1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/divnode.hpp"
40 #include "opto/idealGraphPrinter.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/predicates.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/vectorization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/checkedCast.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //=============================================================================
54 //--------------------------is_cloop_ind_var-----------------------------------
55 // Determine if a node is a counted loop induction variable.
56 // NOTE: The method is declared in "node.hpp".
57 bool Node::is_cloop_ind_var() const {
58 return (is_Phi() &&
59 as_Phi()->region()->is_CountedLoop() &&
60 as_Phi()->region()->as_CountedLoop()->phi() == this);
61 }
62
63 //=============================================================================
64 //------------------------------dump_spec--------------------------------------
65 // Dump special per-node info
66 #ifndef PRODUCT
67 void LoopNode::dump_spec(outputStream *st) const {
68 RegionNode::dump_spec(st);
69 if (is_inner_loop()) st->print( "inner " );
70 if (is_partial_peel_loop()) st->print( "partial_peel " );
71 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
72 }
73 #endif
74
75 //------------------------------is_valid_counted_loop-------------------------
76 bool LoopNode::is_valid_counted_loop(BasicType bt) const {
77 if (is_BaseCountedLoop() && as_BaseCountedLoop()->bt() == bt) {
78 BaseCountedLoopNode* l = as_BaseCountedLoop();
79 BaseCountedLoopEndNode* le = l->loopexit_or_null();
80 if (le != nullptr &&
81 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
82 Node* phi = l->phi();
83 Node* exit = le->proj_out_or_null(0 /* false */);
84 if (exit != nullptr && exit->Opcode() == Op_IfFalse &&
85 phi != nullptr && phi->is_Phi() &&
86 phi->in(LoopNode::LoopBackControl) == l->incr() &&
87 le->loopnode() == l && le->stride_is_con()) {
88 return true;
89 }
90 }
91 }
92 return false;
93 }
94
95 //------------------------------get_early_ctrl---------------------------------
96 // Compute earliest legal control
97 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
98 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
99 uint i;
100 Node *early;
101 if (n->in(0) && !n->is_expensive()) {
102 early = n->in(0);
103 if (!early->is_CFG()) // Might be a non-CFG multi-def
104 early = get_ctrl(early); // So treat input as a straight data input
105 i = 1;
106 } else {
107 early = get_ctrl(n->in(1));
108 i = 2;
109 }
110 uint e_d = dom_depth(early);
111 assert( early, "" );
112 for (; i < n->req(); i++) {
113 Node *cin = get_ctrl(n->in(i));
114 assert( cin, "" );
115 // Keep deepest dominator depth
116 uint c_d = dom_depth(cin);
117 if (c_d > e_d) { // Deeper guy?
118 early = cin; // Keep deepest found so far
119 e_d = c_d;
120 } else if (c_d == e_d && // Same depth?
121 early != cin) { // If not equal, must use slower algorithm
122 // If same depth but not equal, one _must_ dominate the other
123 // and we want the deeper (i.e., dominated) guy.
124 Node *n1 = early;
125 Node *n2 = cin;
126 while (1) {
127 n1 = idom(n1); // Walk up until break cycle
128 n2 = idom(n2);
129 if (n1 == cin || // Walked early up to cin
130 dom_depth(n2) < c_d)
131 break; // early is deeper; keep him
132 if (n2 == early || // Walked cin up to early
133 dom_depth(n1) < c_d) {
134 early = cin; // cin is deeper; keep him
135 break;
136 }
137 }
138 e_d = dom_depth(early); // Reset depth register cache
139 }
140 }
141
142 // Return earliest legal location
143 assert(early == find_non_split_ctrl(early), "unexpected early control");
144
145 if (n->is_expensive() && !_verify_only && !_verify_me) {
146 assert(n->in(0), "should have control input");
147 early = get_early_ctrl_for_expensive(n, early);
148 }
149
150 return early;
151 }
152
153 //------------------------------get_early_ctrl_for_expensive---------------------------------
154 // Move node up the dominator tree as high as legal while still beneficial
155 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
156 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
157 assert(OptimizeExpensiveOps, "optimization off?");
158
159 Node* ctl = n->in(0);
160 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
161 uint min_dom_depth = dom_depth(earliest);
162 #ifdef ASSERT
163 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
164 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
165 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
166 }
167 #endif
168 if (dom_depth(ctl) < min_dom_depth) {
169 return earliest;
170 }
171
172 while (true) {
173 Node* next = ctl;
174 // Moving the node out of a loop on the projection of an If
175 // confuses Loop Predication. So, once we hit a loop in an If branch
176 // that doesn't branch to an UNC, we stop. The code that process
177 // expensive nodes will notice the loop and skip over it to try to
178 // move the node further up.
179 if (ctl->is_CountedLoop() && ctl->in(1) != nullptr && ctl->in(1)->in(0) != nullptr && ctl->in(1)->in(0)->is_If()) {
180 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern()) {
181 break;
182 }
183 next = idom(ctl->in(1)->in(0));
184 } else if (ctl->is_Proj()) {
185 // We only move it up along a projection if the projection is
186 // the single control projection for its parent: same code path,
187 // if it's a If with UNC or fallthrough of a call.
188 Node* parent_ctl = ctl->in(0);
189 if (parent_ctl == nullptr) {
190 break;
191 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != nullptr) {
192 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
193 } else if (parent_ctl->is_If()) {
194 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern()) {
195 break;
196 }
197 assert(idom(ctl) == parent_ctl, "strange");
198 next = idom(parent_ctl);
199 } else if (ctl->is_CatchProj()) {
200 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
201 break;
202 }
203 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
204 next = parent_ctl->in(0)->in(0)->in(0);
205 } else {
206 // Check if parent control has a single projection (this
207 // control is the only possible successor of the parent
208 // control). If so, we can try to move the node above the
209 // parent control.
210 int nb_ctl_proj = 0;
211 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
212 Node *p = parent_ctl->fast_out(i);
213 if (p->is_Proj() && p->is_CFG()) {
214 nb_ctl_proj++;
215 if (nb_ctl_proj > 1) {
216 break;
217 }
218 }
219 }
220
221 if (nb_ctl_proj > 1) {
222 break;
223 }
224 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
225 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
226 assert(idom(ctl) == parent_ctl, "strange");
227 next = idom(parent_ctl);
228 }
229 } else {
230 next = idom(ctl);
231 }
232 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
233 break;
234 }
235 ctl = next;
236 }
237
238 if (ctl != n->in(0)) {
239 _igvn.replace_input_of(n, 0, ctl);
240 _igvn.hash_insert(n);
241 }
242
243 return ctl;
244 }
245
246
247 //------------------------------set_early_ctrl---------------------------------
248 // Set earliest legal control
249 void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {
250 Node *early = get_early_ctrl(n);
251
252 // Record earliest legal location
253 set_ctrl(n, early);
254 IdealLoopTree *loop = get_loop(early);
255 if (update_body && loop->_child == nullptr) {
256 loop->_body.push(n);
257 }
258 }
259
260 //------------------------------set_subtree_ctrl-------------------------------
261 // set missing _ctrl entries on new nodes
262 void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {
263 // Already set? Get out.
264 if (_loop_or_ctrl[n->_idx]) return;
265 // Recursively set _loop_or_ctrl array to indicate where the Node goes
266 uint i;
267 for (i = 0; i < n->req(); ++i) {
268 Node *m = n->in(i);
269 if (m && m != C->root()) {
270 set_subtree_ctrl(m, update_body);
271 }
272 }
273
274 // Fixup self
275 set_early_ctrl(n, update_body);
276 }
277
278 IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {
279 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
280 IdealLoopTree* parent = loop->_parent;
281 IdealLoopTree* sibling = parent->_child;
282 if (sibling == loop) {
283 parent->_child = outer_ilt;
284 } else {
285 while (sibling->_next != loop) {
286 sibling = sibling->_next;
287 }
288 sibling->_next = outer_ilt;
289 }
290 outer_ilt->_next = loop->_next;
291 outer_ilt->_parent = parent;
292 outer_ilt->_child = loop;
293 outer_ilt->_nest = loop->_nest;
294 loop->_parent = outer_ilt;
295 loop->_next = nullptr;
296 loop->_nest++;
297 assert(loop->_nest <= SHRT_MAX, "sanity");
298 return outer_ilt;
299 }
300
301 // Create a skeleton strip mined outer loop: a Loop head before the
302 // inner strip mined loop, a safepoint and an exit condition guarded
303 // by an opaque node after the inner strip mined loop with a backedge
304 // to the loop head. The inner strip mined loop is left as it is. Only
305 // once loop optimizations are over, do we adjust the inner loop exit
306 // condition to limit its number of iterations, set the outer loop
307 // exit condition and add Phis to the outer loop head. Some loop
308 // optimizations that operate on the inner strip mined loop need to be
309 // aware of the outer strip mined loop: loop unswitching needs to
310 // clone the outer loop as well as the inner, unrolling needs to only
311 // clone the inner loop etc. No optimizations need to change the outer
312 // strip mined loop as it is only a skeleton.
313 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
314 IdealLoopTree* loop, float cl_prob, float le_fcnt,
315 Node*& entry_control, Node*& iffalse) {
316 Node* outer_test = _igvn.intcon(0);
317 set_ctrl(outer_test, C->root());
318 Node *orig = iffalse;
319 iffalse = iffalse->clone();
320 _igvn.register_new_node_with_optimizer(iffalse);
321 set_idom(iffalse, idom(orig), dom_depth(orig));
322
323 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
324 Node *outer_ift = new IfTrueNode (outer_le);
325 Node* outer_iff = orig;
326 _igvn.replace_input_of(outer_iff, 0, outer_le);
327
328 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
329 entry_control = outer_l;
330
331 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
332
333 set_loop(iffalse, outer_ilt);
334 // When this code runs, loop bodies have not yet been populated.
335 const bool body_populated = false;
336 register_control(outer_le, outer_ilt, iffalse, body_populated);
337 register_control(outer_ift, outer_ilt, outer_le, body_populated);
338 set_idom(outer_iff, outer_le, dom_depth(outer_le));
339 _igvn.register_new_node_with_optimizer(outer_l);
340 set_loop(outer_l, outer_ilt);
341 set_idom(outer_l, init_control, dom_depth(init_control)+1);
342
343 return outer_ilt;
344 }
345
346 void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
347 Node* cmp_limit, Node* bol) {
348 assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
349 Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
350 Deoptimization::Reason_loop_limit_check,
351 Op_If);
352 Node* iff = new_predicate_proj->in(0);
353 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
354 bol = _igvn.register_new_node_with_optimizer(bol);
355 set_subtree_ctrl(bol, false);
356 _igvn.replace_input_of(iff, 1, bol);
357
358 #ifndef PRODUCT
359 // report that the loop predication has been actually performed
360 // for this loop
361 if (TraceLoopLimitCheck) {
362 tty->print_cr("Counted Loop Limit Check generated:");
363 debug_only( bol->dump(2); )
364 }
365 #endif
366 }
367
368 Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {
369 // Counted loop head must be a good RegionNode with only 3 not null
370 // control input edges: Self, Entry, LoopBack.
371 if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {
372 return nullptr;
373 }
374 Node *init_control = x->in(LoopNode::EntryControl);
375 Node *back_control = x->in(LoopNode::LoopBackControl);
376 if (init_control == nullptr || back_control == nullptr) { // Partially dead
377 return nullptr;
378 }
379 // Must also check for TOP when looking for a dead loop
380 if (init_control->is_top() || back_control->is_top()) {
381 return nullptr;
382 }
383
384 // Allow funny placement of Safepoint
385 if (back_control->Opcode() == Op_SafePoint) {
386 back_control = back_control->in(TypeFunc::Control);
387 }
388
389 // Controlling test for loop
390 Node *iftrue = back_control;
391 uint iftrue_op = iftrue->Opcode();
392 if (iftrue_op != Op_IfTrue &&
393 iftrue_op != Op_IfFalse) {
394 // I have a weird back-control. Probably the loop-exit test is in
395 // the middle of the loop and I am looking at some trailing control-flow
396 // merge point. To fix this I would have to partially peel the loop.
397 return nullptr; // Obscure back-control
398 }
399
400 // Get boolean guarding loop-back test
401 Node *iff = iftrue->in(0);
402 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {
403 return nullptr;
404 }
405 return iftrue;
406 }
407
408 Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {
409 Node* iftrue = back_control;
410 uint iftrue_op = iftrue->Opcode();
411 Node* iff = iftrue->in(0);
412 BoolNode* test = iff->in(1)->as_Bool();
413 bt = test->_test._test;
414 cl_prob = iff->as_If()->_prob;
415 if (iftrue_op == Op_IfFalse) {
416 bt = BoolTest(bt).negate();
417 cl_prob = 1.0 - cl_prob;
418 }
419 // Get backedge compare
420 Node* cmp = test->in(1);
421 if (!cmp->is_Cmp()) {
422 return nullptr;
423 }
424
425 // Find the trip-counter increment & limit. Limit must be loop invariant.
426 incr = cmp->in(1);
427 limit = cmp->in(2);
428
429 // ---------
430 // need 'loop()' test to tell if limit is loop invariant
431 // ---------
432
433 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
434 Node* tmp = incr; // Then reverse order into the CmpI
435 incr = limit;
436 limit = tmp;
437 bt = BoolTest(bt).commute(); // And commute the exit test
438 }
439 if (is_member(loop, get_ctrl(limit))) { // Limit must be loop-invariant
440 return nullptr;
441 }
442 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
443 return nullptr;
444 }
445 return cmp;
446 }
447
448 Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {
449 if (incr->is_Phi()) {
450 if (incr->as_Phi()->region() != x || incr->req() != 3) {
451 return nullptr; // Not simple trip counter expression
452 }
453 phi_incr = incr;
454 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
455 if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant
456 return nullptr;
457 }
458 }
459 return incr;
460 }
461
462 Node* PhaseIdealLoop::loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi) {
463 assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
464 // Get merge point
465 xphi = incr->in(1);
466 Node *stride = incr->in(2);
467 if (!stride->is_Con()) { // Oops, swap these
468 if (!xphi->is_Con()) { // Is the other guy a constant?
469 return nullptr; // Nope, unknown stride, bail out
470 }
471 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
472 xphi = stride;
473 stride = tmp;
474 }
475 return stride;
476 }
477
478 PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop) {
479 if (!xphi->is_Phi()) {
480 return nullptr; // Too much math on the trip counter
481 }
482 if (phi_incr != nullptr && phi_incr != xphi) {
483 return nullptr;
484 }
485 PhiNode *phi = xphi->as_Phi();
486
487 // Phi must be of loop header; backedge must wrap to increment
488 if (phi->region() != x) {
489 return nullptr;
490 }
491 return phi;
492 }
493
494 static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {
495 if (final_correction > 0) {
496 if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {
497 return -1;
498 }
499 if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {
500 return 1;
501 }
502 } else {
503 if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {
504 return -1;
505 }
506 if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {
507 return 1;
508 }
509 }
510 return 0;
511 }
512
513 static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {
514 // If the condition is inverted and we will be rolling
515 // through MININT to MAXINT, then bail out.
516 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
517 // Odd stride
518 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
519 // Count down loop rolls through MAXINT
520 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
521 // Count up loop rolls through MININT
522 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
523 return false; // Bail out
524 }
525 return true;
526 }
527
528 Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
529 BasicType bt) {
530 Node* iv_as_long;
531 if (bt == T_LONG) {
532 iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
533 register_new_node(iv_as_long, inner_head);
534 } else {
535 iv_as_long = inner_iv;
536 }
537 Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
538 register_new_node(iv_replacement, inner_head);
539 for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
540 Node* u = iv_to_replace->last_out(i);
541 #ifdef ASSERT
542 if (!is_dominator(inner_head, ctrl_or_self(u))) {
543 assert(u->is_Phi(), "should be a Phi");
544 for (uint j = 1; j < u->req(); j++) {
545 if (u->in(j) == iv_to_replace) {
546 assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
547 }
548 }
549 }
550 #endif
551 _igvn.rehash_node_delayed(u);
552 int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
553 i -= nb;
554 }
555 return iv_replacement;
556 }
557
558 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
559 void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
560 SafePointNode* sfpt) {
561 if (!C->too_many_traps(reason)) {
562 ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
563 register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
564 Node* if_false = new IfFalseNode(parse_predicate);
565 register_control(if_false, _ltree_root, parse_predicate);
566 Node* if_true = new IfTrueNode(parse_predicate);
567 register_control(if_true, loop, parse_predicate);
568
569 int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
570 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
571 const TypePtr* no_memory_effects = nullptr;
572 JVMState* jvms = sfpt->jvms();
573 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
574 no_memory_effects);
575
576 Node* mem = nullptr;
577 Node* i_o = nullptr;
578 if (sfpt->is_Call()) {
579 mem = sfpt->proj_out(TypeFunc::Memory);
580 i_o = sfpt->proj_out(TypeFunc::I_O);
581 } else {
582 mem = sfpt->memory();
583 i_o = sfpt->i_o();
584 }
585
586 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
587 register_new_node(frame, C->start());
588 Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
589 register_new_node(ret, C->start());
590
591 unc->init_req(TypeFunc::Control, if_false);
592 unc->init_req(TypeFunc::I_O, i_o);
593 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
594 unc->init_req(TypeFunc::FramePtr, frame);
595 unc->init_req(TypeFunc::ReturnAdr, ret);
596 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
597 unc->set_cnt(PROB_UNLIKELY_MAG(4));
598 unc->copy_call_debug_info(&_igvn, sfpt);
599
600 for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
601 set_subtree_ctrl(unc->in(i), false);
602 }
603 register_control(unc, _ltree_root, if_false);
604
605 Node* ctrl = new ProjNode(unc, TypeFunc::Control);
606 register_control(ctrl, _ltree_root, unc);
607 Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
608 register_control(halt, _ltree_root, ctrl);
609 _igvn.add_input_to(C->root(), halt);
610
611 _igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
612 set_idom(inner_head, if_true, dom_depth(inner_head));
613 }
614 }
615
616 // Find a safepoint node that dominates the back edge. We need a
617 // SafePointNode so we can use its jvm state to create empty
618 // predicates.
619 static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
620 SafePointNode* safepoint = nullptr;
621 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
622 Node* u = x->fast_out(i);
623 if (u->is_memory_phi()) {
624 Node* m = u->in(LoopNode::LoopBackControl);
625 if (u->adr_type() == TypePtr::BOTTOM) {
626 if (m->is_MergeMem() && mem->is_MergeMem()) {
627 if (m != mem DEBUG_ONLY(|| true)) {
628 // MergeMemStream can modify m, for example to adjust the length to mem.
629 // This is unfortunate, and probably unnecessary. But as it is, we need
630 // to add m to the igvn worklist, else we may have a modified node that
631 // is not on the igvn worklist.
632 phase->igvn()._worklist.push(m);
633 for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
634 if (!mms.is_empty()) {
635 if (mms.memory() != mms.memory2()) {
636 return false;
637 }
638 #ifdef ASSERT
639 if (mms.alias_idx() != Compile::AliasIdxBot) {
640 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
641 }
642 #endif
643 }
644 }
645 }
646 } else if (mem->is_MergeMem()) {
647 if (m != mem->as_MergeMem()->base_memory()) {
648 return false;
649 }
650 } else {
651 return false;
652 }
653 } else {
654 if (mem->is_MergeMem()) {
655 if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {
656 return false;
657 }
658 #ifdef ASSERT
659 mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
660 #endif
661 } else {
662 if (m != mem) {
663 return false;
664 }
665 }
666 }
667 }
668 }
669 return true;
670 }
671
672 SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
673 IfNode* exit_test = back_control->in(0)->as_If();
674 SafePointNode* safepoint = nullptr;
675 if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
676 safepoint = exit_test->in(0)->as_SafePoint();
677 } else {
678 Node* c = back_control;
679 while (c != x && c->Opcode() != Op_SafePoint) {
680 c = idom(c);
681 }
682
683 if (c->Opcode() == Op_SafePoint) {
684 safepoint = c->as_SafePoint();
685 }
686
687 if (safepoint == nullptr) {
688 return nullptr;
689 }
690
691 Node* mem = safepoint->in(TypeFunc::Memory);
692
693 // We can only use that safepoint if there's no side effect between the backedge and the safepoint.
694
695 // mm is the memory state at the safepoint (when it's a MergeMem)
696 // no_side_effect_since_safepoint() goes over the memory state at the backedge. It resets the mm input for each
697 // component of the memory state it encounters so it points to the base memory. Once no_side_effect_since_safepoint()
698 // is done, if no side effect after the safepoint was found, mm should transform to the base memory: the states at
699 // the backedge and safepoint are the same so all components of the memory state at the safepoint should have been
700 // reset.
701 MergeMemNode* mm = nullptr;
702 #ifdef ASSERT
703 if (mem->is_MergeMem()) {
704 mm = mem->clone()->as_MergeMem();
705 _igvn._worklist.push(mm);
706 for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {
707 // Loop invariant memory state won't be reset by no_side_effect_since_safepoint(). Do it here.
708 // Escape Analysis can add state to mm that it doesn't add to the backedge memory Phis, breaking verification
709 // code that relies on mm. Clear that extra state here.
710 if (mms.alias_idx() != Compile::AliasIdxBot &&
711 (loop != get_loop(ctrl_or_self(mms.memory())) ||
712 (mms.adr_type()->isa_oop_ptr() && mms.adr_type()->is_known_instance()))) {
713 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
714 }
715 }
716 }
717 #endif
718 if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
719 safepoint = nullptr;
720 } else {
721 assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
722 }
723 #ifdef ASSERT
724 if (mm != nullptr) {
725 _igvn.remove_dead_node(mm);
726 }
727 #endif
728 }
729 return safepoint;
730 }
731
732 // If the loop has the shape of a counted loop but with a long
733 // induction variable, transform the loop in a loop nest: an inner
734 // loop that iterates for at most max int iterations with an integer
735 // induction variable and an outer loop that iterates over the full
736 // range of long values from the initial loop in (at most) max int
737 // steps. That is:
738 //
739 // x: for (long phi = init; phi < limit; phi += stride) {
740 // // phi := Phi(L, init, incr)
741 // // incr := AddL(phi, longcon(stride))
742 // long incr = phi + stride;
743 // ... use phi and incr ...
744 // }
745 //
746 // OR:
747 //
748 // x: for (long phi = init; (phi += stride) < limit; ) {
749 // // phi := Phi(L, AddL(init, stride), incr)
750 // // incr := AddL(phi, longcon(stride))
751 // long incr = phi + stride;
752 // ... use phi and (phi + stride) ...
753 // }
754 //
755 // ==transform=>
756 //
757 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
758 // assert(stride <= inner_iters_limit); // else abort transform
759 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
760 // outer_head: for (long outer_phi = init;;) {
761 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
762 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
763 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
764 // assert(inner_iters_actual == (int)inner_iters_actual);
765 // int inner_phi, inner_incr;
766 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
767 // // inner_phi := Phi(x, intcon(0), inner_incr)
768 // // inner_incr := AddI(inner_phi, intcon(stride))
769 // inner_incr = inner_phi + stride;
770 // if (inner_incr < inner_iters_actual) {
771 // ... use phi=>(outer_phi+inner_phi) ...
772 // continue;
773 // }
774 // else break;
775 // }
776 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
777 // continue;
778 // else break;
779 // }
780 //
781 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
782 // loops with long range checks transformed to int range checks in the inner loop.
783 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
784 Node* x = loop->_head;
785 // Only for inner loops
786 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
787 return false;
788 }
789
790 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
791 return false;
792 }
793
794 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
795 BasicType bt = x->as_BaseCountedLoop()->bt();
796
797 check_counted_loop_shape(loop, x, bt);
798
799 #ifndef PRODUCT
800 if (bt == T_LONG) {
801 Atomic::inc(&_long_loop_candidates);
802 }
803 #endif
804
805 jlong stride_con_long = head->stride_con();
806 assert(stride_con_long != 0, "missed some peephole opt");
807 // We can't iterate for more than max int at a time.
808 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
809 assert(bt == T_LONG, "only for long loops");
810 return false;
811 }
812 jint stride_con = checked_cast<jint>(stride_con_long);
813 // The number of iterations for the integer count loop: guarantee no
814 // overflow: max_jint - stride_con max. -1 so there's no need for a
815 // loop limit check if the exit test is <= or >=.
816 int iters_limit = max_jint - ABS(stride_con) - 1;
817 #ifdef ASSERT
818 if (bt == T_LONG && StressLongCountedLoop > 0) {
819 iters_limit = iters_limit / StressLongCountedLoop;
820 }
821 #endif
822 // At least 2 iterations so counted loop construction doesn't fail
823 if (iters_limit/ABS(stride_con) < 2) {
824 return false;
825 }
826
827 PhiNode* phi = head->phi()->as_Phi();
828 Node* incr = head->incr();
829
830 Node* back_control = head->in(LoopNode::LoopBackControl);
831
832 // data nodes on back branch not supported
833 if (back_control->outcnt() > 1) {
834 return false;
835 }
836
837 Node* limit = head->limit();
838 // We'll need to use the loop limit before the inner loop is entered
839 if (!is_dominator(get_ctrl(limit), x)) {
840 return false;
841 }
842
843 IfNode* exit_test = head->loopexit();
844
845 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
846
847 Node_List range_checks;
848 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
849
850 if (bt == T_INT) {
851 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
852 if (range_checks.size() == 0) {
853 return false;
854 }
855 }
856
857 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
858 // the inner loop.
859 const Node* init = head->init_trip();
860 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
861 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
862 if (stride_con < 0) {
863 swap(lo, hi);
864 }
865 if (hi->hi_as_long() <= lo->lo_as_long()) {
866 // not a loop after all
867 return false;
868 }
869
870 if (range_checks.size() > 0) {
871 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
872 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
873 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
874 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
875 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
876 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
877 // logic on loop entry and for the outer loop.
878 loop->compute_trip_count(this);
879 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
880 if (head->as_CountedLoop()->trip_count() <= 3) {
881 return false;
882 }
883 } else {
884 loop->compute_profile_trip_cnt(this);
885 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
886 return false;
887 }
888 }
889 }
890
891 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
892 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
893
894 // We need a safepoint to insert Parse Predicates for the inner loop.
895 SafePointNode* safepoint;
896 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
897 // Loop is strip mined: use the safepoint of the outer strip mined loop
898 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
899 assert(outer_loop != nullptr, "no outer loop");
900 safepoint = outer_loop->outer_safepoint();
901 outer_loop->transform_to_counted_loop(&_igvn, this);
902 exit_test = head->loopexit();
903 } else {
904 safepoint = find_safepoint(back_control, x, loop);
905 }
906
907 Node* exit_branch = exit_test->proj_out(false);
908 Node* entry_control = head->in(LoopNode::EntryControl);
909
910 // Clone the control flow of the loop to build an outer loop
911 Node* outer_back_branch = back_control->clone();
912 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
913 Node* inner_exit_branch = exit_branch->clone();
914
915 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
916 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
917
918 const bool body_populated = true;
919 register_control(outer_head, outer_ilt, entry_control, body_populated);
920
921 _igvn.register_new_node_with_optimizer(inner_exit_branch);
922 set_loop(inner_exit_branch, outer_ilt);
923 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
924
925 outer_exit_test->set_req(0, inner_exit_branch);
926 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
927
928 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
929 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
930
931 outer_back_branch->set_req(0, outer_exit_test);
932 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
933
934 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
935 set_idom(x, outer_head, dom_depth(x));
936
937 // add an iv phi to the outer loop and use it to compute the inner
938 // loop iteration limit
939 Node* outer_phi = phi->clone();
940 outer_phi->set_req(0, outer_head);
941 register_new_node(outer_phi, outer_head);
942
943 Node* inner_iters_max = nullptr;
944 if (stride_con > 0) {
945 inner_iters_max = MaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
946 } else {
947 inner_iters_max = MaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
948 }
949
950 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
951 // inner_iters_max may not fit in a signed integer (iterating from
952 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
953 // min.
954 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
955 Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
956
957 Node* inner_iters_actual_int;
958 if (bt == T_LONG) {
959 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
960 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
961 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
962 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
963 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
964 // opts pass, an accurate range of values for the limits is found.
965 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
966 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::UnconditionalDependency);
967 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
968 } else {
969 inner_iters_actual_int = inner_iters_actual;
970 }
971
972 Node* int_zero = _igvn.intcon(0);
973 set_ctrl(int_zero, C->root());
974 if (stride_con < 0) {
975 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
976 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
977 }
978
979 // Clone the iv data nodes as an integer iv
980 Node* int_stride = _igvn.intcon(stride_con);
981 set_ctrl(int_stride, C->root());
982 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
983 Node* inner_incr = new AddINode(inner_phi, int_stride);
984 Node* inner_cmp = nullptr;
985 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
986 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
987 inner_phi->set_req(LoopNode::EntryControl, int_zero);
988 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
989 register_new_node(inner_phi, x);
990 register_new_node(inner_incr, x);
991 register_new_node(inner_cmp, x);
992 register_new_node(inner_bol, x);
993
994 _igvn.replace_input_of(exit_test, 1, inner_bol);
995
996 // Clone inner loop phis to outer loop
997 for (uint i = 0; i < head->outcnt(); i++) {
998 Node* u = head->raw_out(i);
999 if (u->is_Phi() && u != inner_phi && u != phi) {
1000 assert(u->in(0) == head, "inconsistent");
1001 Node* clone = u->clone();
1002 clone->set_req(0, outer_head);
1003 register_new_node(clone, outer_head);
1004 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1005 }
1006 }
1007
1008 // Replace inner loop long iv phi as inner loop int iv phi + outer
1009 // loop iv phi
1010 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1011
1012 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1013
1014 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1015
1016 // Summary of steps from initial loop to loop nest:
1017 //
1018 // == old IR nodes =>
1019 //
1020 // entry_control: {...}
1021 // x:
1022 // for (long phi = init;;) {
1023 // // phi := Phi(x, init, incr)
1024 // // incr := AddL(phi, longcon(stride))
1025 // exit_test:
1026 // if (phi < limit)
1027 // back_control: fallthrough;
1028 // else
1029 // exit_branch: break;
1030 // long incr = phi + stride;
1031 // ... use phi and incr ...
1032 // phi = incr;
1033 // }
1034 //
1035 // == new IR nodes (just before final peel) =>
1036 //
1037 // entry_control: {...}
1038 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1039 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1040 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1041 // outer_head:
1042 // for (long outer_phi = init;;) {
1043 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1044 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1045 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1046 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1047 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1048 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1049 // inner_head: x: //in(1) := outer_head
1050 // int inner_phi;
1051 // for (inner_phi = 0;;) {
1052 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1053 // int inner_incr = inner_phi + stride;
1054 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1055 // exit_test: //exit_test->in(1) := inner_bol;
1056 // if (inner_bol) // WAS (phi < limit)
1057 // back_control: fallthrough;
1058 // else
1059 // inner_exit_branch: break; //exit_branch->clone()
1060 // ... use phi=>(outer_phi+inner_phi) ...
1061 // inner_phi = inner_phi + stride; // inner_incr
1062 // }
1063 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1064 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1065 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1066 // else
1067 // exit_branch: break; //in(0) := outer_exit_test
1068 // }
1069
1070 if (bt == T_INT) {
1071 outer_phi = new ConvI2LNode(outer_phi);
1072 register_new_node(outer_phi, outer_head);
1073 }
1074
1075 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1076 inner_phi, iv_add, inner_head);
1077 // Peel one iteration of the loop and use the safepoint at the end
1078 // of the peeled iteration to insert Parse Predicates. If no well
1079 // positioned safepoint peel to guarantee a safepoint in the outer
1080 // loop.
1081 if (safepoint != nullptr || !loop->_has_call) {
1082 old_new.clear();
1083 do_peeling(loop, old_new);
1084 } else {
1085 C->set_major_progress();
1086 }
1087
1088 if (safepoint != nullptr) {
1089 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1090
1091 if (UseLoopPredicate) {
1092 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
1093 }
1094 if (UseProfiledLoopPredicate) {
1095 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1096 }
1097 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1098 }
1099
1100 #ifndef PRODUCT
1101 if (bt == T_LONG) {
1102 Atomic::inc(&_long_loop_nests);
1103 }
1104 #endif
1105
1106 inner_head->mark_loop_nest_inner_loop();
1107 outer_head->mark_loop_nest_outer_loop();
1108
1109 return true;
1110 }
1111
1112 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1113 Node_List& range_checks) {
1114 const jlong min_iters = 2;
1115 jlong reduced_iters_limit = iters_limit;
1116 jlong original_iters_limit = iters_limit;
1117 for (uint i = 0; i < loop->_body.size(); i++) {
1118 Node* c = loop->_body.at(i);
1119 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1120 IfProjNode* if_proj = c->as_IfProj();
1121 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1122 if (call != nullptr) {
1123 Node* range = nullptr;
1124 Node* offset = nullptr;
1125 jlong scale = 0;
1126 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1127 loop->is_invariant(range) && loop->is_invariant(offset) &&
1128 scale != min_jlong &&
1129 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1130 assert(scale == (jint)scale, "scale should be an int");
1131 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1132 range_checks.push(c);
1133 }
1134 }
1135 }
1136 }
1137
1138 return checked_cast<int>(reduced_iters_limit);
1139 }
1140
1141 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1142 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1143 // 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
1144 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1145
1146 // 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,
1147 // 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)
1148 // 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].
1149
1150 // 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
1151 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1152 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1153
1154 // 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
1155 // 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
1156 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1157 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1158
1159 // 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
1160 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1161
1162 // 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].
1163 // 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
1164 // 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
1165 // 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
1166 // down (from C and 0 again).
1167
1168 // 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).
1169 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1170 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1171 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1172 // PhaseIdealLoop::add_constraint.)
1173
1174 // 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
1175 // 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
1176 // 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.
1177 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1178 // values for the 32-bit counter j within which the range checks can be eliminated.
1179
1180 // 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
1181 // 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
1182 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1183
1184 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1185
1186 // 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
1187 // 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),
1188 // 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)
1189
1190 // 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).
1191 // 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.
1192 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1193
1194 // Case A: Some Negatives (but no overflow).
1195 // Number line:
1196 // |s64_min . . . 0 . . . s64_max|
1197 // | . Q_min..Q_max . 0 . . . . | s64 negative
1198 // | . . . . R=0 R< R< R< R< | (against R values)
1199 // | . . . Q_min..0..Q_max . . . | small mixed
1200 // | . . . . R R R< R< R< | (against R values)
1201 //
1202 // 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<.
1203 //
1204 // So, if Q_min <s64 0, then use this test:
1205 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1206 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1207 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1208 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1209
1210 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1211
1212 // Case B: No Negatives.
1213 // Number line:
1214 // |s64_min . . . 0 . . . s64_max|
1215 // | . . . . 0 Q_min..Q_max . . | small positive
1216 // | . . . . R> R R R< R< | (against R values)
1217 // | . . . . 0 . Q_min..Q_max . | s64 positive
1218 // | . . . . R> R> R R R< | (against R values)
1219 //
1220 // 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.
1221 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1222 //
1223 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1224 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1225 // More generally:
1226 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1227
1228 // Case C: Overflow in the 64-bit domain
1229 // Number line:
1230 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1231 // | . . . . R> R> R> R> R | (against R values)
1232 //
1233 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1234 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1235 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1236 //
1237 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1238 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1239 // More generally:
1240 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1241 //
1242 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1243 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1244 //
1245 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1246 // clamp(X, L, H) := max(L, min(X, H))
1247 // When degenerately L > H, it returns L not H.
1248 //
1249 // All of the formulas above can be merged into a single one:
1250 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1251 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1252 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1253 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1254 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1255 // replacement R.C.:
1256 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1257 // or equivalently:
1258 // j*K + L_2 <u32 R_2
1259 // where
1260 // L_2 = Q_first - L_clamp
1261 // R_2 = R_clamp - L_clamp
1262 //
1263 // Note on why R is never negative:
1264 //
1265 // Various details of this transformation would break badly if R could be negative, so this transformation only
1266 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1267 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1268 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1269 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1270 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1271 // R.C., but R values from a.length and inside checkIndex are good to go.
1272 //
1273 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1274 Node* inner_iters_actual_int, Node* inner_phi,
1275 Node* iv_add, LoopNode* inner_head) {
1276 Node* long_zero = _igvn.longcon(0);
1277 set_ctrl(long_zero, C->root());
1278 Node* int_zero = _igvn.intcon(0);
1279 set_ctrl(int_zero, this->C->root());
1280 Node* long_one = _igvn.longcon(1);
1281 set_ctrl(long_one, this->C->root());
1282 Node* int_stride = _igvn.intcon(checked_cast<int>(stride_con));
1283 set_ctrl(int_stride, this->C->root());
1284
1285 for (uint i = 0; i < range_checks.size(); i++) {
1286 ProjNode* proj = range_checks.at(i)->as_Proj();
1287 ProjNode* unc_proj = proj->other_if_proj();
1288 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1289 jlong scale = 0;
1290 Node* offset = nullptr;
1291 Node* rc_bol = rc->in(1);
1292 Node* rc_cmp = rc_bol->in(1);
1293 if (rc_cmp->Opcode() == Op_CmpU) {
1294 // could be shared and have already been taken care of
1295 continue;
1296 }
1297 bool short_scale = false;
1298 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1299 assert(ok, "inconsistent: was tested before");
1300 Node* range = rc_cmp->in(2);
1301 Node* c = rc->in(0);
1302 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1303
1304 Node* R = range;
1305 Node* K = _igvn.longcon(scale);
1306 set_ctrl(K, this->C->root());
1307
1308 Node* L = offset;
1309
1310 if (short_scale) {
1311 // This converts:
1312 // (int)i*K + L <u64 R
1313 // with K an int into:
1314 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1315 // to protect against an overflow of (int)i*K
1316 //
1317 // Because if (int)i*K overflows, there are K,L where:
1318 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1319 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1320 // i*(long)K + L <u64 R is true.
1321 //
1322 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1323 //
1324 // It's always true that:
1325 // (int)i*K <u64 (long)max_jint + 1
1326 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1327 // As a consequence:
1328 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1329 // is always false in case of overflow of i*K
1330 //
1331 // Note, there are also K,L where i*K overflows and
1332 // i*K + L <u64 R is true, but
1333 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1334 // So this transformation could cause spurious deoptimizations and failed range check elimination
1335 // (but not incorrect execution) for unlikely corner cases with overflow.
1336 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1337 Node* max_jint_plus_one_long = _igvn.longcon((jlong)max_jint + 1);
1338 set_ctrl(max_jint_plus_one_long, C->root());
1339 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1340 register_new_node(max_range, entry_control);
1341 R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1342 set_subtree_ctrl(R, true);
1343 }
1344
1345 Node* C = outer_phi;
1346
1347 // Start with 64-bit values:
1348 // i*K + L <u64 R
1349 // (C+j)*K + L <u64 R
1350 // j*K + Q <u64 R where Q = Q_first = C*K+L
1351 Node* Q_first = new MulLNode(C, K);
1352 register_new_node(Q_first, entry_control);
1353 Q_first = new AddLNode(Q_first, L);
1354 register_new_node(Q_first, entry_control);
1355
1356 // Compute endpoints of the range of values j*K + Q.
1357 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1358 Node* Q_min = Q_first;
1359
1360 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1361 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1362 register_new_node(B_2, entry_control);
1363 B_2 = new SubINode(B_2, int_stride);
1364 register_new_node(B_2, entry_control);
1365 B_2 = new ConvI2LNode(B_2);
1366 register_new_node(B_2, entry_control);
1367
1368 Node* Q_max = new MulLNode(B_2, K);
1369 register_new_node(Q_max, entry_control);
1370 Q_max = new AddLNode(Q_max, Q_first);
1371 register_new_node(Q_max, entry_control);
1372
1373 if (scale * stride_con < 0) {
1374 swap(Q_min, Q_max);
1375 }
1376 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1377
1378 // L_clamp = Q_min < 0 ? 0 : Q_min
1379 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1380 register_new_node(Q_min_cmp, entry_control);
1381 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1382 register_new_node(Q_min_bool, entry_control);
1383 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1384 register_new_node(L_clamp, entry_control);
1385 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1386
1387 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1388 register_new_node(Q_max_plus_one, entry_control);
1389
1390 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1391 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1392 Node* max_jlong_long = _igvn.longcon(max_jlong);
1393 set_ctrl(max_jlong_long, this->C->root());
1394 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1395 register_new_node(Q_max_cmp, entry_control);
1396 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1397 register_new_node(Q_max_bool, entry_control);
1398 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1399 register_new_node(H_clamp, entry_control);
1400 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1401
1402 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1403 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1404 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1405 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1406 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1407 Node* R_2 = clamp(R, L_clamp, H_clamp);
1408 R_2 = new SubLNode(R_2, L_clamp);
1409 register_new_node(R_2, entry_control);
1410 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1411 register_new_node(R_2, entry_control);
1412
1413 // L_2 = Q_first - L_clamp
1414 // We are subtracting L_clamp from both sides of the <u32 comparison.
1415 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1416 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1417 Node* L_2 = new SubLNode(Q_first, L_clamp);
1418 register_new_node(L_2, entry_control);
1419 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1420 register_new_node(L_2, entry_control);
1421
1422 // Transform the range check using the computed values L_2/R_2
1423 // from: i*K + L <u64 R
1424 // to: j*K + L_2 <u32 R_2
1425 // that is:
1426 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1427 K = _igvn.intcon(checked_cast<int>(scale));
1428 set_ctrl(K, this->C->root());
1429 Node* scaled_iv = new MulINode(inner_phi, K);
1430 register_new_node(scaled_iv, c);
1431 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1432 register_new_node(scaled_iv_plus_offset, c);
1433
1434 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1435 register_new_node(new_rc_cmp, c);
1436
1437 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1438 }
1439 }
1440
1441 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1442 Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1443 set_subtree_ctrl(min, true);
1444 Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1445 set_subtree_ctrl(max, true);
1446 return max;
1447 }
1448
1449 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1450 IfNode* exit_test) {
1451 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1452 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1453 _igvn.register_new_node_with_optimizer(new_inner_head);
1454 _igvn.register_new_node_with_optimizer(new_inner_exit);
1455 loop->_body.push(new_inner_head);
1456 loop->_body.push(new_inner_exit);
1457 loop->_body.yank(head);
1458 loop->_body.yank(exit_test);
1459 set_loop(new_inner_head, loop);
1460 set_loop(new_inner_exit, loop);
1461 set_idom(new_inner_head, idom(head), dom_depth(head));
1462 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1463 lazy_replace(head, new_inner_head);
1464 lazy_replace(exit_test, new_inner_exit);
1465 loop->_head = new_inner_head;
1466 return new_inner_head;
1467 }
1468
1469 #ifdef ASSERT
1470 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1471 Node* back_control = loop_exit_control(x, loop);
1472 assert(back_control != nullptr, "no back control");
1473
1474 BoolTest::mask mask = BoolTest::illegal;
1475 float cl_prob = 0;
1476 Node* incr = nullptr;
1477 Node* limit = nullptr;
1478
1479 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1480 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1481
1482 Node* phi_incr = nullptr;
1483 incr = loop_iv_incr(incr, x, loop, phi_incr);
1484 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1485
1486 Node* xphi = nullptr;
1487 Node* stride = loop_iv_stride(incr, loop, xphi);
1488
1489 assert(stride != nullptr, "no stride");
1490
1491 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1492
1493 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1494
1495 jlong stride_con = stride->get_integer_as_long(bt);
1496
1497 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1498
1499 assert(mask != BoolTest::ne, "unexpected condition");
1500 assert(phi_incr == nullptr, "bad loop shape");
1501 assert(cmp->in(1) == incr, "bad exit test shape");
1502
1503 // Safepoint on backedge not supported
1504 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1505 }
1506 #endif
1507
1508 #ifdef ASSERT
1509 // convert an int counted loop to a long counted to stress handling of
1510 // long counted loops
1511 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1512 Unique_Node_List iv_nodes;
1513 Node_List old_new;
1514 iv_nodes.push(cmp);
1515 bool failed = false;
1516
1517 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1518 Node* n = iv_nodes.at(i);
1519 switch(n->Opcode()) {
1520 case Op_Phi: {
1521 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1522 old_new.map(n->_idx, clone);
1523 break;
1524 }
1525 case Op_CmpI: {
1526 Node* clone = new CmpLNode(nullptr, nullptr);
1527 old_new.map(n->_idx, clone);
1528 break;
1529 }
1530 case Op_AddI: {
1531 Node* clone = new AddLNode(nullptr, nullptr);
1532 old_new.map(n->_idx, clone);
1533 break;
1534 }
1535 case Op_CastII: {
1536 failed = true;
1537 break;
1538 }
1539 default:
1540 DEBUG_ONLY(n->dump());
1541 fatal("unexpected");
1542 }
1543
1544 for (uint i = 1; i < n->req(); i++) {
1545 Node* in = n->in(i);
1546 if (in == nullptr) {
1547 continue;
1548 }
1549 if (loop->is_member(get_loop(get_ctrl(in)))) {
1550 iv_nodes.push(in);
1551 }
1552 }
1553 }
1554
1555 if (failed) {
1556 for (uint i = 0; i < iv_nodes.size(); i++) {
1557 Node* n = iv_nodes.at(i);
1558 Node* clone = old_new[n->_idx];
1559 if (clone != nullptr) {
1560 _igvn.remove_dead_node(clone);
1561 }
1562 }
1563 return false;
1564 }
1565
1566 for (uint i = 0; i < iv_nodes.size(); i++) {
1567 Node* n = iv_nodes.at(i);
1568 Node* clone = old_new[n->_idx];
1569 for (uint i = 1; i < n->req(); i++) {
1570 Node* in = n->in(i);
1571 if (in == nullptr) {
1572 continue;
1573 }
1574 Node* in_clone = old_new[in->_idx];
1575 if (in_clone == nullptr) {
1576 assert(_igvn.type(in)->isa_int(), "");
1577 in_clone = new ConvI2LNode(in);
1578 _igvn.register_new_node_with_optimizer(in_clone);
1579 set_subtree_ctrl(in_clone, false);
1580 }
1581 if (in_clone->in(0) == nullptr) {
1582 in_clone->set_req(0, C->top());
1583 clone->set_req(i, in_clone);
1584 in_clone->set_req(0, nullptr);
1585 } else {
1586 clone->set_req(i, in_clone);
1587 }
1588 }
1589 _igvn.register_new_node_with_optimizer(clone);
1590 }
1591 set_ctrl(old_new[phi->_idx], phi->in(0));
1592
1593 for (uint i = 0; i < iv_nodes.size(); i++) {
1594 Node* n = iv_nodes.at(i);
1595 Node* clone = old_new[n->_idx];
1596 set_subtree_ctrl(clone, false);
1597 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1598 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1599 Node* u = n->fast_out(i);
1600 if (iv_nodes.member(u)) {
1601 continue;
1602 }
1603 if (m == nullptr) {
1604 m = new ConvL2INode(clone);
1605 _igvn.register_new_node_with_optimizer(m);
1606 set_subtree_ctrl(m, false);
1607 }
1608 _igvn.rehash_node_delayed(u);
1609 int nb = u->replace_edge(n, m, &_igvn);
1610 --i, imax -= nb;
1611 }
1612 }
1613 return true;
1614 }
1615 #endif
1616
1617 //------------------------------is_counted_loop--------------------------------
1618 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {
1619 PhaseGVN *gvn = &_igvn;
1620
1621 Node* back_control = loop_exit_control(x, loop);
1622 if (back_control == nullptr) {
1623 return false;
1624 }
1625
1626 BoolTest::mask bt = BoolTest::illegal;
1627 float cl_prob = 0;
1628 Node* incr = nullptr;
1629 Node* limit = nullptr;
1630 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1631 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1632 return false; // Avoid pointer & float & 64-bit compares
1633 }
1634
1635 // Trip-counter increment must be commutative & associative.
1636 if (incr->Opcode() == Op_Cast(iv_bt)) {
1637 incr = incr->in(1);
1638 }
1639
1640 Node* phi_incr = nullptr;
1641 incr = loop_iv_incr(incr, x, loop, phi_incr);
1642 if (incr == nullptr) {
1643 return false;
1644 }
1645
1646 Node* trunc1 = nullptr;
1647 Node* trunc2 = nullptr;
1648 const TypeInteger* iv_trunc_t = nullptr;
1649 Node* orig_incr = incr;
1650 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1651 return false; // Funny increment opcode
1652 }
1653 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1654
1655 Node* xphi = nullptr;
1656 Node* stride = loop_iv_stride(incr, loop, xphi);
1657
1658 if (stride == nullptr) {
1659 return false;
1660 }
1661
1662 if (xphi->Opcode() == Op_Cast(iv_bt)) {
1663 xphi = xphi->in(1);
1664 }
1665
1666 // Stride must be constant
1667 jlong stride_con = stride->get_integer_as_long(iv_bt);
1668 assert(stride_con != 0, "missed some peephole opt");
1669
1670 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1671
1672 if (phi == nullptr ||
1673 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1674 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1675 return false;
1676 }
1677
1678 Node* iftrue = back_control;
1679 uint iftrue_op = iftrue->Opcode();
1680 Node* iff = iftrue->in(0);
1681 BoolNode* test = iff->in(1)->as_Bool();
1682
1683 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1684 if (trunc1 != nullptr) {
1685 // When there is a truncation, we must be sure that after the truncation
1686 // the trip counter will end up higher than the limit, otherwise we are looking
1687 // at an endless loop. Can happen with range checks.
1688
1689 // Example:
1690 // int i = 0;
1691 // while (true)
1692 // sum + = array[i];
1693 // i++;
1694 // i = i && 0x7fff;
1695 // }
1696 //
1697 // If the array is shorter than 0x8000 this exits through a AIOOB
1698 // - Counted loop transformation is ok
1699 // If the array is longer then this is an endless loop
1700 // - No transformation can be done.
1701
1702 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1703 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1704 // if the limit can have a higher value than the increment (before the phi)
1705 return false;
1706 }
1707 }
1708
1709 Node *init_trip = phi->in(LoopNode::EntryControl);
1710
1711 // If iv trunc type is smaller than int, check for possible wrap.
1712 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1713 assert(trunc1 != nullptr, "must have found some truncation");
1714
1715 // Get a better type for the phi (filtered thru if's)
1716 const TypeInteger* phi_ft = filtered_type(phi);
1717
1718 // Can iv take on a value that will wrap?
1719 //
1720 // Ensure iv's limit is not within "stride" of the wrap value.
1721 //
1722 // Example for "short" type
1723 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
1724 // If the stride is +10, then the last value of the induction
1725 // variable before the increment (phi_ft->_hi) must be
1726 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
1727 // ensure no truncation occurs after the increment.
1728
1729 if (stride_con > 0) {
1730 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
1731 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
1732 return false; // truncation may occur
1733 }
1734 } else if (stride_con < 0) {
1735 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
1736 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
1737 return false; // truncation may occur
1738 }
1739 }
1740 // No possibility of wrap so truncation can be discarded
1741 // Promote iv type to Int
1742 } else {
1743 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
1744 }
1745
1746 if (!condition_stride_ok(bt, stride_con)) {
1747 return false;
1748 }
1749
1750 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
1751
1752 if (stride_con > 0) {
1753 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
1754 return false; // cyclic loop
1755 }
1756 } else {
1757 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
1758 return false; // cyclic loop
1759 }
1760 }
1761
1762 if (phi_incr != nullptr && bt != BoolTest::ne) {
1763 // check if there is a possibility of IV overflowing after the first increment
1764 if (stride_con > 0) {
1765 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
1766 return false;
1767 }
1768 } else {
1769 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
1770 return false;
1771 }
1772 }
1773 }
1774
1775 // =================================================
1776 // ---- SUCCESS! Found A Trip-Counted Loop! -----
1777 //
1778
1779 if (x->Opcode() == Op_Region) {
1780 // x has not yet been transformed to Loop or LongCountedLoop.
1781 // This should only happen if we are inside an infinite loop.
1782 // It happens like this:
1783 // build_loop_tree -> do not attach infinite loop and nested loops
1784 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
1785 // build_loop_tree -> find and attach infinite and nested loops
1786 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
1787 assert(x->as_Region()->is_in_infinite_subgraph(),
1788 "x can only be a Region and not Loop if inside infinite loop");
1789 // Come back later when Region is transformed to LoopNode
1790 return false;
1791 }
1792
1793 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
1794 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
1795
1796 // ===================================================
1797 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
1798 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
1799 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
1800 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
1801 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
1802 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
1803 //
1804 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
1805 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
1806 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
1807 //
1808 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
1809 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
1810 //
1811 //
1812 // The loop to be converted does not always need to have the often used shape:
1813 //
1814 // i = init
1815 // i = init loop:
1816 // do { ...
1817 // // ... equivalent i+=stride
1818 // i+=stride <==> if (i < limit)
1819 // } while (i < limit); goto loop
1820 // exit:
1821 // ...
1822 //
1823 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
1824 //
1825 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
1826 // in the loop exit check:
1827 //
1828 // i = init
1829 // loop:
1830 // ...
1831 // if (i <= limit)
1832 // i+=stride
1833 // goto loop
1834 // exit:
1835 // ...
1836 //
1837 // Let's define the following terms:
1838 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
1839 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
1840 //
1841 // The iv_pre_i and iv_post_i have the following relation:
1842 // iv_pre_i + stride = iv_post_i
1843 //
1844 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
1845 // iv_post_i < adjusted_limit
1846 //
1847 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
1848 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
1849 // -> adjusted_limit = limit.
1850 // (LE2) iv_post_i <= limit:
1851 // iv_post_i < limit + 1
1852 // -> adjusted limit = limit + 1
1853 // (LE3) iv_pre_i < limit:
1854 // iv_pre_i + stride < limit + stride
1855 // iv_post_i < limit + stride
1856 // -> adjusted_limit = limit + stride
1857 // (LE4) iv_pre_i <= limit:
1858 // iv_pre_i < limit + 1
1859 // iv_pre_i + stride < limit + stride + 1
1860 // iv_post_i < limit + stride + 1
1861 // -> adjusted_limit = limit + stride + 1
1862 //
1863 // Note that:
1864 // (AL) limit <= adjusted_limit.
1865 //
1866 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
1867 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
1868 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
1869 //
1870 // To prove (INV), we require the following two conditions/assumptions:
1871 // (i): adjusted_limit - 1 + stride <= max_int
1872 // (ii): init < limit
1873 //
1874 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
1875 // induction by assuming (i) and (ii).
1876 //
1877 // Proof by Induction
1878 // ------------------
1879 // > Base case (i = 1): We show that (INV) holds after the first iteration:
1880 // min_int <= iv_post_1 = init + stride <= max_int
1881 // Proof:
1882 // First, we note that (ii) implies
1883 // (iii) init <= limit - 1
1884 // max_int >= adjusted_limit - 1 + stride [using (i)]
1885 // >= limit - 1 + stride [using (AL)]
1886 // >= init + stride [using (iii)]
1887 // >= min_int [using stride > 0, no underflow]
1888 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
1889 //
1890 // Note that to prove the base case we need (i) and (ii).
1891 //
1892 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
1893 // min_int <= iv_post_j <= max_int
1894 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
1895 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
1896 // Proof:
1897 // If iv_post_j >= adjusted_limit:
1898 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
1899 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
1900 // If iv_post_j < adjusted_limit:
1901 // First, we note that:
1902 // (iv) iv_post_j <= adjusted_limit - 1
1903 // max_int >= adjusted_limit - 1 + stride [using (i)]
1904 // >= iv_post_j + stride [using (iv)]
1905 // >= min_int [using stride > 0, no underflow]
1906 //
1907 // Note that to prove the step case we only need (i).
1908 //
1909 // Thus, by assuming (i) and (ii), we proved (INV).
1910 //
1911 //
1912 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
1913 //
1914 // (1) Loop Limit Check Predicate for (i):
1915 // Using (i): adjusted_limit - 1 + stride <= max_int
1916 //
1917 // This condition is now restated to use limit instead of adjusted_limit:
1918 //
1919 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
1920 // max_int - stride + 1 >= adjusted_limit
1921 // We can merge the two constants into
1922 // canonicalized_correction = stride - 1
1923 // which gives us
1924 // max_int - canonicalized_correction >= adjusted_limit
1925 //
1926 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
1927 // adjusted_limit = limit + limit_correction
1928 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
1929 // max_int - canonicalized_correction - limit_correction >= limit
1930 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
1931 // (v) final_correction = canonicalized_correction + limit_correction
1932 //
1933 // which gives us:
1934 //
1935 // Final predicate condition:
1936 // max_int - final_correction >= limit
1937 //
1938 // However, we need to be careful that (v) does not over- or underflow.
1939 // We know that:
1940 // canonicalized_correction = stride - 1
1941 // and
1942 // limit_correction <= stride + 1
1943 // and thus
1944 // canonicalized_correction + limit_correction <= 2 * stride
1945 // To prevent an over- or underflow of (v), we must ensure that
1946 // 2 * stride <= max_int
1947 // which can safely be checked without over- or underflow with
1948 // (vi) stride != min_int AND abs(stride) <= max_int / 2
1949 //
1950 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
1951 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
1952 //
1953 // (2) Loop Limit Check Predicate for (ii):
1954 // Using (ii): init < limit
1955 //
1956 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
1957 // (2.1) type(init) < type(limit)
1958 // In this case, we know:
1959 // all possible values of init < all possible values of limit
1960 // and we can skip the predicate.
1961 //
1962 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
1963 // In this case, we do not need to re-check the condition and can skip the predicate.
1964 // This is often found for while- and for-loops which have the following shape:
1965 //
1966 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
1967 // i = init;
1968 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
1969 // do {
1970 // i += stride;
1971 // } while (i < limit);
1972 // }
1973 //
1974 // (2.3) init + stride <= max_int
1975 // In this case, there is no overflow of the iv phi after the first loop iteration.
1976 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
1977 // init < limit
1978 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
1979 // compile time that init + stride <= max_int then we have trivially proven the base case and that
1980 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
1981 // again and can skip the predicate.
1982
1983 // Check (vi) and bail out if the stride is too big.
1984 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
1985 return false;
1986 }
1987
1988 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
1989 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
1990
1991 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
1992 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
1993 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
1994
1995 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
1996 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
1997 const jlong final_correction = canonicalized_correction + limit_correction;
1998
1999 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
2000 Node* init_control = x->in(LoopNode::EntryControl);
2001
2002 // If sov==0, limit's type always satisfies the condition, for
2003 // example, when it is an array length.
2004 if (sov != 0) {
2005 if (sov < 0) {
2006 return false; // Bailout: integer overflow is certain.
2007 }
2008 // (1) Loop Limit Check Predicate is required because we could not statically prove that
2009 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
2010 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2011 const Predicates predicates(init_control);
2012 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2013 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2014 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2015 #ifdef ASSERT
2016 if (TraceLoopLimitCheck) {
2017 tty->print("Missing Loop Limit Check Parse Predicate:");
2018 loop->dump_head();
2019 x->dump(1);
2020 }
2021 #endif
2022 return false;
2023 }
2024
2025 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2026 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2027 return false;
2028 }
2029
2030 Node* cmp_limit;
2031 Node* bol;
2032
2033 if (stride_con > 0) {
2034 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2035 bol = new BoolNode(cmp_limit, BoolTest::le);
2036 } else {
2037 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2038 bol = new BoolNode(cmp_limit, BoolTest::ge);
2039 }
2040
2041 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2042 }
2043
2044 // (2.3)
2045 const bool init_plus_stride_could_overflow =
2046 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2047 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2048 // (2.1)
2049 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2050 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2051
2052 if (init_gte_limit && // (2.1)
2053 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2054 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2055 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2056 // We use the following condition:
2057 // - stride > 0: init < limit
2058 // - stride < 0: init > limit
2059 //
2060 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2061 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2062 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2063
2064 const Predicates predicates(init_control);
2065 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2066 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2067 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2068 #ifdef ASSERT
2069 if (TraceLoopLimitCheck) {
2070 tty->print("Missing Loop Limit Check Parse Predicate:");
2071 loop->dump_head();
2072 x->dump(1);
2073 }
2074 #endif
2075 return false;
2076 }
2077
2078 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2079 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2080 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2081 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2082 return false;
2083 }
2084
2085 Node* cmp_limit;
2086 Node* bol;
2087
2088 if (stride_con > 0) {
2089 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2090 bol = new BoolNode(cmp_limit, BoolTest::lt);
2091 } else {
2092 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2093 bol = new BoolNode(cmp_limit, BoolTest::gt);
2094 }
2095
2096 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2097 }
2098
2099 if (bt == BoolTest::ne) {
2100 // Now we need to canonicalize the loop condition if it is 'ne'.
2101 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2102 if (stride_con > 0) {
2103 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2104 bt = BoolTest::lt;
2105 } else {
2106 assert(stride_con < 0, "must be");
2107 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2108 bt = BoolTest::gt;
2109 }
2110 }
2111
2112 Node* sfpt = nullptr;
2113 if (loop->_child == nullptr) {
2114 sfpt = find_safepoint(back_control, x, loop);
2115 } else {
2116 sfpt = iff->in(0);
2117 if (sfpt->Opcode() != Op_SafePoint) {
2118 sfpt = nullptr;
2119 }
2120 }
2121
2122 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2123 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2124 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2125 iv_bt == T_LONG) &&
2126 sfpt == nullptr) {
2127 // Leaving the safepoint on the backedge and creating a
2128 // CountedLoop will confuse optimizations. We can't move the
2129 // safepoint around because its jvm state wouldn't match a new
2130 // location. Give up on that loop.
2131 return false;
2132 }
2133 if (is_deleteable_safept(backedge_sfpt)) {
2134 lazy_replace(backedge_sfpt, iftrue);
2135 if (loop->_safepts != nullptr) {
2136 loop->_safepts->yank(backedge_sfpt);
2137 }
2138 loop->_tail = iftrue;
2139 }
2140 }
2141
2142
2143 #ifdef ASSERT
2144 if (iv_bt == T_INT &&
2145 !x->as_Loop()->is_loop_nest_inner_loop() &&
2146 StressLongCountedLoop > 0 &&
2147 trunc1 == nullptr &&
2148 convert_to_long_loop(cmp, phi, loop)) {
2149 return false;
2150 }
2151 #endif
2152
2153 Node* adjusted_limit = limit;
2154 if (phi_incr != nullptr) {
2155 // If compare points directly to the phi we need to adjust
2156 // the compare so that it points to the incr. Limit have
2157 // to be adjusted to keep trip count the same and we
2158 // should avoid int overflow.
2159 //
2160 // i = init; do {} while(i++ < limit);
2161 // is converted to
2162 // i = init; do {} while(++i < limit+1);
2163 //
2164 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2165 }
2166
2167 if (includes_limit) {
2168 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2169 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2170 //
2171 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2172 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2173 if (bt == BoolTest::le)
2174 bt = BoolTest::lt;
2175 else if (bt == BoolTest::ge)
2176 bt = BoolTest::gt;
2177 else
2178 ShouldNotReachHere();
2179 }
2180 set_subtree_ctrl(adjusted_limit, false);
2181
2182 // Build a canonical trip test.
2183 // Clone code, as old values may be in use.
2184 incr = incr->clone();
2185 incr->set_req(1,phi);
2186 incr->set_req(2,stride);
2187 incr = _igvn.register_new_node_with_optimizer(incr);
2188 set_early_ctrl(incr, false);
2189 _igvn.rehash_node_delayed(phi);
2190 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2191
2192 // If phi type is more restrictive than Int, raise to
2193 // Int to prevent (almost) infinite recursion in igvn
2194 // which can only handle integer types for constants or minint..maxint.
2195 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2196 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2197 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2198 nphi = _igvn.register_new_node_with_optimizer(nphi);
2199 set_ctrl(nphi, get_ctrl(phi));
2200 _igvn.replace_node(phi, nphi);
2201 phi = nphi->as_Phi();
2202 }
2203 cmp = cmp->clone();
2204 cmp->set_req(1,incr);
2205 cmp->set_req(2, adjusted_limit);
2206 cmp = _igvn.register_new_node_with_optimizer(cmp);
2207 set_ctrl(cmp, iff->in(0));
2208
2209 test = test->clone()->as_Bool();
2210 (*(BoolTest*)&test->_test)._test = bt;
2211 test->set_req(1,cmp);
2212 _igvn.register_new_node_with_optimizer(test);
2213 set_ctrl(test, iff->in(0));
2214
2215 // Replace the old IfNode with a new LoopEndNode
2216 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2217 IfNode *le = lex->as_If();
2218 uint dd = dom_depth(iff);
2219 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2220 set_loop(le, loop);
2221
2222 // Get the loop-exit control
2223 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2224
2225 // Need to swap loop-exit and loop-back control?
2226 if (iftrue_op == Op_IfFalse) {
2227 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2228 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2229
2230 loop->_tail = back_control = ift2;
2231 set_loop(ift2, loop);
2232 set_loop(iff2, get_loop(iffalse));
2233
2234 // Lazy update of 'get_ctrl' mechanism.
2235 lazy_replace(iffalse, iff2);
2236 lazy_replace(iftrue, ift2);
2237
2238 // Swap names
2239 iffalse = iff2;
2240 iftrue = ift2;
2241 } else {
2242 _igvn.rehash_node_delayed(iffalse);
2243 _igvn.rehash_node_delayed(iftrue);
2244 iffalse->set_req_X( 0, le, &_igvn );
2245 iftrue ->set_req_X( 0, le, &_igvn );
2246 }
2247
2248 set_idom(iftrue, le, dd+1);
2249 set_idom(iffalse, le, dd+1);
2250 assert(iff->outcnt() == 0, "should be dead now");
2251 lazy_replace( iff, le ); // fix 'get_ctrl'
2252
2253 Node* entry_control = init_control;
2254 bool strip_mine_loop = iv_bt == T_INT &&
2255 loop->_child == nullptr &&
2256 sfpt != nullptr &&
2257 !loop->_has_call &&
2258 is_deleteable_safept(sfpt);
2259 IdealLoopTree* outer_ilt = nullptr;
2260 if (strip_mine_loop) {
2261 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
2262 cl_prob, le->_fcnt, entry_control,
2263 iffalse);
2264 }
2265
2266 // Now setup a new CountedLoopNode to replace the existing LoopNode
2267 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2268 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2269 // The following assert is approximately true, and defines the intention
2270 // of can_be_counted_loop. It fails, however, because phase->type
2271 // is not yet initialized for this loop and its parts.
2272 //assert(l->can_be_counted_loop(this), "sanity");
2273 _igvn.register_new_node_with_optimizer(l);
2274 set_loop(l, loop);
2275 loop->_head = l;
2276 // Fix all data nodes placed at the old loop head.
2277 // Uses the lazy-update mechanism of 'get_ctrl'.
2278 lazy_replace( x, l );
2279 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2280
2281 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2282 // Check for immediately preceding SafePoint and remove
2283 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2284 if (strip_mine_loop) {
2285 Node* outer_le = outer_ilt->_tail->in(0);
2286 Node* sfpt_clone = sfpt->clone();
2287 sfpt_clone->set_req(0, iffalse);
2288 outer_le->set_req(0, sfpt_clone);
2289
2290 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2291 if (polladdr != nullptr && polladdr->is_Load()) {
2292 // Polling load should be pinned outside inner loop.
2293 Node* new_polladdr = polladdr->clone();
2294 new_polladdr->set_req(0, iffalse);
2295 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2296 set_ctrl(new_polladdr, iffalse);
2297 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2298 }
2299 // When this code runs, loop bodies have not yet been populated.
2300 const bool body_populated = false;
2301 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2302 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2303 }
2304 lazy_replace(sfpt, sfpt->in(TypeFunc::Control));
2305 if (loop->_safepts != nullptr) {
2306 loop->_safepts->yank(sfpt);
2307 }
2308 }
2309 }
2310
2311 #ifdef ASSERT
2312 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2313 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2314 #endif
2315 #ifndef PRODUCT
2316 if (TraceLoopOpts) {
2317 tty->print("Counted ");
2318 loop->dump_head();
2319 }
2320 #endif
2321
2322 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2323
2324 // Capture bounds of the loop in the induction variable Phi before
2325 // subsequent transformation (iteration splitting) obscures the
2326 // bounds
2327 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2328
2329 if (strip_mine_loop) {
2330 l->mark_strip_mined();
2331 l->verify_strip_mined(1);
2332 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2333 loop = outer_ilt;
2334 }
2335
2336 #ifndef PRODUCT
2337 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2338 Atomic::inc(&_long_loop_counted_loops);
2339 }
2340 #endif
2341 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2342 l->mark_loop_nest_outer_loop();
2343 }
2344
2345 return true;
2346 }
2347
2348 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2349 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2350 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2351 const BasicType iv_bt, Node* loop_entry) {
2352 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2353 // successfully find a dominated test with the If node below.
2354 Node* cmp_limit;
2355 Node* bol;
2356 if (stride_con > 0) {
2357 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2358 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2359 } else {
2360 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2361 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2362 }
2363
2364 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2365 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2366 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2367 IfFalseNode* if_false = new IfFalseNode(iff);
2368 IfTrueNode* if_true = new IfTrueNode(iff);
2369 Node* dominated_iff = _igvn.transform(iff);
2370 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2371 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2372
2373 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2374 _igvn.replace_input_of(iff, 0, C->top());
2375 _igvn.replace_input_of(iff, 1, C->top());
2376 return found_dominating_test;
2377 }
2378
2379 //----------------------exact_limit-------------------------------------------
2380 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2381 assert(loop->_head->is_CountedLoop(), "");
2382 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2383 assert(cl->is_valid_counted_loop(T_INT), "");
2384
2385 if (cl->stride_con() == 1 ||
2386 cl->stride_con() == -1 ||
2387 cl->limit()->Opcode() == Op_LoopLimit) {
2388 // Old code has exact limit (it could be incorrect in case of int overflow).
2389 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2390 return cl->limit();
2391 }
2392 Node *limit = nullptr;
2393 #ifdef ASSERT
2394 BoolTest::mask bt = cl->loopexit()->test_trip();
2395 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2396 #endif
2397 if (cl->has_exact_trip_count()) {
2398 // Simple case: loop has constant boundaries.
2399 // Use jlongs to avoid integer overflow.
2400 int stride_con = cl->stride_con();
2401 jlong init_con = cl->init_trip()->get_int();
2402 jlong limit_con = cl->limit()->get_int();
2403 julong trip_cnt = cl->trip_count();
2404 jlong final_con = init_con + trip_cnt*stride_con;
2405 int final_int = (int)final_con;
2406 // The final value should be in integer range since the loop
2407 // is counted and the limit was checked for overflow.
2408 assert(final_con == (jlong)final_int, "final value should be integer");
2409 limit = _igvn.intcon(final_int);
2410 } else {
2411 // Create new LoopLimit node to get exact limit (final iv value).
2412 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2413 register_new_node(limit, cl->in(LoopNode::EntryControl));
2414 }
2415 assert(limit != nullptr, "sanity");
2416 return limit;
2417 }
2418
2419 //------------------------------Ideal------------------------------------------
2420 // Return a node which is more "ideal" than the current node.
2421 // Attempt to convert into a counted-loop.
2422 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2423 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2424 phase->C->set_major_progress();
2425 }
2426 return RegionNode::Ideal(phase, can_reshape);
2427 }
2428
2429 #ifdef ASSERT
2430 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2431 const OuterStripMinedLoopNode* outer = nullptr;
2432 const CountedLoopNode* inner = nullptr;
2433 if (is_strip_mined()) {
2434 if (!is_valid_counted_loop(T_INT)) {
2435 return; // Skip malformed counted loop
2436 }
2437 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2438 inner = as_CountedLoop();
2439 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2440 } else if (is_OuterStripMinedLoop()) {
2441 outer = this->as_OuterStripMinedLoop();
2442 inner = outer->unique_ctrl_out()->as_CountedLoop();
2443 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2444 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2445 }
2446 if (inner != nullptr || outer != nullptr) {
2447 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2448 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2449 Node* outer_le = outer_tail->in(0);
2450 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2451 Node* sfpt = outer_le->in(0);
2452 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2453 Node* inner_out = sfpt->in(0);
2454 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2455 assert(cle == inner->loopexit_or_null(), "mismatch");
2456 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2457 if (has_skeleton) {
2458 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2459 assert(outer->outcnt() == 2, "only control nodes");
2460 } else {
2461 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2462 uint phis = 0;
2463 uint be_loads = 0;
2464 Node* be = inner->in(LoopNode::LoopBackControl);
2465 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2466 Node* u = inner->fast_out(i);
2467 if (u->is_Phi()) {
2468 phis++;
2469 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2470 Node* n = be->fast_out(j);
2471 if (n->is_Load()) {
2472 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2473 do {
2474 n = n->raw_out(0);
2475 } while (!n->is_Phi());
2476 if (n == u) {
2477 be_loads++;
2478 break;
2479 }
2480 }
2481 }
2482 }
2483 }
2484 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2485 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2486 Node* u = outer->fast_out(i);
2487 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2488 }
2489 uint stores = 0;
2490 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2491 Node* u = inner_out->fast_out(i);
2492 if (u->is_Store()) {
2493 stores++;
2494 }
2495 }
2496 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2497 // not guaranteed to be optimized out.
2498 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2499 }
2500 assert(sfpt->outcnt() == 1, "no data node");
2501 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2502 }
2503 }
2504 #endif
2505
2506 //=============================================================================
2507 //------------------------------Ideal------------------------------------------
2508 // Return a node which is more "ideal" than the current node.
2509 // Attempt to convert into a counted-loop.
2510 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2511 return RegionNode::Ideal(phase, can_reshape);
2512 }
2513
2514 //------------------------------dump_spec--------------------------------------
2515 // Dump special per-node info
2516 #ifndef PRODUCT
2517 void CountedLoopNode::dump_spec(outputStream *st) const {
2518 LoopNode::dump_spec(st);
2519 if (stride_is_con()) {
2520 st->print("stride: %d ",stride_con());
2521 }
2522 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2523 if (is_main_loop()) st->print("main of N%d", _idx);
2524 if (is_post_loop()) st->print("post of N%d", _main_idx);
2525 if (is_strip_mined()) st->print(" strip mined");
2526 }
2527 #endif
2528
2529 //=============================================================================
2530 jlong BaseCountedLoopEndNode::stride_con() const {
2531 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2532 }
2533
2534
2535 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2536 if (bt == T_INT) {
2537 return new CountedLoopEndNode(control, test, prob, cnt);
2538 }
2539 assert(bt == T_LONG, "unsupported");
2540 return new LongCountedLoopEndNode(control, test, prob, cnt);
2541 }
2542
2543 //=============================================================================
2544 //------------------------------Value-----------------------------------------
2545 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2546 const Type* init_t = phase->type(in(Init));
2547 const Type* limit_t = phase->type(in(Limit));
2548 const Type* stride_t = phase->type(in(Stride));
2549 // Either input is TOP ==> the result is TOP
2550 if (init_t == Type::TOP) return Type::TOP;
2551 if (limit_t == Type::TOP) return Type::TOP;
2552 if (stride_t == Type::TOP) return Type::TOP;
2553
2554 int stride_con = stride_t->is_int()->get_con();
2555 if (stride_con == 1)
2556 return bottom_type(); // Identity
2557
2558 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2559 // Use jlongs to avoid integer overflow.
2560 jlong init_con = init_t->is_int()->get_con();
2561 jlong limit_con = limit_t->is_int()->get_con();
2562 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2563 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2564 jlong final_con = init_con + stride_con*trip_count;
2565 int final_int = (int)final_con;
2566 // The final value should be in integer range since the loop
2567 // is counted and the limit was checked for overflow.
2568 // Assert checks for overflow only if all input nodes are ConINodes, as during CCP
2569 // there might be a temporary overflow from PhiNodes see JDK-8309266
2570 assert((in(Init)->is_ConI() && in(Limit)->is_ConI() && in(Stride)->is_ConI()) ? final_con == (jlong)final_int : true, "final value should be integer");
2571 if (final_con == (jlong)final_int) {
2572 return TypeInt::make(final_int);
2573 } else {
2574 return bottom_type();
2575 }
2576 }
2577
2578 return bottom_type(); // TypeInt::INT
2579 }
2580
2581 //------------------------------Ideal------------------------------------------
2582 // Return a node which is more "ideal" than the current node.
2583 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2584 if (phase->type(in(Init)) == Type::TOP ||
2585 phase->type(in(Limit)) == Type::TOP ||
2586 phase->type(in(Stride)) == Type::TOP)
2587 return nullptr; // Dead
2588
2589 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2590 if (stride_con == 1)
2591 return nullptr; // Identity
2592
2593 if (in(Init)->is_Con() && in(Limit)->is_Con())
2594 return nullptr; // Value
2595
2596 // Delay following optimizations until all loop optimizations
2597 // done to keep Ideal graph simple.
2598 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2599 return nullptr;
2600 }
2601
2602 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2603 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2604 jlong stride_p;
2605 jlong lim, ini;
2606 julong max;
2607 if (stride_con > 0) {
2608 stride_p = stride_con;
2609 lim = limit_t->_hi;
2610 ini = init_t->_lo;
2611 max = (julong)max_jint;
2612 } else {
2613 stride_p = -(jlong)stride_con;
2614 lim = init_t->_hi;
2615 ini = limit_t->_lo;
2616 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2617 }
2618 julong range = lim - ini + stride_p;
2619 if (range <= max) {
2620 // Convert to integer expression if it is not overflow.
2621 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2622 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2623 Node *bias = phase->transform(new AddINode(range, stride_m));
2624 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2625 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2626 return new AddINode(span, in(Init)); // exact limit
2627 }
2628
2629 if (is_power_of_2(stride_p) || // divisor is 2^n
2630 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2631 // Convert to long expression to avoid integer overflow
2632 // and let igvn optimizer convert this division.
2633 //
2634 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2635 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2636 Node* stride = phase->longcon(stride_con);
2637 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2638
2639 Node *range = phase->transform(new SubLNode(limit, init));
2640 Node *bias = phase->transform(new AddLNode(range, stride_m));
2641 Node *span;
2642 if (stride_con > 0 && is_power_of_2(stride_p)) {
2643 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2644 // and avoid generating rounding for division. Zero trip guard should
2645 // guarantee that init < limit but sometimes the guard is missing and
2646 // we can get situation when init > limit. Note, for the empty loop
2647 // optimization zero trip guard is generated explicitly which leaves
2648 // only RCE predicate where exact limit is used and the predicate
2649 // will simply fail forcing recompilation.
2650 Node* neg_stride = phase->longcon(-stride_con);
2651 span = phase->transform(new AndLNode(bias, neg_stride));
2652 } else {
2653 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2654 span = phase->transform(new MulLNode(trip, stride));
2655 }
2656 // Convert back to int
2657 Node *span_int = phase->transform(new ConvL2INode(span));
2658 return new AddINode(span_int, in(Init)); // exact limit
2659 }
2660
2661 return nullptr; // No progress
2662 }
2663
2664 //------------------------------Identity---------------------------------------
2665 // If stride == 1 return limit node.
2666 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2667 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2668 if (stride_con == 1 || stride_con == -1)
2669 return in(Limit);
2670 return this;
2671 }
2672
2673 //=============================================================================
2674 //----------------------match_incr_with_optional_truncation--------------------
2675 // Match increment with optional truncation:
2676 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2677 // Return null for failure. Success returns the increment node.
2678 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2679 const TypeInteger** trunc_type,
2680 BasicType bt) {
2681 // Quick cutouts:
2682 if (expr == nullptr || expr->req() != 3) return nullptr;
2683
2684 Node *t1 = nullptr;
2685 Node *t2 = nullptr;
2686 Node* n1 = expr;
2687 int n1op = n1->Opcode();
2688 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2689
2690 if (bt == T_INT) {
2691 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2692 if (n1op == Op_AndI &&
2693 n1->in(2)->is_Con() &&
2694 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2695 // %%% This check should match any mask of 2**K-1.
2696 t1 = n1;
2697 n1 = t1->in(1);
2698 n1op = n1->Opcode();
2699 trunc_t = TypeInt::CHAR;
2700 } else if (n1op == Op_RShiftI &&
2701 n1->in(1) != nullptr &&
2702 n1->in(1)->Opcode() == Op_LShiftI &&
2703 n1->in(2) == n1->in(1)->in(2) &&
2704 n1->in(2)->is_Con()) {
2705 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2706 // %%% This check should match any shift in [1..31].
2707 if (shift == 16 || shift == 8) {
2708 t1 = n1;
2709 t2 = t1->in(1);
2710 n1 = t2->in(1);
2711 n1op = n1->Opcode();
2712 if (shift == 16) {
2713 trunc_t = TypeInt::SHORT;
2714 } else if (shift == 8) {
2715 trunc_t = TypeInt::BYTE;
2716 }
2717 }
2718 }
2719 }
2720
2721 // If (maybe after stripping) it is an AddI, we won:
2722 if (n1op == Op_Add(bt)) {
2723 *trunc1 = t1;
2724 *trunc2 = t2;
2725 *trunc_type = trunc_t;
2726 return n1;
2727 }
2728
2729 // failed
2730 return nullptr;
2731 }
2732
2733 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
2734 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
2735 verify_strip_mined(expect_skeleton);
2736 return in(EntryControl)->as_Loop();
2737 }
2738 return this;
2739 }
2740
2741 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
2742 assert(is_strip_mined(), "not a strip mined loop");
2743 Node* c = in(EntryControl);
2744 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
2745 return nullptr;
2746 }
2747 return c->as_OuterStripMinedLoop();
2748 }
2749
2750 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
2751 Node* c = in(LoopBackControl);
2752 if (c == nullptr || c->is_top()) {
2753 return nullptr;
2754 }
2755 return c->as_IfTrue();
2756 }
2757
2758 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
2759 LoopNode* l = outer_loop();
2760 if (l == nullptr) {
2761 return nullptr;
2762 }
2763 return l->outer_loop_tail();
2764 }
2765
2766 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
2767 IfTrueNode* proj = outer_loop_tail();
2768 if (proj == nullptr) {
2769 return nullptr;
2770 }
2771 Node* c = proj->in(0);
2772 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
2773 return nullptr;
2774 }
2775 return c->as_OuterStripMinedLoopEnd();
2776 }
2777
2778 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
2779 LoopNode* l = outer_loop();
2780 if (l == nullptr) {
2781 return nullptr;
2782 }
2783 return l->outer_loop_end();
2784 }
2785
2786 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
2787 IfNode* le = outer_loop_end();
2788 if (le == nullptr) {
2789 return nullptr;
2790 }
2791 Node* c = le->proj_out_or_null(false);
2792 if (c == nullptr) {
2793 return nullptr;
2794 }
2795 return c->as_IfFalse();
2796 }
2797
2798 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
2799 LoopNode* l = outer_loop();
2800 if (l == nullptr) {
2801 return nullptr;
2802 }
2803 return l->outer_loop_exit();
2804 }
2805
2806 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
2807 IfNode* le = outer_loop_end();
2808 if (le == nullptr) {
2809 return nullptr;
2810 }
2811 Node* c = le->in(0);
2812 if (c == nullptr || c->is_top()) {
2813 return nullptr;
2814 }
2815 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
2816 return c->as_SafePoint();
2817 }
2818
2819 SafePointNode* CountedLoopNode::outer_safepoint() const {
2820 LoopNode* l = outer_loop();
2821 if (l == nullptr) {
2822 return nullptr;
2823 }
2824 return l->outer_safepoint();
2825 }
2826
2827 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
2828 Node* ctrl = in(LoopNode::EntryControl);
2829 if (ctrl == nullptr) {
2830 // Dying loop.
2831 return nullptr;
2832 }
2833 if (is_main_loop()) {
2834 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
2835 }
2836 if (is_main_loop() || is_post_loop()) {
2837 AssertionPredicates assertion_predicates(ctrl);
2838 return assertion_predicates.entry();
2839 }
2840 return ctrl;
2841 }
2842
2843
2844 int CountedLoopNode::stride_con() const {
2845 CountedLoopEndNode* cle = loopexit_or_null();
2846 return cle != nullptr ? cle->stride_con() : 0;
2847 }
2848
2849 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
2850 if (bt == T_INT) {
2851 return new CountedLoopNode(entry, backedge);
2852 }
2853 assert(bt == T_LONG, "unsupported");
2854 return new LongCountedLoopNode(entry, backedge);
2855 }
2856
2857 void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
2858 PhaseIdealLoop* iloop) {
2859 Node* cle_out = inner_cle->proj_out(false);
2860 Node* cle_tail = inner_cle->proj_out(true);
2861 if (cle_out->outcnt() > 1) {
2862 // Look for chains of stores that were sunk
2863 // out of the inner loop and are in the outer loop
2864 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
2865 Node* u = cle_out->fast_out(i);
2866 if (u->is_Store()) {
2867 int alias_idx = igvn->C->get_alias_index(u->adr_type());
2868 Node* first = u;
2869 for (;;) {
2870 Node* next = first->in(MemNode::Memory);
2871 if (!next->is_Store() || next->in(0) != cle_out) {
2872 break;
2873 }
2874 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2875 first = next;
2876 }
2877 Node* last = u;
2878 for (;;) {
2879 Node* next = nullptr;
2880 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
2881 Node* uu = last->fast_out(j);
2882 if (uu->is_Store() && uu->in(0) == cle_out) {
2883 assert(next == nullptr, "only one in the outer loop");
2884 next = uu;
2885 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2886 }
2887 }
2888 if (next == nullptr) {
2889 break;
2890 }
2891 last = next;
2892 }
2893 Node* phi = nullptr;
2894 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2895 Node* uu = inner_cl->fast_out(j);
2896 if (uu->is_Phi()) {
2897 Node* be = uu->in(LoopNode::LoopBackControl);
2898 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
2899 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
2900 }
2901 if (be == last || be == first->in(MemNode::Memory)) {
2902 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
2903 assert(phi == nullptr, "only one phi");
2904 phi = uu;
2905 }
2906 }
2907 }
2908 #ifdef ASSERT
2909 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
2910 Node* uu = inner_cl->fast_out(j);
2911 if (uu->is_memory_phi()) {
2912 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
2913 assert(phi == uu, "what's that phi?");
2914 } else if (uu->adr_type() == TypePtr::BOTTOM) {
2915 Node* n = uu->in(LoopNode::LoopBackControl);
2916 uint limit = igvn->C->live_nodes();
2917 uint i = 0;
2918 while (n != uu) {
2919 i++;
2920 assert(i < limit, "infinite loop");
2921 if (n->is_Proj()) {
2922 n = n->in(0);
2923 } else if (n->is_SafePoint() || n->is_MemBar()) {
2924 n = n->in(TypeFunc::Memory);
2925 } else if (n->is_Phi()) {
2926 n = n->in(1);
2927 } else if (n->is_MergeMem()) {
2928 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
2929 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
2930 n = n->in(MemNode::Memory);
2931 } else {
2932 n->dump();
2933 ShouldNotReachHere();
2934 }
2935 }
2936 }
2937 }
2938 }
2939 #endif
2940 if (phi == nullptr) {
2941 // If an entire chains was sunk, the
2942 // inner loop has no phi for that memory
2943 // slice, create one for the outer loop
2944 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
2945 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
2946 phi->set_req(LoopNode::LoopBackControl, last);
2947 phi = register_new_node(phi, inner_cl, igvn, iloop);
2948 igvn->replace_input_of(first, MemNode::Memory, phi);
2949 } else {
2950 // Or fix the outer loop fix to include
2951 // that chain of stores.
2952 Node* be = phi->in(LoopNode::LoopBackControl);
2953 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
2954 if (be == first->in(MemNode::Memory)) {
2955 if (be == phi->in(LoopNode::LoopBackControl)) {
2956 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
2957 } else {
2958 igvn->replace_input_of(be, MemNode::Memory, last);
2959 }
2960 } else {
2961 #ifdef ASSERT
2962 if (be == phi->in(LoopNode::LoopBackControl)) {
2963 assert(phi->in(LoopNode::LoopBackControl) == last, "");
2964 } else {
2965 assert(be->in(MemNode::Memory) == last, "");
2966 }
2967 #endif
2968 }
2969 }
2970 }
2971 }
2972 }
2973 }
2974
2975 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
2976 // Look for the outer & inner strip mined loop, reduce number of
2977 // iterations of the inner loop, set exit condition of outer loop,
2978 // construct required phi nodes for outer loop.
2979 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
2980 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
2981 if (LoopStripMiningIter == 0) {
2982 remove_outer_loop_and_safepoint(igvn);
2983 return;
2984 }
2985 if (LoopStripMiningIter == 1) {
2986 transform_to_counted_loop(igvn, nullptr);
2987 return;
2988 }
2989 Node* inner_iv_phi = inner_cl->phi();
2990 if (inner_iv_phi == nullptr) {
2991 IfNode* outer_le = outer_loop_end();
2992 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
2993 igvn->replace_node(outer_le, iff);
2994 inner_cl->clear_strip_mined();
2995 return;
2996 }
2997 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
2998
2999 int stride = inner_cl->stride_con();
3000 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
3001 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
3002 // stride, the method is guaranteed to return at the next check below.
3003 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
3004 int scaled_iters = (int)scaled_iters_long;
3005 if ((jlong)scaled_iters != scaled_iters_long) {
3006 // Remove outer loop and safepoint (too few iterations)
3007 remove_outer_loop_and_safepoint(igvn);
3008 return;
3009 }
3010 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3011 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3012 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3013 assert(iter_estimate > 0, "broken");
3014 if (iter_estimate <= short_scaled_iters) {
3015 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3016 remove_outer_loop_and_safepoint(igvn);
3017 return;
3018 }
3019 if (iter_estimate <= scaled_iters_long) {
3020 // We would only go through one iteration of
3021 // the outer loop: drop the outer loop but
3022 // keep the safepoint so we don't run for
3023 // too long without a safepoint
3024 IfNode* outer_le = outer_loop_end();
3025 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3026 igvn->replace_node(outer_le, iff);
3027 inner_cl->clear_strip_mined();
3028 return;
3029 }
3030
3031 Node* cle_tail = inner_cle->proj_out(true);
3032 ResourceMark rm;
3033 Node_List old_new;
3034 if (cle_tail->outcnt() > 1) {
3035 // Look for nodes on backedge of inner loop and clone them
3036 Unique_Node_List backedge_nodes;
3037 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3038 Node* u = cle_tail->fast_out(i);
3039 if (u != inner_cl) {
3040 assert(!u->is_CFG(), "control flow on the backedge?");
3041 backedge_nodes.push(u);
3042 }
3043 }
3044 uint last = igvn->C->unique();
3045 for (uint next = 0; next < backedge_nodes.size(); next++) {
3046 Node* n = backedge_nodes.at(next);
3047 old_new.map(n->_idx, n->clone());
3048 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3049 Node* u = n->fast_out(i);
3050 assert(!u->is_CFG(), "broken");
3051 if (u->_idx >= last) {
3052 continue;
3053 }
3054 if (!u->is_Phi()) {
3055 backedge_nodes.push(u);
3056 } else {
3057 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3058 }
3059 }
3060 }
3061 // Put the clones on the outer loop backedge
3062 Node* le_tail = outer_loop_tail();
3063 for (uint next = 0; next < backedge_nodes.size(); next++) {
3064 Node *n = old_new[backedge_nodes.at(next)->_idx];
3065 for (uint i = 1; i < n->req(); i++) {
3066 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3067 n->set_req(i, old_new[n->in(i)->_idx]);
3068 }
3069 }
3070 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3071 n->set_req(0, le_tail);
3072 }
3073 igvn->register_new_node_with_optimizer(n);
3074 }
3075 }
3076
3077 Node* iv_phi = nullptr;
3078 // Make a clone of each phi in the inner loop
3079 // for the outer loop
3080 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3081 Node* u = inner_cl->raw_out(i);
3082 if (u->is_Phi()) {
3083 assert(u->in(0) == inner_cl, "inconsistent");
3084 Node* phi = u->clone();
3085 phi->set_req(0, this);
3086 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3087 if (be != nullptr) {
3088 phi->set_req(LoopNode::LoopBackControl, be);
3089 }
3090 phi = igvn->transform(phi);
3091 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3092 if (u == inner_iv_phi) {
3093 iv_phi = phi;
3094 }
3095 }
3096 }
3097
3098 if (iv_phi != nullptr) {
3099 // Now adjust the inner loop's exit condition
3100 Node* limit = inner_cl->limit();
3101 // If limit < init for stride > 0 (or limit > init for stride < 0),
3102 // the loop body is run only once. Given limit - init (init - limit resp.)
3103 // would be negative, the unsigned comparison below would cause
3104 // the loop body to be run for LoopStripMiningIter.
3105 Node* max = nullptr;
3106 if (stride > 0) {
3107 max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3108 } else {
3109 max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3110 }
3111 // sub is positive and can be larger than the max signed int
3112 // value. Use an unsigned min.
3113 Node* const_iters = igvn->intcon(scaled_iters);
3114 Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3115 // min is the number of iterations for the next inner loop execution:
3116 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3117 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3118
3119 Node* new_limit = nullptr;
3120 if (stride > 0) {
3121 new_limit = igvn->transform(new AddINode(min, iv_phi));
3122 } else {
3123 new_limit = igvn->transform(new SubINode(iv_phi, min));
3124 }
3125 Node* inner_cmp = inner_cle->cmp_node();
3126 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3127 Node* outer_bol = inner_bol;
3128 // cmp node for inner loop may be shared
3129 inner_cmp = inner_cmp->clone();
3130 inner_cmp->set_req(2, new_limit);
3131 inner_bol = inner_bol->clone();
3132 inner_bol->set_req(1, igvn->transform(inner_cmp));
3133 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3134 // Set the outer loop's exit condition too
3135 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3136 } else {
3137 assert(false, "should be able to adjust outer loop");
3138 IfNode* outer_le = outer_loop_end();
3139 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3140 igvn->replace_node(outer_le, iff);
3141 inner_cl->clear_strip_mined();
3142 }
3143 }
3144
3145 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3146 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3147 CountedLoopEndNode* cle = inner_cl->loopexit();
3148 Node* inner_test = cle->in(1);
3149 IfNode* outer_le = outer_loop_end();
3150 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3151 Node* safepoint = outer_safepoint();
3152
3153 fix_sunk_stores(inner_cle, inner_cl, igvn, iloop);
3154
3155 // make counted loop exit test always fail
3156 ConINode* zero = igvn->intcon(0);
3157 if (iloop != nullptr) {
3158 iloop->set_ctrl(zero, igvn->C->root());
3159 }
3160 igvn->replace_input_of(cle, 1, zero);
3161 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3162 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3163 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3164 if (iloop == nullptr) {
3165 igvn->replace_node(outer_le, new_end);
3166 } else {
3167 iloop->lazy_replace(outer_le, new_end);
3168 }
3169 // the backedge of the inner loop must be rewired to the new loop end
3170 Node* backedge = cle->proj_out(true);
3171 igvn->replace_input_of(backedge, 0, new_end);
3172 if (iloop != nullptr) {
3173 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3174 }
3175 // make the outer loop go away
3176 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3177 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3178 inner_cl->clear_strip_mined();
3179 if (iloop != nullptr) {
3180 Unique_Node_List wq;
3181 wq.push(safepoint);
3182
3183 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3184 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3185
3186 for (uint i = 0; i < wq.size(); i++) {
3187 Node* n = wq.at(i);
3188 for (uint j = 0; j < n->req(); ++j) {
3189 Node* in = n->in(j);
3190 if (in == nullptr || in->is_CFG()) {
3191 continue;
3192 }
3193 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3194 continue;
3195 }
3196 assert(!loop->_body.contains(in), "");
3197 loop->_body.push(in);
3198 wq.push(in);
3199 }
3200 }
3201 iloop->set_loop(safepoint, loop);
3202 loop->_body.push(safepoint);
3203 iloop->set_loop(safepoint->in(0), loop);
3204 loop->_body.push(safepoint->in(0));
3205 outer_loop_ilt->_tail = igvn->C->top();
3206 }
3207 }
3208
3209 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3210 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3211 Node* outer_sfpt = outer_safepoint();
3212 Node* outer_out = outer_loop_exit();
3213 igvn->replace_node(outer_out, outer_sfpt->in(0));
3214 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3215 inner_cl->clear_strip_mined();
3216 }
3217
3218 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3219 if (iloop == nullptr) {
3220 return igvn->transform(node);
3221 }
3222 iloop->register_new_node(node, ctrl);
3223 return node;
3224 }
3225
3226 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3227 PhaseIdealLoop* iloop) {
3228 if (iloop == nullptr) {
3229 return igvn->transform(node);
3230 }
3231 iloop->register_control(node, iloop->get_loop(loop), idom);
3232 return node;
3233 }
3234
3235 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3236 if (!in(0)) return Type::TOP;
3237 if (phase->type(in(0)) == Type::TOP)
3238 return Type::TOP;
3239
3240 // Until expansion, the loop end condition is not set so this should not constant fold.
3241 if (is_expanded(phase)) {
3242 return IfNode::Value(phase);
3243 }
3244
3245 return TypeTuple::IFBOTH;
3246 }
3247
3248 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3249 // The outer strip mined loop head only has Phi uses after expansion
3250 if (phase->is_IterGVN()) {
3251 Node* backedge = proj_out_or_null(true);
3252 if (backedge != nullptr) {
3253 Node* head = backedge->unique_ctrl_out_or_null();
3254 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3255 if (head->find_out_with(Op_Phi) != nullptr) {
3256 return true;
3257 }
3258 }
3259 }
3260 }
3261 return false;
3262 }
3263
3264 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3265 if (remove_dead_region(phase, can_reshape)) return this;
3266
3267 return nullptr;
3268 }
3269
3270 //------------------------------filtered_type--------------------------------
3271 // Return a type based on condition control flow
3272 // A successful return will be a type that is restricted due
3273 // to a series of dominating if-tests, such as:
3274 // if (i < 10) {
3275 // if (i > 0) {
3276 // here: "i" type is [1..10)
3277 // }
3278 // }
3279 // or a control flow merge
3280 // if (i < 10) {
3281 // do {
3282 // phi( , ) -- at top of loop type is [min_int..10)
3283 // i = ?
3284 // } while ( i < 10)
3285 //
3286 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3287 assert(n && n->bottom_type()->is_int(), "must be int");
3288 const TypeInt* filtered_t = nullptr;
3289 if (!n->is_Phi()) {
3290 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3291 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3292
3293 } else {
3294 Node* phi = n->as_Phi();
3295 Node* region = phi->in(0);
3296 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3297 if (region && region != C->top()) {
3298 for (uint i = 1; i < phi->req(); i++) {
3299 Node* val = phi->in(i);
3300 Node* use_c = region->in(i);
3301 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3302 if (val_t != nullptr) {
3303 if (filtered_t == nullptr) {
3304 filtered_t = val_t;
3305 } else {
3306 filtered_t = filtered_t->meet(val_t)->is_int();
3307 }
3308 }
3309 }
3310 }
3311 }
3312 const TypeInt* n_t = _igvn.type(n)->is_int();
3313 if (filtered_t != nullptr) {
3314 n_t = n_t->join(filtered_t)->is_int();
3315 }
3316 return n_t;
3317 }
3318
3319
3320 //------------------------------filtered_type_from_dominators--------------------------------
3321 // Return a possibly more restrictive type for val based on condition control flow of dominators
3322 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3323 if (val->is_Con()) {
3324 return val->bottom_type()->is_int();
3325 }
3326 uint if_limit = 10; // Max number of dominating if's visited
3327 const TypeInt* rtn_t = nullptr;
3328
3329 if (use_ctrl && use_ctrl != C->top()) {
3330 Node* val_ctrl = get_ctrl(val);
3331 uint val_dom_depth = dom_depth(val_ctrl);
3332 Node* pred = use_ctrl;
3333 uint if_cnt = 0;
3334 while (if_cnt < if_limit) {
3335 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3336 if_cnt++;
3337 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3338 if (if_t != nullptr) {
3339 if (rtn_t == nullptr) {
3340 rtn_t = if_t;
3341 } else {
3342 rtn_t = rtn_t->join(if_t)->is_int();
3343 }
3344 }
3345 }
3346 pred = idom(pred);
3347 if (pred == nullptr || pred == C->top()) {
3348 break;
3349 }
3350 // Stop if going beyond definition block of val
3351 if (dom_depth(pred) < val_dom_depth) {
3352 break;
3353 }
3354 }
3355 }
3356 return rtn_t;
3357 }
3358
3359
3360 //------------------------------dump_spec--------------------------------------
3361 // Dump special per-node info
3362 #ifndef PRODUCT
3363 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3364 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3365 BoolTest bt( test_trip()); // Added this for g++.
3366
3367 st->print("[");
3368 bt.dump_on(st);
3369 st->print("]");
3370 }
3371 st->print(" ");
3372 IfNode::dump_spec(st);
3373 }
3374 #endif
3375
3376 //=============================================================================
3377 //------------------------------is_member--------------------------------------
3378 // Is 'l' a member of 'this'?
3379 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3380 while( l->_nest > _nest ) l = l->_parent;
3381 return l == this;
3382 }
3383
3384 //------------------------------set_nest---------------------------------------
3385 // Set loop tree nesting depth. Accumulate _has_call bits.
3386 int IdealLoopTree::set_nest( uint depth ) {
3387 assert(depth <= SHRT_MAX, "sanity");
3388 _nest = depth;
3389 int bits = _has_call;
3390 if( _child ) bits |= _child->set_nest(depth+1);
3391 if( bits ) _has_call = 1;
3392 if( _next ) bits |= _next ->set_nest(depth );
3393 return bits;
3394 }
3395
3396 //------------------------------split_fall_in----------------------------------
3397 // Split out multiple fall-in edges from the loop header. Move them to a
3398 // private RegionNode before the loop. This becomes the loop landing pad.
3399 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3400 PhaseIterGVN &igvn = phase->_igvn;
3401 uint i;
3402
3403 // Make a new RegionNode to be the landing pad.
3404 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3405 phase->set_loop(landing_pad,_parent);
3406 // If _head was irreducible loop entry, landing_pad may now be too
3407 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3408 // Gather all the fall-in control paths into the landing pad
3409 uint icnt = fall_in_cnt;
3410 uint oreq = _head->req();
3411 for( i = oreq-1; i>0; i-- )
3412 if( !phase->is_member( this, _head->in(i) ) )
3413 landing_pad->set_req(icnt--,_head->in(i));
3414
3415 // Peel off PhiNode edges as well
3416 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3417 Node *oj = _head->fast_out(j);
3418 if( oj->is_Phi() ) {
3419 PhiNode* old_phi = oj->as_Phi();
3420 assert( old_phi->region() == _head, "" );
3421 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3422 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3423 uint icnt = fall_in_cnt;
3424 for( i = oreq-1; i>0; i-- ) {
3425 if( !phase->is_member( this, _head->in(i) ) ) {
3426 p->init_req(icnt--, old_phi->in(i));
3427 // Go ahead and clean out old edges from old phi
3428 old_phi->del_req(i);
3429 }
3430 }
3431 // Search for CSE's here, because ZKM.jar does a lot of
3432 // loop hackery and we need to be a little incremental
3433 // with the CSE to avoid O(N^2) node blow-up.
3434 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3435 if( p2 ) { // Found CSE
3436 p->destruct(&igvn); // Recover useless new node
3437 p = p2; // Use old node
3438 } else {
3439 igvn.register_new_node_with_optimizer(p, old_phi);
3440 }
3441 // Make old Phi refer to new Phi.
3442 old_phi->add_req(p);
3443 // Check for the special case of making the old phi useless and
3444 // disappear it. In JavaGrande I have a case where this useless
3445 // Phi is the loop limit and prevents recognizing a CountedLoop
3446 // which in turn prevents removing an empty loop.
3447 Node *id_old_phi = old_phi->Identity(&igvn);
3448 if( id_old_phi != old_phi ) { // Found a simple identity?
3449 // Note that I cannot call 'replace_node' here, because
3450 // that will yank the edge from old_phi to the Region and
3451 // I'm mid-iteration over the Region's uses.
3452 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3453 Node* use = old_phi->last_out(i);
3454 igvn.rehash_node_delayed(use);
3455 uint uses_found = 0;
3456 for (uint j = 0; j < use->len(); j++) {
3457 if (use->in(j) == old_phi) {
3458 if (j < use->req()) use->set_req (j, id_old_phi);
3459 else use->set_prec(j, id_old_phi);
3460 uses_found++;
3461 }
3462 }
3463 i -= uses_found; // we deleted 1 or more copies of this edge
3464 }
3465 }
3466 igvn._worklist.push(old_phi);
3467 }
3468 }
3469 // Finally clean out the fall-in edges from the RegionNode
3470 for( i = oreq-1; i>0; i-- ) {
3471 if( !phase->is_member( this, _head->in(i) ) ) {
3472 _head->del_req(i);
3473 }
3474 }
3475 igvn.rehash_node_delayed(_head);
3476 // Transform landing pad
3477 igvn.register_new_node_with_optimizer(landing_pad, _head);
3478 // Insert landing pad into the header
3479 _head->add_req(landing_pad);
3480 }
3481
3482 //------------------------------split_outer_loop-------------------------------
3483 // Split out the outermost loop from this shared header.
3484 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3485 PhaseIterGVN &igvn = phase->_igvn;
3486
3487 // Find index of outermost loop; it should also be my tail.
3488 uint outer_idx = 1;
3489 while( _head->in(outer_idx) != _tail ) outer_idx++;
3490
3491 // Make a LoopNode for the outermost loop.
3492 Node *ctl = _head->in(LoopNode::EntryControl);
3493 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3494 outer = igvn.register_new_node_with_optimizer(outer, _head);
3495 phase->set_created_loop_node();
3496
3497 // Outermost loop falls into '_head' loop
3498 _head->set_req(LoopNode::EntryControl, outer);
3499 _head->del_req(outer_idx);
3500 // Split all the Phis up between '_head' loop and 'outer' loop.
3501 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3502 Node *out = _head->fast_out(j);
3503 if( out->is_Phi() ) {
3504 PhiNode *old_phi = out->as_Phi();
3505 assert( old_phi->region() == _head, "" );
3506 Node *phi = PhiNode::make_blank(outer, old_phi);
3507 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3508 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3509 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3510 // Make old Phi point to new Phi on the fall-in path
3511 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3512 old_phi->del_req(outer_idx);
3513 }
3514 }
3515
3516 // Use the new loop head instead of the old shared one
3517 _head = outer;
3518 phase->set_loop(_head, this);
3519 }
3520
3521 //------------------------------fix_parent-------------------------------------
3522 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3523 loop->_parent = parent;
3524 if( loop->_child ) fix_parent( loop->_child, loop );
3525 if( loop->_next ) fix_parent( loop->_next , parent );
3526 }
3527
3528 //------------------------------estimate_path_freq-----------------------------
3529 static float estimate_path_freq( Node *n ) {
3530 // Try to extract some path frequency info
3531 IfNode *iff;
3532 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3533 uint nop = n->Opcode();
3534 if( nop == Op_SafePoint ) { // Skip any safepoint
3535 n = n->in(0);
3536 continue;
3537 }
3538 if( nop == Op_CatchProj ) { // Get count from a prior call
3539 // Assume call does not always throw exceptions: means the call-site
3540 // count is also the frequency of the fall-through path.
3541 assert( n->is_CatchProj(), "" );
3542 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3543 return 0.0f; // Assume call exception path is rare
3544 Node *call = n->in(0)->in(0)->in(0);
3545 assert( call->is_Call(), "expect a call here" );
3546 const JVMState *jvms = ((CallNode*)call)->jvms();
3547 ciMethodData* methodData = jvms->method()->method_data();
3548 if (!methodData->is_mature()) return 0.0f; // No call-site data
3549 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3550 if ((data == nullptr) || !data->is_CounterData()) {
3551 // no call profile available, try call's control input
3552 n = n->in(0);
3553 continue;
3554 }
3555 return data->as_CounterData()->count()/FreqCountInvocations;
3556 }
3557 // See if there's a gating IF test
3558 Node *n_c = n->in(0);
3559 if( !n_c->is_If() ) break; // No estimate available
3560 iff = n_c->as_If();
3561 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3562 // Compute how much count comes on this path
3563 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3564 // Have no count info. Skip dull uncommon-trap like branches.
3565 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3566 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3567 break;
3568 // Skip through never-taken branch; look for a real loop exit.
3569 n = iff->in(0);
3570 }
3571 return 0.0f; // No estimate available
3572 }
3573
3574 //------------------------------merge_many_backedges---------------------------
3575 // Merge all the backedges from the shared header into a private Region.
3576 // Feed that region as the one backedge to this loop.
3577 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3578 uint i;
3579
3580 // Scan for the top 2 hottest backedges
3581 float hotcnt = 0.0f;
3582 float warmcnt = 0.0f;
3583 uint hot_idx = 0;
3584 // Loop starts at 2 because slot 1 is the fall-in path
3585 for( i = 2; i < _head->req(); i++ ) {
3586 float cnt = estimate_path_freq(_head->in(i));
3587 if( cnt > hotcnt ) { // Grab hottest path
3588 warmcnt = hotcnt;
3589 hotcnt = cnt;
3590 hot_idx = i;
3591 } else if( cnt > warmcnt ) { // And 2nd hottest path
3592 warmcnt = cnt;
3593 }
3594 }
3595
3596 // See if the hottest backedge is worthy of being an inner loop
3597 // by being much hotter than the next hottest backedge.
3598 if( hotcnt <= 0.0001 ||
3599 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3600
3601 // Peel out the backedges into a private merge point; peel
3602 // them all except optionally hot_idx.
3603 PhaseIterGVN &igvn = phase->_igvn;
3604
3605 Node *hot_tail = nullptr;
3606 // Make a Region for the merge point
3607 Node *r = new RegionNode(1);
3608 for( i = 2; i < _head->req(); i++ ) {
3609 if( i != hot_idx )
3610 r->add_req( _head->in(i) );
3611 else hot_tail = _head->in(i);
3612 }
3613 igvn.register_new_node_with_optimizer(r, _head);
3614 // Plug region into end of loop _head, followed by hot_tail
3615 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3616 igvn.replace_input_of(_head, 2, r);
3617 if( hot_idx ) _head->add_req(hot_tail);
3618
3619 // Split all the Phis up between '_head' loop and the Region 'r'
3620 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3621 Node *out = _head->fast_out(j);
3622 if( out->is_Phi() ) {
3623 PhiNode* n = out->as_Phi();
3624 igvn.hash_delete(n); // Delete from hash before hacking edges
3625 Node *hot_phi = nullptr;
3626 Node *phi = new PhiNode(r, n->type(), n->adr_type());
3627 // Check all inputs for the ones to peel out
3628 uint j = 1;
3629 for( uint i = 2; i < n->req(); i++ ) {
3630 if( i != hot_idx )
3631 phi->set_req( j++, n->in(i) );
3632 else hot_phi = n->in(i);
3633 }
3634 // Register the phi but do not transform until whole place transforms
3635 igvn.register_new_node_with_optimizer(phi, n);
3636 // Add the merge phi to the old Phi
3637 while( n->req() > 3 ) n->del_req( n->req()-1 );
3638 igvn.replace_input_of(n, 2, phi);
3639 if( hot_idx ) n->add_req(hot_phi);
3640 }
3641 }
3642
3643
3644 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
3645 // of self loop tree. Turn self into a loop headed by _head and with
3646 // tail being the new merge point.
3647 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
3648 phase->set_loop(_tail,ilt); // Adjust tail
3649 _tail = r; // Self's tail is new merge point
3650 phase->set_loop(r,this);
3651 ilt->_child = _child; // New guy has my children
3652 _child = ilt; // Self has new guy as only child
3653 ilt->_parent = this; // new guy has self for parent
3654 ilt->_nest = _nest; // Same nesting depth (for now)
3655
3656 // Starting with 'ilt', look for child loop trees using the same shared
3657 // header. Flatten these out; they will no longer be loops in the end.
3658 IdealLoopTree **pilt = &_child;
3659 while( ilt ) {
3660 if( ilt->_head == _head ) {
3661 uint i;
3662 for( i = 2; i < _head->req(); i++ )
3663 if( _head->in(i) == ilt->_tail )
3664 break; // Still a loop
3665 if( i == _head->req() ) { // No longer a loop
3666 // Flatten ilt. Hang ilt's "_next" list from the end of
3667 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
3668 IdealLoopTree **cp = &ilt->_child;
3669 while( *cp ) cp = &(*cp)->_next; // Find end of child list
3670 *cp = ilt->_next; // Hang next list at end of child list
3671 *pilt = ilt->_child; // Move child up to replace ilt
3672 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
3673 ilt = ilt->_child; // Repeat using new ilt
3674 continue; // do not advance over ilt->_child
3675 }
3676 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
3677 phase->set_loop(_head,ilt);
3678 }
3679 pilt = &ilt->_child; // Advance to next
3680 ilt = *pilt;
3681 }
3682
3683 if( _child ) fix_parent( _child, this );
3684 }
3685
3686 //------------------------------beautify_loops---------------------------------
3687 // Split shared headers and insert loop landing pads.
3688 // Insert a LoopNode to replace the RegionNode.
3689 // Return TRUE if loop tree is structurally changed.
3690 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
3691 bool result = false;
3692 // Cache parts in locals for easy
3693 PhaseIterGVN &igvn = phase->_igvn;
3694
3695 igvn.hash_delete(_head); // Yank from hash before hacking edges
3696
3697 // Check for multiple fall-in paths. Peel off a landing pad if need be.
3698 int fall_in_cnt = 0;
3699 for( uint i = 1; i < _head->req(); i++ )
3700 if( !phase->is_member( this, _head->in(i) ) )
3701 fall_in_cnt++;
3702 assert( fall_in_cnt, "at least 1 fall-in path" );
3703 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
3704 split_fall_in( phase, fall_in_cnt );
3705
3706 // Swap inputs to the _head and all Phis to move the fall-in edge to
3707 // the left.
3708 fall_in_cnt = 1;
3709 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
3710 fall_in_cnt++;
3711 if( fall_in_cnt > 1 ) {
3712 // Since I am just swapping inputs I do not need to update def-use info
3713 Node *tmp = _head->in(1);
3714 igvn.rehash_node_delayed(_head);
3715 _head->set_req( 1, _head->in(fall_in_cnt) );
3716 _head->set_req( fall_in_cnt, tmp );
3717 // Swap also all Phis
3718 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
3719 Node* phi = _head->fast_out(i);
3720 if( phi->is_Phi() ) {
3721 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
3722 tmp = phi->in(1);
3723 phi->set_req( 1, phi->in(fall_in_cnt) );
3724 phi->set_req( fall_in_cnt, tmp );
3725 }
3726 }
3727 }
3728 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
3729 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
3730
3731 // If I am a shared header (multiple backedges), peel off the many
3732 // backedges into a private merge point and use the merge point as
3733 // the one true backedge.
3734 if (_head->req() > 3) {
3735 // Merge the many backedges into a single backedge but leave
3736 // the hottest backedge as separate edge for the following peel.
3737 if (!_irreducible) {
3738 merge_many_backedges( phase );
3739 }
3740
3741 // When recursively beautify my children, split_fall_in can change
3742 // loop tree structure when I am an irreducible loop. Then the head
3743 // of my children has a req() not bigger than 3. Here we need to set
3744 // result to true to catch that case in order to tell the caller to
3745 // rebuild loop tree. See issue JDK-8244407 for details.
3746 result = true;
3747 }
3748
3749 // If I have one hot backedge, peel off myself loop.
3750 // I better be the outermost loop.
3751 if (_head->req() > 3 && !_irreducible) {
3752 split_outer_loop( phase );
3753 result = true;
3754
3755 } else if (!_head->is_Loop() && !_irreducible) {
3756 // Make a new LoopNode to replace the old loop head
3757 Node *l = new LoopNode( _head->in(1), _head->in(2) );
3758 l = igvn.register_new_node_with_optimizer(l, _head);
3759 phase->set_created_loop_node();
3760 // Go ahead and replace _head
3761 phase->_igvn.replace_node( _head, l );
3762 _head = l;
3763 phase->set_loop(_head, this);
3764 }
3765
3766 // Now recursively beautify nested loops
3767 if( _child ) result |= _child->beautify_loops( phase );
3768 if( _next ) result |= _next ->beautify_loops( phase );
3769 return result;
3770 }
3771
3772 //------------------------------allpaths_check_safepts----------------------------
3773 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
3774 // safepoint encountered. Helper for check_safepts.
3775 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
3776 assert(stack.size() == 0, "empty stack");
3777 stack.push(_head);
3778 visited.clear();
3779 visited.set(_head->_idx);
3780 while (stack.size() > 0) {
3781 Node* n = stack.pop();
3782 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3783 // Terminate this path
3784 } else if (n->Opcode() == Op_SafePoint) {
3785 if (_phase->get_loop(n) != this) {
3786 if (_required_safept == nullptr) _required_safept = new Node_List();
3787 // save the first we run into on that path: closest to the tail if the head has a single backedge
3788 _required_safept->push(n);
3789 }
3790 // Terminate this path
3791 } else {
3792 uint start = n->is_Region() ? 1 : 0;
3793 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
3794 for (uint i = start; i < end; i++) {
3795 Node* in = n->in(i);
3796 assert(in->is_CFG(), "must be");
3797 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
3798 stack.push(in);
3799 }
3800 }
3801 }
3802 }
3803 }
3804
3805 //------------------------------check_safepts----------------------------
3806 // Given dominators, try to find loops with calls that must always be
3807 // executed (call dominates loop tail). These loops do not need non-call
3808 // safepoints (ncsfpt).
3809 //
3810 // A complication is that a safepoint in a inner loop may be needed
3811 // by an outer loop. In the following, the inner loop sees it has a
3812 // call (block 3) on every path from the head (block 2) to the
3813 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
3814 // in block 2, _but_ this leaves the outer loop without a safepoint.
3815 //
3816 // entry 0
3817 // |
3818 // v
3819 // outer 1,2 +->1
3820 // | |
3821 // | v
3822 // | 2<---+ ncsfpt in 2
3823 // |_/|\ |
3824 // | v |
3825 // inner 2,3 / 3 | call in 3
3826 // / | |
3827 // v +--+
3828 // exit 4
3829 //
3830 //
3831 // This method creates a list (_required_safept) of ncsfpt nodes that must
3832 // be protected is created for each loop. When a ncsfpt maybe deleted, it
3833 // is first looked for in the lists for the outer loops of the current loop.
3834 //
3835 // The insights into the problem:
3836 // A) counted loops are okay
3837 // B) innermost loops are okay (only an inner loop can delete
3838 // a ncsfpt needed by an outer loop)
3839 // C) a loop is immune from an inner loop deleting a safepoint
3840 // if the loop has a call on the idom-path
3841 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
3842 // idom-path that is not in a nested loop
3843 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
3844 // loop needs to be prevented from deletion by an inner loop
3845 //
3846 // There are two analyses:
3847 // 1) The first, and cheaper one, scans the loop body from
3848 // tail to head following the idom (immediate dominator)
3849 // chain, looking for the cases (C,D,E) above.
3850 // Since inner loops are scanned before outer loops, there is summary
3851 // information about inner loops. Inner loops can be skipped over
3852 // when the tail of an inner loop is encountered.
3853 //
3854 // 2) The second, invoked if the first fails to find a call or ncsfpt on
3855 // the idom path (which is rare), scans all predecessor control paths
3856 // from the tail to the head, terminating a path when a call or sfpt
3857 // is encountered, to find the ncsfpt's that are closest to the tail.
3858 //
3859 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
3860 // Bottom up traversal
3861 IdealLoopTree* ch = _child;
3862 if (_child) _child->check_safepts(visited, stack);
3863 if (_next) _next ->check_safepts(visited, stack);
3864
3865 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
3866 bool has_call = false; // call on dom-path
3867 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
3868 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
3869 if (!_irreducible) {
3870 // Scan the dom-path nodes from tail to head
3871 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
3872 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
3873 has_call = true;
3874 _has_sfpt = 1; // Then no need for a safept!
3875 break;
3876 } else if (n->Opcode() == Op_SafePoint) {
3877 if (_phase->get_loop(n) == this) {
3878 has_local_ncsfpt = true;
3879 break;
3880 }
3881 if (nonlocal_ncsfpt == nullptr) {
3882 nonlocal_ncsfpt = n; // save the one closest to the tail
3883 }
3884 } else {
3885 IdealLoopTree* nlpt = _phase->get_loop(n);
3886 if (this != nlpt) {
3887 // If at an inner loop tail, see if the inner loop has already
3888 // recorded seeing a call on the dom-path (and stop.) If not,
3889 // jump to the head of the inner loop.
3890 assert(is_member(nlpt), "nested loop");
3891 Node* tail = nlpt->_tail;
3892 if (tail->in(0)->is_If()) tail = tail->in(0);
3893 if (n == tail) {
3894 // If inner loop has call on dom-path, so does outer loop
3895 if (nlpt->_has_sfpt) {
3896 has_call = true;
3897 _has_sfpt = 1;
3898 break;
3899 }
3900 // Skip to head of inner loop
3901 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
3902 n = nlpt->_head;
3903 if (_head == n) {
3904 // this and nlpt (inner loop) have the same loop head. This should not happen because
3905 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
3906 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
3907 // but then attached in the build_loop_tree afterwards, and so still have unmerged
3908 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
3909 // since we have reached the loop head of this.
3910 assert(_head->as_Region()->is_in_infinite_subgraph(),
3911 "only expect unmerged backedges in infinite loops");
3912 break;
3913 }
3914 }
3915 }
3916 }
3917 }
3918 }
3919 // Record safept's that this loop needs preserved when an
3920 // inner loop attempts to delete it's safepoints.
3921 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
3922 if (nonlocal_ncsfpt != nullptr) {
3923 if (_required_safept == nullptr) _required_safept = new Node_List();
3924 _required_safept->push(nonlocal_ncsfpt);
3925 } else {
3926 // Failed to find a suitable safept on the dom-path. Now use
3927 // an all paths walk from tail to head, looking for safepoints to preserve.
3928 allpaths_check_safepts(visited, stack);
3929 }
3930 }
3931 }
3932 }
3933
3934 //---------------------------is_deleteable_safept----------------------------
3935 // Is safept not required by an outer loop?
3936 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
3937 assert(sfpt->Opcode() == Op_SafePoint, "");
3938 IdealLoopTree* lp = get_loop(sfpt)->_parent;
3939 while (lp != nullptr) {
3940 Node_List* sfpts = lp->_required_safept;
3941 if (sfpts != nullptr) {
3942 for (uint i = 0; i < sfpts->size(); i++) {
3943 if (sfpt == sfpts->at(i))
3944 return false;
3945 }
3946 }
3947 lp = lp->_parent;
3948 }
3949 return true;
3950 }
3951
3952 //---------------------------replace_parallel_iv-------------------------------
3953 // Replace parallel induction variable (parallel to trip counter)
3954 // This optimization looks for patterns similar to:
3955 //
3956 // int a = init2;
3957 // for (int iv = init; iv < limit; iv += stride_con) {
3958 // a += stride_con2;
3959 // }
3960 //
3961 // and transforms it to:
3962 //
3963 // int iv2 = init2
3964 // int iv = init
3965 // loop:
3966 // if (iv >= limit) goto exit
3967 // iv += stride_con
3968 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
3969 // goto loop
3970 // exit:
3971 // ...
3972 //
3973 // Such transformation introduces more optimization opportunities. In this
3974 // particular example, the loop can be eliminated entirely given that
3975 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
3976 // place to only perform this optimization if such a division is exact. This
3977 // example will be transformed into its semantic equivalence:
3978 //
3979 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
3980 //
3981 // which corresponds to the structure of transformed subgraph.
3982 //
3983 // However, if there is a mismatch between types of the loop and the parallel
3984 // induction variable (e.g., a long-typed IV in an int-typed loop), type
3985 // conversions are required:
3986 //
3987 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
3988 //
3989 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
3990 assert(loop->_head->is_CountedLoop(), "");
3991 CountedLoopNode *cl = loop->_head->as_CountedLoop();
3992 if (!cl->is_valid_counted_loop(T_INT)) {
3993 return; // skip malformed counted loop
3994 }
3995 Node *incr = cl->incr();
3996 if (incr == nullptr) {
3997 return; // Dead loop?
3998 }
3999 Node *init = cl->init_trip();
4000 Node *phi = cl->phi();
4001 jlong stride_con = cl->stride_con();
4002
4003 // Visit all children, looking for Phis
4004 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
4005 Node *out = cl->out(i);
4006 // Look for other phis (secondary IVs). Skip dead ones
4007 if (!out->is_Phi() || out == phi || !has_node(out)) {
4008 continue;
4009 }
4010
4011 PhiNode* phi2 = out->as_Phi();
4012 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4013 // Look for induction variables of the form: X += constant
4014 if (phi2->region() != loop->_head ||
4015 incr2->req() != 3 ||
4016 incr2->in(1)->uncast() != phi2 ||
4017 incr2 == incr ||
4018 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4019 !incr2->in(2)->is_Con()) {
4020 continue;
4021 }
4022
4023 if (incr2->in(1)->is_ConstraintCast() &&
4024 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4025 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4026 continue;
4027 }
4028 // Check for parallel induction variable (parallel to trip counter)
4029 // via an affine function. In particular, count-down loops with
4030 // count-up array indices are common. We only RCE references off
4031 // the trip-counter, so we need to convert all these to trip-counter
4032 // expressions.
4033 Node* init2 = phi2->in(LoopNode::EntryControl);
4034
4035 // Determine the basic type of the stride constant (and the iv being incremented).
4036 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4037 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4038
4039 // The ratio of the two strides cannot be represented as an int
4040 // if stride_con2 is min_jint (or min_jlong, respectively) and
4041 // stride_con is -1.
4042 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4043 continue;
4044 }
4045
4046 // The general case here gets a little tricky. We want to find the
4047 // GCD of all possible parallel IV's and make a new IV using this
4048 // GCD for the loop. Then all possible IVs are simple multiples of
4049 // the GCD. In practice, this will cover very few extra loops.
4050 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4051 // where +/-1 is the common case, but other integer multiples are
4052 // also easy to handle.
4053 jlong ratio_con = stride_con2 / stride_con;
4054
4055 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4056 continue;
4057 }
4058
4059 #ifndef PRODUCT
4060 if (TraceLoopOpts) {
4061 tty->print("Parallel IV: %d ", phi2->_idx);
4062 loop->dump_head();
4063 }
4064 #endif
4065
4066 // Convert to using the trip counter. The parallel induction
4067 // variable differs from the trip counter by a loop-invariant
4068 // amount, the difference between their respective initial values.
4069 // It is scaled by the 'ratio_con'.
4070 Node* ratio = _igvn.integercon(ratio_con, stride_con2_bt);
4071 set_ctrl(ratio, C->root());
4072
4073 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4074 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4075
4076 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4077 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4078 set_early_ctrl(ratio_init, false);
4079
4080 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4081 _igvn.register_new_node_with_optimizer(diff, init2);
4082 set_early_ctrl(diff, false);
4083
4084 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4085 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4086 set_ctrl(ratio_idx, cl);
4087
4088 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4089 _igvn.register_new_node_with_optimizer(add);
4090 set_ctrl(add, cl);
4091
4092 _igvn.replace_node( phi2, add );
4093 // Sometimes an induction variable is unused
4094 if (add->outcnt() == 0) {
4095 _igvn.remove_dead_node(add);
4096 }
4097 --i; // deleted this phi; rescan starting with next position
4098 }
4099 }
4100
4101 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4102 BasicType source = _igvn.type(input)->basic_type();
4103 if (source == target) {
4104 return input;
4105 }
4106
4107 Node* converted = ConvertNode::create_convert(source, target, input);
4108 _igvn.register_new_node_with_optimizer(converted, input);
4109 set_early_ctrl(converted, false);
4110
4111 return converted;
4112 }
4113
4114 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4115 Node* keep = nullptr;
4116 if (keep_one) {
4117 // Look for a safepoint on the idom-path.
4118 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4119 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4120 keep = i;
4121 break; // Found one
4122 }
4123 }
4124 }
4125
4126 // Don't remove any safepoints if it is requested to keep a single safepoint and
4127 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4128 // in this case since there's no safepoint dominating all paths in the loop body.
4129 bool prune = !keep_one || keep != nullptr;
4130
4131 // Delete other safepoints in this loop.
4132 Node_List* sfpts = _safepts;
4133 if (prune && sfpts != nullptr) {
4134 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4135 for (uint i = 0; i < sfpts->size(); i++) {
4136 Node* n = sfpts->at(i);
4137 assert(phase->get_loop(n) == this, "");
4138 if (n != keep && phase->is_deleteable_safept(n)) {
4139 phase->lazy_replace(n, n->in(TypeFunc::Control));
4140 }
4141 }
4142 }
4143 }
4144
4145 //------------------------------counted_loop-----------------------------------
4146 // Convert to counted loops where possible
4147 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4148
4149 // For grins, set the inner-loop flag here
4150 if (!_child) {
4151 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4152 }
4153
4154 IdealLoopTree* loop = this;
4155 if (_head->is_CountedLoop() ||
4156 phase->is_counted_loop(_head, loop, T_INT)) {
4157
4158 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4159 // Indicate we do not need a safepoint here
4160 _has_sfpt = 1;
4161 }
4162
4163 // Remove safepoints
4164 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4165 remove_safepoints(phase, keep_one_sfpt);
4166
4167 // Look for induction variables
4168 phase->replace_parallel_iv(this);
4169 } else if (_head->is_LongCountedLoop() ||
4170 phase->is_counted_loop(_head, loop, T_LONG)) {
4171 remove_safepoints(phase, true);
4172 } else {
4173 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4174 if (_parent != nullptr && !_irreducible) {
4175 // Not a counted loop. Keep one safepoint.
4176 bool keep_one_sfpt = true;
4177 remove_safepoints(phase, keep_one_sfpt);
4178 }
4179 }
4180
4181 // Recursively
4182 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4183 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4184 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4185 if (loop->_next) loop->_next ->counted_loop(phase);
4186 }
4187
4188
4189 // The Estimated Loop Clone Size:
4190 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4191 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4192 // respectively, used to ensure that the node usage estimates made are on the
4193 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4194 // possible additional/excessive nodes generated due to data and control flow
4195 // merging, for edges reaching outside the loop.
4196 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4197
4198 precond(0 < factor && factor < 16);
4199
4200 uint const bc = 13;
4201 uint const cc = 17;
4202 uint const sz = _body.size() + (_body.size() + 7) / 2;
4203 uint estimate = factor * (sz + bc) + cc;
4204
4205 assert((estimate - cc) / factor == sz + bc, "overflow");
4206
4207 return estimate + est_loop_flow_merge_sz();
4208 }
4209
4210 // The Estimated Loop (full-) Unroll Size:
4211 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4212 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4213 // node usage estimates made are on the safe side, for the most part. This is
4214 // a "light" version of the loop clone size calculation (above), based on the
4215 // assumption that most of the loop-construct overhead will be unraveled when
4216 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4217 // including an overflow check and returning UINT_MAX in case of an overflow.
4218 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4219
4220 precond(factor > 0);
4221
4222 // Take into account that after unroll conjoined heads and tails will fold.
4223 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4224 uint const cc = 7;
4225 uint const sz = b0 + (b0 + 15) / 16;
4226 uint estimate = factor * sz + cc;
4227
4228 if ((estimate - cc) / factor != sz) {
4229 return UINT_MAX;
4230 }
4231
4232 return estimate + est_loop_flow_merge_sz();
4233 }
4234
4235 // Estimate the growth effect (in nodes) of merging control and data flow when
4236 // cloning a loop body, based on the amount of control and data flow reaching
4237 // outside of the (current) loop body.
4238 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4239
4240 uint ctrl_edge_out_cnt = 0;
4241 uint data_edge_out_cnt = 0;
4242
4243 for (uint i = 0; i < _body.size(); i++) {
4244 Node* node = _body.at(i);
4245 uint outcnt = node->outcnt();
4246
4247 for (uint k = 0; k < outcnt; k++) {
4248 Node* out = node->raw_out(k);
4249 if (out == nullptr) continue;
4250 if (out->is_CFG()) {
4251 if (!is_member(_phase->get_loop(out))) {
4252 ctrl_edge_out_cnt++;
4253 }
4254 } else if (_phase->has_ctrl(out)) {
4255 Node* ctrl = _phase->get_ctrl(out);
4256 assert(ctrl != nullptr, "must be");
4257 assert(ctrl->is_CFG(), "must be");
4258 if (!is_member(_phase->get_loop(ctrl))) {
4259 data_edge_out_cnt++;
4260 }
4261 }
4262 }
4263 }
4264 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4265 // a rather pessimistic estimate for the most part, in particular for some
4266 // complex loops, but still not enough to capture all loops.
4267 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4268 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4269 }
4270 return 0;
4271 }
4272
4273 #ifndef PRODUCT
4274 //------------------------------dump_head--------------------------------------
4275 // Dump 1 liner for loop header info
4276 void IdealLoopTree::dump_head() {
4277 tty->sp(2 * _nest);
4278 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4279 if (_irreducible) tty->print(" IRREDUCIBLE");
4280 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4281 : _head->in(LoopNode::EntryControl);
4282 const Predicates predicates(entry);
4283 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4284 tty->print(" limit_check");
4285 }
4286 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4287 tty->print(" profile_predicated");
4288 }
4289 if (UseLoopPredicate && predicates.loop_predicate_block()->is_non_empty()) {
4290 tty->print(" predicated");
4291 }
4292 if (_head->is_CountedLoop()) {
4293 CountedLoopNode *cl = _head->as_CountedLoop();
4294 tty->print(" counted");
4295
4296 Node* init_n = cl->init_trip();
4297 if (init_n != nullptr && init_n->is_Con())
4298 tty->print(" [%d,", cl->init_trip()->get_int());
4299 else
4300 tty->print(" [int,");
4301 Node* limit_n = cl->limit();
4302 if (limit_n != nullptr && limit_n->is_Con())
4303 tty->print("%d),", cl->limit()->get_int());
4304 else
4305 tty->print("int),");
4306 int stride_con = cl->stride_con();
4307 if (stride_con > 0) tty->print("+");
4308 tty->print("%d", stride_con);
4309
4310 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4311
4312 if (cl->is_pre_loop ()) tty->print(" pre" );
4313 if (cl->is_main_loop()) tty->print(" main");
4314 if (cl->is_post_loop()) tty->print(" post");
4315 if (cl->is_vectorized_loop()) tty->print(" vector");
4316 if (range_checks_present()) tty->print(" rc ");
4317 }
4318 if (_has_call) tty->print(" has_call");
4319 if (_has_sfpt) tty->print(" has_sfpt");
4320 if (_rce_candidate) tty->print(" rce");
4321 if (_safepts != nullptr && _safepts->size() > 0) {
4322 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4323 }
4324 if (_required_safept != nullptr && _required_safept->size() > 0) {
4325 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4326 }
4327 if (Verbose) {
4328 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4329 }
4330 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4331 tty->print(" strip_mined");
4332 }
4333 tty->cr();
4334 }
4335
4336 //------------------------------dump-------------------------------------------
4337 // Dump loops by loop tree
4338 void IdealLoopTree::dump() {
4339 dump_head();
4340 if (_child) _child->dump();
4341 if (_next) _next ->dump();
4342 }
4343
4344 #endif
4345
4346 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4347 if (loop == root) {
4348 if (loop->_child != nullptr) {
4349 log->begin_head("loop_tree");
4350 log->end_head();
4351 log_loop_tree_helper(root, loop->_child, log);
4352 log->tail("loop_tree");
4353 assert(loop->_next == nullptr, "what?");
4354 }
4355 } else if (loop != nullptr) {
4356 Node* head = loop->_head;
4357 log->begin_head("loop");
4358 log->print(" idx='%d' ", head->_idx);
4359 if (loop->_irreducible) log->print("irreducible='1' ");
4360 if (head->is_Loop()) {
4361 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4362 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4363 } else if (head->is_CountedLoop()) {
4364 CountedLoopNode* cl = head->as_CountedLoop();
4365 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4366 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4367 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4368 }
4369 log->end_head();
4370 log_loop_tree_helper(root, loop->_child, log);
4371 log->tail("loop");
4372 log_loop_tree_helper(root, loop->_next, log);
4373 }
4374 }
4375
4376 void PhaseIdealLoop::log_loop_tree() {
4377 if (C->log() != nullptr) {
4378 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4379 }
4380 }
4381
4382 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4383 // predicates will be removed during the next round of IGVN.
4384 void PhaseIdealLoop::eliminate_useless_predicates() {
4385 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4386 return; // No predicates left.
4387 }
4388
4389 eliminate_useless_parse_predicates();
4390 eliminate_useless_template_assertion_predicates();
4391 }
4392
4393 // Eliminate all Parse Predicates that do not belong to a loop anymore by marking them useless. These will be removed
4394 // during the next round of IGVN.
4395 void PhaseIdealLoop::eliminate_useless_parse_predicates() {
4396 mark_all_parse_predicates_useless();
4397 if (C->has_loops()) {
4398 mark_loop_associated_parse_predicates_useful();
4399 }
4400 add_useless_parse_predicates_to_igvn_worklist();
4401 }
4402
4403 void PhaseIdealLoop::mark_all_parse_predicates_useless() const {
4404 for (int i = 0; i < C->parse_predicate_count(); i++) {
4405 C->parse_predicate(i)->mark_useless();
4406 }
4407 }
4408
4409 void PhaseIdealLoop::mark_loop_associated_parse_predicates_useful() {
4410 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4411 IdealLoopTree* loop = iterator.current();
4412 if (loop->can_apply_loop_predication()) {
4413 mark_useful_parse_predicates_for_loop(loop);
4414 }
4415 }
4416 }
4417
4418 // This visitor marks all visited Parse Predicates useful.
4419 class ParsePredicateUsefulMarker : public PredicateVisitor {
4420 public:
4421 using PredicateVisitor::visit;
4422
4423 void visit(const ParsePredicate& parse_predicate) override {
4424 parse_predicate.head()->mark_useful();
4425 }
4426 };
4427
4428 void PhaseIdealLoop::mark_useful_parse_predicates_for_loop(IdealLoopTree* loop) {
4429 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4430 const PredicateIterator predicate_iterator(entry);
4431 ParsePredicateUsefulMarker useful_marker;
4432 predicate_iterator.for_each(useful_marker);
4433 }
4434
4435 void PhaseIdealLoop::add_useless_parse_predicates_to_igvn_worklist() {
4436 for (int i = 0; i < C->parse_predicate_count(); i++) {
4437 ParsePredicateNode* parse_predicate_node = C->parse_predicate(i);
4438 if (parse_predicate_node->is_useless()) {
4439 _igvn._worklist.push(parse_predicate_node);
4440 }
4441 }
4442 }
4443
4444
4445 // Eliminate all Template Assertion Predicates that do not belong to their originally associated loop anymore by
4446 // replacing the OpaqueTemplateAssertionPredicate node of the If node with true. These nodes will be removed during the
4447 // next round of IGVN.
4448 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates() {
4449 Unique_Node_List useful_predicates;
4450 if (C->has_loops()) {
4451 collect_useful_template_assertion_predicates(useful_predicates);
4452 }
4453 eliminate_useless_template_assertion_predicates(useful_predicates);
4454 }
4455
4456 void PhaseIdealLoop::collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4457 for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {
4458 IdealLoopTree* loop = iterator.current();
4459 if (loop->can_apply_loop_predication()) {
4460 collect_useful_template_assertion_predicates_for_loop(loop, useful_predicates);
4461 }
4462 }
4463 }
4464
4465 void PhaseIdealLoop::collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop,
4466 Unique_Node_List &useful_predicates) {
4467 Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4468 const Predicates predicates(entry);
4469 if (UseProfiledLoopPredicate) {
4470 const PredicateBlock* profiled_loop_predicate_block = predicates.profiled_loop_predicate_block();
4471 if (profiled_loop_predicate_block->has_parse_predicate()) {
4472 ParsePredicateSuccessProj* parse_predicate_proj = profiled_loop_predicate_block->parse_predicate_success_proj();
4473 get_template_assertion_predicates(parse_predicate_proj, useful_predicates, true);
4474 }
4475 }
4476
4477 if (UseLoopPredicate) {
4478 const PredicateBlock* loop_predicate_block = predicates.loop_predicate_block();
4479 if (loop_predicate_block->has_parse_predicate()) {
4480 ParsePredicateSuccessProj* parse_predicate_proj = loop_predicate_block->parse_predicate_success_proj();
4481 get_template_assertion_predicates(parse_predicate_proj, useful_predicates, true);
4482 }
4483 }
4484 }
4485
4486 void PhaseIdealLoop::eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates) {
4487 for (int i = C->template_assertion_predicate_count(); i > 0; i--) {
4488 OpaqueTemplateAssertionPredicateNode* opaque_node =
4489 C->template_assertion_predicate_opaq_node(i - 1)->as_OpaqueTemplateAssertionPredicate();
4490 if (!useful_predicates.member(opaque_node)) { // not in the useful list
4491 ConINode* one = _igvn.intcon(1);
4492 set_ctrl(one, C->root());
4493 _igvn.replace_node(opaque_node, one);
4494 }
4495 }
4496 }
4497
4498 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4499 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4500 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4501 return;
4502 }
4503 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4504 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4505 IdealLoopTree* lpt = iter.current();
4506 if (lpt->_child == nullptr && lpt->is_counted()) {
4507 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4508 Node* opaque = head->is_canonical_loop_entry();
4509 if (opaque != nullptr) {
4510 useful_zero_trip_guard_opaques_nodes.push(opaque);
4511 }
4512 }
4513 }
4514 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4515 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4516 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4517 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4518 IfNode* iff = opaque->if_node();
4519 IdealLoopTree* loop = get_loop(iff);
4520 while (loop != _ltree_root && loop != nullptr) {
4521 loop = loop->_parent;
4522 }
4523 if (loop == nullptr) {
4524 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4525 // We can't tell if the opaque node is useful or not
4526 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4527 } else {
4528 assert(guarded_loop == nullptr, "");
4529 this->_igvn.replace_node(opaque, opaque->in(1));
4530 }
4531 } else {
4532 assert(guarded_loop != nullptr, "");
4533 }
4534 }
4535 }
4536
4537 //------------------------process_expensive_nodes-----------------------------
4538 // Expensive nodes have their control input set to prevent the GVN
4539 // from commoning them and as a result forcing the resulting node to
4540 // be in a more frequent path. Use CFG information here, to change the
4541 // control inputs so that some expensive nodes can be commoned while
4542 // not executed more frequently.
4543 bool PhaseIdealLoop::process_expensive_nodes() {
4544 assert(OptimizeExpensiveOps, "optimization off?");
4545
4546 // Sort nodes to bring similar nodes together
4547 C->sort_expensive_nodes();
4548
4549 bool progress = false;
4550
4551 for (int i = 0; i < C->expensive_count(); ) {
4552 Node* n = C->expensive_node(i);
4553 int start = i;
4554 // Find nodes similar to n
4555 i++;
4556 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4557 int end = i;
4558 // And compare them two by two
4559 for (int j = start; j < end; j++) {
4560 Node* n1 = C->expensive_node(j);
4561 if (is_node_unreachable(n1)) {
4562 continue;
4563 }
4564 for (int k = j+1; k < end; k++) {
4565 Node* n2 = C->expensive_node(k);
4566 if (is_node_unreachable(n2)) {
4567 continue;
4568 }
4569
4570 assert(n1 != n2, "should be pair of nodes");
4571
4572 Node* c1 = n1->in(0);
4573 Node* c2 = n2->in(0);
4574
4575 Node* parent_c1 = c1;
4576 Node* parent_c2 = c2;
4577
4578 // The call to get_early_ctrl_for_expensive() moves the
4579 // expensive nodes up but stops at loops that are in a if
4580 // branch. See whether we can exit the loop and move above the
4581 // If.
4582 if (c1->is_Loop()) {
4583 parent_c1 = c1->in(1);
4584 }
4585 if (c2->is_Loop()) {
4586 parent_c2 = c2->in(1);
4587 }
4588
4589 if (parent_c1 == parent_c2) {
4590 _igvn._worklist.push(n1);
4591 _igvn._worklist.push(n2);
4592 continue;
4593 }
4594
4595 // Look for identical expensive node up the dominator chain.
4596 if (is_dominator(c1, c2)) {
4597 c2 = c1;
4598 } else if (is_dominator(c2, c1)) {
4599 c1 = c2;
4600 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4601 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4602 // Both branches have the same expensive node so move it up
4603 // before the if.
4604 c1 = c2 = idom(parent_c1->in(0));
4605 }
4606 // Do the actual moves
4607 if (n1->in(0) != c1) {
4608 _igvn.replace_input_of(n1, 0, c1);
4609 progress = true;
4610 }
4611 if (n2->in(0) != c2) {
4612 _igvn.replace_input_of(n2, 0, c2);
4613 progress = true;
4614 }
4615 }
4616 }
4617 }
4618
4619 return progress;
4620 }
4621
4622 //=============================================================================
4623 //----------------------------build_and_optimize-------------------------------
4624 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4625 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4626 void PhaseIdealLoop::build_and_optimize() {
4627 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4628
4629 bool do_split_ifs = (_mode == LoopOptsDefault);
4630 bool skip_loop_opts = (_mode == LoopOptsNone);
4631 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4632
4633
4634 int old_progress = C->major_progress();
4635 uint orig_worklist_size = _igvn._worklist.size();
4636
4637 // Reset major-progress flag for the driver's heuristics
4638 C->clear_major_progress();
4639
4640 #ifndef PRODUCT
4641 // Capture for later assert
4642 uint unique = C->unique();
4643 _loop_invokes++;
4644 _loop_work += unique;
4645 #endif
4646
4647 // True if the method has at least 1 irreducible loop
4648 _has_irreducible_loops = false;
4649
4650 _created_loop_node = false;
4651
4652 VectorSet visited;
4653 // Pre-grow the mapping from Nodes to IdealLoopTrees.
4654 _loop_or_ctrl.map(C->unique(), nullptr);
4655 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
4656
4657 // Pre-build the top-level outermost loop tree entry
4658 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
4659 // Do not need a safepoint at the top level
4660 _ltree_root->_has_sfpt = 1;
4661
4662 // Initialize Dominators.
4663 // Checked in clone_loop_predicate() during beautify_loops().
4664 _idom_size = 0;
4665 _idom = nullptr;
4666 _dom_depth = nullptr;
4667 _dom_stk = nullptr;
4668
4669 // Empty pre-order array
4670 allocate_preorders();
4671
4672 // Build a loop tree on the fly. Build a mapping from CFG nodes to
4673 // IdealLoopTree entries. Data nodes are NOT walked.
4674 build_loop_tree();
4675 // Check for bailout, and return
4676 if (C->failing()) {
4677 return;
4678 }
4679
4680 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
4681 // is an exception edge, parsing doesn't set has_loops().
4682 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
4683 // No loops after all
4684 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
4685
4686 // There should always be an outer loop containing the Root and Return nodes.
4687 // If not, we have a degenerate empty program. Bail out in this case.
4688 if (!has_node(C->root())) {
4689 if (!_verify_only) {
4690 C->clear_major_progress();
4691 assert(false, "empty program detected during loop optimization");
4692 C->record_method_not_compilable("empty program detected during loop optimization");
4693 }
4694 return;
4695 }
4696
4697 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4698 // Nothing to do, so get out
4699 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
4700 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
4701 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
4702 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
4703 if (stop_early && !do_expensive_nodes) {
4704 return;
4705 }
4706
4707 // Set loop nesting depth
4708 _ltree_root->set_nest( 0 );
4709
4710 // Split shared headers and insert loop landing pads.
4711 // Do not bother doing this on the Root loop of course.
4712 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
4713 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
4714 if( _ltree_root->_child->beautify_loops( this ) ) {
4715 // Re-build loop tree!
4716 _ltree_root->_child = nullptr;
4717 _loop_or_ctrl.clear();
4718 reallocate_preorders();
4719 build_loop_tree();
4720 // Check for bailout, and return
4721 if (C->failing()) {
4722 return;
4723 }
4724 // Reset loop nesting depth
4725 _ltree_root->set_nest( 0 );
4726
4727 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
4728 }
4729 }
4730
4731 // Build Dominators for elision of null checks & loop finding.
4732 // Since nodes do not have a slot for immediate dominator, make
4733 // a persistent side array for that info indexed on node->_idx.
4734 _idom_size = C->unique();
4735 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
4736 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
4737 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
4738 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
4739
4740 Dominators();
4741
4742 if (!_verify_only) {
4743 // As a side effect, Dominators removed any unreachable CFG paths
4744 // into RegionNodes. It doesn't do this test against Root, so
4745 // we do it here.
4746 for( uint i = 1; i < C->root()->req(); i++ ) {
4747 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
4748 _igvn.delete_input_of(C->root(), i);
4749 i--; // Rerun same iteration on compressed edges
4750 }
4751 }
4752
4753 // Given dominators, try to find inner loops with calls that must
4754 // always be executed (call dominates loop tail). These loops do
4755 // not need a separate safepoint.
4756 Node_List cisstack;
4757 _ltree_root->check_safepts(visited, cisstack);
4758 }
4759
4760 // Walk the DATA nodes and place into loops. Find earliest control
4761 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
4762 // holding the associated IdealLoopTree pointer. For DATA nodes, the
4763 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
4764
4765 // Allocate stack with enough space to avoid frequent realloc
4766 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
4767 Node_Stack nstack(stack_size);
4768
4769 visited.clear();
4770 Node_List worklist;
4771 // Don't need C->root() on worklist since
4772 // it will be processed among C->top() inputs
4773 worklist.push(C->top());
4774 visited.set(C->top()->_idx); // Set C->top() as visited now
4775 build_loop_early( visited, worklist, nstack );
4776
4777 // Given early legal placement, try finding counted loops. This placement
4778 // is good enough to discover most loop invariants.
4779 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
4780 _ltree_root->counted_loop( this );
4781 }
4782
4783 // Find latest loop placement. Find ideal loop placement.
4784 visited.clear();
4785 init_dom_lca_tags();
4786 // Need C->root() on worklist when processing outs
4787 worklist.push(C->root());
4788 NOT_PRODUCT( C->verify_graph_edges(); )
4789 worklist.push(C->top());
4790 build_loop_late( visited, worklist, nstack );
4791 if (C->failing()) { return; }
4792
4793 if (_verify_only) {
4794 C->restore_major_progress(old_progress);
4795 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
4796 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
4797 return;
4798 }
4799
4800 // clear out the dead code after build_loop_late
4801 while (_deadlist.size()) {
4802 _igvn.remove_globally_dead_node(_deadlist.pop());
4803 }
4804
4805 eliminate_useless_zero_trip_guard();
4806
4807 if (stop_early) {
4808 assert(do_expensive_nodes, "why are we here?");
4809 if (process_expensive_nodes()) {
4810 // If we made some progress when processing expensive nodes then
4811 // the IGVN may modify the graph in a way that will allow us to
4812 // make some more progress: we need to try processing expensive
4813 // nodes again.
4814 C->set_major_progress();
4815 }
4816 return;
4817 }
4818
4819 // Some parser-inserted loop predicates could never be used by loop
4820 // predication or they were moved away from loop during some optimizations.
4821 // For example, peeling. Eliminate them before next loop optimizations.
4822 eliminate_useless_predicates();
4823
4824 #ifndef PRODUCT
4825 C->verify_graph_edges();
4826 if (_verify_me) { // Nested verify pass?
4827 // Check to see if the verify _mode is broken
4828 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
4829 return;
4830 }
4831 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4832 if (TraceLoopOpts && C->has_loops()) {
4833 _ltree_root->dump();
4834 }
4835 #endif
4836
4837 if (skip_loop_opts) {
4838 C->restore_major_progress(old_progress);
4839 return;
4840 }
4841
4842 if (do_max_unroll) {
4843 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4844 IdealLoopTree* lpt = iter.current();
4845 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
4846 lpt->compute_trip_count(this);
4847 if (!lpt->do_one_iteration_loop(this) &&
4848 !lpt->do_remove_empty_loop(this)) {
4849 AutoNodeBudget node_budget(this);
4850 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
4851 lpt->policy_maximally_unroll(this)) {
4852 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4853 do_maximally_unroll(lpt, worklist);
4854 }
4855 }
4856 }
4857 }
4858
4859 C->restore_major_progress(old_progress);
4860 return;
4861 }
4862
4863 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
4864 return;
4865 }
4866
4867 if (ReassociateInvariants && !C->major_progress()) {
4868 // Reassociate invariants and prep for split_thru_phi
4869 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4870 IdealLoopTree* lpt = iter.current();
4871 if (!lpt->is_loop()) {
4872 continue;
4873 }
4874 Node* head = lpt->_head;
4875 if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
4876
4877 // check for vectorized loops, any reassociation of invariants was already done
4878 if (head->is_CountedLoop()) {
4879 if (head->as_CountedLoop()->is_unroll_only()) {
4880 continue;
4881 } else {
4882 AutoNodeBudget node_budget(this);
4883 lpt->reassociate_invariants(this);
4884 }
4885 }
4886 // Because RCE opportunities can be masked by split_thru_phi,
4887 // look for RCE candidates and inhibit split_thru_phi
4888 // on just their loop-phi's for this pass of loop opts
4889 if (SplitIfBlocks && do_split_ifs &&
4890 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
4891 (lpt->policy_range_check(this, true, T_LONG) ||
4892 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
4893 lpt->_rce_candidate = 1; // = true
4894 }
4895 }
4896 }
4897
4898 // Check for aggressive application of split-if and other transforms
4899 // that require basic-block info (like cloning through Phi's)
4900 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
4901 visited.clear();
4902 split_if_with_blocks( visited, nstack);
4903 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
4904 }
4905
4906 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
4907 C->set_major_progress();
4908 }
4909
4910 // Perform loop predication before iteration splitting
4911 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
4912 _ltree_root->_child->loop_predication(this);
4913 }
4914
4915 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
4916 if (do_intrinsify_fill()) {
4917 C->set_major_progress();
4918 }
4919 }
4920
4921 // Perform iteration-splitting on inner loops. Split iterations to avoid
4922 // range checks or one-shot null checks.
4923
4924 // If split-if's didn't hack the graph too bad (no CFG changes)
4925 // then do loop opts.
4926 if (C->has_loops() && !C->major_progress()) {
4927 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4928 _ltree_root->_child->iteration_split( this, worklist );
4929 // No verify after peeling! GCM has hoisted code out of the loop.
4930 // After peeling, the hoisted code could sink inside the peeled area.
4931 // The peeling code does not try to recompute the best location for
4932 // all the code before the peeled area, so the verify pass will always
4933 // complain about it.
4934 }
4935
4936 // Check for bailout, and return
4937 if (C->failing()) {
4938 return;
4939 }
4940
4941 // Do verify graph edges in any case
4942 NOT_PRODUCT( C->verify_graph_edges(); );
4943
4944 if (!do_split_ifs) {
4945 // We saw major progress in Split-If to get here. We forced a
4946 // pass with unrolling and not split-if, however more split-if's
4947 // might make progress. If the unrolling didn't make progress
4948 // then the major-progress flag got cleared and we won't try
4949 // another round of Split-If. In particular the ever-common
4950 // instance-of/check-cast pattern requires at least 2 rounds of
4951 // Split-If to clear out.
4952 C->set_major_progress();
4953 }
4954
4955 // Repeat loop optimizations if new loops were seen
4956 if (created_loop_node()) {
4957 C->set_major_progress();
4958 }
4959
4960 // Keep loop predicates and perform optimizations with them
4961 // until no more loop optimizations could be done.
4962 // After that switch predicates off and do more loop optimizations.
4963 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
4964 C->mark_parse_predicate_nodes_useless(_igvn);
4965 assert(C->parse_predicate_count() == 0, "should be zero now");
4966 if (TraceLoopOpts) {
4967 tty->print_cr("PredicatesOff");
4968 }
4969 C->set_major_progress();
4970 }
4971
4972 // Auto-vectorize main-loop
4973 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
4974 Compile::TracePhase tp(_t_autoVectorize);
4975
4976 // Shared data structures for all AutoVectorizations, to reduce allocations
4977 // of large arrays.
4978 VSharedData vshared;
4979 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4980 IdealLoopTree* lpt = iter.current();
4981 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
4982
4983 if (status == AutoVectorizeStatus::TriedAndFailed) {
4984 // We tried vectorization, but failed. From now on only unroll the loop.
4985 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4986 if (cl->has_passed_slp()) {
4987 C->set_major_progress();
4988 cl->set_notpassed_slp();
4989 cl->mark_do_unroll_only();
4990 }
4991 }
4992 }
4993 }
4994
4995 // Move UnorderedReduction out of counted loop. Can be introduced by AutoVectorization.
4996 if (C->has_loops() && !C->major_progress()) {
4997 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4998 IdealLoopTree* lpt = iter.current();
4999 if (lpt->is_counted() && lpt->is_innermost()) {
5000 move_unordered_reduction_out_of_loop(lpt);
5001 }
5002 }
5003 }
5004 }
5005
5006 #ifndef PRODUCT
5007 //------------------------------print_statistics-------------------------------
5008 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
5009 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5010 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5011 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5012 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5013 void PhaseIdealLoop::print_statistics() {
5014 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);
5015 }
5016 #endif
5017
5018 #ifdef ASSERT
5019 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5020 void PhaseIdealLoop::verify() const {
5021 ResourceMark rm;
5022 int old_progress = C->major_progress();
5023 bool success = true;
5024
5025 PhaseIdealLoop phase_verify(_igvn, this);
5026 if (C->failing_internal()) {
5027 return;
5028 }
5029
5030 // Verify ctrl and idom of every node.
5031 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5032
5033 // Verify loop-tree.
5034 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5035
5036 assert(success, "VerifyLoopOptimizations failed");
5037
5038 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5039 C->restore_major_progress(old_progress);
5040 }
5041
5042 // Perform a BFS starting at n, through all inputs.
5043 // Call verify_idom and verify_node on all nodes of BFS traversal.
5044 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5045 Unique_Node_List worklist;
5046 worklist.push(root);
5047 bool success = true;
5048 for (uint i = 0; i < worklist.size(); i++) {
5049 Node* n = worklist.at(i);
5050 // process node
5051 success &= verify_idom(n, phase_verify);
5052 success &= verify_loop_ctrl(n, phase_verify);
5053 // visit inputs
5054 for (uint j = 0; j < n->req(); j++) {
5055 if (n->in(j) != nullptr) {
5056 worklist.push(n->in(j));
5057 }
5058 }
5059 }
5060 return success;
5061 }
5062
5063 // Verify dominator structure (IDOM).
5064 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5065 // Verify IDOM for all CFG nodes (except root).
5066 if (!n->is_CFG() || n->is_Root()) {
5067 return true; // pass
5068 }
5069
5070 if (n->_idx >= _idom_size) {
5071 tty->print("CFG Node with no idom: ");
5072 n->dump();
5073 return false; // fail
5074 }
5075
5076 Node* id = idom_no_update(n);
5077 Node* id_verify = phase_verify->idom_no_update(n);
5078 if (id != id_verify) {
5079 tty->print("Mismatching idom for node: ");
5080 n->dump();
5081 tty->print(" We have idom: ");
5082 id->dump();
5083 tty->print(" Verify has idom: ");
5084 id_verify->dump();
5085 tty->cr();
5086 return false; // fail
5087 }
5088 return true; // pass
5089 }
5090
5091 // Verify "_loop_or_ctrl": control and loop membership.
5092 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5093 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5094 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5095 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5096 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5097 const uint i = n->_idx;
5098 // The loop-tree was built from def to use (top-down).
5099 // The verification happens from use to def (bottom-up).
5100 // We may thus find nodes during verification that are not in the loop-tree.
5101 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5102 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5103 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5104 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5105 n->dump();
5106 return false; // fail
5107 }
5108 // Not reachable for both.
5109 return true; // pass
5110 }
5111
5112 if (n->is_CFG() == has_ctrl(n)) {
5113 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5114 n->dump();
5115 return false; // fail
5116 }
5117
5118 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5119 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5120 n->dump();
5121 return false; // fail
5122 } else if (has_ctrl(n)) {
5123 assert(phase_verify->has_ctrl(n), "sanity");
5124 // n is a data node.
5125 // Verify that its ctrl is the same.
5126
5127 // Broken part of VerifyLoopOptimizations (A)
5128 // Reason:
5129 // BUG, wrong control set for example in
5130 // PhaseIdealLoop::split_if_with_blocks
5131 // at "set_ctrl(x, new_ctrl);"
5132 /*
5133 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5134 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5135 tty->print("Mismatched control setting for: ");
5136 n->dump();
5137 if( fail++ > 10 ) return;
5138 Node *c = get_ctrl_no_update(n);
5139 tty->print("We have it as: ");
5140 if( c->in(0) ) c->dump();
5141 else tty->print_cr("N%d",c->_idx);
5142 tty->print("Verify thinks: ");
5143 if( loop_verify->has_ctrl(n) )
5144 loop_verify->get_ctrl_no_update(n)->dump();
5145 else
5146 loop_verify->get_loop_idx(n)->dump();
5147 tty->cr();
5148 }
5149 */
5150 return true; // pass
5151 } else {
5152 assert(!phase_verify->has_ctrl(n), "sanity");
5153 // n is a ctrl node.
5154 // Verify that not has_ctrl, and that get_loop_idx is the same.
5155
5156 // Broken part of VerifyLoopOptimizations (B)
5157 // Reason:
5158 // NeverBranch node for example is added to loop outside its scope.
5159 // Once we run build_loop_tree again, it is added to the correct loop.
5160 /*
5161 if (!C->major_progress()) {
5162 // Loop selection can be messed up if we did a major progress
5163 // operation, like split-if. Do not verify in that case.
5164 IdealLoopTree *us = get_loop_idx(n);
5165 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5166 if( us->_head != them->_head || us->_tail != them->_tail ) {
5167 tty->print("Unequals loops for: ");
5168 n->dump();
5169 if( fail++ > 10 ) return;
5170 tty->print("We have it as: ");
5171 us->dump();
5172 tty->print("Verify thinks: ");
5173 them->dump();
5174 tty->cr();
5175 }
5176 }
5177 */
5178 return true; // pass
5179 }
5180 }
5181
5182 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5183 assert(a != nullptr && b != nullptr, "must be");
5184 return a->_head->_idx - b->_head->_idx;
5185 }
5186
5187 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5188 GrowableArray<IdealLoopTree*> children;
5189 IdealLoopTree* child = _child;
5190 while (child != nullptr) {
5191 assert(child->_parent == this, "all must be children of this");
5192 children.insert_sorted<compare_tree>(child);
5193 child = child->_next;
5194 }
5195 return children;
5196 }
5197
5198 // Verify that tree structures match. Because the CFG can change, siblings
5199 // within the loop tree can be reordered. We attempt to deal with that by
5200 // reordering the verify's loop tree if possible.
5201 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5202 assert(_head == loop_verify->_head, "mismatched loop head");
5203 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5204
5205 // Collect the children
5206 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5207 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5208
5209 bool success = true;
5210
5211 // Compare the two children lists
5212 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5213 IdealLoopTree* child = nullptr;
5214 IdealLoopTree* child_verify = nullptr;
5215 // Read from both lists, if possible.
5216 if (i < children.length()) {
5217 child = children.at(i);
5218 }
5219 if (j < children_verify.length()) {
5220 child_verify = children_verify.at(j);
5221 }
5222 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5223 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5224 // We found two non-equal children. Select the smaller one.
5225 if (child->_head->_idx < child_verify->_head->_idx) {
5226 child_verify = nullptr;
5227 } else {
5228 child = nullptr;
5229 }
5230 }
5231 // Process the two children, or potentially log the failure if we only found one.
5232 if (child_verify == nullptr) {
5233 if (child->_irreducible && Compile::current()->major_progress()) {
5234 // Irreducible loops can pick a different header (one of its entries).
5235 } else {
5236 tty->print_cr("We have a loop that verify does not have");
5237 child->dump();
5238 success = false;
5239 }
5240 i++; // step for this
5241 } else if (child == nullptr) {
5242 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5243 // Irreducible loops can pick a different header (one of its entries).
5244 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5245 // Infinite loops do not get attached to the loop-tree on their first visit.
5246 // "this" runs before "loop_verify". It is thus possible that we find the
5247 // infinite loop only for "child_verify". Only finding it with "child" would
5248 // mean that we lost it, which is not ok.
5249 } else {
5250 tty->print_cr("Verify has a loop that we do not have");
5251 child_verify->dump();
5252 success = false;
5253 }
5254 j++; // step for verify
5255 } else {
5256 assert(child->_head == child_verify->_head, "We have both and they are equal");
5257 success &= child->verify_tree(child_verify); // Recursion
5258 i++; // step for this
5259 j++; // step for verify
5260 }
5261 }
5262
5263 // Broken part of VerifyLoopOptimizations (D)
5264 // Reason:
5265 // split_if has to update the _tail, if it is modified. But that is done by
5266 // checking to what loop the iff belongs to. That info can be wrong, and then
5267 // we do not update the _tail correctly.
5268 /*
5269 Node *tail = _tail; // Inline a non-updating version of
5270 while( !tail->in(0) ) // the 'tail()' call.
5271 tail = tail->in(1);
5272 assert( tail == loop->_tail, "mismatched loop tail" );
5273 */
5274
5275 if (_head->is_CountedLoop()) {
5276 CountedLoopNode *cl = _head->as_CountedLoop();
5277
5278 Node* ctrl = cl->init_control();
5279 Node* back = cl->back_control();
5280 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5281 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5282 cl->loopexit(); // assert implied
5283 }
5284
5285 // Broken part of VerifyLoopOptimizations (E)
5286 // Reason:
5287 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5288 // to the loop body. Or maybe they are not added to the correct loop.
5289 // at "Node* x = n->clone();"
5290 /*
5291 // Innermost loops need to verify loop bodies,
5292 // but only if no 'major_progress'
5293 int fail = 0;
5294 if (!Compile::current()->major_progress() && _child == nullptr) {
5295 for( uint i = 0; i < _body.size(); i++ ) {
5296 Node *n = _body.at(i);
5297 if (n->outcnt() == 0) continue; // Ignore dead
5298 uint j;
5299 for( j = 0; j < loop->_body.size(); j++ )
5300 if( loop->_body.at(j) == n )
5301 break;
5302 if( j == loop->_body.size() ) { // Not found in loop body
5303 // Last ditch effort to avoid assertion: Its possible that we
5304 // have some users (so outcnt not zero) but are still dead.
5305 // Try to find from root.
5306 if (Compile::current()->root()->find(n->_idx)) {
5307 fail++;
5308 tty->print("We have that verify does not: ");
5309 n->dump();
5310 }
5311 }
5312 }
5313 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5314 Node *n = loop->_body.at(i2);
5315 if (n->outcnt() == 0) continue; // Ignore dead
5316 uint j;
5317 for( j = 0; j < _body.size(); j++ )
5318 if( _body.at(j) == n )
5319 break;
5320 if( j == _body.size() ) { // Not found in loop body
5321 // Last ditch effort to avoid assertion: Its possible that we
5322 // have some users (so outcnt not zero) but are still dead.
5323 // Try to find from root.
5324 if (Compile::current()->root()->find(n->_idx)) {
5325 fail++;
5326 tty->print("Verify has that we do not: ");
5327 n->dump();
5328 }
5329 }
5330 }
5331 assert( !fail, "loop body mismatch" );
5332 }
5333 */
5334 return success;
5335 }
5336 #endif
5337
5338 //------------------------------set_idom---------------------------------------
5339 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5340 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5341 uint idx = d->_idx;
5342 if (idx >= _idom_size) {
5343 uint newsize = next_power_of_2(idx);
5344 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5345 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5346 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5347 _idom_size = newsize;
5348 }
5349 _idom[idx] = n;
5350 _dom_depth[idx] = dom_depth;
5351 }
5352
5353 //------------------------------recompute_dom_depth---------------------------------------
5354 // The dominator tree is constructed with only parent pointers.
5355 // This recomputes the depth in the tree by first tagging all
5356 // nodes as "no depth yet" marker. The next pass then runs up
5357 // the dom tree from each node marked "no depth yet", and computes
5358 // the depth on the way back down.
5359 void PhaseIdealLoop::recompute_dom_depth() {
5360 uint no_depth_marker = C->unique();
5361 uint i;
5362 // Initialize depth to "no depth yet" and realize all lazy updates
5363 for (i = 0; i < _idom_size; i++) {
5364 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5365 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5366 _dom_depth[i] = no_depth_marker;
5367
5368 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5369 if (_idom[i]->in(0) == nullptr) {
5370 idom(i);
5371 }
5372 }
5373 }
5374 if (_dom_stk == nullptr) {
5375 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5376 if (init_size < 10) init_size = 10;
5377 _dom_stk = new GrowableArray<uint>(init_size);
5378 }
5379 // Compute new depth for each node.
5380 for (i = 0; i < _idom_size; i++) {
5381 uint j = i;
5382 // Run up the dom tree to find a node with a depth
5383 while (_dom_depth[j] == no_depth_marker) {
5384 _dom_stk->push(j);
5385 j = _idom[j]->_idx;
5386 }
5387 // Compute the depth on the way back down this tree branch
5388 uint dd = _dom_depth[j] + 1;
5389 while (_dom_stk->length() > 0) {
5390 uint j = _dom_stk->pop();
5391 _dom_depth[j] = dd;
5392 dd++;
5393 }
5394 }
5395 }
5396
5397 //------------------------------sort-------------------------------------------
5398 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5399 // loop tree, not the root.
5400 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5401 if( !innermost ) return loop; // New innermost loop
5402
5403 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5404 assert( loop_preorder, "not yet post-walked loop" );
5405 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5406 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5407
5408 // Insert at start of list
5409 while( l ) { // Insertion sort based on pre-order
5410 if( l == loop ) return innermost; // Already on list!
5411 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5412 assert( l_preorder, "not yet post-walked l" );
5413 // Check header pre-order number to figure proper nesting
5414 if( loop_preorder > l_preorder )
5415 break; // End of insertion
5416 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5417 // Since I split shared headers, you'd think this could not happen.
5418 // BUT: I must first do the preorder numbering before I can discover I
5419 // have shared headers, so the split headers all get the same preorder
5420 // number as the RegionNode they split from.
5421 if( loop_preorder == l_preorder &&
5422 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5423 break; // Also check for shared headers (same pre#)
5424 pp = &l->_parent; // Chain up list
5425 l = *pp;
5426 }
5427 // Link into list
5428 // Point predecessor to me
5429 *pp = loop;
5430 // Point me to successor
5431 IdealLoopTree *p = loop->_parent;
5432 loop->_parent = l; // Point me to successor
5433 if( p ) sort( p, innermost ); // Insert my parents into list as well
5434 return innermost;
5435 }
5436
5437 //------------------------------build_loop_tree--------------------------------
5438 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5439 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5440 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5441 // tightest enclosing IdealLoopTree for post-walked.
5442 //
5443 // During my forward walk I do a short 1-layer lookahead to see if I can find
5444 // a loop backedge with that doesn't have any work on the backedge. This
5445 // helps me construct nested loops with shared headers better.
5446 //
5447 // Once I've done the forward recursion, I do the post-work. For each child
5448 // I check to see if there is a backedge. Backedges define a loop! I
5449 // insert an IdealLoopTree at the target of the backedge.
5450 //
5451 // During the post-work I also check to see if I have several children
5452 // belonging to different loops. If so, then this Node is a decision point
5453 // where control flow can choose to change loop nests. It is at this
5454 // decision point where I can figure out how loops are nested. At this
5455 // time I can properly order the different loop nests from my children.
5456 // Note that there may not be any backedges at the decision point!
5457 //
5458 // Since the decision point can be far removed from the backedges, I can't
5459 // order my loops at the time I discover them. Thus at the decision point
5460 // I need to inspect loop header pre-order numbers to properly nest my
5461 // loops. This means I need to sort my childrens' loops by pre-order.
5462 // The sort is of size number-of-control-children, which generally limits
5463 // it to size 2 (i.e., I just choose between my 2 target loops).
5464 void PhaseIdealLoop::build_loop_tree() {
5465 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5466 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5467 Node *n = C->root();
5468 bltstack.push(n);
5469 int pre_order = 1;
5470 int stack_size;
5471
5472 while ( ( stack_size = bltstack.length() ) != 0 ) {
5473 n = bltstack.top(); // Leave node on stack
5474 if ( !is_visited(n) ) {
5475 // ---- Pre-pass Work ----
5476 // Pre-walked but not post-walked nodes need a pre_order number.
5477
5478 set_preorder_visited( n, pre_order ); // set as visited
5479
5480 // ---- Scan over children ----
5481 // Scan first over control projections that lead to loop headers.
5482 // This helps us find inner-to-outer loops with shared headers better.
5483
5484 // Scan children's children for loop headers.
5485 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5486 Node* m = n->raw_out(i); // Child
5487 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5488 // Scan over children's children to find loop
5489 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5490 Node* l = m->fast_out(j);
5491 if( is_visited(l) && // Been visited?
5492 !is_postvisited(l) && // But not post-visited
5493 get_preorder(l) < pre_order ) { // And smaller pre-order
5494 // Found! Scan the DFS down this path before doing other paths
5495 bltstack.push(m);
5496 break;
5497 }
5498 }
5499 }
5500 }
5501 pre_order++;
5502 }
5503 else if ( !is_postvisited(n) ) {
5504 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5505 // such as com.sun.rsasign.am::a.
5506 // For non-recursive version, first, process current children.
5507 // On next iteration, check if additional children were added.
5508 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5509 Node* u = n->raw_out(k);
5510 if ( u->is_CFG() && !is_visited(u) ) {
5511 bltstack.push(u);
5512 }
5513 }
5514 if ( bltstack.length() == stack_size ) {
5515 // There were no additional children, post visit node now
5516 (void)bltstack.pop(); // Remove node from stack
5517 pre_order = build_loop_tree_impl(n, pre_order);
5518 // Check for bailout
5519 if (C->failing()) {
5520 return;
5521 }
5522 // Check to grow _preorders[] array for the case when
5523 // build_loop_tree_impl() adds new nodes.
5524 check_grow_preorders();
5525 }
5526 }
5527 else {
5528 (void)bltstack.pop(); // Remove post-visited node from stack
5529 }
5530 }
5531 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5532 }
5533
5534 //------------------------------build_loop_tree_impl---------------------------
5535 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5536 // ---- Post-pass Work ----
5537 // Pre-walked but not post-walked nodes need a pre_order number.
5538
5539 // Tightest enclosing loop for this Node
5540 IdealLoopTree *innermost = nullptr;
5541
5542 // For all children, see if any edge is a backedge. If so, make a loop
5543 // for it. Then find the tightest enclosing loop for the self Node.
5544 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5545 Node* m = n->fast_out(i); // Child
5546 if (n == m) continue; // Ignore control self-cycles
5547 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5548
5549 IdealLoopTree *l; // Child's loop
5550 if (!is_postvisited(m)) { // Child visited but not post-visited?
5551 // Found a backedge
5552 assert(get_preorder(m) < pre_order, "should be backedge");
5553 // Check for the RootNode, which is already a LoopNode and is allowed
5554 // to have multiple "backedges".
5555 if (m == C->root()) { // Found the root?
5556 l = _ltree_root; // Root is the outermost LoopNode
5557 } else { // Else found a nested loop
5558 // Insert a LoopNode to mark this loop.
5559 l = new IdealLoopTree(this, m, n);
5560 } // End of Else found a nested loop
5561 if (!has_loop(m)) { // If 'm' does not already have a loop set
5562 set_loop(m, l); // Set loop header to loop now
5563 }
5564 } else { // Else not a nested loop
5565 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5566 IdealLoopTree* m_loop = get_loop(m);
5567 l = m_loop; // Get previously determined loop
5568 // If successor is header of a loop (nest), move up-loop till it
5569 // is a member of some outer enclosing loop. Since there are no
5570 // shared headers (I've split them already) I only need to go up
5571 // at most 1 level.
5572 while (l && l->_head == m) { // Successor heads loop?
5573 l = l->_parent; // Move up 1 for me
5574 }
5575 // If this loop is not properly parented, then this loop
5576 // has no exit path out, i.e. its an infinite loop.
5577 if (!l) {
5578 // Make loop "reachable" from root so the CFG is reachable. Basically
5579 // insert a bogus loop exit that is never taken. 'm', the loop head,
5580 // points to 'n', one (of possibly many) fall-in paths. There may be
5581 // many backedges as well.
5582
5583 if (!_verify_only) {
5584 // Insert the NeverBranch between 'm' and it's control user.
5585 NeverBranchNode *iff = new NeverBranchNode( m );
5586 _igvn.register_new_node_with_optimizer(iff);
5587 set_loop(iff, m_loop);
5588 Node *if_t = new CProjNode( iff, 0 );
5589 _igvn.register_new_node_with_optimizer(if_t);
5590 set_loop(if_t, m_loop);
5591
5592 Node* cfg = nullptr; // Find the One True Control User of m
5593 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5594 Node* x = m->fast_out(j);
5595 if (x->is_CFG() && x != m && x != iff)
5596 { cfg = x; break; }
5597 }
5598 assert(cfg != nullptr, "must find the control user of m");
5599 uint k = 0; // Probably cfg->in(0)
5600 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5601 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5602
5603 // Now create the never-taken loop exit
5604 Node *if_f = new CProjNode( iff, 1 );
5605 _igvn.register_new_node_with_optimizer(if_f);
5606 set_loop(if_f, _ltree_root);
5607 // Find frame ptr for Halt. Relies on the optimizer
5608 // V-N'ing. Easier and quicker than searching through
5609 // the program structure.
5610 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5611 _igvn.register_new_node_with_optimizer(frame);
5612 // Halt & Catch Fire
5613 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5614 _igvn.register_new_node_with_optimizer(halt);
5615 set_loop(halt, _ltree_root);
5616 _igvn.add_input_to(C->root(), halt);
5617 }
5618 set_loop(C->root(), _ltree_root);
5619 // move to outer most loop with same header
5620 l = m_loop;
5621 while (true) {
5622 IdealLoopTree* next = l->_parent;
5623 if (next == nullptr || next->_head != m) {
5624 break;
5625 }
5626 l = next;
5627 }
5628 // properly insert infinite loop in loop tree
5629 sort(_ltree_root, l);
5630 // fix child link from parent
5631 IdealLoopTree* p = l->_parent;
5632 l->_next = p->_child;
5633 p->_child = l;
5634 // code below needs enclosing loop
5635 l = l->_parent;
5636 }
5637 }
5638 if (is_postvisited(l->_head)) {
5639 // We are currently visiting l, but its head has already been post-visited.
5640 // l is irreducible: we just found a second entry m.
5641 _has_irreducible_loops = true;
5642 RegionNode* secondary_entry = m->as_Region();
5643 DEBUG_ONLY(secondary_entry->verify_can_be_irreducible_entry();)
5644
5645 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
5646 // Since m is a secondary entry to them all.
5647 while( is_postvisited(l->_head) ) {
5648 l->_irreducible = 1; // = true
5649 RegionNode* head = l->_head->as_Region();
5650 DEBUG_ONLY(head->verify_can_be_irreducible_entry();)
5651 l = l->_parent;
5652 // Check for bad CFG here to prevent crash, and bailout of compile
5653 if (l == nullptr) {
5654 #ifndef PRODUCT
5655 if (TraceLoopOpts) {
5656 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
5657 m->dump();
5658 }
5659 #endif
5660 // This is a rare case that we do not want to handle in C2.
5661 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
5662 return pre_order;
5663 }
5664 }
5665 }
5666 if (!_verify_only) {
5667 C->set_has_irreducible_loop(_has_irreducible_loops);
5668 }
5669
5670 // This Node might be a decision point for loops. It is only if
5671 // it's children belong to several different loops. The sort call
5672 // does a trivial amount of work if there is only 1 child or all
5673 // children belong to the same loop. If however, the children
5674 // belong to different loops, the sort call will properly set the
5675 // _parent pointers to show how the loops nest.
5676 //
5677 // In any case, it returns the tightest enclosing loop.
5678 innermost = sort( l, innermost );
5679 }
5680
5681 // Def-use info will have some dead stuff; dead stuff will have no
5682 // loop decided on.
5683
5684 // Am I a loop header? If so fix up my parent's child and next ptrs.
5685 if( innermost && innermost->_head == n ) {
5686 assert( get_loop(n) == innermost, "" );
5687 IdealLoopTree *p = innermost->_parent;
5688 IdealLoopTree *l = innermost;
5689 while (p && l->_head == n) {
5690 l->_next = p->_child; // Put self on parents 'next child'
5691 p->_child = l; // Make self as first child of parent
5692 l = p; // Now walk up the parent chain
5693 p = l->_parent;
5694 }
5695 } else {
5696 // Note that it is possible for a LoopNode to reach here, if the
5697 // backedge has been made unreachable (hence the LoopNode no longer
5698 // denotes a Loop, and will eventually be removed).
5699
5700 // Record tightest enclosing loop for self. Mark as post-visited.
5701 set_loop(n, innermost);
5702 // Also record has_call flag early on
5703 if (innermost) {
5704 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
5705 // Do not count uncommon calls
5706 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
5707 Node *iff = n->in(0)->in(0);
5708 // No any calls for vectorized loops.
5709 if (C->do_superword() ||
5710 !iff->is_If() ||
5711 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
5712 iff->as_If()->_prob >= 0.01) {
5713 innermost->_has_call = 1;
5714 }
5715 }
5716 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
5717 // Disable loop optimizations if the loop has a scalar replaceable
5718 // allocation. This disabling may cause a potential performance lost
5719 // if the allocation is not eliminated for some reason.
5720 innermost->_allow_optimizations = false;
5721 innermost->_has_call = 1; // = true
5722 } else if (n->Opcode() == Op_SafePoint) {
5723 // Record all safepoints in this loop.
5724 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
5725 innermost->_safepts->push(n);
5726 }
5727 }
5728 }
5729
5730 // Flag as post-visited now
5731 set_postvisited(n);
5732 return pre_order;
5733 }
5734
5735 #ifdef ASSERT
5736 //--------------------------verify_regions_in_irreducible_loops----------------
5737 // Iterate down from Root through CFG, verify for every region:
5738 // if it is in an irreducible loop it must be marked as such
5739 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
5740 ResourceMark rm;
5741 if (!_has_irreducible_loops) {
5742 // last build_loop_tree has not found any irreducible loops
5743 // hence no region has to be marked is_in_irreduible_loop
5744 return;
5745 }
5746
5747 RootNode* root = C->root();
5748 Unique_Node_List worklist; // visit all nodes once
5749 worklist.push(root);
5750 bool failure = false;
5751 for (uint i = 0; i < worklist.size(); i++) {
5752 Node* n = worklist.at(i);
5753 if (n->is_Region()) {
5754 RegionNode* region = n->as_Region();
5755 if (is_in_irreducible_loop(region) &&
5756 region->loop_status() == RegionNode::LoopStatus::Reducible) {
5757 failure = true;
5758 tty->print("irreducible! ");
5759 region->dump();
5760 }
5761 }
5762 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
5763 Node* use = n->fast_out(j);
5764 if (use->is_CFG()) {
5765 worklist.push(use); // push if was not pushed before
5766 }
5767 }
5768 }
5769 assert(!failure, "region in irreducible loop was marked as reducible");
5770 }
5771
5772 //---------------------------is_in_irreducible_loop-------------------------
5773 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
5774 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
5775 if (!_has_irreducible_loops) {
5776 return false; // no irreducible loop in graph
5777 }
5778 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
5779 do {
5780 if (l->_irreducible) {
5781 return true; // found it
5782 }
5783 if (l == _ltree_root) {
5784 return false; // reached root, terimnate
5785 }
5786 l = l->_parent;
5787 } while (l != nullptr);
5788 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
5789 // We have "l->_parent == nullptr", which happens only for infinite loops,
5790 // where no parent is attached to the loop. We did not find any irreducible
5791 // loop from this block out to lp. Thus lp only has one entry, and no exit
5792 // (it is infinite and reducible). We can always rewrite an infinite loop
5793 // that is nested inside other loops:
5794 // while(condition) { infinite_loop; }
5795 // with an equivalent program where the infinite loop is an outermost loop
5796 // that is not nested in any loop:
5797 // while(condition) { break; } infinite_loop;
5798 // Thus, we can understand lp as an outermost loop, and can terminate and
5799 // conclude: this block is in no irreducible loop.
5800 return false;
5801 }
5802 #endif
5803
5804 //------------------------------build_loop_early-------------------------------
5805 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
5806 // First pass computes the earliest controlling node possible. This is the
5807 // controlling input with the deepest dominating depth.
5808 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
5809 while (worklist.size() != 0) {
5810 // Use local variables nstack_top_n & nstack_top_i to cache values
5811 // on nstack's top.
5812 Node *nstack_top_n = worklist.pop();
5813 uint nstack_top_i = 0;
5814 //while_nstack_nonempty:
5815 while (true) {
5816 // Get parent node and next input's index from stack's top.
5817 Node *n = nstack_top_n;
5818 uint i = nstack_top_i;
5819 uint cnt = n->req(); // Count of inputs
5820 if (i == 0) { // Pre-process the node.
5821 if( has_node(n) && // Have either loop or control already?
5822 !has_ctrl(n) ) { // Have loop picked out already?
5823 // During "merge_many_backedges" we fold up several nested loops
5824 // into a single loop. This makes the members of the original
5825 // loop bodies pointing to dead loops; they need to move up
5826 // to the new UNION'd larger loop. I set the _head field of these
5827 // dead loops to null and the _parent field points to the owning
5828 // loop. Shades of UNION-FIND algorithm.
5829 IdealLoopTree *ilt;
5830 while( !(ilt = get_loop(n))->_head ) {
5831 // Normally I would use a set_loop here. But in this one special
5832 // case, it is legal (and expected) to change what loop a Node
5833 // belongs to.
5834 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
5835 }
5836 // Remove safepoints ONLY if I've already seen I don't need one.
5837 // (the old code here would yank a 2nd safepoint after seeing a
5838 // first one, even though the 1st did not dominate in the loop body
5839 // and thus could be avoided indefinitely)
5840 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
5841 is_deleteable_safept(n)) {
5842 Node *in = n->in(TypeFunc::Control);
5843 lazy_replace(n,in); // Pull safepoint now
5844 if (ilt->_safepts != nullptr) {
5845 ilt->_safepts->yank(n);
5846 }
5847 // Carry on with the recursion "as if" we are walking
5848 // only the control input
5849 if( !visited.test_set( in->_idx ) ) {
5850 worklist.push(in); // Visit this guy later, using worklist
5851 }
5852 // Get next node from nstack:
5853 // - skip n's inputs processing by setting i > cnt;
5854 // - we also will not call set_early_ctrl(n) since
5855 // has_node(n) == true (see the condition above).
5856 i = cnt + 1;
5857 }
5858 }
5859 } // if (i == 0)
5860
5861 // Visit all inputs
5862 bool done = true; // Assume all n's inputs will be processed
5863 while (i < cnt) {
5864 Node *in = n->in(i);
5865 ++i;
5866 if (in == nullptr) continue;
5867 if (in->pinned() && !in->is_CFG())
5868 set_ctrl(in, in->in(0));
5869 int is_visited = visited.test_set( in->_idx );
5870 if (!has_node(in)) { // No controlling input yet?
5871 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
5872 assert( !is_visited, "visit only once" );
5873 nstack.push(n, i); // Save parent node and next input's index.
5874 nstack_top_n = in; // Process current input now.
5875 nstack_top_i = 0;
5876 done = false; // Not all n's inputs processed.
5877 break; // continue while_nstack_nonempty;
5878 } else if (!is_visited) {
5879 // This guy has a location picked out for him, but has not yet
5880 // been visited. Happens to all CFG nodes, for instance.
5881 // Visit him using the worklist instead of recursion, to break
5882 // cycles. Since he has a location already we do not need to
5883 // find his location before proceeding with the current Node.
5884 worklist.push(in); // Visit this guy later, using worklist
5885 }
5886 }
5887 if (done) {
5888 // All of n's inputs have been processed, complete post-processing.
5889
5890 // Compute earliest point this Node can go.
5891 // CFG, Phi, pinned nodes already know their controlling input.
5892 if (!has_node(n)) {
5893 // Record earliest legal location
5894 set_early_ctrl(n, false);
5895 }
5896 if (nstack.is_empty()) {
5897 // Finished all nodes on stack.
5898 // Process next node on the worklist.
5899 break;
5900 }
5901 // Get saved parent node and next input's index.
5902 nstack_top_n = nstack.node();
5903 nstack_top_i = nstack.index();
5904 nstack.pop();
5905 }
5906 } // while (true)
5907 }
5908 }
5909
5910 //------------------------------dom_lca_internal--------------------------------
5911 // Pair-wise LCA
5912 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
5913 if( !n1 ) return n2; // Handle null original LCA
5914 assert( n1->is_CFG(), "" );
5915 assert( n2->is_CFG(), "" );
5916 // find LCA of all uses
5917 uint d1 = dom_depth(n1);
5918 uint d2 = dom_depth(n2);
5919 while (n1 != n2) {
5920 if (d1 > d2) {
5921 n1 = idom(n1);
5922 d1 = dom_depth(n1);
5923 } else if (d1 < d2) {
5924 n2 = idom(n2);
5925 d2 = dom_depth(n2);
5926 } else {
5927 // Here d1 == d2. Due to edits of the dominator-tree, sections
5928 // of the tree might have the same depth. These sections have
5929 // to be searched more carefully.
5930
5931 // Scan up all the n1's with equal depth, looking for n2.
5932 Node *t1 = idom(n1);
5933 while (dom_depth(t1) == d1) {
5934 if (t1 == n2) return n2;
5935 t1 = idom(t1);
5936 }
5937 // Scan up all the n2's with equal depth, looking for n1.
5938 Node *t2 = idom(n2);
5939 while (dom_depth(t2) == d2) {
5940 if (t2 == n1) return n1;
5941 t2 = idom(t2);
5942 }
5943 // Move up to a new dominator-depth value as well as up the dom-tree.
5944 n1 = t1;
5945 n2 = t2;
5946 d1 = dom_depth(n1);
5947 d2 = dom_depth(n2);
5948 }
5949 }
5950 return n1;
5951 }
5952
5953 //------------------------------compute_idom-----------------------------------
5954 // Locally compute IDOM using dom_lca call. Correct only if the incoming
5955 // IDOMs are correct.
5956 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
5957 assert( region->is_Region(), "" );
5958 Node *LCA = nullptr;
5959 for( uint i = 1; i < region->req(); i++ ) {
5960 if( region->in(i) != C->top() )
5961 LCA = dom_lca( LCA, region->in(i) );
5962 }
5963 return LCA;
5964 }
5965
5966 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
5967 bool had_error = false;
5968 #ifdef ASSERT
5969 if (early != C->root()) {
5970 // Make sure that there's a dominance path from LCA to early
5971 Node* d = LCA;
5972 while (d != early) {
5973 if (d == C->root()) {
5974 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
5975 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
5976 had_error = true;
5977 break;
5978 }
5979 d = idom(d);
5980 }
5981 }
5982 #endif
5983 return had_error;
5984 }
5985
5986
5987 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
5988 // Compute LCA over list of uses
5989 bool had_error = false;
5990 Node *LCA = nullptr;
5991 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
5992 Node* c = n->fast_out(i);
5993 if (_loop_or_ctrl[c->_idx] == nullptr)
5994 continue; // Skip the occasional dead node
5995 if( c->is_Phi() ) { // For Phis, we must land above on the path
5996 for( uint j=1; j<c->req(); j++ ) {// For all inputs
5997 if( c->in(j) == n ) { // Found matching input?
5998 Node *use = c->in(0)->in(j);
5999 if (_verify_only && use->is_top()) continue;
6000 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6001 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6002 }
6003 }
6004 } else {
6005 // For CFG data-users, use is in the block just prior
6006 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
6007 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6008 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6009 }
6010 }
6011 assert(!had_error, "bad dominance");
6012 return LCA;
6013 }
6014
6015 // Check the shape of the graph at the loop entry. In some cases,
6016 // the shape of the graph does not match the shape outlined below.
6017 // That is caused by the Opaque1 node "protecting" the shape of
6018 // the graph being removed by, for example, the IGVN performed
6019 // in PhaseIdealLoop::build_and_optimize().
6020 //
6021 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6022 // loop unswitching, and IGVN, or a combination of them) can freely change
6023 // the graph's shape. As a result, the graph shape outlined below cannot
6024 // be guaranteed anymore.
6025 Node* CountedLoopNode::is_canonical_loop_entry() {
6026 if (!is_main_loop() && !is_post_loop()) {
6027 return nullptr;
6028 }
6029 Node* ctrl = skip_assertion_predicates_with_halt();
6030
6031 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6032 return nullptr;
6033 }
6034 Node* iffm = ctrl->in(0);
6035 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6036 return nullptr;
6037 }
6038 Node* bolzm = iffm->in(1);
6039 if (bolzm == nullptr || !bolzm->is_Bool()) {
6040 return nullptr;
6041 }
6042 Node* cmpzm = bolzm->in(1);
6043 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6044 return nullptr;
6045 }
6046
6047 uint input = is_main_loop() ? 2 : 1;
6048 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6049 return nullptr;
6050 }
6051 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6052 #ifdef ASSERT
6053 bool found_opaque = false;
6054 for (uint i = 1; i < cmpzm->req(); i++) {
6055 Node* opnd = cmpzm->in(i);
6056 if (opnd && opnd->is_Opaque1()) {
6057 found_opaque = true;
6058 break;
6059 }
6060 }
6061 assert(found_opaque == res, "wrong pattern");
6062 #endif
6063 return res ? cmpzm->in(input) : nullptr;
6064 }
6065
6066 // Find pre loop end from main loop. Returns nullptr if none.
6067 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6068 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6069 // The loop cannot be optimized if the graph shape at the loop entry is
6070 // inappropriate.
6071 if (is_canonical_loop_entry() == nullptr) {
6072 return nullptr;
6073 }
6074
6075 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6076 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6077 return nullptr;
6078 }
6079 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6080 CountedLoopNode* loop_node = pre_end->loopnode();
6081 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6082 return nullptr;
6083 }
6084 return pre_end;
6085 }
6086
6087 //------------------------------get_late_ctrl----------------------------------
6088 // Compute latest legal control.
6089 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6090 assert(early != nullptr, "early control should not be null");
6091
6092 Node* LCA = compute_lca_of_uses(n, early);
6093 #ifdef ASSERT
6094 if (LCA == C->root() && LCA != early) {
6095 // def doesn't dominate uses so print some useful debugging output
6096 compute_lca_of_uses(n, early, true);
6097 }
6098 #endif
6099
6100 if (n->is_Load() && LCA != early) {
6101 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6102 }
6103
6104 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6105 return LCA;
6106 }
6107
6108 // if this is a load, check for anti-dependent stores
6109 // We use a conservative algorithm to identify potential interfering
6110 // instructions and for rescheduling the load. The users of the memory
6111 // input of this load are examined. Any use which is not a load and is
6112 // dominated by early is considered a potentially interfering store.
6113 // This can produce false positives.
6114 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6115 int load_alias_idx = C->get_alias_index(n->adr_type());
6116 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6117 Unique_Node_List worklist;
6118
6119 Node* mem = n->in(MemNode::Memory);
6120 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6121 Node* s = mem->fast_out(i);
6122 worklist.push(s);
6123 }
6124 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6125 Node* s = worklist.at(i);
6126 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6127 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6128 s->is_Phi()) {
6129 continue;
6130 } else if (s->is_MergeMem()) {
6131 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6132 Node* s1 = s->fast_out(i);
6133 worklist.push(s1);
6134 }
6135 } else {
6136 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6137 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6138 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6139 const TypePtr* adr_type = s->adr_type();
6140 if (s->is_ArrayCopy()) {
6141 // Copy to known instance needs destination type to test for aliasing
6142 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6143 if (dest_type != TypeOopPtr::BOTTOM) {
6144 adr_type = dest_type;
6145 }
6146 }
6147 if (C->can_alias(adr_type, load_alias_idx)) {
6148 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6149 } else if (s->is_CFG() && s->is_Multi()) {
6150 // Look for the memory use of s (that is the use of its memory projection)
6151 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6152 Node* s1 = s->fast_out(i);
6153 assert(s1->is_Proj(), "projection expected");
6154 if (_igvn.type(s1) == Type::MEMORY) {
6155 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6156 Node* s2 = s1->fast_out(j);
6157 worklist.push(s2);
6158 }
6159 }
6160 }
6161 }
6162 }
6163 }
6164 }
6165 // For Phis only consider Region's inputs that were reached by following the memory edges
6166 if (LCA != early) {
6167 for (uint i = 0; i < worklist.size(); i++) {
6168 Node* s = worklist.at(i);
6169 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6170 Node* r = s->in(0);
6171 for (uint j = 1; j < s->req(); j++) {
6172 Node* in = s->in(j);
6173 Node* r_in = r->in(j);
6174 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6175 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6176 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6177 }
6178 }
6179 }
6180 }
6181 }
6182 }
6183 return LCA;
6184 }
6185
6186 // Is CFG node 'dominator' dominating node 'n'?
6187 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6188 if (dominator == n) {
6189 return true;
6190 }
6191 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6192 uint dd = dom_depth(dominator);
6193 while (dom_depth(n) >= dd) {
6194 if (n == dominator) {
6195 return true;
6196 }
6197 n = idom(n);
6198 }
6199 return false;
6200 }
6201
6202 // Is CFG node 'dominator' strictly dominating node 'n'?
6203 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6204 return dominator != n && is_dominator(dominator, n);
6205 }
6206
6207 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6208 // Pair-wise LCA with tags.
6209 // Tag each index with the node 'tag' currently being processed
6210 // before advancing up the dominator chain using idom().
6211 // Later calls that find a match to 'tag' know that this path has already
6212 // been considered in the current LCA (which is input 'n1' by convention).
6213 // Since get_late_ctrl() is only called once for each node, the tag array
6214 // does not need to be cleared between calls to get_late_ctrl().
6215 // Algorithm trades a larger constant factor for better asymptotic behavior
6216 //
6217 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6218 uint d1 = dom_depth(n1);
6219 uint d2 = dom_depth(n2);
6220 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6221
6222 do {
6223 if (d1 > d2) {
6224 // current lca is deeper than n2
6225 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6226 n1 = idom(n1);
6227 d1 = dom_depth(n1);
6228 } else if (d1 < d2) {
6229 // n2 is deeper than current lca
6230 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6231 if (memo == tag) {
6232 return n1; // Return the current LCA
6233 }
6234 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6235 n2 = idom(n2);
6236 d2 = dom_depth(n2);
6237 } else {
6238 // Here d1 == d2. Due to edits of the dominator-tree, sections
6239 // of the tree might have the same depth. These sections have
6240 // to be searched more carefully.
6241
6242 // Scan up all the n1's with equal depth, looking for n2.
6243 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6244 Node *t1 = idom(n1);
6245 while (dom_depth(t1) == d1) {
6246 if (t1 == n2) return n2;
6247 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6248 t1 = idom(t1);
6249 }
6250 // Scan up all the n2's with equal depth, looking for n1.
6251 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6252 Node *t2 = idom(n2);
6253 while (dom_depth(t2) == d2) {
6254 if (t2 == n1) return n1;
6255 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6256 t2 = idom(t2);
6257 }
6258 // Move up to a new dominator-depth value as well as up the dom-tree.
6259 n1 = t1;
6260 n2 = t2;
6261 d1 = dom_depth(n1);
6262 d2 = dom_depth(n2);
6263 }
6264 } while (n1 != n2);
6265 return n1;
6266 }
6267
6268 //------------------------------init_dom_lca_tags------------------------------
6269 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6270 // Intended use does not involve any growth for the array, so it could
6271 // be of fixed size.
6272 void PhaseIdealLoop::init_dom_lca_tags() {
6273 uint limit = C->unique() + 1;
6274 _dom_lca_tags.at_grow(limit, 0);
6275 _dom_lca_tags_round = 0;
6276 #ifdef ASSERT
6277 for (uint i = 0; i < limit; ++i) {
6278 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6279 }
6280 #endif // ASSERT
6281 }
6282
6283 //------------------------------build_loop_late--------------------------------
6284 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6285 // Second pass finds latest legal placement, and ideal loop placement.
6286 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6287 while (worklist.size() != 0) {
6288 Node *n = worklist.pop();
6289 // Only visit once
6290 if (visited.test_set(n->_idx)) continue;
6291 uint cnt = n->outcnt();
6292 uint i = 0;
6293 while (true) {
6294 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6295 // Visit all children
6296 if (i < cnt) {
6297 Node* use = n->raw_out(i);
6298 ++i;
6299 // Check for dead uses. Aggressively prune such junk. It might be
6300 // dead in the global sense, but still have local uses so I cannot
6301 // easily call 'remove_dead_node'.
6302 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6303 // Due to cycles, we might not hit the same fixed point in the verify
6304 // pass as we do in the regular pass. Instead, visit such phis as
6305 // simple uses of the loop head.
6306 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6307 if( !visited.test(use->_idx) )
6308 worklist.push(use);
6309 } else if( !visited.test_set(use->_idx) ) {
6310 nstack.push(n, i); // Save parent and next use's index.
6311 n = use; // Process all children of current use.
6312 cnt = use->outcnt();
6313 i = 0;
6314 }
6315 } else {
6316 // Do not visit around the backedge of loops via data edges.
6317 // push dead code onto a worklist
6318 _deadlist.push(use);
6319 }
6320 } else {
6321 // All of n's children have been processed, complete post-processing.
6322 build_loop_late_post(n);
6323 if (C->failing()) { return; }
6324 if (nstack.is_empty()) {
6325 // Finished all nodes on stack.
6326 // Process next node on the worklist.
6327 break;
6328 }
6329 // Get saved parent node and next use's index. Visit the rest of uses.
6330 n = nstack.node();
6331 cnt = n->outcnt();
6332 i = nstack.index();
6333 nstack.pop();
6334 }
6335 }
6336 }
6337 }
6338
6339 // Verify that no data node is scheduled in the outer loop of a strip
6340 // mined loop.
6341 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6342 #ifdef ASSERT
6343 if (get_loop(least)->_nest == 0) {
6344 return;
6345 }
6346 IdealLoopTree* loop = get_loop(least);
6347 Node* head = loop->_head;
6348 if (head->is_OuterStripMinedLoop() &&
6349 // Verification can't be applied to fully built strip mined loops
6350 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6351 Node* sfpt = head->as_Loop()->outer_safepoint();
6352 ResourceMark rm;
6353 Unique_Node_List wq;
6354 wq.push(sfpt);
6355 for (uint i = 0; i < wq.size(); i++) {
6356 Node *m = wq.at(i);
6357 for (uint i = 1; i < m->req(); i++) {
6358 Node* nn = m->in(i);
6359 if (nn == n) {
6360 return;
6361 }
6362 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6363 wq.push(nn);
6364 }
6365 }
6366 }
6367 ShouldNotReachHere();
6368 }
6369 #endif
6370 }
6371
6372
6373 //------------------------------build_loop_late_post---------------------------
6374 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6375 // Second pass finds latest legal placement, and ideal loop placement.
6376 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6377 build_loop_late_post_work(n, true);
6378 }
6379
6380 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6381 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6382 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6383 // Predication or creating a Loop Limit Check Predicate later.
6384 class DominatedPredicates : public UnifiedPredicateVisitor {
6385 Node* const _dominator;
6386 Node* _earliest_dominated_predicate_entry;
6387 bool _should_continue;
6388 PhaseIdealLoop* const _phase;
6389
6390 public:
6391 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6392 : _dominator(dominator),
6393 _earliest_dominated_predicate_entry(start_node),
6394 _should_continue(true),
6395 _phase(phase) {}
6396 NONCOPYABLE(DominatedPredicates);
6397
6398 bool should_continue() const override {
6399 return _should_continue;
6400 }
6401
6402 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6403 Node* earliest_dominated_predicate_entry() const {
6404 return _earliest_dominated_predicate_entry;
6405 }
6406
6407 void visit_predicate(const Predicate& predicate) override {
6408 Node* entry = predicate.entry();
6409 if (_phase->is_strict_dominator(entry, _dominator)) {
6410 _should_continue = false;
6411 } else {
6412 _earliest_dominated_predicate_entry = entry;
6413 }
6414 }
6415 };
6416
6417 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6418
6419 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6420 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6421 }
6422
6423 #ifdef ASSERT
6424 if (_verify_only && !n->is_CFG()) {
6425 // Check def-use domination.
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_ModF:
6444 case Op_ModD:
6445 case Op_LoadB: // Same with Loads; they can sink
6446 case Op_LoadUB: // during loop optimizations.
6447 case Op_LoadUS:
6448 case Op_LoadD:
6449 case Op_LoadF:
6450 case Op_LoadI:
6451 case Op_LoadKlass:
6452 case Op_LoadNKlass:
6453 case Op_LoadL:
6454 case Op_LoadS:
6455 case Op_LoadP:
6456 case Op_LoadN:
6457 case Op_LoadRange:
6458 case Op_LoadD_unaligned:
6459 case Op_LoadL_unaligned:
6460 case Op_StrComp: // Does a bunch of load-like effects
6461 case Op_StrEquals:
6462 case Op_StrIndexOf:
6463 case Op_StrIndexOfChar:
6464 case Op_AryEq:
6465 case Op_VectorizedHashCode:
6466 case Op_CountPositives:
6467 pinned = false;
6468 }
6469 if (n->is_CMove() || n->is_ConstraintCast()) {
6470 pinned = false;
6471 }
6472 if( pinned ) {
6473 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6474 if( !chosen_loop->_child ) // Inner loop?
6475 chosen_loop->_body.push(n); // Collect inner loops
6476 return;
6477 }
6478 } else { // No slot zero
6479 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6480 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6481 return;
6482 }
6483 assert(!n->is_CFG() || n->outcnt() == 0, "");
6484 }
6485
6486 // Do I have a "safe range" I can select over?
6487 Node *early = get_ctrl(n);// Early location already computed
6488
6489 // Compute latest point this Node can go
6490 Node *LCA = get_late_ctrl( n, early );
6491 // LCA is null due to uses being dead
6492 if( LCA == nullptr ) {
6493 #ifdef ASSERT
6494 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6495 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6496 }
6497 #endif
6498 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6499 _deadlist.push(n);
6500 return;
6501 }
6502 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6503
6504 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6505 Node *least = legal; // Best legal position so far
6506 while( early != legal ) { // While not at earliest legal
6507 if (legal->is_Start() && !early->is_Root()) {
6508 #ifdef ASSERT
6509 // Bad graph. Print idom path and fail.
6510 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6511 assert(false, "Bad graph detected in build_loop_late");
6512 #endif
6513 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6514 return;
6515 }
6516 // Find least loop nesting depth
6517 legal = idom(legal); // Bump up the IDOM tree
6518 // Check for lower nesting depth
6519 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6520 least = legal;
6521 }
6522 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6523
6524 if (least != early) {
6525 // Move the node above predicates as far up as possible so a
6526 // following pass of Loop Predication doesn't hoist a predicate
6527 // that depends on it above that node.
6528 const PredicateIterator predicate_iterator(least);
6529 DominatedPredicates dominated_predicates(early, least, this);
6530 predicate_iterator.for_each(dominated_predicates);
6531 least = dominated_predicates.earliest_dominated_predicate_entry();
6532 }
6533 // Try not to place code on a loop entry projection
6534 // which can inhibit range check elimination.
6535 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6536 Node* ctrl_out = least->unique_ctrl_out_or_null();
6537 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6538 least == ctrl_out->in(LoopNode::EntryControl) &&
6539 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6540 Node* least_dom = idom(least);
6541 if (get_loop(least_dom)->is_member(get_loop(least))) {
6542 least = least_dom;
6543 }
6544 }
6545 }
6546 // Don't extend live ranges of raw oops
6547 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6548 !n->bottom_type()->isa_rawptr()) {
6549 least = early;
6550 }
6551
6552 #ifdef ASSERT
6553 // Broken part of VerifyLoopOptimizations (F)
6554 // Reason:
6555 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6556 /*
6557 // If verifying, verify that 'verify_me' has a legal location
6558 // and choose it as our location.
6559 if( _verify_me ) {
6560 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6561 Node *legal = LCA;
6562 while( early != legal ) { // While not at earliest legal
6563 if( legal == v_ctrl ) break; // Check for prior good location
6564 legal = idom(legal) ;// Bump up the IDOM tree
6565 }
6566 // Check for prior good location
6567 if( legal == v_ctrl ) least = legal; // Keep prior if found
6568 }
6569 */
6570 #endif
6571
6572 // Assign discovered "here or above" point
6573 least = find_non_split_ctrl(least);
6574 verify_strip_mined_scheduling(n, least);
6575 set_ctrl(n, least);
6576
6577 // Collect inner loop bodies
6578 IdealLoopTree *chosen_loop = get_loop(least);
6579 if( !chosen_loop->_child ) // Inner loop?
6580 chosen_loop->_body.push(n);// Collect inner loops
6581
6582 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6583 _zero_trip_guard_opaque_nodes.push(n);
6584 }
6585
6586 }
6587
6588 #ifdef ASSERT
6589 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6590 tty->print_cr("%s", msg);
6591 tty->print("n: "); n->dump();
6592 tty->print("early(n): "); early->dump();
6593 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6594 n->in(0) != early && !n->in(0)->is_Root()) {
6595 tty->print("n->in(0): "); n->in(0)->dump();
6596 }
6597 for (uint i = 1; i < n->req(); i++) {
6598 Node* in1 = n->in(i);
6599 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6600 tty->print("n->in(%d): ", i); in1->dump();
6601 Node* in1_early = get_ctrl(in1);
6602 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6603 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6604 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6605 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6606 }
6607 for (uint j = 1; j < in1->req(); j++) {
6608 Node* in2 = in1->in(j);
6609 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6610 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6611 Node* in2_early = get_ctrl(in2);
6612 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6613 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6614 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
6615 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
6616 }
6617 }
6618 }
6619 }
6620 }
6621 tty->cr();
6622 tty->print("LCA(n): "); LCA->dump();
6623 for (uint i = 0; i < n->outcnt(); i++) {
6624 Node* u1 = n->raw_out(i);
6625 if (u1 == n)
6626 continue;
6627 tty->print("n->out(%d): ", i); u1->dump();
6628 if (u1->is_CFG()) {
6629 for (uint j = 0; j < u1->outcnt(); j++) {
6630 Node* u2 = u1->raw_out(j);
6631 if (u2 != u1 && u2 != n && u2->is_CFG()) {
6632 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6633 }
6634 }
6635 } else {
6636 Node* u1_later = get_ctrl(u1);
6637 tty->print("later(n->out(%d)): ", i); u1_later->dump();
6638 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
6639 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
6640 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
6641 }
6642 for (uint j = 0; j < u1->outcnt(); j++) {
6643 Node* u2 = u1->raw_out(j);
6644 if (u2 == n || u2 == u1)
6645 continue;
6646 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
6647 if (!u2->is_CFG()) {
6648 Node* u2_later = get_ctrl(u2);
6649 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
6650 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
6651 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
6652 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
6653 }
6654 }
6655 }
6656 }
6657 }
6658 dump_idoms(early, LCA);
6659 tty->cr();
6660 }
6661
6662 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
6663 class RealLCA {
6664 const PhaseIdealLoop* _phase;
6665 Node* _early;
6666 Node* _wrong_lca;
6667 uint _early_index;
6668 int _wrong_lca_index;
6669
6670 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
6671 // is different: Return the previously visited node which must be the real LCA.
6672 // The node lists also contain _early and _wrong_lca, respectively.
6673 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
6674 int early_index = early_with_idoms.size() - 1;
6675 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
6676 bool found_difference = false;
6677 do {
6678 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
6679 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
6680 found_difference = true;
6681 break;
6682 }
6683 early_index--;
6684 wrong_lca_index--;
6685 } while (wrong_lca_index >= 0);
6686
6687 assert(early_index >= 0, "must always find an LCA - cannot be early");
6688 _early_index = early_index;
6689 _wrong_lca_index = wrong_lca_index;
6690 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
6691 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
6692 return real_lca;
6693 }
6694
6695 void dump(Node* real_lca) {
6696 tty->cr();
6697 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
6698 _phase->dump_idom(_early, _early_index + 1);
6699
6700 tty->cr();
6701 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
6702 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
6703
6704 tty->cr();
6705 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
6706 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
6707 if (_wrong_lca_index >= 0) {
6708 tty->print(" (idom[%d])", _wrong_lca_index);
6709 }
6710 tty->print_cr(":");
6711 real_lca->dump();
6712 }
6713
6714 public:
6715 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
6716 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
6717 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
6718 }
6719
6720 void compute_and_dump() {
6721 ResourceMark rm;
6722 Unique_Node_List early_with_idoms;
6723 Unique_Node_List wrong_lca_with_idoms;
6724 early_with_idoms.push(_early);
6725 wrong_lca_with_idoms.push(_wrong_lca);
6726 _phase->get_idoms(_early, 10000, early_with_idoms);
6727 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
6728 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
6729 dump(real_lca);
6730 }
6731 };
6732
6733 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
6734 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
6735 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
6736 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
6737
6738 RealLCA real_lca(this, early, wrong_lca);
6739 real_lca.compute_and_dump();
6740 }
6741 #endif // ASSERT
6742
6743 #ifndef PRODUCT
6744 //------------------------------dump-------------------------------------------
6745 void PhaseIdealLoop::dump() const {
6746 ResourceMark rm;
6747 Node_Stack stack(C->live_nodes() >> 2);
6748 Node_List rpo_list;
6749 VectorSet visited;
6750 visited.set(C->top()->_idx);
6751 rpo(C->root(), stack, visited, rpo_list);
6752 // Dump root loop indexed by last element in PO order
6753 dump(_ltree_root, rpo_list.size(), rpo_list);
6754 }
6755
6756 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
6757 loop->dump_head();
6758
6759 // Now scan for CFG nodes in the same loop
6760 for (uint j = idx; j > 0; j--) {
6761 Node* n = rpo_list[j-1];
6762 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
6763 continue;
6764
6765 if (get_loop(n) != loop) { // Wrong loop nest
6766 if (get_loop(n)->_head == n && // Found nested loop?
6767 get_loop(n)->_parent == loop)
6768 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
6769 continue;
6770 }
6771
6772 // Dump controlling node
6773 tty->sp(2 * loop->_nest);
6774 tty->print("C");
6775 if (n == C->root()) {
6776 n->dump();
6777 } else {
6778 Node* cached_idom = idom_no_update(n);
6779 Node* computed_idom = n->in(0);
6780 if (n->is_Region()) {
6781 computed_idom = compute_idom(n);
6782 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
6783 // any MultiBranch ctrl node), so apply a similar transform to
6784 // the cached idom returned from idom_no_update.
6785 cached_idom = find_non_split_ctrl(cached_idom);
6786 }
6787 tty->print(" ID:%d", computed_idom->_idx);
6788 n->dump();
6789 if (cached_idom != computed_idom) {
6790 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
6791 computed_idom->_idx, cached_idom->_idx);
6792 }
6793 }
6794 // Dump nodes it controls
6795 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
6796 // (k < C->unique() && get_ctrl(find(k)) == n)
6797 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
6798 Node* m = C->root()->find(k);
6799 if (m && m->outcnt() > 0) {
6800 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
6801 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
6802 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
6803 }
6804 tty->sp(2 * loop->_nest + 1);
6805 m->dump();
6806 }
6807 }
6808 }
6809 }
6810 }
6811
6812 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
6813 if (has_ctrl(n)) {
6814 tty->print_cr("No idom for data nodes");
6815 } else {
6816 ResourceMark rm;
6817 Unique_Node_List idoms;
6818 get_idoms(n, count, idoms);
6819 dump_idoms_in_reverse(n, idoms);
6820 }
6821 }
6822
6823 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
6824 Node* next = n;
6825 for (uint i = 0; !next->is_Start() && i < count; i++) {
6826 next = idom(next);
6827 assert(!idoms.member(next), "duplicated idom is not possible");
6828 idoms.push(next);
6829 }
6830 }
6831
6832 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
6833 Node* next;
6834 uint padding = 3;
6835 uint node_index_padding_width = static_cast<int>(log10(static_cast<double>(C->unique()))) + 1;
6836 for (int i = idom_list.size() - 1; i >= 0; i--) {
6837 if (i == 9 || i == 99) {
6838 padding++;
6839 }
6840 next = idom_list[i];
6841 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
6842 }
6843 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
6844 }
6845 #endif // NOT PRODUCT
6846
6847 // Collect a R-P-O for the whole CFG.
6848 // Result list is in post-order (scan backwards for RPO)
6849 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
6850 stk.push(start, 0);
6851 visited.set(start->_idx);
6852
6853 while (stk.is_nonempty()) {
6854 Node* m = stk.node();
6855 uint idx = stk.index();
6856 if (idx < m->outcnt()) {
6857 stk.set_index(idx + 1);
6858 Node* n = m->raw_out(idx);
6859 if (n->is_CFG() && !visited.test_set(n->_idx)) {
6860 stk.push(n, 0);
6861 }
6862 } else {
6863 rpo_list.push(m);
6864 stk.pop();
6865 }
6866 }
6867 }
6868
6869
6870 //=============================================================================
6871 //------------------------------LoopTreeIterator-------------------------------
6872
6873 // Advance to next loop tree using a preorder, left-to-right traversal.
6874 void LoopTreeIterator::next() {
6875 assert(!done(), "must not be done.");
6876 if (_curnt->_child != nullptr) {
6877 _curnt = _curnt->_child;
6878 } else if (_curnt->_next != nullptr) {
6879 _curnt = _curnt->_next;
6880 } else {
6881 while (_curnt != _root && _curnt->_next == nullptr) {
6882 _curnt = _curnt->_parent;
6883 }
6884 if (_curnt == _root) {
6885 _curnt = nullptr;
6886 assert(done(), "must be done.");
6887 } else {
6888 assert(_curnt->_next != nullptr, "must be more to do");
6889 _curnt = _curnt->_next;
6890 }
6891 }
6892 }