outer loop.
2969 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
2970 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
2971 if (LoopStripMiningIter == 0) {
2972 remove_outer_loop_and_safepoint(igvn);
2973 return;
2974 }
2975 if (LoopStripMiningIter == 1) {
2976 transform_to_counted_loop(igvn, nullptr);
2977 return;
2978 }
2979 Node* inner_iv_phi = inner_cl->phi();
2980 if (inner_iv_phi == nullptr) {
2981 IfNode* outer_le = outer_loop_end();
2982 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
2983 igvn->replace_node(outer_le, iff);
2984 inner_cl->clear_strip_mined();
2985 return;
2986 }
2987 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
2988
2989 int stride = inner_cl->stride_con();
2990 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
2991 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
2992 // stride, the method is guaranteed to return at the next check below.
2993 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
2994 int scaled_iters = (int)scaled_iters_long;
2995 if ((jlong)scaled_iters != scaled_iters_long) {
2996 // Remove outer loop and safepoint (too few iterations)
2997 remove_outer_loop_and_safepoint(igvn);
2998 return;
2999 }
3000 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3001 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3002 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3003 assert(iter_estimate > 0, "broken");
3004 if (iter_estimate <= short_scaled_iters) {
3005 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3006 remove_outer_loop_and_safepoint(igvn);
3007 return;
3008 }
3009 if (iter_estimate <= scaled_iters_long) {
3010 // We would only go through one iteration of
3011 // the outer loop: drop the outer loop but
3012 // keep the safepoint so we don't run for
3013 // too long without a safepoint
3014 IfNode* outer_le = outer_loop_end();
3015 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3016 igvn->replace_node(outer_le, iff);
3017 inner_cl->clear_strip_mined();
3018 return;
3019 }
3020
3021 Node* cle_tail = inner_cle->proj_out(true);
3022 ResourceMark rm;
3023 Node_List old_new;
3024 if (cle_tail->outcnt() > 1) {
3025 // Look for nodes on backedge of inner loop and clone them
3026 Unique_Node_List backedge_nodes;
3027 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3028 Node* u = cle_tail->fast_out(i);
3029 if (u != inner_cl) {
3030 assert(!u->is_CFG(), "control flow on the backedge?");
3031 backedge_nodes.push(u);
3032 }
3033 }
3034 uint last = igvn->C->unique();
3035 for (uint next = 0; next < backedge_nodes.size(); next++) {
3036 Node* n = backedge_nodes.at(next);
3037 old_new.map(n->_idx, n->clone());
3038 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3039 Node* u = n->fast_out(i);
3040 assert(!u->is_CFG(), "broken");
3041 if (u->_idx >= last) {
3042 continue;
3043 }
3044 if (!u->is_Phi()) {
3045 backedge_nodes.push(u);
3046 } else {
3047 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3048 }
3049 }
3050 }
3051 // Put the clones on the outer loop backedge
3052 Node* le_tail = outer_loop_tail();
3053 for (uint next = 0; next < backedge_nodes.size(); next++) {
3054 Node *n = old_new[backedge_nodes.at(next)->_idx];
3055 for (uint i = 1; i < n->req(); i++) {
3056 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3057 n->set_req(i, old_new[n->in(i)->_idx]);
3058 }
3059 }
3060 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3061 n->set_req(0, le_tail);
3062 }
3063 igvn->register_new_node_with_optimizer(n);
3064 }
3065 }
3066
3067 Node* iv_phi = nullptr;
3068 // Make a clone of each phi in the inner loop
3069 // for the outer loop
3070 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3071 Node* u = inner_cl->raw_out(i);
3072 if (u->is_Phi()) {
3073 assert(u->in(0) == inner_cl, "inconsistent");
3074 Node* phi = u->clone();
3075 phi->set_req(0, this);
3076 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3077 if (be != nullptr) {
3078 phi->set_req(LoopNode::LoopBackControl, be);
3079 }
3080 phi = igvn->transform(phi);
3081 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3082 if (u == inner_iv_phi) {
3083 iv_phi = phi;
3084 }
3085 }
3086 }
3087
3088 if (iv_phi != nullptr) {
3089 // Now adjust the inner loop's exit condition
3090 Node* limit = inner_cl->limit();
3091 // If limit < init for stride > 0 (or limit > init for stride < 0),
3092 // the loop body is run only once. Given limit - init (init - limit resp.)
3093 // would be negative, the unsigned comparison below would cause
3094 // the loop body to be run for LoopStripMiningIter.
3095 Node* max = nullptr;
3096 if (stride > 0) {
3097 max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3098 } else {
3099 max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3100 }
3101 // sub is positive and can be larger than the max signed int
3102 // value. Use an unsigned min.
3103 Node* const_iters = igvn->intcon(scaled_iters);
3104 Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3105 // min is the number of iterations for the next inner loop execution:
3106 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3107 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3108
3109 Node* new_limit = nullptr;
3110 if (stride > 0) {
3111 new_limit = igvn->transform(new AddINode(min, iv_phi));
3112 } else {
3113 new_limit = igvn->transform(new SubINode(iv_phi, min));
3114 }
3115 Node* inner_cmp = inner_cle->cmp_node();
3116 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3117 Node* outer_bol = inner_bol;
3118 // cmp node for inner loop may be shared
3119 inner_cmp = inner_cmp->clone();
3120 inner_cmp->set_req(2, new_limit);
3121 inner_bol = inner_bol->clone();
3122 inner_bol->set_req(1, igvn->transform(inner_cmp));
3123 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3124 // Set the outer loop's exit condition too
3125 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3126 } else {
3127 assert(false, "should be able to adjust outer loop");
3128 IfNode* outer_le = outer_loop_end();
3129 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3130 igvn->replace_node(outer_le, iff);
3131 inner_cl->clear_strip_mined();
3132 }
3133 }
3134
3135 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3136 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3137 CountedLoopEndNode* cle = inner_cl->loopexit();
3138 Node* inner_test = cle->in(1);
3139 IfNode* outer_le = outer_loop_end();
3140 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3141 Node* safepoint = outer_safepoint();
3142
3143 fix_sunk_stores(inner_cle, inner_cl, igvn, iloop);
3144
3145 // make counted loop exit test always fail
3146 ConINode* zero = igvn->intcon(0);
3147 if (iloop != nullptr) {
3148 iloop->set_root_as_ctrl(zero);
3149 }
3150 igvn->replace_input_of(cle, 1, zero);
3151 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3152 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3153 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3154 if (iloop == nullptr) {
3155 igvn->replace_node(outer_le, new_end);
3156 } else {
3157 iloop->lazy_replace(outer_le, new_end);
3158 }
3159 // the backedge of the inner loop must be rewired to the new loop end
3160 Node* backedge = cle->proj_out(true);
3161 igvn->replace_input_of(backedge, 0, new_end);
3162 if (iloop != nullptr) {
3163 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3164 }
3165 // make the outer loop go away
3166 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3167 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3168 inner_cl->clear_strip_mined();
3169 if (iloop != nullptr) {
3170 Unique_Node_List wq;
3171 wq.push(safepoint);
3172
3173 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3174 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3175
3176 for (uint i = 0; i < wq.size(); i++) {
3177 Node* n = wq.at(i);
3178 for (uint j = 0; j < n->req(); ++j) {
3179 Node* in = n->in(j);
3180 if (in == nullptr || in->is_CFG()) {
3181 continue;
3182 }
3183 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3184 continue;
3185 }
3186 assert(!loop->_body.contains(in), "");
3187 loop->_body.push(in);
3188 wq.push(in);
3189 }
3190 }
3191 iloop->set_loop(safepoint, loop);
3192 loop->_body.push(safepoint);
3193 iloop->set_loop(safepoint->in(0), loop);
3194 loop->_body.push(safepoint->in(0));
3195 outer_loop_ilt->_tail = igvn->C->top();
3196 }
3197 }
3198
3199 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3200 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3201 Node* outer_sfpt = outer_safepoint();
3202 Node* outer_out = outer_loop_exit();
3203 igvn->replace_node(outer_out, outer_sfpt->in(0));
3204 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3205 inner_cl->clear_strip_mined();
3206 }
3207
3208 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3209 if (iloop == nullptr) {
3210 return igvn->transform(node);
3211 }
3212 iloop->register_new_node(node, ctrl);
3213 return node;
3214 }
3215
3216 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3217 PhaseIdealLoop* iloop) {
3218 if (iloop == nullptr) {
3219 return igvn->transform(node);
3220 }
3221 iloop->register_control(node, iloop->get_loop(loop), idom);
3222 return node;
3223 }
3224
3225 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3226 if (!in(0)) return Type::TOP;
3227 if (phase->type(in(0)) == Type::TOP)
3228 return Type::TOP;
3229
3230 // Until expansion, the loop end condition is not set so this should not constant fold.
3231 if (is_expanded(phase)) {
3232 return IfNode::Value(phase);
3233 }
3234
3235 return TypeTuple::IFBOTH;
3236 }
3237
3238 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3239 // The outer strip mined loop head only has Phi uses after expansion
3240 if (phase->is_IterGVN()) {
3241 Node* backedge = proj_out_or_null(true);
3242 if (backedge != nullptr) {
3243 Node* head = backedge->unique_ctrl_out_or_null();
3244 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3245 if (head->find_out_with(Op_Phi) != nullptr) {
3246 return true;
3247 }
3248 }
3249 }
3250 }
3251 return false;
3252 }
3253
3254 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3255 if (remove_dead_region(phase, can_reshape)) return this;
3256
3257 return nullptr;
3258 }
3259
3260 //------------------------------filtered_type--------------------------------
3261 // Return a type based on condition control flow
3262 // A successful return will be a type that is restricted due
3263 // to a series of dominating if-tests, such as:
3264 // if (i < 10) {
3265 // if (i > 0) {
3266 // here: "i" type is [1..10)
3267 // }
3268 // }
3269 // or a control flow merge
3270 // if (i < 10) {
3271 // do {
3272 // phi( , ) -- at top of loop type is [min_int..10)
3273 // i = ?
3274 // } while ( i < 10)
3275 //
3276 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3277 assert(n && n->bottom_type()->is_int(), "must be int");
3278 const TypeInt* filtered_t = nullptr;
3279 if (!n->is_Phi()) {
3280 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3281 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3282
3283 } else {
3284 Node* phi = n->as_Phi();
3285 Node* region = phi->in(0);
3286 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3287 if (region && region != C->top()) {
3288 for (uint i = 1; i < phi->req(); i++) {
3289 Node* val = phi->in(i);
3290 Node* use_c = region->in(i);
3291 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3292 if (val_t != nullptr) {
3293 if (filtered_t == nullptr) {
3294 filtered_t = val_t;
3295 } else {
3296 filtered_t = filtered_t->meet(val_t)->is_int();
3297 }
3298 }
3299 }
3300 }
3301 }
3302 const TypeInt* n_t = _igvn.type(n)->is_int();
3303 if (filtered_t != nullptr) {
3304 n_t = n_t->join(filtered_t)->is_int();
3305 }
3306 return n_t;
3307 }
3308
3309
3310 //------------------------------filtered_type_from_dominators--------------------------------
3311 // Return a possibly more restrictive type for val based on condition control flow of dominators
3312 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3313 if (val->is_Con()) {
3314 return val->bottom_type()->is_int();
3315 }
3316 uint if_limit = 10; // Max number of dominating if's visited
3317 const TypeInt* rtn_t = nullptr;
3318
3319 if (use_ctrl && use_ctrl != C->top()) {
3320 Node* val_ctrl = get_ctrl(val);
3321 uint val_dom_depth = dom_depth(val_ctrl);
3322 Node* pred = use_ctrl;
3323 uint if_cnt = 0;
3324 while (if_cnt < if_limit) {
3325 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3326 if_cnt++;
3327 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3328 if (if_t != nullptr) {
3329 if (rtn_t == nullptr) {
3330 rtn_t = if_t;
3331 } else {
3332 rtn_t = rtn_t->join(if_t)->is_int();
3333 }
3334 }
3335 }
3336 pred = idom(pred);
3337 if (pred == nullptr || pred == C->top()) {
3338 break;
3339 }
3340 // Stop if going beyond definition block of val
3341 if (dom_depth(pred) < val_dom_depth) {
3342 break;
3343 }
3344 }
3345 }
3346 return rtn_t;
3347 }
3348
3349
3350 //------------------------------dump_spec--------------------------------------
3351 // Dump special per-node info
3352 #ifndef PRODUCT
3353 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3354 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3355 BoolTest bt( test_trip()); // Added this for g++.
3356
3357 st->print("[");
3358 bt.dump_on(st);
3359 st->print("]");
3360 }
3361 st->print(" ");
3362 IfNode::dump_spec(st);
3363 }
3364 #endif
3365
3366 //=============================================================================
3367 //------------------------------is_member--------------------------------------
3368 // Is 'l' a member of 'this'?
3369 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3370 while( l->_nest > _nest ) l = l->_parent;
3371 return l == this;
3372 }
3373
3374 //------------------------------set_nest---------------------------------------
3375 // Set loop tree nesting depth. Accumulate _has_call bits.
3376 int IdealLoopTree::set_nest( uint depth ) {
3377 assert(depth <= SHRT_MAX, "sanity");
3378 _nest = depth;
3379 int bits = _has_call;
3380 if( _child ) bits |= _child->set_nest(depth+1);
3381 if( bits ) _has_call = 1;
3382 if( _next ) bits |= _next ->set_nest(depth );
3383 return bits;
3384 }
3385
3386 //------------------------------split_fall_in----------------------------------
3387 // Split out multiple fall-in edges from the loop header. Move them to a
3388 // private RegionNode before the loop. This becomes the loop landing pad.
3389 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3390 PhaseIterGVN &igvn = phase->_igvn;
3391 uint i;
3392
3393 // Make a new RegionNode to be the landing pad.
3394 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3395 phase->set_loop(landing_pad,_parent);
3396 // If _head was irreducible loop entry, landing_pad may now be too
3397 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3398 // Gather all the fall-in control paths into the landing pad
3399 uint icnt = fall_in_cnt;
3400 uint oreq = _head->req();
3401 for( i = oreq-1; i>0; i-- )
3402 if( !phase->is_member( this, _head->in(i) ) )
3403 landing_pad->set_req(icnt--,_head->in(i));
3404
3405 // Peel off PhiNode edges as well
3406 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3407 Node *oj = _head->fast_out(j);
3408 if( oj->is_Phi() ) {
3409 PhiNode* old_phi = oj->as_Phi();
3410 assert( old_phi->region() == _head, "" );
3411 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3412 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3413 uint icnt = fall_in_cnt;
3414 for( i = oreq-1; i>0; i-- ) {
3415 if( !phase->is_member( this, _head->in(i) ) ) {
3416 p->init_req(icnt--, old_phi->in(i));
3417 // Go ahead and clean out old edges from old phi
3418 old_phi->del_req(i);
3419 }
3420 }
3421 // Search for CSE's here, because ZKM.jar does a lot of
3422 // loop hackery and we need to be a little incremental
3423 // with the CSE to avoid O(N^2) node blow-up.
3424 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3425 if( p2 ) { // Found CSE
3426 p->destruct(&igvn); // Recover useless new node
3427 p = p2; // Use old node
3428 } else {
3429 igvn.register_new_node_with_optimizer(p, old_phi);
3430 }
3431 // Make old Phi refer to new Phi.
3432 old_phi->add_req(p);
3433 // Check for the special case of making the old phi useless and
3434 // disappear it. In JavaGrande I have a case where this useless
3435 // Phi is the loop limit and prevents recognizing a CountedLoop
3436 // which in turn prevents removing an empty loop.
3437 Node *id_old_phi = old_phi->Identity(&igvn);
3438 if( id_old_phi != old_phi ) { // Found a simple identity?
3439 // Note that I cannot call 'replace_node' here, because
3440 // that will yank the edge from old_phi to the Region and
3441 // I'm mid-iteration over the Region's uses.
3442 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3443 Node* use = old_phi->last_out(i);
3444 igvn.rehash_node_delayed(use);
3445 uint uses_found = 0;
3446 for (uint j = 0; j < use->len(); j++) {
3447 if (use->in(j) == old_phi) {
3448 if (j < use->req()) use->set_req (j, id_old_phi);
3449 else use->set_prec(j, id_old_phi);
3450 uses_found++;
3451 }
3452 }
3453 i -= uses_found; // we deleted 1 or more copies of this edge
3454 }
3455 }
3456 igvn._worklist.push(old_phi);
3457 }
3458 }
3459 // Finally clean out the fall-in edges from the RegionNode
3460 for( i = oreq-1; i>0; i-- ) {
3461 if( !phase->is_member( this, _head->in(i) ) ) {
3462 _head->del_req(i);
3463 }
3464 }
3465 igvn.rehash_node_delayed(_head);
3466 // Transform landing pad
3467 igvn.register_new_node_with_optimizer(landing_pad, _head);
3468 // Insert landing pad into the header
3469 _head->add_req(landing_pad);
3470 }
3471
3472 //------------------------------split_outer_loop-------------------------------
3473 // Split out the outermost loop from this shared header.
3474 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3475 PhaseIterGVN &igvn = phase->_igvn;
3476
3477 // Find index of outermost loop; it should also be my tail.
3478 uint outer_idx = 1;
3479 while( _head->in(outer_idx) != _tail ) outer_idx++;
3480
3481 // Make a LoopNode for the outermost loop.
3482 Node *ctl = _head->in(LoopNode::EntryControl);
3483 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3484 outer = igvn.register_new_node_with_optimizer(outer, _head);
3485 phase->set_created_loop_node();
3486
3487 // Outermost loop falls into '_head' loop
3488 _head->set_req(LoopNode::EntryControl, outer);
3489 _head->del_req(outer_idx);
3490 // Split all the Phis up between '_head' loop and 'outer' loop.
3491 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3492 Node *out = _head->fast_out(j);
3493 if( out->is_Phi() ) {
3494 PhiNode *old_phi = out->as_Phi();
3495 assert( old_phi->region() == _head, "" );
3496 Node *phi = PhiNode::make_blank(outer, old_phi);
3497 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3498 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3499 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3500 // Make old Phi point to new Phi on the fall-in path
3501 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3502 old_phi->del_req(outer_idx);
3503 }
3504 }
3505
3506 // Use the new loop head instead of the old shared one
3507 _head = outer;
3508 phase->set_loop(_head, this);
3509 }
3510
3511 //------------------------------fix_parent-------------------------------------
3512 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3513 loop->_parent = parent;
3514 if( loop->_child ) fix_parent( loop->_child, loop );
3515 if( loop->_next ) fix_parent( loop->_next , parent );
3516 }
3517
3518 //------------------------------estimate_path_freq-----------------------------
3519 static float estimate_path_freq( Node *n ) {
3520 // Try to extract some path frequency info
3521 IfNode *iff;
3522 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3523 uint nop = n->Opcode();
3524 if( nop == Op_SafePoint ) { // Skip any safepoint
3525 n = n->in(0);
3526 continue;
3527 }
3528 if( nop == Op_CatchProj ) { // Get count from a prior call
3529 // Assume call does not always throw exceptions: means the call-site
3530 // count is also the frequency of the fall-through path.
3531 assert( n->is_CatchProj(), "" );
3532 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3533 return 0.0f; // Assume call exception path is rare
3534 Node *call = n->in(0)->in(0)->in(0);
3535 assert( call->is_Call(), "expect a call here" );
3536 const JVMState *jvms = ((CallNode*)call)->jvms();
3537 ciMethodData* methodData = jvms->method()->method_data();
3538 if (!methodData->is_mature()) return 0.0f; // No call-site data
3539 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3540 if ((data == nullptr) || !data->is_CounterData()) {
3541 // no call profile available, try call's control input
3542 n = n->in(0);
3543 continue;
3544 }
3545 return data->as_CounterData()->count()/FreqCountInvocations;
3546 }
3547 // See if there's a gating IF test
3548 Node *n_c = n->in(0);
3549 if( !n_c->is_If() ) break; // No estimate available
3550 iff = n_c->as_If();
3551 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3552 // Compute how much count comes on this path
3553 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3554 // Have no count info. Skip dull uncommon-trap like branches.
3555 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3556 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3557 break;
3558 // Skip through never-taken branch; look for a real loop exit.
3559 n = iff->in(0);
3560 }
3561 return 0.0f; // No estimate available
3562 }
3563
3564 //------------------------------merge_many_backedges---------------------------
3565 // Merge all the backedges from the shared header into a private Region.
3566 // Feed that region as the one backedge to this loop.
3567 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3568 uint i;
3569
3570 // Scan for the top 2 hottest backedges
3571 float hotcnt = 0.0f;
3572 float warmcnt = 0.0f;
3573 uint hot_idx = 0;
3574 // Loop starts at 2 because slot 1 is the fall-in path
3575 for( i = 2; i < _head->req(); i++ ) {
3576 float cnt = estimate_path_freq(_head->in(i));
3577 if( cnt > hotcnt ) { // Grab hottest path
3578 warmcnt = hotcnt;
3579 hotcnt = cnt;
3580 hot_idx = i;
3581 } else if( cnt > warmcnt ) { // And 2nd hottest path
3582 warmcnt = cnt;
3583 }
3584 }
3585
3586 // See if the hottest backedge is worthy of being an inner loop
3587 // by being much hotter than the next hottest backedge.
3588 if( hotcnt <= 0.0001 ||
3589 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3590
3591 // Peel out the backedges into a private merge point; peel
3592 // them all except optionally hot_idx.
3593 PhaseIterGVN &igvn = phase->_igvn;
3594
3595 Node *hot_tail = nullptr;
3596 // Make a Region for the merge point
3597 Node *r = new RegionNode(1);
3598 for( i = 2; i < _head->req(); i++ ) {
3599 if( i != hot_idx )
3600 r->add_req( _head->in(i) );
3601 else hot_tail = _head->in(i);
3602 }
3603 igvn.register_new_node_with_optimizer(r, _head);
3604 // Plug region into end of loop _head, followed by hot_tail
3605 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3606 igvn.replace_input_of(_head, 2, r);
3607 if( hot_idx ) _head->add_req(hot_tail);
3608
3609 // Split all the Phis up between '_head' loop and the Region 'r'
3610 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3611 Node *out = _head->fast_out(j);
3612 if( out->is_Phi() ) {
3613 PhiNode* n = out->as_Phi();
3614 igvn.hash_delete(n); // Delete from hash before hacking edges
3615 Node *hot_phi = nullptr;
3616 Node *phi = new PhiNode(r, n->type(), n->adr_type());
3617 // Check all inputs for the ones to peel out
3618 uint j = 1;
3619 for( uint i = 2; i < n->req(); i++ ) {
3620 if( i != hot_idx )
3621 phi->set_req( j++, n->in(i) );
3622 else hot_phi = n->in(i);
3623 }
3624 // Register the phi but do not transform until whole place transforms
3625 igvn.register_new_node_with_optimizer(phi, n);
3626 // Add the merge phi to the old Phi
3627 while( n->req() > 3 ) n->del_req( n->req()-1 );
3628 igvn.replace_input_of(n, 2, phi);
3629 if( hot_idx ) n->add_req(hot_phi);
3630 }
3631 }
3632
3633
3634 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
3635 // of self loop tree. Turn self into a loop headed by _head and with
3636 // tail being the new merge point.
3637 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
3638 phase->set_loop(_tail,ilt); // Adjust tail
3639 _tail = r; // Self's tail is new merge point
3640 phase->set_loop(r,this);
3641 ilt->_child = _child; // New guy has my children
3642 _child = ilt; // Self has new guy as only child
3643 ilt->_parent = this; // new guy has self for parent
3644 ilt->_nest = _nest; // Same nesting depth (for now)
3645
3646 // Starting with 'ilt', look for child loop trees using the same shared
3647 // header. Flatten these out; they will no longer be loops in the end.
3648 IdealLoopTree **pilt = &_child;
3649 while( ilt ) {
3650 if( ilt->_head == _head ) {
3651 uint i;
3652 for( i = 2; i < _head->req(); i++ )
3653 if( _head->in(i) == ilt->_tail )
3654 break; // Still a loop
3655 if( i == _head->req() ) { // No longer a loop
3656 // Flatten ilt. Hang ilt's "_next" list from the end of
3657 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
3658 IdealLoopTree **cp = &ilt->_child;
3659 while( *cp ) cp = &(*cp)->_next; // Find end of child list
3660 *cp = ilt->_next; // Hang next list at end of child list
3661 *pilt = ilt->_child; // Move child up to replace ilt
3662 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
3663 ilt = ilt->_child; // Repeat using new ilt
3664 continue; // do not advance over ilt->_child
3665 }
3666 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
3667 phase->set_loop(_head,ilt);
3668 }
3669 pilt = &ilt->_child; // Advance to next
3670 ilt = *pilt;
3671 }
3672
3673 if( _child ) fix_parent( _child, this );
3674 }
3675
3676 //------------------------------beautify_loops---------------------------------
3677 // Split shared headers and insert loop landing pads.
3678 // Insert a LoopNode to replace the RegionNode.
3679 // Return TRUE if loop tree is structurally changed.
3680 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
3681 bool result = false;
3682 // Cache parts in locals for easy
3683 PhaseIterGVN &igvn = phase->_igvn;
3684
3685 igvn.hash_delete(_head); // Yank from hash before hacking edges
3686
3687 // Check for multiple fall-in paths. Peel off a landing pad if need be.
3688 int fall_in_cnt = 0;
3689 for( uint i = 1; i < _head->req(); i++ )
3690 if( !phase->is_member( this, _head->in(i) ) )
3691 fall_in_cnt++;
3692 assert( fall_in_cnt, "at least 1 fall-in path" );
3693 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
3694 split_fall_in( phase, fall_in_cnt );
3695
3696 // Swap inputs to the _head and all Phis to move the fall-in edge to
3697 // the left.
3698 fall_in_cnt = 1;
3699 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
3700 fall_in_cnt++;
3701 if( fall_in_cnt > 1 ) {
3702 // Since I am just swapping inputs I do not need to update def-use info
3703 Node *tmp = _head->in(1);
3704 igvn.rehash_node_delayed(_head);
3705 _head->set_req( 1, _head->in(fall_in_cnt) );
3706 _head->set_req( fall_in_cnt, tmp );
3707 // Swap also all Phis
3708 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
3709 Node* phi = _head->fast_out(i);
3710 if( phi->is_Phi() ) {
3711 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
3712 tmp = phi->in(1);
3713 phi->set_req( 1, phi->in(fall_in_cnt) );
3714 phi->set_req( fall_in_cnt, tmp );
3715 }
3716 }
3717 }
3718 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
3719 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
3720
3721 // If I am a shared header (multiple backedges), peel off the many
3722 // backedges into a private merge point and use the merge point as
3723 // the one true backedge.
3724 if (_head->req() > 3) {
3725 // Merge the many backedges into a single backedge but leave
3726 // the hottest backedge as separate edge for the following peel.
3727 if (!_irreducible) {
3728 merge_many_backedges( phase );
3729 }
3730
3731 // When recursively beautify my children, split_fall_in can change
3732 // loop tree structure when I am an irreducible loop. Then the head
3733 // of my children has a req() not bigger than 3. Here we need to set
3734 // result to true to catch that case in order to tell the caller to
3735 // rebuild loop tree. See issue JDK-8244407 for details.
3736 result = true;
3737 }
3738
3739 // If I have one hot backedge, peel off myself loop.
3740 // I better be the outermost loop.
3741 if (_head->req() > 3 && !_irreducible) {
3742 split_outer_loop( phase );
3743 result = true;
3744
3745 } else if (!_head->is_Loop() && !_irreducible) {
3746 // Make a new LoopNode to replace the old loop head
3747 Node *l = new LoopNode( _head->in(1), _head->in(2) );
3748 l = igvn.register_new_node_with_optimizer(l, _head);
3749 phase->set_created_loop_node();
3750 // Go ahead and replace _head
3751 phase->_igvn.replace_node( _head, l );
3752 _head = l;
3753 phase->set_loop(_head, this);
3754 }
3755
3756 // Now recursively beautify nested loops
3757 if( _child ) result |= _child->beautify_loops( phase );
3758 if( _next ) result |= _next ->beautify_loops( phase );
3759 return result;
3760 }
3761
3762 //------------------------------allpaths_check_safepts----------------------------
3763 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
3764 // safepoint encountered. Helper for check_safepts.
3765 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
3766 assert(stack.size() == 0, "empty stack");
3767 stack.push(_head);
3768 visited.clear();
3769 visited.set(_head->_idx);
3770 while (stack.size() > 0) {
3771 Node* n = stack.pop();
3772 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3773 // Terminate this path
3774 } else if (n->Opcode() == Op_SafePoint) {
3775 if (_phase->get_loop(n) != this) {
3776 if (_required_safept == nullptr) _required_safept = new Node_List();
3777 // save the first we run into on that path: closest to the tail if the head has a single backedge
3778 _required_safept->push(n);
3779 }
3780 // Terminate this path
3781 } else {
3782 uint start = n->is_Region() ? 1 : 0;
3783 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
3784 for (uint i = start; i < end; i++) {
3785 Node* in = n->in(i);
3786 assert(in->is_CFG(), "must be");
3787 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
3788 stack.push(in);
3789 }
3790 }
3791 }
3792 }
3793 }
3794
3795 //------------------------------check_safepts----------------------------
3796 // Given dominators, try to find loops with calls that must always be
3797 // executed (call dominates loop tail). These loops do not need non-call
3798 // safepoints (ncsfpt).
3799 //
3800 // A complication is that a safepoint in a inner loop may be needed
3801 // by an outer loop. In the following, the inner loop sees it has a
3802 // call (block 3) on every path from the head (block 2) to the
3803 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
3804 // in block 2, _but_ this leaves the outer loop without a safepoint.
3805 //
3806 // entry 0
3807 // |
3808 // v
3809 // outer 1,2 +->1
3810 // | |
3811 // | v
3812 // | 2<---+ ncsfpt in 2
3813 // |_/|\ |
3814 // | v |
3815 // inner 2,3 / 3 | call in 3
3816 // / | |
3817 // v +--+
3818 // exit 4
3819 //
3820 //
3821 // This method creates a list (_required_safept) of ncsfpt nodes that must
3822 // be protected is created for each loop. When a ncsfpt maybe deleted, it
3823 // is first looked for in the lists for the outer loops of the current loop.
3824 //
3825 // The insights into the problem:
3826 // A) counted loops are okay
3827 // B) innermost loops are okay (only an inner loop can delete
3828 // a ncsfpt needed by an outer loop)
3829 // C) a loop is immune from an inner loop deleting a safepoint
3830 // if the loop has a call on the idom-path
3831 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
3832 // idom-path that is not in a nested loop
3833 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
3834 // loop needs to be prevented from deletion by an inner loop
3835 //
3836 // There are two analyses:
3837 // 1) The first, and cheaper one, scans the loop body from
3838 // tail to head following the idom (immediate dominator)
3839 // chain, looking for the cases (C,D,E) above.
3840 // Since inner loops are scanned before outer loops, there is summary
3841 // information about inner loops. Inner loops can be skipped over
3842 // when the tail of an inner loop is encountered.
3843 //
3844 // 2) The second, invoked if the first fails to find a call or ncsfpt on
3845 // the idom path (which is rare), scans all predecessor control paths
3846 // from the tail to the head, terminating a path when a call or sfpt
3847 // is encountered, to find the ncsfpt's that are closest to the tail.
3848 //
3849 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
3850 // Bottom up traversal
3851 IdealLoopTree* ch = _child;
3852 if (_child) _child->check_safepts(visited, stack);
3853 if (_next) _next ->check_safepts(visited, stack);
3854
3855 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
3856 bool has_call = false; // call on dom-path
3857 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
3858 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
3859 if (!_irreducible) {
3860 // Scan the dom-path nodes from tail to head
3861 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
3862 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3863 has_call = true;
3864 _has_sfpt = 1; // Then no need for a safept!
3865 break;
3866 } else if (n->Opcode() == Op_SafePoint) {
3867 if (_phase->get_loop(n) == this) {
3868 has_local_ncsfpt = true;
3869 break;
3870 }
3871 if (nonlocal_ncsfpt == nullptr) {
3872 nonlocal_ncsfpt = n; // save the one closest to the tail
3873 }
3874 } else {
3875 IdealLoopTree* nlpt = _phase->get_loop(n);
3876 if (this != nlpt) {
3877 // If at an inner loop tail, see if the inner loop has already
3878 // recorded seeing a call on the dom-path (and stop.) If not,
3879 // jump to the head of the inner loop.
3880 assert(is_member(nlpt), "nested loop");
3881 Node* tail = nlpt->_tail;
3882 if (tail->in(0)->is_If()) tail = tail->in(0);
3883 if (n == tail) {
3884 // If inner loop has call on dom-path, so does outer loop
3885 if (nlpt->_has_sfpt) {
3886 has_call = true;
3887 _has_sfpt = 1;
3888 break;
3889 }
3890 // Skip to head of inner loop
3891 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
3892 n = nlpt->_head;
3893 if (_head == n) {
3894 // this and nlpt (inner loop) have the same loop head. This should not happen because
3895 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
3896 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
3897 // but then attached in the build_loop_tree afterwards, and so still have unmerged
3898 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
3899 // since we have reached the loop head of this.
3900 assert(_head->as_Region()->is_in_infinite_subgraph(),
3901 "only expect unmerged backedges in infinite loops");
3902 break;
3903 }
3904 }
3905 }
3906 }
3907 }
3908 }
3909 // Record safept's that this loop needs preserved when an
3910 // inner loop attempts to delete it's safepoints.
3911 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
3912 if (nonlocal_ncsfpt != nullptr) {
3913 if (_required_safept == nullptr) _required_safept = new Node_List();
3914 _required_safept->push(nonlocal_ncsfpt);
3915 } else {
3916 // Failed to find a suitable safept on the dom-path. Now use
3917 // an all paths walk from tail to head, looking for safepoints to preserve.
3918 allpaths_check_safepts(visited, stack);
3919 }
3920 }
3921 }
3922 }
3923
3924 //---------------------------is_deleteable_safept----------------------------
3925 // Is safept not required by an outer loop?
3926 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
3927 assert(sfpt->Opcode() == Op_SafePoint, "");
3928 IdealLoopTree* lp = get_loop(sfpt)->_parent;
3929 while (lp != nullptr) {
3930 Node_List* sfpts = lp->_required_safept;
3931 if (sfpts != nullptr) {
3932 for (uint i = 0; i < sfpts->size(); i++) {
3933 if (sfpt == sfpts->at(i))
3934 return false;
3935 }
3936 }
3937 lp = lp->_parent;
3938 }
3939 return true;
3940 }
3941
3942 //---------------------------replace_parallel_iv-------------------------------
3943 // Replace parallel induction variable (parallel to trip counter)
3944 // This optimization looks for patterns similar to:
3945 //
3946 // int a = init2;
3947 // for (int iv = init; iv < limit; iv += stride_con) {
3948 // a += stride_con2;
3949 // }
3950 //
3951 // and transforms it to:
3952 //
3953 // int iv2 = init2
3954 // int iv = init
3955 // loop:
3956 // if (iv >= limit) goto exit
3957 // iv += stride_con
3958 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
3959 // goto loop
3960 // exit:
3961 // ...
3962 //
3963 // Such transformation introduces more optimization opportunities. In this
3964 // particular example, the loop can be eliminated entirely given that
3965 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
3966 // place to only perform this optimization if such a division is exact. This
3967 // example will be transformed into its semantic equivalence:
3968 //
3969 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
3970 //
3971 // which corresponds to the structure of transformed subgraph.
3972 //
3973 // However, if there is a mismatch between types of the loop and the parallel
3974 // induction variable (e.g., a long-typed IV in an int-typed loop), type
3975 // conversions are required:
3976 //
3977 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
3978 //
3979 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
3980 assert(loop->_head->is_CountedLoop(), "");
3981 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3982 if (!cl->is_valid_counted_loop(T_INT)) {
3983 return; // skip malformed counted loop
3984 }
3985 Node *incr = cl->incr();
3986 if (incr == nullptr) {
3987 return; // Dead loop?
3988 }
3989 Node *init = cl->init_trip();
3990 Node *phi = cl->phi();
3991 jlong stride_con = cl->stride_con();
3992
3993 // Visit all children, looking for Phis
3994 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
3995 Node *out = cl->out(i);
3996 // Look for other phis (secondary IVs). Skip dead ones
3997 if (!out->is_Phi() || out == phi || !has_node(out)) {
3998 continue;
3999 }
4000
4001 PhiNode* phi2 = out->as_Phi();
4002 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4003 // Look for induction variables of the form: X += constant
4004 if (phi2->region() != loop->_head ||
4005 incr2->req() != 3 ||
4006 incr2->in(1)->uncast() != phi2 ||
4007 incr2 == incr ||
4008 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4009 !incr2->in(2)->is_Con()) {
4010 continue;
4011 }
4012
4013 if (incr2->in(1)->is_ConstraintCast() &&
4014 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4015 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4016 continue;
4017 }
4018 // Check for parallel induction variable (parallel to trip counter)
4019 // via an affine function. In particular, count-down loops with
4020 // count-up array indices are common. We only RCE references off
4021 // the trip-counter, so we need to convert all these to trip-counter
4022 // expressions.
4023 Node* init2 = phi2->in(LoopNode::EntryControl);
4024
4025 // Determine the basic type of the stride constant (and the iv being incremented).
4026 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4027 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4028
4029 // The ratio of the two strides cannot be represented as an int
4030 // if stride_con2 is min_jint (or min_jlong, respectively) and
4031 // stride_con is -1.
4032 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4033 continue;
4034 }
4035
4036 // The general case here gets a little tricky. We want to find the
4037 // GCD of all possible parallel IV's and make a new IV using this
4038 // GCD for the loop. Then all possible IVs are simple multiples of
4039 // the GCD. In practice, this will cover very few extra loops.
4040 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4041 // where +/-1 is the common case, but other integer multiples are
4042 // also easy to handle.
4043 jlong ratio_con = stride_con2 / stride_con;
4044
4045 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4046 continue;
4047 }
4048
4049 #ifndef PRODUCT
4050 if (TraceLoopOpts) {
4051 tty->print("Parallel IV: %d ", phi2->_idx);
4052 loop->dump_head();
4053 }
4054 #endif
4055
4056 // Convert to using the trip counter. The parallel induction
4057 // variable differs from the trip counter by a loop-invariant
4058 // amount, the difference between their respective initial values.
4059 // It is scaled by the 'ratio_con'.
4060 Node* ratio = integercon(ratio_con, stride_con2_bt);
4061
4062 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4063 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4064
4065 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4066 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4067 set_early_ctrl(ratio_init, false);
4068
4069 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4070 _igvn.register_new_node_with_optimizer(diff, init2);
4071 set_early_ctrl(diff, false);
4072
4073 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4074 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4075 set_ctrl(ratio_idx, cl);
4076
4077 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4078 _igvn.register_new_node_with_optimizer(add);
4079 set_ctrl(add, cl);
4080
4081 _igvn.replace_node( phi2, add );
4082 // Sometimes an induction variable is unused
4083 if (add->outcnt() == 0) {
4084 _igvn.remove_dead_node(add);
4085 }
4086 --i; // deleted this phi; rescan starting with next position
4087 }
4088 }
4089
4090 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4091 BasicType source = _igvn.type(input)->basic_type();
4092 if (source == target) {
4093 return input;
4094 }
4095
4096 Node* converted = ConvertNode::create_convert(source, target, input);
4097 _igvn.register_new_node_with_optimizer(converted, input);
4098 set_early_ctrl(converted, false);
4099
4100 return converted;
4101 }
4102
4103 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4104 Node* keep = nullptr;
4105 if (keep_one) {
4106 // Look for a safepoint on the idom-path.
4107 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4108 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4109 keep = i;
4110 break; // Found one
4111 }
4112 }
4113 }
4114
4115 // Don't remove any safepoints if it is requested to keep a single safepoint and
4116 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4117 // in this case since there's no safepoint dominating all paths in the loop body.
4118 bool prune = !keep_one || keep != nullptr;
4119
4120 // Delete other safepoints in this loop.
4121 Node_List* sfpts = _safepts;
4122 if (prune && sfpts != nullptr) {
4123 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4124 for (uint i = 0; i < sfpts->size(); i++) {
4125 Node* n = sfpts->at(i);
4126 assert(phase->get_loop(n) == this, "");
4127 if (n != keep && phase->is_deleteable_safept(n)) {
4128 phase->lazy_replace(n, n->in(TypeFunc::Control));
4129 }
4130 }
4131 }
4132 }
4133
4134 //------------------------------counted_loop-----------------------------------
4135 // Convert to counted loops where possible
4136 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4137
4138 // For grins, set the inner-loop flag here
4139 if (!_child) {
4140 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4141 }
4142
4143 IdealLoopTree* loop = this;
4144 if (_head->is_CountedLoop() ||
4145 phase->is_counted_loop(_head, loop, T_INT)) {
4146
4147 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4148 // Indicate we do not need a safepoint here
4149 _has_sfpt = 1;
4150 }
4151
4152 // Remove safepoints
4153 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4154 remove_safepoints(phase, keep_one_sfpt);
4155
4156 // Look for induction variables
4157 phase->replace_parallel_iv(this);
4158 } else if (_head->is_LongCountedLoop() ||
4159 phase->is_counted_loop(_head, loop, T_LONG)) {
4160 remove_safepoints(phase, true);
4161 } else {
4162 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4163 if (_parent != nullptr && !_irreducible) {
4164 // Not a counted loop. Keep one safepoint.
4165 bool keep_one_sfpt = true;
4166 remove_safepoints(phase, keep_one_sfpt);
4167 }
4168 }
4169
4170 // Recursively
4171 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4172 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4173 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4174 if (loop->_next) loop->_next ->counted_loop(phase);
4175 }
4176
4177
4178 // The Estimated Loop Clone Size:
4179 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4180 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4181 // respectively, used to ensure that the node usage estimates made are on the
4182 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4183 // possible additional/excessive nodes generated due to data and control flow
4184 // merging, for edges reaching outside the loop.
4185 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4186
4187 precond(0 < factor && factor < 16);
4188
4189 uint const bc = 13;
4190 uint const cc = 17;
4191 uint const sz = _body.size() + (_body.size() + 7) / 2;
4192 uint estimate = factor * (sz + bc) + cc;
4193
4194 assert((estimate - cc) / factor == sz + bc, "overflow");
4195
4196 return estimate + est_loop_flow_merge_sz();
4197 }
4198
4199 // The Estimated Loop (full-) Unroll Size:
4200 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4201 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4202 // node usage estimates made are on the safe side, for the most part. This is
4203 // a "light" version of the loop clone size calculation (above), based on the
4204 // assumption that most of the loop-construct overhead will be unraveled when
4205 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4206 // including an overflow check and returning UINT_MAX in case of an overflow.
4207 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4208
4209 precond(factor > 0);
4210
4211 // Take into account that after unroll conjoined heads and tails will fold.
4212 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4213 uint const cc = 7;
4214 uint const sz = b0 + (b0 + 15) / 16;
4215 uint estimate = factor * sz + cc;
4216
4217 if ((estimate - cc) / factor != sz) {
4218 return UINT_MAX;
4219 }
4220
4221 return estimate + est_loop_flow_merge_sz();
4222 }
4223
4224 // Estimate the growth effect (in nodes) of merging control and data flow when
4225 // cloning a loop body, based on the amount of control and data flow reaching
4226 // outside of the (current) loop body.
4227 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4228
4229 uint ctrl_edge_out_cnt = 0;
4230 uint data_edge_out_cnt = 0;
4231
4232 for (uint i = 0; i < _body.size(); i++) {
4233 Node* node = _body.at(i);
4234 uint outcnt = node->outcnt();
4235
4236 for (uint k = 0; k < outcnt; k++) {
4237 Node* out = node->raw_out(k);
4238 if (out == nullptr) continue;
4239 if (out->is_CFG()) {
4240 if (!is_member(_phase->get_loop(out))) {
4241 ctrl_edge_out_cnt++;
4242 }
4243 } else if (_phase->has_ctrl(out)) {
4244 Node* ctrl = _phase->get_ctrl(out);
4245 assert(ctrl != nullptr, "must be");
4246 assert(ctrl->is_CFG(), "must be");
4247 if (!is_member(_phase->get_loop(ctrl))) {
4248 data_edge_out_cnt++;
4249 }
4250 }
4251 }
4252 }
4253 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4254 // a rather pessimistic estimate for the most part, in particular for some
4255 // complex loops, but still not enough to capture all loops.
4256 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4257 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4258 }
4259 return 0;
4260 }
4261
4262 #ifndef PRODUCT
4263 //------------------------------dump_head--------------------------------------
4264 // Dump 1 liner for loop header info
4265 void IdealLoopTree::dump_head() {
4266 tty->sp(2 * _nest);
4267 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4268 if (_irreducible) tty->print(" IRREDUCIBLE");
4269 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4270 : _head->in(LoopNode::EntryControl);
4271 const Predicates predicates(entry);
4272 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4273 tty->print(" limit_check");
4274 }
4275 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4276 tty->print(" profile_predicated");
4277 }
4278 if (UseLoopPredicate && predicates.loop_predicate_block()->is_non_empty()) {
4279 tty->print(" predicated");
4280 }
4281 if (_head->is_CountedLoop()) {
4282 CountedLoopNode *cl = _head->as_CountedLoop();
4283 tty->print(" counted");
4284
4285 Node* init_n = cl->init_trip();
4286 if (init_n != nullptr && init_n->is_Con())
4287 tty->print(" [%d,", cl->init_trip()->get_int());
4288 else
4289 tty->print(" [int,");
4290 Node* limit_n = cl->limit();
4291 if (limit_n != nullptr && limit_n->is_Con())
4292 tty->print("%d),", cl->limit()->get_int());
4293 else
4294 tty->print("int),");
4295 int stride_con = cl->stride_con();
4296 if (stride_con > 0) tty->print("+");
4297 tty->print("%d", stride_con);
4298
4299 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4300
4301 if (cl->is_pre_loop ()) tty->print(" pre" );
4302 if (cl->is_main_loop()) tty->print(" main");
4303 if (cl->is_post_loop()) tty->print(" post");
4304 if (cl->is_vectorized_loop()) tty->print(" vector");
4305 if (range_checks_present()) tty->print(" rc ");
4306 }
4307 if (_has_call) tty->print(" has_call");
4308 if (_has_sfpt) tty->print(" has_sfpt");
4309 if (_rce_candidate) tty->print(" rce");
4310 if (_safepts != nullptr && _safepts->size() > 0) {
4311 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4312 }
4313 if (_required_safept != nullptr && _required_safept->size() > 0) {
4314 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4315 }
4316 if (Verbose) {
4317 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4318 }
4319 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4320 tty->print(" strip_mined");
4321 }
4322 tty->cr();
4323 }
4324
4325 //------------------------------dump-------------------------------------------
4326 // Dump loops by loop tree
4327 void IdealLoopTree::dump() {
4328 dump_head();
4329 if (_child) _child->dump();
4330 if (_next) _next ->dump();
4331 }
4332
4333 #endif
4334
4335 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4336 if (loop == root) {
4337 if (loop->_child != nullptr) {
4338 log->begin_head("loop_tree");
4339 log->end_head();
4340 log_loop_tree_helper(root, loop->_child, log);
4341 log->tail("loop_tree");
4342 assert(loop->_next == nullptr, "what?");
4343 }
4344 } else if (loop != nullptr) {
4345 Node* head = loop->_head;
4346 log->begin_head("loop idx='%d'", head->_idx);
4347 if (loop->_irreducible) log->print(" irreducible='1'");
4348 if (head->is_Loop()) {
4349 if (head->as_Loop()->is_inner_loop()) log->print(" inner_loop='1'");
4350 if (head->as_Loop()->is_partial_peel_loop()) log->print(" partial_peel_loop='1'");
4351 } else if (head->is_CountedLoop()) {
4352 CountedLoopNode* cl = head->as_CountedLoop();
4353 if (cl->is_pre_loop()) log->print(" pre_loop='%d'", cl->main_idx());
4354 if (cl->is_main_loop()) log->print(" main_loop='%d'", cl->_idx);
4355 if (cl->is_post_loop()) log->print(" post_loop='%d'", cl->main_idx());
4356 }
4357 log->end_head();
4358 log_loop_tree_helper(root, loop->_child, log);
4359 log->tail("loop");
4360 log_loop_tree_helper(root, loop->_next, log);
4361 }
4362 }
4363
4364 void PhaseIdealLoop::log_loop_tree() {
4365 if (C->log() != nullptr) {
4366 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4367 }
4368 }
4369
4370 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4371 // predicates will be removed during the next round of IGVN.
4372 void PhaseIdealLoop::eliminate_useless_predicates() {
4373 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4374 return; // No predicates left.
4375 }
4376
4377 eliminate_useless_parse_predicates();
4378 eliminate_useless_template_assertion_predicates();
4379 }
4380
4381 // Eliminate all Parse Predicates that do not belong to a loop anymore by marking them useless. These will be removed
4382 // during the next round of IGVN.
4383 void PhaseIdealLoop::eliminate_useless_parse_predicates() {
4384 mark_all_parse_predicates_useless();
4385 if (C->has_loops()) {
4386 mark_loop_associated_parse_predicates_useful();
4387 }
4388 add_useless_parse_predicates_to_igvn_worklist();
4389 }
4390
4391 void PhaseIdealLoop::mark_all_parse_predicates_useless() const {
4392 for (int i = 0; i < C->parse_predicate_count(); i++) {
4393 C->parse_predicate(i)->mark_useless();
4394 }
4395 }
4396
4397 void PhaseIdealLoop::mark_loop_associated_parse_predicates_useful() {
4398 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4399 IdealLoopTree* loop = iterator.current();
4400 if (loop->can_apply_loop_predication()) {
4401 mark_useful_parse_predicates_for_loop(loop);
4402 }
4403 }
4404 }
4405
4406 // This visitor marks all visited Parse Predicates useful.
4407 class ParsePredicateUsefulMarker : public PredicateVisitor {
4408 public:
4409 using PredicateVisitor::visit;
4410
4411 void visit(const ParsePredicate& parse_predicate) override {
4412 parse_predicate.head()->mark_useful();
4413 }
4414 };
4415
4416 void PhaseIdealLoop::mark_useful_parse_predicates_for_loop(IdealLoopTree* loop) {
4417 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4418 const PredicateIterator predicate_iterator(entry);
4419 ParsePredicateUsefulMarker useful_marker;
4420 predicate_iterator.for_each(useful_marker);
4421 }
4422
4423 void PhaseIdealLoop::add_useless_parse_predicates_to_igvn_worklist() {
4424 for (int i = 0; i < C->parse_predicate_count(); i++) {
4425 ParsePredicateNode* parse_predicate_node = C->parse_predicate(i);
4426 if (parse_predicate_node->is_useless()) {
4427 _igvn._worklist.push(parse_predicate_node);
4428 }
4429 }
4430 }
4431
4432
4433 // Eliminate all Template Assertion Predicates that do not belong to their originally associated loop anymore by
4434 // replacing the OpaqueTemplateAssertionPredicate node of the If node with true. These nodes will be removed during the
4435 // next round of IGVN.
4436 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates() {
4437 Unique_Node_List useful_predicates;
4438 if (C->has_loops()) {
4439 collect_useful_template_assertion_predicates(useful_predicates);
4440 }
4441 eliminate_useless_template_assertion_predicates(useful_predicates);
4442 }
4443
4444 void PhaseIdealLoop::collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4445 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4446 IdealLoopTree* loop = iterator.current();
4447 if (loop->can_apply_loop_predication()) {
4448 collect_useful_template_assertion_predicates_for_loop(loop, useful_predicates);
4449 }
4450 }
4451 }
4452
4453 void PhaseIdealLoop::collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop,
4454 Unique_Node_List &useful_predicates) {
4455 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4456 const Predicates predicates(entry);
4457 if (UseProfiledLoopPredicate) {
4458 const PredicateBlock* profiled_loop_predicate_block = predicates.profiled_loop_predicate_block();
4459 if (profiled_loop_predicate_block->has_parse_predicate()) {
4460 ParsePredicateSuccessProj* parse_predicate_proj = profiled_loop_predicate_block->parse_predicate_success_proj();
4461 get_opaque_template_assertion_predicate_nodes(parse_predicate_proj, useful_predicates);
4462 }
4463 }
4464
4465 if (UseLoopPredicate) {
4466 const PredicateBlock* loop_predicate_block = predicates.loop_predicate_block();
4467 if (loop_predicate_block->has_parse_predicate()) {
4468 ParsePredicateSuccessProj* parse_predicate_proj = loop_predicate_block->parse_predicate_success_proj();
4469 get_opaque_template_assertion_predicate_nodes(parse_predicate_proj, useful_predicates);
4470 }
4471 }
4472 }
4473
4474 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4475 for (int i = C->template_assertion_predicate_count(); i > 0; i--) {
4476 OpaqueTemplateAssertionPredicateNode* opaque_node =
4477 C->template_assertion_predicate_opaq_node(i - 1)->as_OpaqueTemplateAssertionPredicate();
4478 if (!useful_predicates.member(opaque_node)) { // not in the useful list
4479 ConINode* one = intcon(1);
4480 _igvn.replace_node(opaque_node, one);
4481 }
4482 }
4483 }
4484
4485 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4486 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4487 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4488 return;
4489 }
4490 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4491 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4492 IdealLoopTree* lpt = iter.current();
4493 if (lpt->_child == nullptr && lpt->is_counted()) {
4494 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4495 Node* opaque = head->is_canonical_loop_entry();
4496 if (opaque != nullptr) {
4497 useful_zero_trip_guard_opaques_nodes.push(opaque);
4498 }
4499 }
4500 }
4501 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4502 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4503 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4504 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4505 IfNode* iff = opaque->if_node();
4506 IdealLoopTree* loop = get_loop(iff);
4507 while (loop != _ltree_root && loop != nullptr) {
4508 loop = loop->_parent;
4509 }
4510 if (loop == nullptr) {
4511 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4512 // We can't tell if the opaque node is useful or not
4513 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4514 } else {
4515 assert(guarded_loop == nullptr, "");
4516 this->_igvn.replace_node(opaque, opaque->in(1));
4517 }
4518 } else {
4519 assert(guarded_loop != nullptr, "");
4520 }
4521 }
4522 }
4523
4524 //------------------------process_expensive_nodes-----------------------------
4525 // Expensive nodes have their control input set to prevent the GVN
4526 // from commoning them and as a result forcing the resulting node to
4527 // be in a more frequent path. Use CFG information here, to change the
4528 // control inputs so that some expensive nodes can be commoned while
4529 // not executed more frequently.
4530 bool PhaseIdealLoop::process_expensive_nodes() {
4531 assert(OptimizeExpensiveOps, "optimization off?");
4532
4533 // Sort nodes to bring similar nodes together
4534 C->sort_expensive_nodes();
4535
4536 bool progress = false;
4537
4538 for (int i = 0; i < C->expensive_count(); ) {
4539 Node* n = C->expensive_node(i);
4540 int start = i;
4541 // Find nodes similar to n
4542 i++;
4543 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4544 int end = i;
4545 // And compare them two by two
4546 for (int j = start; j < end; j++) {
4547 Node* n1 = C->expensive_node(j);
4548 if (is_node_unreachable(n1)) {
4549 continue;
4550 }
4551 for (int k = j+1; k < end; k++) {
4552 Node* n2 = C->expensive_node(k);
4553 if (is_node_unreachable(n2)) {
4554 continue;
4555 }
4556
4557 assert(n1 != n2, "should be pair of nodes");
4558
4559 Node* c1 = n1->in(0);
4560 Node* c2 = n2->in(0);
4561
4562 Node* parent_c1 = c1;
4563 Node* parent_c2 = c2;
4564
4565 // The call to get_early_ctrl_for_expensive() moves the
4566 // expensive nodes up but stops at loops that are in a if
4567 // branch. See whether we can exit the loop and move above the
4568 // If.
4569 if (c1->is_Loop()) {
4570 parent_c1 = c1->in(1);
4571 }
4572 if (c2->is_Loop()) {
4573 parent_c2 = c2->in(1);
4574 }
4575
4576 if (parent_c1 == parent_c2) {
4577 _igvn._worklist.push(n1);
4578 _igvn._worklist.push(n2);
4579 continue;
4580 }
4581
4582 // Look for identical expensive node up the dominator chain.
4583 if (is_dominator(c1, c2)) {
4584 c2 = c1;
4585 } else if (is_dominator(c2, c1)) {
4586 c1 = c2;
4587 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4588 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4589 // Both branches have the same expensive node so move it up
4590 // before the if.
4591 c1 = c2 = idom(parent_c1->in(0));
4592 }
4593 // Do the actual moves
4594 if (n1->in(0) != c1) {
4595 _igvn.replace_input_of(n1, 0, c1);
4596 progress = true;
4597 }
4598 if (n2->in(0) != c2) {
4599 _igvn.replace_input_of(n2, 0, c2);
4600 progress = true;
4601 }
4602 }
4603 }
4604 }
4605
4606 return progress;
4607 }
4608
4609 //=============================================================================
4610 //----------------------------build_and_optimize-------------------------------
4611 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4612 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4613 void PhaseIdealLoop::build_and_optimize() {
4614 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4615
4616 bool do_split_ifs = (_mode == LoopOptsDefault);
4617 bool skip_loop_opts = (_mode == LoopOptsNone);
4618 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4619
4620
4621 int old_progress = C->major_progress();
4622 uint orig_worklist_size = _igvn._worklist.size();
4623
4624 // Reset major-progress flag for the driver's heuristics
4625 C->clear_major_progress();
4626
4627 #ifndef PRODUCT
4628 // Capture for later assert
4629 uint unique = C->unique();
4630 _loop_invokes++;
4631 _loop_work += unique;
4632 #endif
4633
4634 // True if the method has at least 1 irreducible loop
4635 _has_irreducible_loops = false;
4636
4637 _created_loop_node = false;
4638
4639 VectorSet visited;
4640 // Pre-grow the mapping from Nodes to IdealLoopTrees.
4641 _loop_or_ctrl.map(C->unique(), nullptr);
4642 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
4643
4644 // Pre-build the top-level outermost loop tree entry
4645 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
4646 // Do not need a safepoint at the top level
4647 _ltree_root->_has_sfpt = 1;
4648
4649 // Initialize Dominators.
4650 // Checked in clone_loop_predicate() during beautify_loops().
4651 _idom_size = 0;
4652 _idom = nullptr;
4653 _dom_depth = nullptr;
4654 _dom_stk = nullptr;
4655
4656 // Empty pre-order array
4657 allocate_preorders();
4658
4659 // Build a loop tree on the fly. Build a mapping from CFG nodes to
4660 // IdealLoopTree entries. Data nodes are NOT walked.
4661 build_loop_tree();
4662 // Check for bailout, and return
4663 if (C->failing()) {
4664 return;
4665 }
4666
4667 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
4668 // is an exception edge, parsing doesn't set has_loops().
4669 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
4670 // No loops after all
4671 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
4672
4673 // There should always be an outer loop containing the Root and Return nodes.
4674 // If not, we have a degenerate empty program. Bail out in this case.
4675 if (!has_node(C->root())) {
4676 if (!_verify_only) {
4677 C->clear_major_progress();
4678 assert(false, "empty program detected during loop optimization");
4679 C->record_method_not_compilable("empty program detected during loop optimization");
4680 }
4681 return;
4682 }
4683
4684 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4685 // Nothing to do, so get out
4686 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
4687 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
4688 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
4689 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
4690 if (stop_early && !do_expensive_nodes) {
4691 return;
4692 }
4693
4694 // Set loop nesting depth
4695 _ltree_root->set_nest( 0 );
4696
4697 // Split shared headers and insert loop landing pads.
4698 // Do not bother doing this on the Root loop of course.
4699 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
4700 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
4701 if( _ltree_root->_child->beautify_loops( this ) ) {
4702 // Re-build loop tree!
4703 _ltree_root->_child = nullptr;
4704 _loop_or_ctrl.clear();
4705 reallocate_preorders();
4706 build_loop_tree();
4707 // Check for bailout, and return
4708 if (C->failing()) {
4709 return;
4710 }
4711 // Reset loop nesting depth
4712 _ltree_root->set_nest( 0 );
4713
4714 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
4715 }
4716 }
4717
4718 // Build Dominators for elision of null checks & loop finding.
4719 // Since nodes do not have a slot for immediate dominator, make
4720 // a persistent side array for that info indexed on node->_idx.
4721 _idom_size = C->unique();
4722 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
4723 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
4724 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
4725 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
4726
4727 Dominators();
4728
4729 if (!_verify_only) {
4730 // As a side effect, Dominators removed any unreachable CFG paths
4731 // into RegionNodes. It doesn't do this test against Root, so
4732 // we do it here.
4733 for( uint i = 1; i < C->root()->req(); i++ ) {
4734 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
4735 _igvn.delete_input_of(C->root(), i);
4736 i--; // Rerun same iteration on compressed edges
4737 }
4738 }
4739
4740 // Given dominators, try to find inner loops with calls that must
4741 // always be executed (call dominates loop tail). These loops do
4742 // not need a separate safepoint.
4743 Node_List cisstack;
4744 _ltree_root->check_safepts(visited, cisstack);
4745 }
4746
4747 // Walk the DATA nodes and place into loops. Find earliest control
4748 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
4749 // holding the associated IdealLoopTree pointer. For DATA nodes, the
4750 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
4751
4752 // Allocate stack with enough space to avoid frequent realloc
4753 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
4754 Node_Stack nstack(stack_size);
4755
4756 visited.clear();
4757 Node_List worklist;
4758 // Don't need C->root() on worklist since
4759 // it will be processed among C->top() inputs
4760 worklist.push(C->top());
4761 visited.set(C->top()->_idx); // Set C->top() as visited now
4762 build_loop_early( visited, worklist, nstack );
4763
4764 // Given early legal placement, try finding counted loops. This placement
4765 // is good enough to discover most loop invariants.
4766 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
4767 _ltree_root->counted_loop( this );
4768 }
4769
4770 // Find latest loop placement. Find ideal loop placement.
4771 visited.clear();
4772 init_dom_lca_tags();
4773 // Need C->root() on worklist when processing outs
4774 worklist.push(C->root());
4775 NOT_PRODUCT( C->verify_graph_edges(); )
4776 worklist.push(C->top());
4777 build_loop_late( visited, worklist, nstack );
4778 if (C->failing()) { return; }
4779
4780 if (_verify_only) {
4781 C->restore_major_progress(old_progress);
4782 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
4783 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
4784 return;
4785 }
4786
4787 // clear out the dead code after build_loop_late
4788 while (_deadlist.size()) {
4789 _igvn.remove_globally_dead_node(_deadlist.pop());
4790 }
4791
4792 eliminate_useless_zero_trip_guard();
4793
4794 if (stop_early) {
4795 assert(do_expensive_nodes, "why are we here?");
4796 if (process_expensive_nodes()) {
4797 // If we made some progress when processing expensive nodes then
4798 // the IGVN may modify the graph in a way that will allow us to
4799 // make some more progress: we need to try processing expensive
4800 // nodes again.
4801 C->set_major_progress();
4802 }
4803 return;
4804 }
4805
4806 // Some parser-inserted loop predicates could never be used by loop
4807 // predication or they were moved away from loop during some optimizations.
4808 // For example, peeling. Eliminate them before next loop optimizations.
4809 eliminate_useless_predicates();
4810
4811 #ifndef PRODUCT
4812 C->verify_graph_edges();
4813 if (_verify_me) { // Nested verify pass?
4814 // Check to see if the verify _mode is broken
4815 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
4816 return;
4817 }
4818 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4819 if (TraceLoopOpts && C->has_loops()) {
4820 _ltree_root->dump();
4821 }
4822 #endif
4823
4824 if (skip_loop_opts) {
4825 C->restore_major_progress(old_progress);
4826 return;
4827 }
4828
4829 if (do_max_unroll) {
4830 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4831 IdealLoopTree* lpt = iter.current();
4832 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
4833 lpt->compute_trip_count(this);
4834 if (!lpt->do_one_iteration_loop(this) &&
4835 !lpt->do_remove_empty_loop(this)) {
4836 AutoNodeBudget node_budget(this);
4837 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
4838 lpt->policy_maximally_unroll(this)) {
4839 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4840 do_maximally_unroll(lpt, worklist);
4841 }
4842 }
4843 }
4844 }
4845
4846 C->restore_major_progress(old_progress);
4847 return;
4848 }
4849
4850 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
4851 return;
4852 }
4853
4854 if (ReassociateInvariants && !C->major_progress()) {
4855 // Reassociate invariants and prep for split_thru_phi
4856 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4857 IdealLoopTree* lpt = iter.current();
4858 if (!lpt->is_loop()) {
4859 continue;
4860 }
4861 Node* head = lpt->_head;
4862 if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
4863
4864 // check for vectorized loops, any reassociation of invariants was already done
4865 if (head->is_CountedLoop()) {
4866 if (head->as_CountedLoop()->is_unroll_only()) {
4867 continue;
4868 } else {
4869 AutoNodeBudget node_budget(this);
4870 lpt->reassociate_invariants(this);
4871 }
4872 }
4873 // Because RCE opportunities can be masked by split_thru_phi,
4874 // look for RCE candidates and inhibit split_thru_phi
4875 // on just their loop-phi's for this pass of loop opts
4876 if (SplitIfBlocks && do_split_ifs &&
4877 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
4878 (lpt->policy_range_check(this, true, T_LONG) ||
4879 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
4880 lpt->_rce_candidate = 1; // = true
4881 }
4882 }
4883 }
4884
4885 // Check for aggressive application of split-if and other transforms
4886 // that require basic-block info (like cloning through Phi's)
4887 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
4888 visited.clear();
4889 split_if_with_blocks(visited, nstack);
4890 if (C->failing()) {
4891 return;
4892 }
4893 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4894 }
4895
4896 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
4897 C->set_major_progress();
4898 }
4899
4900 // Perform loop predication before iteration splitting
4901 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
4902 _ltree_root->_child->loop_predication(this);
4903 }
4904
4905 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
4906 if (do_intrinsify_fill()) {
4907 C->set_major_progress();
4908 }
4909 }
4910
4911 // Perform iteration-splitting on inner loops. Split iterations to avoid
4912 // range checks or one-shot null checks.
4913
4914 // If split-if's didn't hack the graph too bad (no CFG changes)
4915 // then do loop opts.
4916 if (C->has_loops() && !C->major_progress()) {
4917 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4918 _ltree_root->_child->iteration_split( this, worklist );
4919 // No verify after peeling! GCM has hoisted code out of the loop.
4920 // After peeling, the hoisted code could sink inside the peeled area.
4921 // The peeling code does not try to recompute the best location for
4922 // all the code before the peeled area, so the verify pass will always
4923 // complain about it.
4924 }
4925
4926 // Check for bailout, and return
4927 if (C->failing()) {
4928 return;
4929 }
4930
4931 // Do verify graph edges in any case
4932 NOT_PRODUCT( C->verify_graph_edges(); );
4933
4934 if (!do_split_ifs) {
4935 // We saw major progress in Split-If to get here. We forced a
4936 // pass with unrolling and not split-if, however more split-if's
4937 // might make progress. If the unrolling didn't make progress
4938 // then the major-progress flag got cleared and we won't try
4939 // another round of Split-If. In particular the ever-common
4940 // instance-of/check-cast pattern requires at least 2 rounds of
4941 // Split-If to clear out.
4942 C->set_major_progress();
4943 }
4944
4945 // Repeat loop optimizations if new loops were seen
4946 if (created_loop_node()) {
4947 C->set_major_progress();
4948 }
4949
4950 // Keep loop predicates and perform optimizations with them
4951 // until no more loop optimizations could be done.
4952 // After that switch predicates off and do more loop optimizations.
4953 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
4954 C->mark_parse_predicate_nodes_useless(_igvn);
4955 assert(C->parse_predicate_count() == 0, "should be zero now");
4956 if (TraceLoopOpts) {
4957 tty->print_cr("PredicatesOff");
4958 }
4959 C->set_major_progress();
4960 }
4961
4962 // Auto-vectorize main-loop
4963 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
4964 Compile::TracePhase tp(_t_autoVectorize);
4965
4966 // Shared data structures for all AutoVectorizations, to reduce allocations
4967 // of large arrays.
4968 VSharedData vshared;
4969 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4970 IdealLoopTree* lpt = iter.current();
4971 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
4972
4973 if (status == AutoVectorizeStatus::TriedAndFailed) {
4974 // We tried vectorization, but failed. From now on only unroll the loop.
4975 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4976 if (cl->has_passed_slp()) {
4977 C->set_major_progress();
4978 cl->set_notpassed_slp();
4979 cl->mark_do_unroll_only();
4980 }
4981 }
4982 }
4983 }
4984
4985 // Move UnorderedReduction out of counted loop. Can be introduced by AutoVectorization.
4986 if (C->has_loops() && !C->major_progress()) {
4987 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4988 IdealLoopTree* lpt = iter.current();
4989 if (lpt->is_counted() && lpt->is_innermost()) {
4990 move_unordered_reduction_out_of_loop(lpt);
4991 }
4992 }
4993 }
4994 }
4995
4996 #ifndef PRODUCT
4997 //------------------------------print_statistics-------------------------------
4998 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
4999 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5000 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5001 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5002 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5003 void PhaseIdealLoop::print_statistics() {
5004 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);
5005 }
5006 #endif
5007
5008 #ifdef ASSERT
5009 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5010 void PhaseIdealLoop::verify() const {
5011 ResourceMark rm;
5012 int old_progress = C->major_progress();
5013 bool success = true;
5014
5015 PhaseIdealLoop phase_verify(_igvn, this);
5016 if (C->failing_internal()) {
5017 return;
5018 }
5019
5020 // Verify ctrl and idom of every node.
5021 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5022
5023 // Verify loop-tree.
5024 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5025
5026 assert(success, "VerifyLoopOptimizations failed");
5027
5028 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5029 C->restore_major_progress(old_progress);
5030 }
5031
5032 // Perform a BFS starting at n, through all inputs.
5033 // Call verify_idom and verify_node on all nodes of BFS traversal.
5034 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5035 Unique_Node_List worklist;
5036 worklist.push(root);
5037 bool success = true;
5038 for (uint i = 0; i < worklist.size(); i++) {
5039 Node* n = worklist.at(i);
5040 // process node
5041 success &= verify_idom(n, phase_verify);
5042 success &= verify_loop_ctrl(n, phase_verify);
5043 // visit inputs
5044 for (uint j = 0; j < n->req(); j++) {
5045 if (n->in(j) != nullptr) {
5046 worklist.push(n->in(j));
5047 }
5048 }
5049 }
5050 return success;
5051 }
5052
5053 // Verify dominator structure (IDOM).
5054 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5055 // Verify IDOM for all CFG nodes (except root).
5056 if (!n->is_CFG() || n->is_Root()) {
5057 return true; // pass
5058 }
5059
5060 if (n->_idx >= _idom_size) {
5061 tty->print("CFG Node with no idom: ");
5062 n->dump();
5063 return false; // fail
5064 }
5065
5066 Node* id = idom_no_update(n);
5067 Node* id_verify = phase_verify->idom_no_update(n);
5068 if (id != id_verify) {
5069 tty->print("Mismatching idom for node: ");
5070 n->dump();
5071 tty->print(" We have idom: ");
5072 id->dump();
5073 tty->print(" Verify has idom: ");
5074 id_verify->dump();
5075 tty->cr();
5076 return false; // fail
5077 }
5078 return true; // pass
5079 }
5080
5081 // Verify "_loop_or_ctrl": control and loop membership.
5082 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5083 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5084 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5085 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5086 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5087 const uint i = n->_idx;
5088 // The loop-tree was built from def to use (top-down).
5089 // The verification happens from use to def (bottom-up).
5090 // We may thus find nodes during verification that are not in the loop-tree.
5091 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5092 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5093 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5094 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5095 n->dump();
5096 return false; // fail
5097 }
5098 // Not reachable for both.
5099 return true; // pass
5100 }
5101
5102 if (n->is_CFG() == has_ctrl(n)) {
5103 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5104 n->dump();
5105 return false; // fail
5106 }
5107
5108 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5109 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5110 n->dump();
5111 return false; // fail
5112 } else if (has_ctrl(n)) {
5113 assert(phase_verify->has_ctrl(n), "sanity");
5114 // n is a data node.
5115 // Verify that its ctrl is the same.
5116
5117 // Broken part of VerifyLoopOptimizations (A)
5118 // Reason:
5119 // BUG, wrong control set for example in
5120 // PhaseIdealLoop::split_if_with_blocks
5121 // at "set_ctrl(x, new_ctrl);"
5122 /*
5123 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5124 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5125 tty->print("Mismatched control setting for: ");
5126 n->dump();
5127 if( fail++ > 10 ) return;
5128 Node *c = get_ctrl_no_update(n);
5129 tty->print("We have it as: ");
5130 if( c->in(0) ) c->dump();
5131 else tty->print_cr("N%d",c->_idx);
5132 tty->print("Verify thinks: ");
5133 if( loop_verify->has_ctrl(n) )
5134 loop_verify->get_ctrl_no_update(n)->dump();
5135 else
5136 loop_verify->get_loop_idx(n)->dump();
5137 tty->cr();
5138 }
5139 */
5140 return true; // pass
5141 } else {
5142 assert(!phase_verify->has_ctrl(n), "sanity");
5143 // n is a ctrl node.
5144 // Verify that not has_ctrl, and that get_loop_idx is the same.
5145
5146 // Broken part of VerifyLoopOptimizations (B)
5147 // Reason:
5148 // NeverBranch node for example is added to loop outside its scope.
5149 // Once we run build_loop_tree again, it is added to the correct loop.
5150 /*
5151 if (!C->major_progress()) {
5152 // Loop selection can be messed up if we did a major progress
5153 // operation, like split-if. Do not verify in that case.
5154 IdealLoopTree *us = get_loop_idx(n);
5155 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5156 if( us->_head != them->_head || us->_tail != them->_tail ) {
5157 tty->print("Unequals loops for: ");
5158 n->dump();
5159 if( fail++ > 10 ) return;
5160 tty->print("We have it as: ");
5161 us->dump();
5162 tty->print("Verify thinks: ");
5163 them->dump();
5164 tty->cr();
5165 }
5166 }
5167 */
5168 return true; // pass
5169 }
5170 }
5171
5172 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5173 assert(a != nullptr && b != nullptr, "must be");
5174 return a->_head->_idx - b->_head->_idx;
5175 }
5176
5177 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5178 GrowableArray<IdealLoopTree*> children;
5179 IdealLoopTree* child = _child;
5180 while (child != nullptr) {
5181 assert(child->_parent == this, "all must be children of this");
5182 children.insert_sorted<compare_tree>(child);
5183 child = child->_next;
5184 }
5185 return children;
5186 }
5187
5188 // Verify that tree structures match. Because the CFG can change, siblings
5189 // within the loop tree can be reordered. We attempt to deal with that by
5190 // reordering the verify's loop tree if possible.
5191 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5192 assert(_head == loop_verify->_head, "mismatched loop head");
5193 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5194
5195 // Collect the children
5196 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5197 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5198
5199 bool success = true;
5200
5201 // Compare the two children lists
5202 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5203 IdealLoopTree* child = nullptr;
5204 IdealLoopTree* child_verify = nullptr;
5205 // Read from both lists, if possible.
5206 if (i < children.length()) {
5207 child = children.at(i);
5208 }
5209 if (j < children_verify.length()) {
5210 child_verify = children_verify.at(j);
5211 }
5212 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5213 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5214 // We found two non-equal children. Select the smaller one.
5215 if (child->_head->_idx < child_verify->_head->_idx) {
5216 child_verify = nullptr;
5217 } else {
5218 child = nullptr;
5219 }
5220 }
5221 // Process the two children, or potentially log the failure if we only found one.
5222 if (child_verify == nullptr) {
5223 if (child->_irreducible && Compile::current()->major_progress()) {
5224 // Irreducible loops can pick a different header (one of its entries).
5225 } else {
5226 tty->print_cr("We have a loop that verify does not have");
5227 child->dump();
5228 success = false;
5229 }
5230 i++; // step for this
5231 } else if (child == nullptr) {
5232 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5233 // Irreducible loops can pick a different header (one of its entries).
5234 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5235 // Infinite loops do not get attached to the loop-tree on their first visit.
5236 // "this" runs before "loop_verify". It is thus possible that we find the
5237 // infinite loop only for "child_verify". Only finding it with "child" would
5238 // mean that we lost it, which is not ok.
5239 } else {
5240 tty->print_cr("Verify has a loop that we do not have");
5241 child_verify->dump();
5242 success = false;
5243 }
5244 j++; // step for verify
5245 } else {
5246 assert(child->_head == child_verify->_head, "We have both and they are equal");
5247 success &= child->verify_tree(child_verify); // Recursion
5248 i++; // step for this
5249 j++; // step for verify
5250 }
5251 }
5252
5253 // Broken part of VerifyLoopOptimizations (D)
5254 // Reason:
5255 // split_if has to update the _tail, if it is modified. But that is done by
5256 // checking to what loop the iff belongs to. That info can be wrong, and then
5257 // we do not update the _tail correctly.
5258 /*
5259 Node *tail = _tail; // Inline a non-updating version of
5260 while( !tail->in(0) ) // the 'tail()' call.
5261 tail = tail->in(1);
5262 assert( tail == loop->_tail, "mismatched loop tail" );
5263 */
5264
5265 if (_head->is_CountedLoop()) {
5266 CountedLoopNode *cl = _head->as_CountedLoop();
5267
5268 Node* ctrl = cl->init_control();
5269 Node* back = cl->back_control();
5270 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5271 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5272 cl->loopexit(); // assert implied
5273 }
5274
5275 // Broken part of VerifyLoopOptimizations (E)
5276 // Reason:
5277 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5278 // to the loop body. Or maybe they are not added to the correct loop.
5279 // at "Node* x = n->clone();"
5280 /*
5281 // Innermost loops need to verify loop bodies,
5282 // but only if no 'major_progress'
5283 int fail = 0;
5284 if (!Compile::current()->major_progress() && _child == nullptr) {
5285 for( uint i = 0; i < _body.size(); i++ ) {
5286 Node *n = _body.at(i);
5287 if (n->outcnt() == 0) continue; // Ignore dead
5288 uint j;
5289 for( j = 0; j < loop->_body.size(); j++ )
5290 if( loop->_body.at(j) == n )
5291 break;
5292 if( j == loop->_body.size() ) { // Not found in loop body
5293 // Last ditch effort to avoid assertion: Its possible that we
5294 // have some users (so outcnt not zero) but are still dead.
5295 // Try to find from root.
5296 if (Compile::current()->root()->find(n->_idx)) {
5297 fail++;
5298 tty->print("We have that verify does not: ");
5299 n->dump();
5300 }
5301 }
5302 }
5303 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5304 Node *n = loop->_body.at(i2);
5305 if (n->outcnt() == 0) continue; // Ignore dead
5306 uint j;
5307 for( j = 0; j < _body.size(); j++ )
5308 if( _body.at(j) == n )
5309 break;
5310 if( j == _body.size() ) { // Not found in loop body
5311 // Last ditch effort to avoid assertion: Its possible that we
5312 // have some users (so outcnt not zero) but are still dead.
5313 // Try to find from root.
5314 if (Compile::current()->root()->find(n->_idx)) {
5315 fail++;
5316 tty->print("Verify has that we do not: ");
5317 n->dump();
5318 }
5319 }
5320 }
5321 assert( !fail, "loop body mismatch" );
5322 }
5323 */
5324 return success;
5325 }
5326 #endif
5327
5328 //------------------------------set_idom---------------------------------------
5329 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5330 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5331 uint idx = d->_idx;
5332 if (idx >= _idom_size) {
5333 uint newsize = next_power_of_2(idx);
5334 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5335 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5336 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5337 _idom_size = newsize;
5338 }
5339 _idom[idx] = n;
5340 _dom_depth[idx] = dom_depth;
5341 }
5342
5343 //------------------------------recompute_dom_depth---------------------------------------
5344 // The dominator tree is constructed with only parent pointers.
5345 // This recomputes the depth in the tree by first tagging all
5346 // nodes as "no depth yet" marker. The next pass then runs up
5347 // the dom tree from each node marked "no depth yet", and computes
5348 // the depth on the way back down.
5349 void PhaseIdealLoop::recompute_dom_depth() {
5350 uint no_depth_marker = C->unique();
5351 uint i;
5352 // Initialize depth to "no depth yet" and realize all lazy updates
5353 for (i = 0; i < _idom_size; i++) {
5354 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5355 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5356 _dom_depth[i] = no_depth_marker;
5357
5358 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5359 if (_idom[i]->in(0) == nullptr) {
5360 idom(i);
5361 }
5362 }
5363 }
5364 if (_dom_stk == nullptr) {
5365 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5366 if (init_size < 10) init_size = 10;
5367 _dom_stk = new GrowableArray<uint>(init_size);
5368 }
5369 // Compute new depth for each node.
5370 for (i = 0; i < _idom_size; i++) {
5371 uint j = i;
5372 // Run up the dom tree to find a node with a depth
5373 while (_dom_depth[j] == no_depth_marker) {
5374 _dom_stk->push(j);
5375 j = _idom[j]->_idx;
5376 }
5377 // Compute the depth on the way back down this tree branch
5378 uint dd = _dom_depth[j] + 1;
5379 while (_dom_stk->length() > 0) {
5380 uint j = _dom_stk->pop();
5381 _dom_depth[j] = dd;
5382 dd++;
5383 }
5384 }
5385 }
5386
5387 //------------------------------sort-------------------------------------------
5388 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5389 // loop tree, not the root.
5390 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5391 if( !innermost ) return loop; // New innermost loop
5392
5393 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5394 assert( loop_preorder, "not yet post-walked loop" );
5395 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5396 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5397
5398 // Insert at start of list
5399 while( l ) { // Insertion sort based on pre-order
5400 if( l == loop ) return innermost; // Already on list!
5401 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5402 assert( l_preorder, "not yet post-walked l" );
5403 // Check header pre-order number to figure proper nesting
5404 if( loop_preorder > l_preorder )
5405 break; // End of insertion
5406 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5407 // Since I split shared headers, you'd think this could not happen.
5408 // BUT: I must first do the preorder numbering before I can discover I
5409 // have shared headers, so the split headers all get the same preorder
5410 // number as the RegionNode they split from.
5411 if( loop_preorder == l_preorder &&
5412 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5413 break; // Also check for shared headers (same pre#)
5414 pp = &l->_parent; // Chain up list
5415 l = *pp;
5416 }
5417 // Link into list
5418 // Point predecessor to me
5419 *pp = loop;
5420 // Point me to successor
5421 IdealLoopTree *p = loop->_parent;
5422 loop->_parent = l; // Point me to successor
5423 if( p ) sort( p, innermost ); // Insert my parents into list as well
5424 return innermost;
5425 }
5426
5427 //------------------------------build_loop_tree--------------------------------
5428 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5429 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5430 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5431 // tightest enclosing IdealLoopTree for post-walked.
5432 //
5433 // During my forward walk I do a short 1-layer lookahead to see if I can find
5434 // a loop backedge with that doesn't have any work on the backedge. This
5435 // helps me construct nested loops with shared headers better.
5436 //
5437 // Once I've done the forward recursion, I do the post-work. For each child
5438 // I check to see if there is a backedge. Backedges define a loop! I
5439 // insert an IdealLoopTree at the target of the backedge.
5440 //
5441 // During the post-work I also check to see if I have several children
5442 // belonging to different loops. If so, then this Node is a decision point
5443 // where control flow can choose to change loop nests. It is at this
5444 // decision point where I can figure out how loops are nested. At this
5445 // time I can properly order the different loop nests from my children.
5446 // Note that there may not be any backedges at the decision point!
5447 //
5448 // Since the decision point can be far removed from the backedges, I can't
5449 // order my loops at the time I discover them. Thus at the decision point
5450 // I need to inspect loop header pre-order numbers to properly nest my
5451 // loops. This means I need to sort my childrens' loops by pre-order.
5452 // The sort is of size number-of-control-children, which generally limits
5453 // it to size 2 (i.e., I just choose between my 2 target loops).
5454 void PhaseIdealLoop::build_loop_tree() {
5455 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5456 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5457 Node *n = C->root();
5458 bltstack.push(n);
5459 int pre_order = 1;
5460 int stack_size;
5461
5462 while ( ( stack_size = bltstack.length() ) != 0 ) {
5463 n = bltstack.top(); // Leave node on stack
5464 if ( !is_visited(n) ) {
5465 // ---- Pre-pass Work ----
5466 // Pre-walked but not post-walked nodes need a pre_order number.
5467
5468 set_preorder_visited( n, pre_order ); // set as visited
5469
5470 // ---- Scan over children ----
5471 // Scan first over control projections that lead to loop headers.
5472 // This helps us find inner-to-outer loops with shared headers better.
5473
5474 // Scan children's children for loop headers.
5475 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5476 Node* m = n->raw_out(i); // Child
5477 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5478 // Scan over children's children to find loop
5479 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5480 Node* l = m->fast_out(j);
5481 if( is_visited(l) && // Been visited?
5482 !is_postvisited(l) && // But not post-visited
5483 get_preorder(l) < pre_order ) { // And smaller pre-order
5484 // Found! Scan the DFS down this path before doing other paths
5485 bltstack.push(m);
5486 break;
5487 }
5488 }
5489 }
5490 }
5491 pre_order++;
5492 }
5493 else if ( !is_postvisited(n) ) {
5494 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5495 // such as com.sun.rsasign.am::a.
5496 // For non-recursive version, first, process current children.
5497 // On next iteration, check if additional children were added.
5498 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5499 Node* u = n->raw_out(k);
5500 if ( u->is_CFG() && !is_visited(u) ) {
5501 bltstack.push(u);
5502 }
5503 }
5504 if ( bltstack.length() == stack_size ) {
5505 // There were no additional children, post visit node now
5506 (void)bltstack.pop(); // Remove node from stack
5507 pre_order = build_loop_tree_impl(n, pre_order);
5508 // Check for bailout
5509 if (C->failing()) {
5510 return;
5511 }
5512 // Check to grow _preorders[] array for the case when
5513 // build_loop_tree_impl() adds new nodes.
5514 check_grow_preorders();
5515 }
5516 }
5517 else {
5518 (void)bltstack.pop(); // Remove post-visited node from stack
5519 }
5520 }
5521 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5522 }
5523
5524 //------------------------------build_loop_tree_impl---------------------------
5525 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5526 // ---- Post-pass Work ----
5527 // Pre-walked but not post-walked nodes need a pre_order number.
5528
5529 // Tightest enclosing loop for this Node
5530 IdealLoopTree *innermost = nullptr;
5531
5532 // For all children, see if any edge is a backedge. If so, make a loop
5533 // for it. Then find the tightest enclosing loop for the self Node.
5534 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5535 Node* m = n->fast_out(i); // Child
5536 if (n == m) continue; // Ignore control self-cycles
5537 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5538
5539 IdealLoopTree *l; // Child's loop
5540 if (!is_postvisited(m)) { // Child visited but not post-visited?
5541 // Found a backedge
5542 assert(get_preorder(m) < pre_order, "should be backedge");
5543 // Check for the RootNode, which is already a LoopNode and is allowed
5544 // to have multiple "backedges".
5545 if (m == C->root()) { // Found the root?
5546 l = _ltree_root; // Root is the outermost LoopNode
5547 } else { // Else found a nested loop
5548 // Insert a LoopNode to mark this loop.
5549 l = new IdealLoopTree(this, m, n);
5550 } // End of Else found a nested loop
5551 if (!has_loop(m)) { // If 'm' does not already have a loop set
5552 set_loop(m, l); // Set loop header to loop now
5553 }
5554 } else { // Else not a nested loop
5555 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5556 IdealLoopTree* m_loop = get_loop(m);
5557 l = m_loop; // Get previously determined loop
5558 // If successor is header of a loop (nest), move up-loop till it
5559 // is a member of some outer enclosing loop. Since there are no
5560 // shared headers (I've split them already) I only need to go up
5561 // at most 1 level.
5562 while (l && l->_head == m) { // Successor heads loop?
5563 l = l->_parent; // Move up 1 for me
5564 }
5565 // If this loop is not properly parented, then this loop
5566 // has no exit path out, i.e. its an infinite loop.
5567 if (!l) {
5568 // Make loop "reachable" from root so the CFG is reachable. Basically
5569 // insert a bogus loop exit that is never taken. 'm', the loop head,
5570 // points to 'n', one (of possibly many) fall-in paths. There may be
5571 // many backedges as well.
5572
5573 if (!_verify_only) {
5574 // Insert the NeverBranch between 'm' and it's control user.
5575 NeverBranchNode *iff = new NeverBranchNode( m );
5576 _igvn.register_new_node_with_optimizer(iff);
5577 set_loop(iff, m_loop);
5578 Node *if_t = new CProjNode( iff, 0 );
5579 _igvn.register_new_node_with_optimizer(if_t);
5580 set_loop(if_t, m_loop);
5581
5582 Node* cfg = nullptr; // Find the One True Control User of m
5583 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5584 Node* x = m->fast_out(j);
5585 if (x->is_CFG() && x != m && x != iff)
5586 { cfg = x; break; }
5587 }
5588 assert(cfg != nullptr, "must find the control user of m");
5589 uint k = 0; // Probably cfg->in(0)
5590 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5591 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5592
5593 // Now create the never-taken loop exit
5594 Node *if_f = new CProjNode( iff, 1 );
5595 _igvn.register_new_node_with_optimizer(if_f);
5596 set_loop(if_f, _ltree_root);
5597 // Find frame ptr for Halt. Relies on the optimizer
5598 // V-N'ing. Easier and quicker than searching through
5599 // the program structure.
5600 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5601 _igvn.register_new_node_with_optimizer(frame);
5602 // Halt & Catch Fire
5603 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5604 _igvn.register_new_node_with_optimizer(halt);
5605 set_loop(halt, _ltree_root);
5606 _igvn.add_input_to(C->root(), halt);
5607 }
5608 set_loop(C->root(), _ltree_root);
5609 // move to outer most loop with same header
5610 l = m_loop;
5611 while (true) {
5612 IdealLoopTree* next = l->_parent;
5613 if (next == nullptr || next->_head != m) {
5614 break;
5615 }
5616 l = next;
5617 }
5618 // properly insert infinite loop in loop tree
5619 sort(_ltree_root, l);
5620 // fix child link from parent
5621 IdealLoopTree* p = l->_parent;
5622 l->_next = p->_child;
5623 p->_child = l;
5624 // code below needs enclosing loop
5625 l = l->_parent;
5626 }
5627 }
5628 if (is_postvisited(l->_head)) {
5629 // We are currently visiting l, but its head has already been post-visited.
5630 // l is irreducible: we just found a second entry m.
5631 _has_irreducible_loops = true;
5632 RegionNode* secondary_entry = m->as_Region();
5633
5634 if (!secondary_entry->can_be_irreducible_entry()) {
5635 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
5636 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
5637 return pre_order;
5638 }
5639
5640 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
5641 // Since m is a secondary entry to them all.
5642 while( is_postvisited(l->_head) ) {
5643 l->_irreducible = 1; // = true
5644 RegionNode* head = l->_head->as_Region();
5645 if (!head->can_be_irreducible_entry()) {
5646 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
5647 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
5648 return pre_order;
5649 }
5650 l = l->_parent;
5651 // Check for bad CFG here to prevent crash, and bailout of compile
5652 if (l == nullptr) {
5653 #ifndef PRODUCT
5654 if (TraceLoopOpts) {
5655 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
5656 m->dump();
5657 }
5658 #endif
5659 // This is a rare case that we do not want to handle in C2.
5660 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
5661 return pre_order;
5662 }
5663 }
5664 }
5665 if (!_verify_only) {
5666 C->set_has_irreducible_loop(_has_irreducible_loops);
5667 }
5668
5669 // This Node might be a decision point for loops. It is only if
5670 // it's children belong to several different loops. The sort call
5671 // does a trivial amount of work if there is only 1 child or all
5672 // children belong to the same loop. If however, the children
5673 // belong to different loops, the sort call will properly set the
5674 // _parent pointers to show how the loops nest.
5675 //
5676 // In any case, it returns the tightest enclosing loop.
5677 innermost = sort( l, innermost );
5678 }
5679
5680 // Def-use info will have some dead stuff; dead stuff will have no
5681 // loop decided on.
5682
5683 // Am I a loop header? If so fix up my parent's child and next ptrs.
5684 if( innermost && innermost->_head == n ) {
5685 assert( get_loop(n) == innermost, "" );
5686 IdealLoopTree *p = innermost->_parent;
5687 IdealLoopTree *l = innermost;
5688 while (p && l->_head == n) {
5689 l->_next = p->_child; // Put self on parents 'next child'
5690 p->_child = l; // Make self as first child of parent
5691 l = p; // Now walk up the parent chain
5692 p = l->_parent;
5693 }
5694 } else {
5695 // Note that it is possible for a LoopNode to reach here, if the
5696 // backedge has been made unreachable (hence the LoopNode no longer
5697 // denotes a Loop, and will eventually be removed).
5698
5699 // Record tightest enclosing loop for self. Mark as post-visited.
5700 set_loop(n, innermost);
5701 // Also record has_call flag early on
5702 if (innermost) {
5703 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
5704 // Do not count uncommon calls
5705 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
5706 Node *iff = n->in(0)->in(0);
5707 // No any calls for vectorized loops.
5708 if (C->do_superword() ||
5709 !iff->is_If() ||
5710 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
5711 iff->as_If()->_prob >= 0.01) {
5712 innermost->_has_call = 1;
5713 }
5714 }
5715 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
5716 // Disable loop optimizations if the loop has a scalar replaceable
5717 // allocation. This disabling may cause a potential performance lost
5718 // if the allocation is not eliminated for some reason.
5719 innermost->_allow_optimizations = false;
5720 innermost->_has_call = 1; // = true
5721 } else if (n->Opcode() == Op_SafePoint) {
5722 // Record all safepoints in this loop.
5723 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
5724 innermost->_safepts->push(n);
5725 }
5726 }
5727 }
5728
5729 // Flag as post-visited now
5730 set_postvisited(n);
5731 return pre_order;
5732 }
5733
5734 #ifdef ASSERT
5735 //--------------------------verify_regions_in_irreducible_loops----------------
5736 // Iterate down from Root through CFG, verify for every region:
5737 // if it is in an irreducible loop it must be marked as such
5738 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
5739 ResourceMark rm;
5740 if (!_has_irreducible_loops) {
5741 // last build_loop_tree has not found any irreducible loops
5742 // hence no region has to be marked is_in_irreduible_loop
5743 return;
5744 }
5745
5746 RootNode* root = C->root();
5747 Unique_Node_List worklist; // visit all nodes once
5748 worklist.push(root);
5749 bool failure = false;
5750 for (uint i = 0; i < worklist.size(); i++) {
5751 Node* n = worklist.at(i);
5752 if (n->is_Region()) {
5753 RegionNode* region = n->as_Region();
5754 if (is_in_irreducible_loop(region) &&
5755 region->loop_status() == RegionNode::LoopStatus::Reducible) {
5756 failure = true;
5757 tty->print("irreducible! ");
5758 region->dump();
5759 }
5760 }
5761 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
5762 Node* use = n->fast_out(j);
5763 if (use->is_CFG()) {
5764 worklist.push(use); // push if was not pushed before
5765 }
5766 }
5767 }
5768 assert(!failure, "region in irreducible loop was marked as reducible");
5769 }
5770
5771 //---------------------------is_in_irreducible_loop-------------------------
5772 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
5773 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
5774 if (!_has_irreducible_loops) {
5775 return false; // no irreducible loop in graph
5776 }
5777 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
5778 do {
5779 if (l->_irreducible) {
5780 return true; // found it
5781 }
5782 if (l == _ltree_root) {
5783 return false; // reached root, terimnate
5784 }
5785 l = l->_parent;
5786 } while (l != nullptr);
5787 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
5788 // We have "l->_parent == nullptr", which happens only for infinite loops,
5789 // where no parent is attached to the loop. We did not find any irreducible
5790 // loop from this block out to lp. Thus lp only has one entry, and no exit
5791 // (it is infinite and reducible). We can always rewrite an infinite loop
5792 // that is nested inside other loops:
5793 // while(condition) { infinite_loop; }
5794 // with an equivalent program where the infinite loop is an outermost loop
5795 // that is not nested in any loop:
5796 // while(condition) { break; } infinite_loop;
5797 // Thus, we can understand lp as an outermost loop, and can terminate and
5798 // conclude: this block is in no irreducible loop.
5799 return false;
5800 }
5801 #endif
5802
5803 //------------------------------build_loop_early-------------------------------
5804 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
5805 // First pass computes the earliest controlling node possible. This is the
5806 // controlling input with the deepest dominating depth.
5807 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
5808 while (worklist.size() != 0) {
5809 // Use local variables nstack_top_n & nstack_top_i to cache values
5810 // on nstack's top.
5811 Node *nstack_top_n = worklist.pop();
5812 uint nstack_top_i = 0;
5813 //while_nstack_nonempty:
5814 while (true) {
5815 // Get parent node and next input's index from stack's top.
5816 Node *n = nstack_top_n;
5817 uint i = nstack_top_i;
5818 uint cnt = n->req(); // Count of inputs
5819 if (i == 0) { // Pre-process the node.
5820 if( has_node(n) && // Have either loop or control already?
5821 !has_ctrl(n) ) { // Have loop picked out already?
5822 // During "merge_many_backedges" we fold up several nested loops
5823 // into a single loop. This makes the members of the original
5824 // loop bodies pointing to dead loops; they need to move up
5825 // to the new UNION'd larger loop. I set the _head field of these
5826 // dead loops to null and the _parent field points to the owning
5827 // loop. Shades of UNION-FIND algorithm.
5828 IdealLoopTree *ilt;
5829 while( !(ilt = get_loop(n))->_head ) {
5830 // Normally I would use a set_loop here. But in this one special
5831 // case, it is legal (and expected) to change what loop a Node
5832 // belongs to.
5833 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
5834 }
5835 // Remove safepoints ONLY if I've already seen I don't need one.
5836 // (the old code here would yank a 2nd safepoint after seeing a
5837 // first one, even though the 1st did not dominate in the loop body
5838 // and thus could be avoided indefinitely)
5839 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
5840 is_deleteable_safept(n)) {
5841 Node *in = n->in(TypeFunc::Control);
5842 lazy_replace(n,in); // Pull safepoint now
5843 if (ilt->_safepts != nullptr) {
5844 ilt->_safepts->yank(n);
5845 }
5846 // Carry on with the recursion "as if" we are walking
5847 // only the control input
5848 if( !visited.test_set( in->_idx ) ) {
5849 worklist.push(in); // Visit this guy later, using worklist
5850 }
5851 // Get next node from nstack:
5852 // - skip n's inputs processing by setting i > cnt;
5853 // - we also will not call set_early_ctrl(n) since
5854 // has_node(n) == true (see the condition above).
5855 i = cnt + 1;
5856 }
5857 }
5858 } // if (i == 0)
5859
5860 // Visit all inputs
5861 bool done = true; // Assume all n's inputs will be processed
5862 while (i < cnt) {
5863 Node *in = n->in(i);
5864 ++i;
5865 if (in == nullptr) continue;
5866 if (in->pinned() && !in->is_CFG())
5867 set_ctrl(in, in->in(0));
5868 int is_visited = visited.test_set( in->_idx );
5869 if (!has_node(in)) { // No controlling input yet?
5870 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
5871 assert( !is_visited, "visit only once" );
5872 nstack.push(n, i); // Save parent node and next input's index.
5873 nstack_top_n = in; // Process current input now.
5874 nstack_top_i = 0;
5875 done = false; // Not all n's inputs processed.
5876 break; // continue while_nstack_nonempty;
5877 } else if (!is_visited) {
5878 // This guy has a location picked out for him, but has not yet
5879 // been visited. Happens to all CFG nodes, for instance.
5880 // Visit him using the worklist instead of recursion, to break
5881 // cycles. Since he has a location already we do not need to
5882 // find his location before proceeding with the current Node.
5883 worklist.push(in); // Visit this guy later, using worklist
5884 }
5885 }
5886 if (done) {
5887 // All of n's inputs have been processed, complete post-processing.
5888
5889 // Compute earliest point this Node can go.
5890 // CFG, Phi, pinned nodes already know their controlling input.
5891 if (!has_node(n)) {
5892 // Record earliest legal location
5893 set_early_ctrl(n, false);
5894 }
5895 if (nstack.is_empty()) {
5896 // Finished all nodes on stack.
5897 // Process next node on the worklist.
5898 break;
5899 }
5900 // Get saved parent node and next input's index.
5901 nstack_top_n = nstack.node();
5902 nstack_top_i = nstack.index();
5903 nstack.pop();
5904 }
5905 } // while (true)
5906 }
5907 }
5908
5909 //------------------------------dom_lca_internal--------------------------------
5910 // Pair-wise LCA
5911 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
5912 if( !n1 ) return n2; // Handle null original LCA
5913 assert( n1->is_CFG(), "" );
5914 assert( n2->is_CFG(), "" );
5915 // find LCA of all uses
5916 uint d1 = dom_depth(n1);
5917 uint d2 = dom_depth(n2);
5918 while (n1 != n2) {
5919 if (d1 > d2) {
5920 n1 = idom(n1);
5921 d1 = dom_depth(n1);
5922 } else if (d1 < d2) {
5923 n2 = idom(n2);
5924 d2 = dom_depth(n2);
5925 } else {
5926 // Here d1 == d2. Due to edits of the dominator-tree, sections
5927 // of the tree might have the same depth. These sections have
5928 // to be searched more carefully.
5929
5930 // Scan up all the n1's with equal depth, looking for n2.
5931 Node *t1 = idom(n1);
5932 while (dom_depth(t1) == d1) {
5933 if (t1 == n2) return n2;
5934 t1 = idom(t1);
5935 }
5936 // Scan up all the n2's with equal depth, looking for n1.
5937 Node *t2 = idom(n2);
5938 while (dom_depth(t2) == d2) {
5939 if (t2 == n1) return n1;
5940 t2 = idom(t2);
5941 }
5942 // Move up to a new dominator-depth value as well as up the dom-tree.
5943 n1 = t1;
5944 n2 = t2;
5945 d1 = dom_depth(n1);
5946 d2 = dom_depth(n2);
5947 }
5948 }
5949 return n1;
5950 }
5951
5952 //------------------------------compute_idom-----------------------------------
5953 // Locally compute IDOM using dom_lca call. Correct only if the incoming
5954 // IDOMs are correct.
5955 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
5956 assert( region->is_Region(), "" );
5957 Node *LCA = nullptr;
5958 for( uint i = 1; i < region->req(); i++ ) {
5959 if( region->in(i) != C->top() )
5960 LCA = dom_lca( LCA, region->in(i) );
5961 }
5962 return LCA;
5963 }
5964
5965 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
5966 bool had_error = false;
5967 #ifdef ASSERT
5968 if (early != C->root()) {
5969 // Make sure that there's a dominance path from LCA to early
5970 Node* d = LCA;
5971 while (d != early) {
5972 if (d == C->root()) {
5973 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
5974 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
5975 had_error = true;
5976 break;
5977 }
5978 d = idom(d);
5979 }
5980 }
5981 #endif
5982 return had_error;
5983 }
5984
5985
5986 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
5987 // Compute LCA over list of uses
5988 bool had_error = false;
5989 Node *LCA = nullptr;
5990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
5991 Node* c = n->fast_out(i);
5992 if (_loop_or_ctrl[c->_idx] == nullptr)
5993 continue; // Skip the occasional dead node
5994 if( c->is_Phi() ) { // For Phis, we must land above on the path
5995 for( uint j=1; j<c->req(); j++ ) {// For all inputs
5996 if( c->in(j) == n ) { // Found matching input?
5997 Node *use = c->in(0)->in(j);
5998 if (_verify_only && use->is_top()) continue;
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 } else {
6004 // For CFG data-users, use is in the block just prior
6005 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
6006 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6007 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6008 }
6009 }
6010 assert(!had_error, "bad dominance");
6011 return LCA;
6012 }
6013
6014 // Check the shape of the graph at the loop entry. In some cases,
6015 // the shape of the graph does not match the shape outlined below.
6016 // That is caused by the Opaque1 node "protecting" the shape of
6017 // the graph being removed by, for example, the IGVN performed
6018 // in PhaseIdealLoop::build_and_optimize().
6019 //
6020 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6021 // loop unswitching, and IGVN, or a combination of them) can freely change
6022 // the graph's shape. As a result, the graph shape outlined below cannot
6023 // be guaranteed anymore.
6024 Node* CountedLoopNode::is_canonical_loop_entry() {
6025 if (!is_main_loop() && !is_post_loop()) {
6026 return nullptr;
6027 }
6028 Node* ctrl = skip_assertion_predicates_with_halt();
6029
6030 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6031 return nullptr;
6032 }
6033 Node* iffm = ctrl->in(0);
6034 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6035 return nullptr;
6036 }
6037 Node* bolzm = iffm->in(1);
6038 if (bolzm == nullptr || !bolzm->is_Bool()) {
6039 return nullptr;
6040 }
6041 Node* cmpzm = bolzm->in(1);
6042 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6043 return nullptr;
6044 }
6045
6046 uint input = is_main_loop() ? 2 : 1;
6047 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6048 return nullptr;
6049 }
6050 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6051 #ifdef ASSERT
6052 bool found_opaque = false;
6053 for (uint i = 1; i < cmpzm->req(); i++) {
6054 Node* opnd = cmpzm->in(i);
6055 if (opnd && opnd->is_Opaque1()) {
6056 found_opaque = true;
6057 break;
6058 }
6059 }
6060 assert(found_opaque == res, "wrong pattern");
6061 #endif
6062 return res ? cmpzm->in(input) : nullptr;
6063 }
6064
6065 // Find pre loop end from main loop. Returns nullptr if none.
6066 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6067 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6068 // The loop cannot be optimized if the graph shape at the loop entry is
6069 // inappropriate.
6070 if (is_canonical_loop_entry() == nullptr) {
6071 return nullptr;
6072 }
6073
6074 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6075 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6076 return nullptr;
6077 }
6078 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6079 CountedLoopNode* loop_node = pre_end->loopnode();
6080 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6081 return nullptr;
6082 }
6083 return pre_end;
6084 }
6085
6086 //------------------------------get_late_ctrl----------------------------------
6087 // Compute latest legal control.
6088 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6089 assert(early != nullptr, "early control should not be null");
6090
6091 Node* LCA = compute_lca_of_uses(n, early);
6092 #ifdef ASSERT
6093 if (LCA == C->root() && LCA != early) {
6094 // def doesn't dominate uses so print some useful debugging output
6095 compute_lca_of_uses(n, early, true);
6096 }
6097 #endif
6098
6099 if (n->is_Load() && LCA != early) {
6100 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6101 }
6102
6103 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6104 return LCA;
6105 }
6106
6107 // if this is a load, check for anti-dependent stores
6108 // We use a conservative algorithm to identify potential interfering
6109 // instructions and for rescheduling the load. The users of the memory
6110 // input of this load are examined. Any use which is not a load and is
6111 // dominated by early is considered a potentially interfering store.
6112 // This can produce false positives.
6113 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6114 int load_alias_idx = C->get_alias_index(n->adr_type());
6115 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6116 Unique_Node_List worklist;
6117
6118 Node* mem = n->in(MemNode::Memory);
6119 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6120 Node* s = mem->fast_out(i);
6121 worklist.push(s);
6122 }
6123 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6124 Node* s = worklist.at(i);
6125 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6126 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6127 s->is_Phi()) {
6128 continue;
6129 } else if (s->is_MergeMem()) {
6130 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6131 Node* s1 = s->fast_out(i);
6132 worklist.push(s1);
6133 }
6134 } else {
6135 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6136 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6137 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6138 const TypePtr* adr_type = s->adr_type();
6139 if (s->is_ArrayCopy()) {
6140 // Copy to known instance needs destination type to test for aliasing
6141 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6142 if (dest_type != TypeOopPtr::BOTTOM) {
6143 adr_type = dest_type;
6144 }
6145 }
6146 if (C->can_alias(adr_type, load_alias_idx)) {
6147 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6148 } else if (s->is_CFG() && s->is_Multi()) {
6149 // Look for the memory use of s (that is the use of its memory projection)
6150 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6151 Node* s1 = s->fast_out(i);
6152 assert(s1->is_Proj(), "projection expected");
6153 if (_igvn.type(s1) == Type::MEMORY) {
6154 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6155 Node* s2 = s1->fast_out(j);
6156 worklist.push(s2);
6157 }
6158 }
6159 }
6160 }
6161 }
6162 }
6163 }
6164 // For Phis only consider Region's inputs that were reached by following the memory edges
6165 if (LCA != early) {
6166 for (uint i = 0; i < worklist.size(); i++) {
6167 Node* s = worklist.at(i);
6168 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6169 Node* r = s->in(0);
6170 for (uint j = 1; j < s->req(); j++) {
6171 Node* in = s->in(j);
6172 Node* r_in = r->in(j);
6173 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6174 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6175 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6176 }
6177 }
6178 }
6179 }
6180 }
6181 }
6182 return LCA;
6183 }
6184
6185 // Is CFG node 'dominator' dominating node 'n'?
6186 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6187 if (dominator == n) {
6188 return true;
6189 }
6190 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6191 uint dd = dom_depth(dominator);
6192 while (dom_depth(n) >= dd) {
6193 if (n == dominator) {
6194 return true;
6195 }
6196 n = idom(n);
6197 }
6198 return false;
6199 }
6200
6201 // Is CFG node 'dominator' strictly dominating node 'n'?
6202 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6203 return dominator != n && is_dominator(dominator, n);
6204 }
6205
6206 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6207 // Pair-wise LCA with tags.
6208 // Tag each index with the node 'tag' currently being processed
6209 // before advancing up the dominator chain using idom().
6210 // Later calls that find a match to 'tag' know that this path has already
6211 // been considered in the current LCA (which is input 'n1' by convention).
6212 // Since get_late_ctrl() is only called once for each node, the tag array
6213 // does not need to be cleared between calls to get_late_ctrl().
6214 // Algorithm trades a larger constant factor for better asymptotic behavior
6215 //
6216 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6217 uint d1 = dom_depth(n1);
6218 uint d2 = dom_depth(n2);
6219 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6220
6221 do {
6222 if (d1 > d2) {
6223 // current lca is deeper than n2
6224 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6225 n1 = idom(n1);
6226 d1 = dom_depth(n1);
6227 } else if (d1 < d2) {
6228 // n2 is deeper than current lca
6229 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6230 if (memo == tag) {
6231 return n1; // Return the current LCA
6232 }
6233 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6234 n2 = idom(n2);
6235 d2 = dom_depth(n2);
6236 } else {
6237 // Here d1 == d2. Due to edits of the dominator-tree, sections
6238 // of the tree might have the same depth. These sections have
6239 // to be searched more carefully.
6240
6241 // Scan up all the n1's with equal depth, looking for n2.
6242 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6243 Node *t1 = idom(n1);
6244 while (dom_depth(t1) == d1) {
6245 if (t1 == n2) return n2;
6246 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6247 t1 = idom(t1);
6248 }
6249 // Scan up all the n2's with equal depth, looking for n1.
6250 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6251 Node *t2 = idom(n2);
6252 while (dom_depth(t2) == d2) {
6253 if (t2 == n1) return n1;
6254 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6255 t2 = idom(t2);
6256 }
6257 // Move up to a new dominator-depth value as well as up the dom-tree.
6258 n1 = t1;
6259 n2 = t2;
6260 d1 = dom_depth(n1);
6261 d2 = dom_depth(n2);
6262 }
6263 } while (n1 != n2);
6264 return n1;
6265 }
6266
6267 //------------------------------init_dom_lca_tags------------------------------
6268 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6269 // Intended use does not involve any growth for the array, so it could
6270 // be of fixed size.
6271 void PhaseIdealLoop::init_dom_lca_tags() {
6272 uint limit = C->unique() + 1;
6273 _dom_lca_tags.at_grow(limit, 0);
6274 _dom_lca_tags_round = 0;
6275 #ifdef ASSERT
6276 for (uint i = 0; i < limit; ++i) {
6277 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6278 }
6279 #endif // ASSERT
6280 }
6281
6282 //------------------------------build_loop_late--------------------------------
6283 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6284 // Second pass finds latest legal placement, and ideal loop placement.
6285 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6286 while (worklist.size() != 0) {
6287 Node *n = worklist.pop();
6288 // Only visit once
6289 if (visited.test_set(n->_idx)) continue;
6290 uint cnt = n->outcnt();
6291 uint i = 0;
6292 while (true) {
6293 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6294 // Visit all children
6295 if (i < cnt) {
6296 Node* use = n->raw_out(i);
6297 ++i;
6298 // Check for dead uses. Aggressively prune such junk. It might be
6299 // dead in the global sense, but still have local uses so I cannot
6300 // easily call 'remove_dead_node'.
6301 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6302 // Due to cycles, we might not hit the same fixed point in the verify
6303 // pass as we do in the regular pass. Instead, visit such phis as
6304 // simple uses of the loop head.
6305 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6306 if( !visited.test(use->_idx) )
6307 worklist.push(use);
6308 } else if( !visited.test_set(use->_idx) ) {
6309 nstack.push(n, i); // Save parent and next use's index.
6310 n = use; // Process all children of current use.
6311 cnt = use->outcnt();
6312 i = 0;
6313 }
6314 } else {
6315 // Do not visit around the backedge of loops via data edges.
6316 // push dead code onto a worklist
6317 _deadlist.push(use);
6318 }
6319 } else {
6320 // All of n's children have been processed, complete post-processing.
6321 build_loop_late_post(n);
6322 if (C->failing()) { return; }
6323 if (nstack.is_empty()) {
6324 // Finished all nodes on stack.
6325 // Process next node on the worklist.
6326 break;
6327 }
6328 // Get saved parent node and next use's index. Visit the rest of uses.
6329 n = nstack.node();
6330 cnt = n->outcnt();
6331 i = nstack.index();
6332 nstack.pop();
6333 }
6334 }
6335 }
6336 }
6337
6338 // Verify that no data node is scheduled in the outer loop of a strip
6339 // mined loop.
6340 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6341 #ifdef ASSERT
6342 if (get_loop(least)->_nest == 0) {
6343 return;
6344 }
6345 IdealLoopTree* loop = get_loop(least);
6346 Node* head = loop->_head;
6347 if (head->is_OuterStripMinedLoop() &&
6348 // Verification can't be applied to fully built strip mined loops
6349 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6350 Node* sfpt = head->as_Loop()->outer_safepoint();
6351 ResourceMark rm;
6352 Unique_Node_List wq;
6353 wq.push(sfpt);
6354 for (uint i = 0; i < wq.size(); i++) {
6355 Node *m = wq.at(i);
6356 for (uint i = 1; i < m->req(); i++) {
6357 Node* nn = m->in(i);
6358 if (nn == n) {
6359 return;
6360 }
6361 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6362 wq.push(nn);
6363 }
6364 }
6365 }
6366 ShouldNotReachHere();
6367 }
6368 #endif
6369 }
6370
6371
6372 //------------------------------build_loop_late_post---------------------------
6373 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6374 // Second pass finds latest legal placement, and ideal loop placement.
6375 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6376 build_loop_late_post_work(n, true);
6377 }
6378
6379 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6380 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6381 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6382 // Predication or creating a Loop Limit Check Predicate later.
6383 class DominatedPredicates : public UnifiedPredicateVisitor {
6384 Node* const _dominator;
6385 Node* _earliest_dominated_predicate_entry;
6386 bool _should_continue;
6387 PhaseIdealLoop* const _phase;
6388
6389 public:
6390 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6391 : _dominator(dominator),
6392 _earliest_dominated_predicate_entry(start_node),
6393 _should_continue(true),
6394 _phase(phase) {}
6395 NONCOPYABLE(DominatedPredicates);
6396
6397 bool should_continue() const override {
6398 return _should_continue;
6399 }
6400
6401 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6402 Node* earliest_dominated_predicate_entry() const {
6403 return _earliest_dominated_predicate_entry;
6404 }
6405
6406 void visit_predicate(const Predicate& predicate) override {
6407 Node* entry = predicate.entry();
6408 if (_phase->is_strict_dominator(entry, _dominator)) {
6409 _should_continue = false;
6410 } else {
6411 _earliest_dominated_predicate_entry = entry;
6412 }
6413 }
6414 };
6415
6416 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6417
6418 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6419 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6420 }
6421
6422 #ifdef ASSERT
6423 if (_verify_only && !n->is_CFG()) {
6424 // Check def-use domination.
6425 // We would like to expose this check in product but it appears to be expensive.
6426 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6427 }
6428 #endif
6429
6430 // CFG and pinned nodes already handled
6431 if( n->in(0) ) {
6432 if( n->in(0)->is_top() ) return; // Dead?
6433
6434 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6435 // _must_ be pinned (they have to observe their control edge of course).
6436 // Unlike Stores (which modify an unallocable resource, the memory
6437 // state), Mods/Loads can float around. So free them up.
6438 switch( n->Opcode() ) {
6439 case Op_DivI:
6440 case Op_DivF:
6441 case Op_DivD:
6442 case Op_ModI:
6443 case Op_LoadB: // Same with Loads; they can sink
6444 case Op_LoadUB: // during loop optimizations.
6445 case Op_LoadUS:
6446 case Op_LoadD:
6447 case Op_LoadF:
6448 case Op_LoadI:
6449 case Op_LoadKlass:
6450 case Op_LoadNKlass:
6451 case Op_LoadL:
6452 case Op_LoadS:
6453 case Op_LoadP:
6454 case Op_LoadN:
6455 case Op_LoadRange:
6456 case Op_LoadD_unaligned:
6457 case Op_LoadL_unaligned:
6458 case Op_StrComp: // Does a bunch of load-like effects
6459 case Op_StrEquals:
6460 case Op_StrIndexOf:
6461 case Op_StrIndexOfChar:
6462 case Op_AryEq:
6463 case Op_VectorizedHashCode:
6464 case Op_CountPositives:
6465 pinned = false;
6466 }
6467 if (n->is_CMove() || n->is_ConstraintCast()) {
6468 pinned = false;
6469 }
6470 if( pinned ) {
6471 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6472 if( !chosen_loop->_child ) // Inner loop?
6473 chosen_loop->_body.push(n); // Collect inner loops
6474 return;
6475 }
6476 } else { // No slot zero
6477 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6478 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6479 return;
6480 }
6481 assert(!n->is_CFG() || n->outcnt() == 0, "");
6482 }
6483
6484 // Do I have a "safe range" I can select over?
6485 Node *early = get_ctrl(n);// Early location already computed
6486
6487 // Compute latest point this Node can go
6488 Node *LCA = get_late_ctrl( n, early );
6489 // LCA is null due to uses being dead
6490 if( LCA == nullptr ) {
6491 #ifdef ASSERT
6492 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6493 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6494 }
6495 #endif
6496 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6497 _deadlist.push(n);
6498 return;
6499 }
6500 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6501
6502 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6503 Node *least = legal; // Best legal position so far
6504 while( early != legal ) { // While not at earliest legal
6505 if (legal->is_Start() && !early->is_Root()) {
6506 #ifdef ASSERT
6507 // Bad graph. Print idom path and fail.
6508 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6509 assert(false, "Bad graph detected in build_loop_late");
6510 #endif
6511 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6512 return;
6513 }
6514 // Find least loop nesting depth
6515 legal = idom(legal); // Bump up the IDOM tree
6516 // Check for lower nesting depth
6517 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6518 least = legal;
6519 }
6520 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6521
6522 if (least != early) {
6523 // Move the node above predicates as far up as possible so a
6524 // following pass of Loop Predication doesn't hoist a predicate
6525 // that depends on it above that node.
6526 const PredicateIterator predicate_iterator(least);
6527 DominatedPredicates dominated_predicates(early, least, this);
6528 predicate_iterator.for_each(dominated_predicates);
6529 least = dominated_predicates.earliest_dominated_predicate_entry();
6530 }
6531 // Try not to place code on a loop entry projection
6532 // which can inhibit range check elimination.
6533 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6534 Node* ctrl_out = least->unique_ctrl_out_or_null();
6535 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6536 least == ctrl_out->in(LoopNode::EntryControl) &&
6537 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6538 Node* least_dom = idom(least);
6539 if (get_loop(least_dom)->is_member(get_loop(least))) {
6540 least = least_dom;
6541 }
6542 }
6543 }
6544 // Don't extend live ranges of raw oops
6545 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6546 !n->bottom_type()->isa_rawptr()) {
6547 least = early;
6548 }
6549
6550 #ifdef ASSERT
6551 // Broken part of VerifyLoopOptimizations (F)
6552 // Reason:
6553 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6554 /*
6555 // If verifying, verify that 'verify_me' has a legal location
6556 // and choose it as our location.
6557 if( _verify_me ) {
6558 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6559 Node *legal = LCA;
6560 while( early != legal ) { // While not at earliest legal
6561 if( legal == v_ctrl ) break; // Check for prior good location
6562 legal = idom(legal) ;// Bump up the IDOM tree
6563 }
6564 // Check for prior good location
6565 if( legal == v_ctrl ) least = legal; // Keep prior if found
6566 }
6567 */
6568 #endif
6569
6570 // Assign discovered "here or above" point
6571 least = find_non_split_ctrl(least);
6572 verify_strip_mined_scheduling(n, least);
6573 set_ctrl(n, least);
6574
6575 // Collect inner loop bodies
6576 IdealLoopTree *chosen_loop = get_loop(least);
6577 if( !chosen_loop->_child ) // Inner loop?
6578 chosen_loop->_body.push(n);// Collect inner loops
6579
6580 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6581 _zero_trip_guard_opaque_nodes.push(n);
6582 }
6583
6584 }
6585
6586 #ifdef ASSERT
6587 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6588 tty->print_cr("%s", msg);
6589 tty->print("n: "); n->dump();
6590 tty->print("early(n): "); early->dump();
6591 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6592 n->in(0) != early && !n->in(0)->is_Root()) {
6593 tty->print("n->in(0): "); n->in(0)->dump();
6594 }
6595 for (uint i = 1; i < n->req(); i++) {
6596 Node* in1 = n->in(i);
6597 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6598 tty->print("n->in(%d): ", i); in1->dump();
6599 Node* in1_early = get_ctrl(in1);
6600 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6601 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6602 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6603 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6604 }
6605 for (uint j = 1; j < in1->req(); j++) {
6606 Node* in2 = in1->in(j);
6607 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6608 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6609 Node* in2_early = get_ctrl(in2);
6610 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6611 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6612 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
6613 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
6614 }
6615 }
6616 }
6617 }
6618 }
6619 tty->cr();
6620 tty->print("LCA(n): "); LCA->dump();
6621 for (uint i = 0; i < n->outcnt(); i++) {
6622 Node* u1 = n->raw_out(i);
6623 if (u1 == n)
6624 continue;
6625 tty->print("n->out(%d): ", i); u1->dump();
6626 if (u1->is_CFG()) {
6627 for (uint j = 0; j < u1->outcnt(); j++) {
6628 Node* u2 = u1->raw_out(j);
6629 if (u2 != u1 && u2 != n && u2->is_CFG()) {
6630 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6631 }
6632 }
6633 } else {
6634 Node* u1_later = get_ctrl(u1);
6635 tty->print("later(n->out(%d)): ", i); u1_later->dump();
6636 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
6637 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
6638 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
6639 }
6640 for (uint j = 0; j < u1->outcnt(); j++) {
6641 Node* u2 = u1->raw_out(j);
6642 if (u2 == n || u2 == u1)
6643 continue;
6644 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6645 if (!u2->is_CFG()) {
6646 Node* u2_later = get_ctrl(u2);
6647 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
6648 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
6649 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
6650 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
6651 }
6652 }
6653 }
6654 }
6655 }
6656 dump_idoms(early, LCA);
6657 tty->cr();
6658 }
6659
6660 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
6661 class RealLCA {
6662 const PhaseIdealLoop* _phase;
6663 Node* _early;
6664 Node* _wrong_lca;
6665 uint _early_index;
6666 int _wrong_lca_index;
6667
6668 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
6669 // is different: Return the previously visited node which must be the real LCA.
6670 // The node lists also contain _early and _wrong_lca, respectively.
6671 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
6672 int early_index = early_with_idoms.size() - 1;
6673 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
6674 bool found_difference = false;
6675 do {
6676 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
6677 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
6678 found_difference = true;
6679 break;
6680 }
6681 early_index--;
6682 wrong_lca_index--;
6683 } while (wrong_lca_index >= 0);
6684
6685 assert(early_index >= 0, "must always find an LCA - cannot be early");
6686 _early_index = early_index;
6687 _wrong_lca_index = wrong_lca_index;
6688 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
6689 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
6690 return real_lca;
6691 }
6692
6693 void dump(Node* real_lca) {
6694 tty->cr();
6695 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
6696 _phase->dump_idom(_early, _early_index + 1);
6697
6698 tty->cr();
6699 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
6700 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
6701
6702 tty->cr();
6703 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
6704 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
6705 if (_wrong_lca_index >= 0) {
6706 tty->print(" (idom[%d])", _wrong_lca_index);
6707 }
6708 tty->print_cr(":");
6709 real_lca->dump();
6710 }
6711
6712 public:
6713 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
6714 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
6715 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
6716 }
6717
6718 void compute_and_dump() {
6719 ResourceMark rm;
6720 Unique_Node_List early_with_idoms;
6721 Unique_Node_List wrong_lca_with_idoms;
6722 early_with_idoms.push(_early);
6723 wrong_lca_with_idoms.push(_wrong_lca);
6724 _phase->get_idoms(_early, 10000, early_with_idoms);
6725 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
6726 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
6727 dump(real_lca);
6728 }
6729 };
6730
6731 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
6732 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
6733 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
6734 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
6735
6736 RealLCA real_lca(this, early, wrong_lca);
6737 real_lca.compute_and_dump();
6738 }
6739 #endif // ASSERT
6740
6741 #ifndef PRODUCT
6742 //------------------------------dump-------------------------------------------
6743 void PhaseIdealLoop::dump() const {
6744 ResourceMark rm;
6745 Node_Stack stack(C->live_nodes() >> 2);
6746 Node_List rpo_list;
6747 VectorSet visited;
6748 visited.set(C->top()->_idx);
6749 rpo(C->root(), stack, visited, rpo_list);
6750 // Dump root loop indexed by last element in PO order
6751 dump(_ltree_root, rpo_list.size(), rpo_list);
6752 }
6753
6754 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
6755 loop->dump_head();
6756
6757 // Now scan for CFG nodes in the same loop
6758 for (uint j = idx; j > 0; j--) {
6759 Node* n = rpo_list[j-1];
6760 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
6761 continue;
6762
6763 if (get_loop(n) != loop) { // Wrong loop nest
6764 if (get_loop(n)->_head == n && // Found nested loop?
6765 get_loop(n)->_parent == loop)
6766 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
6767 continue;
6768 }
6769
6770 // Dump controlling node
6771 tty->sp(2 * loop->_nest);
6772 tty->print("C");
6773 if (n == C->root()) {
6774 n->dump();
6775 } else {
6776 Node* cached_idom = idom_no_update(n);
6777 Node* computed_idom = n->in(0);
6778 if (n->is_Region()) {
6779 computed_idom = compute_idom(n);
6780 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
6781 // any MultiBranch ctrl node), so apply a similar transform to
6782 // the cached idom returned from idom_no_update.
6783 cached_idom = find_non_split_ctrl(cached_idom);
6784 }
6785 tty->print(" ID:%d", computed_idom->_idx);
6786 n->dump();
6787 if (cached_idom != computed_idom) {
6788 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
6789 computed_idom->_idx, cached_idom->_idx);
6790 }
6791 }
6792 // Dump nodes it controls
6793 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
6794 // (k < C->unique() && get_ctrl(find(k)) == n)
6795 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
6796 Node* m = C->root()->find(k);
6797 if (m && m->outcnt() > 0) {
6798 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
6799 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
6800 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
6801 }
6802 tty->sp(2 * loop->_nest + 1);
6803 m->dump();
6804 }
6805 }
6806 }
6807 }
6808 }
6809
6810 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
6811 if (has_ctrl(n)) {
6812 tty->print_cr("No idom for data nodes");
6813 } else {
6814 ResourceMark rm;
6815 Unique_Node_List idoms;
6816 get_idoms(n, count, idoms);
6817 dump_idoms_in_reverse(n, idoms);
6818 }
6819 }
6820
6821 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
6822 Node* next = n;
6823 for (uint i = 0; !next->is_Start() && i < count; i++) {
6824 next = idom(next);
6825 assert(!idoms.member(next), "duplicated idom is not possible");
6826 idoms.push(next);
6827 }
6828 }
6829
6830 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
6831 Node* next;
6832 uint padding = 3;
6833 uint node_index_padding_width = (C->unique() == 0 ? 0 : static_cast<int>(log10(static_cast<double>(C->unique())))) + 1;
6834 for (int i = idom_list.size() - 1; i >= 0; i--) {
6835 if (i == 9 || i == 99) {
6836 padding++;
6837 }
6838 next = idom_list[i];
6839 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
6840 }
6841 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
6842 }
6843 #endif // NOT PRODUCT
6844
6845 // Collect a R-P-O for the whole CFG.
6846 // Result list is in post-order (scan backwards for RPO)
6847 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
6848 stk.push(start, 0);
6849 visited.set(start->_idx);
6850
6851 while (stk.is_nonempty()) {
6852 Node* m = stk.node();
6853 uint idx = stk.index();
6854 if (idx < m->outcnt()) {
6855 stk.set_index(idx + 1);
6856 Node* n = m->raw_out(idx);
6857 if (n->is_CFG() && !visited.test_set(n->_idx)) {
6858 stk.push(n, 0);
6859 }
6860 } else {
6861 rpo_list.push(m);
6862 stk.pop();
6863 }
6864 }
6865 }
6866
6867 ConINode* PhaseIdealLoop::intcon(jint i) {
6868 ConINode* node = _igvn.intcon(i);
6869 set_root_as_ctrl(node);
6870 return node;
6871 }
6872
6873 ConLNode* PhaseIdealLoop::longcon(jlong i) {
6874 ConLNode* node = _igvn.longcon(i);
6875 set_root_as_ctrl(node);
6876 return node;
6877 }
6878
6879 ConNode* PhaseIdealLoop::makecon(const Type* t) {
6880 ConNode* node = _igvn.makecon(t);
6881 set_root_as_ctrl(node);
6882 return node;
6883 }
6884
6885 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
6886 ConNode* node = _igvn.integercon(l, bt);
6887 set_root_as_ctrl(node);
6888 return node;
6889 }
6890
6891 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
6892 ConNode* node = _igvn.zerocon(bt);
6893 set_root_as_ctrl(node);
6894 return node;
6895 }
6896
6897
6898 //=============================================================================
6899 //------------------------------LoopTreeIterator-------------------------------
6900
6901 // Advance to next loop tree using a preorder, left-to-right traversal.
6902 void LoopTreeIterator::next() {
6903 assert(!done(), "must not be done.");
6904 if (_curnt->_child != nullptr) {
6905 _curnt = _curnt->_child;
6906 } else if (_curnt->_next != nullptr) {
6907 _curnt = _curnt->_next;
6908 } else {
6909 while (_curnt != _root && _curnt->_next == nullptr) {
6910 _curnt = _curnt->_parent;
6911 }
6912 if (_curnt == _root) {
6913 _curnt = nullptr;
6914 assert(done(), "must be done.");
6915 } else {
6916 assert(_curnt->_next != nullptr, "must be more to do");
6917 _curnt = _curnt->_next;
6918 }
6919 }
6920 }